flash floods: understanding the runoff generation

FLASH FLOODS: UNDERSTANDING THE RUNOFF GENERATION PROCESSES AND THE USE OF SATELLITE-RAINFALL IN HYDROLOGIC SIMULATIONS

EFTHYMIOS IOANNIS NIKOLOPOULOS UNIVERSITY OF CONNECTICUT, 2010

Flash floods pose a threat to human lives and economies worldwide.

The

tremendous societal and economical impact of this natural hazard necessitates the development of accurate warning systems that can help to mitigate the flash flood risk. The inadequate understanding of the hydrologic processes leading to flash floods and the lack of observations associated with flash flood inducing storms are the main reasons that hamper the development of effective monitoring systems. The objectives of this thesis are to a) advance our understanding on the physical mechanisms and controls of runoff generation during flash floods and b) investigate the use of space-based precipitation observations as a way of improving our current monitoring strategies over remote and complex environments. The study takes place over the alpine region of northeastern Italy, an area that suffers from frequent flash floods. The backbone of this research is the integration of detailed ground and remote sensing observations with a physics-based distributed hydrologic model for simulating a series of major flash flood events. Results show that runoff generation mechanisms during flash floods follow a similar pattern with intense type of floods. An interesting and counterintuitive finding is that initial soil moisture conditions can play an important role in flash flood evolution and magnitude for i

the case of basins with high soil moisture capacity. It is shown that the error in rainfall derived from remote sensing magnifies as it propagates through the nonlinear rainfall-torunoff transformation and exhibits a definite dependence on basin scale. Results also suggest that the rainfall-to-runoff error magnification is greater for drier soil moisture conditions. The hydrologic simulations based on satellite-rainfall forcing revealed some potential, but the predicted hydrographs are generally associated with large uncertainties that depend, among other factors, on the relationship between satellite product’s resolution and the scale of application. Overall, the findings suggest that the current state of satellite-based flood prediction suffers from the inability of satellite precipitation observations to accurately estimate the magnitude of high rainfall rates and that improvement of current precipitation products is needed to allow us to invest towards that direction.

ii

FLASH FLOODS: UNDERSTANDING THE RUNOFF GENERATION PROCESSES AND THE USE OF SATELLITERAINFALL IN HYDROLOGIC SIMULATIONS

EFTHYMIOS IOANNIS NIKOLOPOULOS

B.S. TECHNICAL UNIVERSITY OF CRETE, 2002 M.S. THE UNIVERSITY OF IOWA, 2004

A DISSERTATION SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY AT THE

UNIVERSITY OF CONNECTICUT 2010

iii

Copyright by Eythymios Ioannis Nikolopoulos

2010 All Rights Reserved

iv

APPROVAL PAGE

Doctor of Philosophy Dissertation

FLASH FLOODS: UNDERSTANDING THE RUNOFF GENERATION PROCESSES AND THE USE OF SATELLITE-RAINFALL IN HYDROLOGIC SIMULATIONS

by

Efthymios Ioannis Nikolopoulos

Major Advisor_____________________________________________ Emmanouil N. Anagnostou Associate Advisor__________________________________________ Amvrossios C. Bagtzoglou Associate Advisor__________________________________________ Mekonnen Gebremichael Associate Advisor__________________________________________ Guiling Wang Associate Advisor__________________________________________ Glenn Warner

University of Connecticut 2010

v

ACKNOWLEDGMENTS PhD for me was a journey and now that I am reaching the end of it, I finally understand why it is all about the journey and not the destination. It has been a very interesting and extremely challenging “trip”, which was made possible to complete only due to the guidance and support I had from a number of people. This section is dedicated to them. My acknowledgments of course go first to my adviser, Prof. Manos Anagnostou. Manos have been an excellent adviser and I feel lucky for being his student. I was always taking advantage of the fact that he kept his office door open, to jump in and release all my frustration and agony about my progress. He was always very supportive and he patiently and successfully guided me through all these years. A person who is very dear to me and has an indirect but very significant contribution to the completion of this PhD is my former adviser Prof. Witek Krajewski. My two years working with Witek for my MS degree were the most important years because under Witek’s direction and training (pretty tough one to be honest) I created my scientific foundation. Besides a great mentor, Witek has been a great friend, and I feel honored to be able to say that I was his student but most importantly that I am his friend. My deepest gratitude belongs to my father, Ioannis, my mother, Sofia, and my two sisters Theano and Vaso for their love and support all these years. My family is the most sacred thing to me and without them I would have never been able to endure all the difficulties associated with PhD studies in a foreign country. Having mentioned that, I have to admit that I was lucky to have a second family in US, the Lengas family, that treated me like their son and made me feel at home. Costa, Stacia and their daughter Tina are my second family, and I will never forget all the things they did for me. I feel blessed for having met such wonderful people; moreover, I hope I made them proud. I also wish to acknowledge my best friends Alex Ntelekos, Platonas Panagopoulos, Artemis Mamas, Aristotelis Pagoulatos, Giorgos Lykokanelos, Thanasis Zakopoulos, vi

Vassilis Christopoulos, and Andreas Panousis because they are an important part of my life, and each one of them has contributed in his own way to this achievement. Last but definitely not least I want to acknowledge my fiancée, Mary, for constantly being at my side and for tolerating my stress during all this time. She has been a source of inspiration and a reason for me to continuously try to become a better person. I dedicate this dissertation to her, and I promise that now that my PhD is completed I will finally concentrate on our wedding preparations…

vii

To my love, Mary

viii

CONTENTS CHAPTER 1: INTRODUCTION ..................................................................................................1 CHAPTER 2: EXAMINING RUNOFF RESPONSE AND MECHANISMS OF TWO CONTRASTING FLOODS IN A MOUNTAINOUS BASIN IN ITALY 2.1 Introduction.................................................................................................8 2.2 Study Area and Data.................................................................................11 2.3 The Flood Events......................................................................................14 2.3.1. Rainfall Analysis ...........................................................................15 2.3.2. Hydrologic Analysis......................................................................16 2.4 Hydrologic Modeling................................................................................20 2.4.1 Model Description ..........................................................................21 2.4.2. Model Setup...................................................................................22 2.4.3 Model Calibration and Validation ..................................................24 2.4.4. Runoff Generation Mechanisms....................................................28 2.5 Conclusions...............................................................................................33 CHAPTER 3: SENSITIVITY OF A MOUNTAIN BASIN FLASH FLOOD TO INITIAL SOIL MOISTURE AND RAINFALL VARIABILITY 3.1 Introduction...............................................................................................36 3.2 Study Area and Data.................................................................................39 3.3 The 2003 Fella Flash Flood Event............................................................42 3.4 Hydrologic Modeling................................................................................44 3.4.1 Model Description ..........................................................................44 ix

3.4.2. Model Setup...................................................................................45 3.4.3 Model Calibration and Validation ..................................................47 3.5 Sensitivity to Initial Soil Moisture............................................................50 3.6 Sensitivity to Rainfall Variability.............................................................55 3.7 Combined Sensitivity to Initial Soil Moisture and Rainfall Variability .......................................................................................................62 3.8 Conclusions...............................................................................................64 CHAPTER 4: UNDERSTANDING THE SCALE RELATIONSHIPS OF UNCERTAINTY PROPAGATION OF SATELLITE RAINFALL THROUGH A DISTRIBUTED HYDROLOGIC MODEL 4.1 Introduction...............................................................................................67 4.2 Study Area and Data.................................................................................72 4.3 Satellite-Rainfall Ensembles.....................................................................74 4.4 Hydrologic Simulations ............................................................................80 4.6 Analysis of Error Propagation ..................................................................84 4.7 Conclusions...............................................................................................90 CHAPTER 5: ON THE USE OF HIGH RESOLUTION SATELLITE-RAINFALL PRODUCTS FOR FLASH FLOOD SIMULATIONS

5.1 Introduction...............................................................................................93 5.2. Study Area and Data................................................................................97 5.3. Rainfall Analysis .....................................................................................98 5.4. Hydrologic Simulations .........................................................................103 5.5. Conclusions............................................................................................109 x

CHAPTER 6: CONCLUSIONS AND FUTURE RESEARCH 6.1 Conclusions Summary ............................................................................110 6.2 Future Research Directions.....................................................................113 APPENDIX ......................................................................................................................116 BIBLIOGRAPHY .................................................................................................................117

xi

LIST OF TABLES Table 2.1. Topographic characteristics of Posina and its two sub-basins..........................11 Table 2.2. Rainfall and runoff characteristics for the 1992 and 1999 event......................20 Table 2.3. Error metrics, correlation and efficiency score between observed and simulated hydrographs. Note that positive sign means overestimation of the model. .......................................................................................................................26 Table 3.1. Characteristics of the Fella basin and its sub-basins.........................................41 Table 3.2. Results from the calibration/validation exercise for the 2003 flash flood event. Note that minus denote underestimation of the model...................................49 Table 3.3. Runoff contribution of each generation mechanism for different initial soil moisture conditions. ...........................................................................................54 Table 3.4. Total basin-averaged rainfall for different aggregation levels..........................56 Table 3.5. Same as Table 3.3 but for different rainfall resolutions. ..................................61 Table 4.1. Nominal spatial and temporal resolution of the precipitation data used...........74 Table 4.2. List of size and topography slope information of the basins used in this study..........................................................................................................................78

xii

Table 4.3. Calibrated parameters of saturated hydraulic conductivity, conductivity decay coefficient and anisotropy ratio for the three soil classes of Bacchglione basin. .........................................................................................................................82 Table 5.1. Nominal spatial and temporal resolution of radar and satellite precipitation data used in this study..........................................................................97 Table 5.2. Bias and correlation coefficient between radar (reference) and satellite products before and after adjustment for mean field bias. Mean field bias adjustment coefficients derived from quantile-quantile plots are also shown........100 Table 5.3. Error metrics based on the comparison of radar simulated and satellite simulated hydrographs before and after the mean field bias adjustment. Note that the simulated hydrographs were based on the radar-calibrated parameters. ...105 Table 5.4. Same as Table 5.3 but for parameters calibrated separately for each satellite product before and after the adjustment. ...................................................106 Table 5.5. Values for saturated hydraulic conductivity obtained after individual calibration for each product. Note that the values shown correspond to the dominant soil class that occupies over 80% of the basin area. ...............................107

xiii

LIST OF FIGURES Figure 2.1. Map showing the location of Posina River basin in Northeastern Italy. The 20 m digital elevation model shows the high differences in elevation within the basin. Rain and stream gauge locations are also shown in the map. Note that the stream gauge numbers (1,2,3) correspond to the outlet at Stancari, Bazzoni and Rio Freddo at Valoje respectively. .......................................12 Figure 2.2. Land use (left) and soil type (right) map for the Posina basin. .......................14 Figure 2.3. Temporal (left) and spatial (right) distribution of rainfall during the flood events of 1992 (upper panel) and 1999 (lower panel). Left panel shows both the hourly mean areal rainfall accumulation time series (black bars) and the total mean areal rainfall accumulation series (grey line) for Posina basin. Right panel shows the spatial map of total rainfall accumulation derived from the rain gauge observations by applying an inverse distance interpolation. Green circles corresponds to the locations of the rain gauges used to derive the rainfall field. Black triangles correspond to the location of the stream gauges. ......................................................................................................................16 Figure 2.4. Observed rainfall-runoff time series for 1992 (left) and 1999 (right) event. The flood hydrograph (grey) and the mean areal rainfall (black) is shown for Stancari (a,d), Bazzoni (b,e) and Rio Freddo (c,f)...................................18 Figure 2.5. Flow duration curve derived from 13 years of half hourly averaged data from Posina’s outlet station. The inset magnifies a part of the curve to help xiv

display the initial baseflow for 1992 and 1999 event. Note that peak flows for 1992 and 1999 are associated with probability of exceedance equal to 0.003 and < 0.001 percent respectively. .............................................................................19 Figure 2.6. Simulated hydrographs based on the calibrated parameters (solid) versus observed (dashed) for the 1992 (left) and 1999 (right) flood. Results are shown for the outlet at Stancari (a,b), Bazzoni (c,d) and Rio Freddo (e,f). Note that only the 1999 hydrograph at Stancari (b) was used for the calibration of the model. ...........................................................................................27 Figure 2.7. Time series of simulated runoff components for the 1992 (left) and 1999 (right) flood. The different runoff types include infiltration excess (solid black), saturation excess (solid grey), perched return flow (dashed black) and groundwater exfiltration (dashed grey)...................................................30 Figure 2.8. Occurrence (percentage of total simulation time) of different runoff generation mechanism as a function of the topographic index ln(Ac/tan!). .............31 Figure 2.9. Spatial map of the calculated topographic index ln(Ac/tan!) over Posina basin. .........................................................................................................................32 Figure 3.1. Location map of the Fella river basin. The topographic representation is based on 20-m digital elevation model. Numbers 1-4 correspond to the outlets of the basins examined in this study (see Table 3.1). ...............................................40 Figure 3.2. Left: Time series of basin-averaged half-hourly rainfall accumulation for the Fella basin and sub-basins. Right: Total rainfall accumulation map for

xv

Fella basin. Both plots are based on radar-rainfall data during the 2003 flash flood inducing storm.................................................................................................43 Figure 3.3. Groundwater rating curve for the Fella basin expressed as a relation between baseflow and basin-averaged depth to groundwater table (measured from the surface).

The curve was constructed from a simulation-based

drainage experiment..................................................................................................46 Figure 3.4. Observed versus simulated hydrographs at the outlet of each basin. Note that calibration was based on (a) and validation on (b,c,d). .....................................49 Figure 3.5. Left panel: Peak discharge (top) and runoff ratio (bottom) versus basin scale for different initial wetness conditions. Right panel: Difference of peak discharge (top) and runoff ratio (bottom) relative to control simulation (Qb=2) versus basin scale......................................................................................................51 Figure 3.6. Relative difference in peak discharge (left) and runoff ratio (right) versus relative difference in average depth to water table (used to express the difference in saturation level) for each basin. Note that fitted lines were based on least-squares linear regression and aim only in aiding visualization of the scale effect (different slopes) on the sensitivity to initial soil moisture conditions..................................................................................................................53 Figure 3.7. Hourly radar-rainfall accumulation map at the original spatial resolution (0.5 km) and at 1, 2, 4, 8, 16 km resolutions after aggregation. Data correspond to 17:00 UTC during the flash flood event. ...........................................57

xvi

Figure 3.8. Left panel: Peak discharge (top) and total runoff volume (bottom) versus basin scale for different rainfall aggregation resolutions. Right panel: Difference of peak discharge (top) and total runoff volume (bottom) relative to control simulation (0.5km resolution) versus basin scale.....................................58 Figure 3.9. Effect of bias in basin-averaged rainfall volume due to aggregation (left) and effect of variability smoothing due to coarsening of resolution (right) expressed as relative differences of peak discharge (top) and runoff volume (bottom) for different basin scales. ..............................................................60 Figure 3.10. Relative difference of peak discharge for different aggregation levels and different initial wetness conditions for each basin examined. Note that white, grey and black symbols correspond to initial wetness conditions of Qb =2, 20 and 120 respectively. .....................................................................................62 Figure 3.11. Same as Fig. 3.10 but for total runoff volume...............................................63 Figure 4.1. Map showing the locations of the Posina and Bacchiglione basins in the northeastern Italian Alps. Note that the thin (4km) and thick (25km) grid shown, provide a visual comparison between the spatial resolution of the satellite products used and the basin scales. .............................................................73 Figure 4.2. Top: Mean areal rainfall accumulation curves for Bacchiglione basin calculated from radar (black) and SREM2D ensembles for 3B42 (left), KIDD-25km (middle) and KIDD-4km (right). Bottom: Bias (left), relative

xvii

RMSE (middle) and Nash-Sutcliffe score (right) between MAP time series derived from SREM2D ensembles and the reference rainfall (radar). .....................76 Figure 4.3. Same as Figure 4.2 but for Posina basin..........................................................77 Figure 4.4. Stream network of Bacchiglione basin and the locations (black dots) of the basin outlets analyzed in this study. Note that the numeric IDs correspond to the IDs presented in Table 4.2. .............................................................................81 Figure 4.5. Top: Observed (black) and simulated (blue) hydrographs for the Posina basin during the Oct. 1996 flood event. Bottom: Mean areal precipitation over Posina basin based on radar-rainfall data. ................................................................83 Figure 4.6. Top: Simulated hydrographs based on radar (black) and SREM2D (grey) rainfall ensembles for 3B42 (left), KIDD-25km (middle) and KIDD4km (right), for Bacchiglione basin. Bottom: Bias (left), rel. RMSE (middle) and N-S score (right) between reference (radar) and SREM2D hydrographs..........85 Figure 4.7. Same as Figure 4.6 but for Posina basin..........................................................86 Figure 4.8. Error propagation metrics: Top panel shows the relative error in total runoff versus relative error in total rainfall for Bacchiglione (left) and Posina (right) basin. Middle panel shows relative RMSE in discharge versus relative RMSE in rainfall, and bottom panel shows relative error in peak runoff versus relative error in total rainfall respectively. Errors were calculated between SREM2D ensembles and the reference (radar). Note that blue black and red

xviii

triangle corresponds to the ensemble average of KIDD 4km, KIDD 25km and 3B42 respectively. ....................................................................................................88 Figure 4.9. Ratios of relative error in i) total runoff (left) and ii) peak runoff (right) over relative error in total rainfall. Ratios of relative RMSE in runoff (middle) over relative RMSE in rainfall. All ratios are shown for different basin scales. Blue, black and red circles correspond to the average of the 20 realizations for the KIDD 4km, KIDD 25km and 3B42 respectively. Error bars are equal to plus/minus one standard deviation............................................................................89 Figure 5.1. Elevation map of North-Eastern Italian region (highlighted in the inset map of Italy) derived from the NASA Shuttle Radar Topographic Mission (SRTM) 90m digital elevation dataset. Map also shows the location of the OSMER radar and the boundaries of Fella river basin.............................................98 Figure 5.2. Left: Time series of basin-averaged hourly rainfall over Fella basin based on radar, CMORPH and PERSIANN data. Right: Time series of basinaveraged 3-hourly rainfall over Fella basin based on radar and 3B42. All data were collected during the 2003 flash flood induced storm.......................................99 Figure 5.3. Radar-rainfall quantiles vs satellite-rainfall quantiles for CMORPH (left), PERSIANN (middle) and 3B42 (right). Solid line corresponds to the slope =1 line and dash line corresponds to the linear regression applied. ..............102 Figure 5.4. Same as Figure 5.2 but adjusted based on the mean-field bias factor derived from quantile-quantile plots.......................................................................102

xix

Figure 5.5. Observed and simulated hydrographs at the outlet of Fella basin based on radar (reference) and satellite rainfall input before and after the bias adjustment...............................................................................................................104 Figure 5.6. Simulated hydrographs based on individual calibration of radar input and each satellite product before (left) and after (right) bias adjustment. ..............106 Figure 5.7. Same as Figure 5.6 (right) but for fully saturated initial soil moisture conditions................................................................................................................108

xx

CHAPTER 1 INTRODUCTION A flash flood is defined by the American Meteorological Society as “a flood that rises and falls quite rapidly with little or no advance warning, usually as the result of intense rainfall over a relatively small area”. It is considered as one of the most devastating natural hazards, responsible for significant loss of life and property worldwide. The tremendous societal and economical impact of this hazard necessitates the development of effective monitoring systems in order to mitigate the flash flood risk. In spite of all the scientific work and progress done so far, flood prediction still poses a great challenge for hydrologists. Predicting the amount of runoff generated after rainfall, traces back to the fundamental question “where water goes when it rains”.

This

simplified question, involves all the complex physical mechanisms that govern water flow on the surface and subsurface. Over the years hydrologists have improved their understanding on the physical processes that drive runoff generation and along with the simultaneous computational advancements have developed a number of modeling tools to describe and predict the rainfall transformation to runoff. Hydrological models have served as the main tool for rainfall-to-runoff prediction and currently form the basis for flash flood warning and forecasting systems, which include the integration of quantitative precipitation estimates (QPE) with hydrological models that simulate the hydrologic processes at watershed scale. Successful modeling of flash floods relates to the understanding of the hydrological processes that govern these phenomena.

However, flash floods develop at space and time scales that

1

conventional observational systems of rainfall and river discharges are not able to capture (Creutin and Borga, 2003), thus the runoff generation mechanisms during flash floods are still poorly understood. As a result, the inadequate understanding of those mechanisms is embedded in the current modeling systems and limits their ability to efficiently represent the hydrologic response during flash floods.

Recent efforts that aim at improving

techniques for predicting flash flood have revealed the potential for the use of physicallybased distributed hydrologic models (DHM) (Ogden et al., 2000; Vivoni et al., 2007; Borga et al., 2007, among others). Distributed hydrologic modeling, which is the state of the art in hydrologic modeling, is considered the most accurate in terms of describing the physical processes because it involves the solution of the continuity equations of mass and energy as opposed to the currently widely used lumped conceptual models that are based on the concept of interrelated storages to represent the physical elements in a catchment. Furthermore, due to the distributed nature, DHM can provide information on the spatial dynamics of the hydrologic variables, which is essential for successful flash flood prediction due to the high spatial variability that this hazard exhibits (Gaume et al., 2004). Nevertheless, even if we assume that we possess a perfect modeling system that could describe with certainty all the physical watershed processes, the outcome of the prediction of such model would still depend on the accuracy of the forcing variables applied to the system. Consequently, an important issue for the effective application of DHM in flood warning systems is the apprehension of the model’s sensitivity to the different sources of input uncertainty.

Soil moisture and rainfall are the two most

important variables that control runoff generation. Thus, assessing the effect of their

2

uncertainty in hydrologic prediction is necessary for designing modeling strategies and decision making procedures. The sensitivity of runoff response to rainfall and initial soil moisture conditions is a subject that has been well recognized by the research community. Most of the studies so far have focused on the sensitivity effect of each variable independently (Woods and Sivapalan 1999; Bell and Moore 2000; Castillo et al. 2003; Nicótina et al. 2008, among others) and only few have analyzed the combined effect of rainfall and antecedent wetness to runoff generation (Zehe et al. 2005; Vivoni et al. 2007, Noto et al. 2008). The influence of rainfall representation on modeling the hydrologic response during flash floods is of high importance due to the high space-time variability that characterizes the storms causing those floods (Creutin and Borga, 2003). Regarding the effect of initial soil moisture conditions, it is generally recognized that antecedent soil moisture is of little importance in determining the magnitude of extreme floods (Wood et al. 1990), but other studies have provided counterexamples of the possible role of antecedent soil moisture conditions when combined with high soil moisture capacity (Borga et al. 2007). Despite the extensive literature on this subject, results do not converge to a unified conclusion and sometimes are even contradictory, as Segond et al. (2007) points out, which highlight the complexity of the problem and the need for more systematic investigations. Besides the requirement for improvement in the modeling aspect, effectiveness of flash flood warning systems is highly dependent on the advancement of QPE systems. Limitations of current QPE sensors (radars and gauges) to provide information at large spatial scales and especially over mountainous regions (which are prone to flash floods) necessitate the integration of new observational systems that can overcome these

3

obstacles.

Satellite sensors can offer unique advantages (e.g., global coverage and

observations in regions where in situ data are inexistent or sparse) that open up new horizons in hydrological applications at global scale. Integration of satellite QPE with hydrological models may have sounded utopia in a few decades ago, but in recent years the vast development of space-based precipitation estimation has allowed researchers to use global-scale satellite-rainfall products in various hydrologic applications (Guetter et al., 1996; Tsintikidis et al., 1999; Grimes and Diop, 2003; Wilk et al., 2006; Artan et al., 2007; Su et al., 2008; Collischonn et al., 2008). Given that space-based observations will continue to develop and that future missions, like the Global Precipitation Measurement (GPM) mission (Smith et al. 2007), are expected to provide coherent and more accurate satellite precipitation product, one can argue that the future of flood prediction lies on satellite precipitation estimation (Hossain and Lettenmaier, 2006). However, the use of satellite QPE for flood prediction applications is hampered by several deficiencies that relate mainly to (i) the error in satellite QPE and (ii) the propagation of error in simulating the flood response. Estimation of the error in satelliterainfall products is mandatory in order to i) assess the efficiency of satellite products and ii) provide the end-user (i.e. hydrologists) with an estimate of the uncertainty that can be incorporated in the decision making procedures. Recognizing the need for quantifying the uncertainty in satellite QPE, several recent studies have been initiated to compare the accuracy of various satellite rainfall products over land (McCollum et al., 2002; Gebremichael and Krajewski, 2004; Ebert et al., 2007; Dinku et al., 2007; Sapiano and Arkin, 2009). Although these studies have helped to improve our understanding on the various error characteristics of satellite-rainfall estimates, they involve investigations at

4

large spatial scales (e.g. regional) thus do not provide sufficient evidence for the error metrics in smaller scales, which are more relevant to flood prediction applications. Recent work by Bellerby and Sun (2005) and Hossain and Anagnostou (2006a,b) have focused towards that direction and have tried to characterize satellite error structure at scales relevant to flood processes. In addition to the significance in evaluating the error in satellite QPE, assessment of the error propagation in the estimation of hydrologic variables is of great importance since it directly relates to the primary objective of a flood warning system (i.e. prediction of river discharge). Evaluation of error propagation in satellite-rainfall from a hydrologic end is a very challenging task because it involves many factors such as i) specifications of the satellite-rainfall product (resolution etc.), ii) scale of the application, iii) regional climatic characteristics, iv) land surface characteristics and v) storm morphology (spatiotemporal distribution), that interact in a complex way. A realistic approach to this multidimensional problem is through simulation-based experiments performed for various scenarios of i) hydrologic setup (basin scales, initial conditions, land surface characteristics) and ii) different satellite-rainfall products. A few studies recently (Hossain and Anagnostou 2004; Hong et al. 2006) have approached this issue on a similar context using error models to simulate ensembles of satellite-rainfall fields and propagating them through a hydrologic model to assess the error propagation. The

work

presented

in

this

thesis

aims

at

establishing

a

hydrologic

investigation/validation framework to address the issues outlined above and shed light to the multiple questions that rise. In Chapter 2, runoff response and mechanisms are examined during two contrasting flood events. The analysis of runoff generation is based

5

on in-situ observations and application of a physically-based distributed hydrologic model. The use of DHM addresses two distinct aspects that include a) evaluation of distributed model’s ability to represent hydrologic response during flash floods and b) investigation of the spatial dynamics of different runoff generation mechanisms. Comparative analysis of the hydrologic response of the two contrasting storm event is carried out in order to identify similarities and differences in runoff generation with respect to differences in rainfall spatiotemporal distribution. Chapter 3 investigates the sensitivities of runoff generation to rainfall variability and initial soil moisture conditions for flash flood simulations. The distributed hydrologic model used in the previous study is used to simulate the hydrologic response over a range of sub-basins during a major flash flood event. A series of hydrologic simulations are performed for different initial soil moisture conditions and rainfall forcing resolutions in order to evaluate the sensitivity of runoff generation to those variables. Furthermore, the contribution of individual runoff generation mechanisms for the different conditions (initial soil moisture and rainfall aggregation) is examined in order to understand how the physical processes relate to those controlling variables. In Chapter 4, an evaluation framework based on a data-based numerical experiment is presented to address the issue of satellite error propagation in flood prediction.

A

satellite rainfall error model is devised to generate rainfall ensembles based on the characteristics of two satellite products with different retrieval accuracies and spatiotemporal resolutions. The generated ensembles are propagated through the distributed hydrologic model used in the previous studies to simulate the rainfall-runoff processes at different basin scales. The simulated hydrographs are compared against a “reference”

6

hydrograph (obtained by using high-resolution radar-rainfall as input) and the error propagation is evaluated as a function of basin scale and satellite QPE characteristics (resolution, retrieval error). Finally, the use of actual satellite data for simulating a flash flood event is examined in Chapter 5.

Three different global-scale high resolution

satellite products are used to force the distributed hydrologic model for simulating the major flash flood event analyzed in Chapter 3. Comparison between radar-rainfall fields and satellite-rainfall fields is carried out to determine the mean-field bias of the different satellite rainfall products (relative to rain gauge-calibrated radar rainfall). Comparative analysis of the simulated hydrographs from all products is used to examine the adequacy of satellite QPE for flash flood related applications and highlight current strengths and limitations that can serve as a valuable reference to both hydrologists and the satellite product developers. Chapter 6 summarizes the general conclusions derived from the work presented in this dissertation and discusses future research steps that can potentially provide new directions to the solution of the flash flood prediction problem.

7

CHAPTER2 EXAMINING RUNOFF RESPONSE AND MECHANISMS OF TWO CONTRASTING FLOODS IN A MOUNTAINOUS BASIN IN ITALY

2.1 Introduction The sudden increase of a flood wave with relatively high peak discharge is the attribute for which flash floods are considered the most dangerous type of floods and leads to a high ranking as one of the most devastating hazards.

Flash floods are

responsible for significant loss of life and property worldwide and these numbers are expected to increase due to increasing population in flood-prone areas (Gruntfest and Handmer, 2001; Pielke et al., 2002). An indicative example is the 2002 flash flood event in the Gard area in France, with estimated damages at !1.2 billion (Huet et al., 2003). Despite its tremendous societal and economical impact, this natural hazard is still poorly understood because of the high complexity of the meteorological and hydrological processes leading to flash floods and the lack of detailed, spatio-temporal observations associated with those extreme events. The localized character and short duration of flash flood-inducing storms make their observation challenging, especially over complex terrain areas where few or no in-situ observations are available (e.g., Creutin and Borga 2003; Vivoni et al., 2006a).

Furthermore, flash floods occur usually over small,

ungauged mountainous basins which imply that records of the flood hydrograph are typically unavailable.

To overcome this obstacle, recent efforts have focused in

8

collecting existing records on flash flood cases from several regions and compiling them in a unified database (Gaume et al., 2009). Similar efforts must be encouraged because such databases provide an excellent opportunity for scientists and engineers to extend and advance their understanding. Over the last decades several studies have aimed at advancing understanding and techniques to predict floods.

Some have focused primarily on analyzing the

hydrometeorological patterns of flash flood-inducing storms (e.g., Zhang et al., 2001; Giannoni et al., 2003; Delrieu et al., 2005; Vivoni et al., 2006b; Borga et al., 2007), while others have tried to identify and analyze the dominant hydrological processes during flash flood events (e.g., Gaume et al., 2003, 2004; Desilets et al., 2008). Understanding the runoff mechanisms associated with intense floods and flash floods for different types and scales of watersheds is essential in order to advance the predictability of this hazard. This is an extremely difficult task due to: (a) the complexity of the inherent watershed processes (e.g. nonlinear runoff generation due to variability in initial soil moisture conditions, Vivoni et al. 2007), (b) the short space-time scales of occurrence (typically 70%) is forested (both conifer and broad-leaf forests), while the rest is mainly pastures and mixed agriculture. The predominant geological

11

features of the basin are metamorphic and sedimentary rocks that are characterized by low-to-medium permeability. Figure 2.2 shows the spatial distribution of vegetation and soil patterns. The environment is characterized as humid and the mean annual rainfall exceeds 1600 mm (Norbiato et al., 2008a). The high precipitation amounts along with the steep, irregular terrain are responsible for the “flashy” character of the basin that makes it prone to flash floods. Several researchers have identified the value of this basin as a physical hydrologic laboratory and have focused on Posina for flood related studies (Borga et al., 2000; Hossain and Anagnostou, 2004; Norbiato et al., 2008a).

Figure2.1. Map showing the location of Posina River basin in Northeastern Italy. The 20 m digital elevation model shows the high differences in elevation within the basin. Rain and stream gauge locations are also shown in the map. Note that the stream gauge numbers (1,2,3) correspond to the outlet at Stancari, Bazzoni and Rio Freddo at Valoje respectively.

12

The data utilized in this study consist mainly of in-situ observations of rainfall and river discharges. Specifically, three stream gauges located at Stancari (the outlet of Posina), Rio Freddo and Bazzoni provided half-hourly measurements of discharges, and six rain gauges located within and around the Posina basin (Figure 2.1) provided halfhourly rainfall accumulations. All the data regarding topography, land use/cover, soil type, rainfall and discharge were provided by ARPAV (Agenzia Regionale per la Prevenzione e Protezione Ambientale del Veneto). Other meteorological variables (e.g. temperature, radiation, humidity) that were needed to force the hydrologic model were produced with the aid of a numerical weather forecasting system (POSEIDON; Papadopoulos et al., 2002). The system, which is fully operational since 1999, is based on a modified version of the Eta/NCEP mesoscale meteorological model (Mesinger et al. 1988; Janjic 1994). For the specific experiment, the model was integrated over a domain covering the Mediterranean Sea and the surrounding countries. In the vertical direction, 32 levels were used stretching from the ground to the model top (15,800 m). In the horizontal direction, a grid increment of 1/10 of degree was applied. The European Center for Medium range Weather Forecasts (ECMWF) global analysis gridded data on a 1/2° ! 1/2° horizontal grid increment were used for the initial and boundary conditions. The lateral boundaries of the model domain were updated at each time step from the ECMWF data that are available on 6-hourly intervals. The same data source was used to define the initial state of the sea surface temperature (SST) field, and the soil temperature and soil moisture availability at the six soil layers of the meteorological model.

13

Figure 2.2 Land use (left) and soil type (right) map for the Posina basin.

Although there were no in-situ meteorological observations available for the periods of interest, a dataset was available for a three month period (Aug.-Oct.) in 2000 from a station located close to the mid-northern boundary of the basin (45.96o latitude, 11.28o longitude), which was used to assess the accuracy of the meteorological model predictions. Comparisons of model simulations with station observations (not shown here) gave very high linear correlations (>0.8) in terms of the temperature, wind and net radiation variables, which gave us confidence about the use of these variables as meteorological forcing in the DHM.

2.3 The Flood Events Two events that caused major flooding and led to significant damages are analyzed in this study. The first event was an intense flood that occurred during October 4, 1992 and

14

the second event was a flash flood the occurred during September 20, 1999. The rainfall patterns associated with these events and their hydrologic response are analyzed in this section based on the available in-situ observations.

2.3.1. Rainfall Analysis The two flood-inducing storms of 1992 and 1999 have very distinct differences in terms of total rainfall amounts and spatio-temporal structures. The 1992 storm was a long-lived system that lasted 140 hrs and resulted in more than 400 mm of mean areal rainfall accumulation. From the time series shown in Figure 2.3a, it is apparent that the event consists of two major rainfall pulses that caused two different flood peaks (see Figure 2.4a). The first rainfall pulse lasted for approximately 75 hrs and resulted in a mean total rainfall accumulation of ~260 mm, while the second lasted for 65 hours and resulted in ~160 mm. In the case of 1999, the total mean areal rainfall was significantly less (250 mm; i.e., nearly half of the total rainfall of 1992), but the duration of the storm was only 30 hrs (~1/5th of the duration of the 1992 storm), which is translated to a higher intensity event (Figure 2.3c). Both storms exhibit high spatial variability over the basin with maximum difference (maximum – minimum value) of total rainfall reaching ~200 (100) mm, and coefficient of variation of 0.1 (0.06) for the event of 1992 (1999). An interesting feature in the spatial pattern of the two storms is that during the 1992 event the highest rainfall amounts fell over the central-west part of the basin and particularly at the western part of Bazzoni sub-basin, while in the 1999 event the eastern part of Posina and specifically the area close to the outlet received the maximum rainfall amounts.

15

Figure 2.3. Temporal (left) and spatial (right) distribution of rainfall during the flood events of 1992 (upper panel) and 1999 (lower panel). Left panel shows both the hourly mean areal rainfall accumulation time series (black bars) and the total mean areal rainfall accumulation series (grey line) for Posina basin. Right panel shows the spatial map of total rainfall accumulation derived from the rain gauge observations by applying an inverse distance interpolation. Green circles corresponds to the locations of the rain gauges used to derive the rainfall field. Black triangles correspond to the location of the stream gauges.

2.3.2. Hydrologic Analysis Both events initiated under dry antecedent soil moisture conditions. A thirteen-year record of half-hourly discharge data at the Posina outlet was used to construct the flow duration curve (Figure 2.5). According to the derived curve, the initial baseflow of the

16

1992 event corresponds to the 87th quantile, which is nearly half of the 1999 event baseflow (corresponding to the 72nd quantile). The mean areal rainfall forcing and the corresponding hydrograph for each basin are presented in Figure 2.4 for both flood events. Areal rainfall was derived from the six rain-gauge observations based on a simple inverse-distance interpolation scheme that was done at 1-km resolution and without taking into account the topography. Although this can prove a potential source of error in estimating areal rainfall, especially over complex terrain, it is assumed that even with this simple method, the basin-averaged rainfall pattern would be captured adequately for the purpose of the comparison presented below. During the 1992 event, rainfall started falling with a low to moderate intensity (~5-10 mm/hr) and streamflow started rising very slowly until after approximately thirty hours from the beginning of the event (equivalent to ~100 mm of rainfall accumulation, see Figure 2.3) where a fast rising of the flow began and the first flood peak (250 m3/s at Stancari) occurred. After the peak, rainfall reduced to very low intensity (almost paused) for a few hours and the recession limb began. A new intense rainfall pulse interrupted the falling limb causing a second flood peak (~200 m3/s at Stancari) that was followed by a sharp falling limb with a prolonged recession (>40hrs). While this describes the general hydrologic response of the whole Posina, the hydrographs of Rio Freddo and Bazzoni exhibited significant differences. For example, the two flood peaks appear equal in terms of magnitude for the Stancari and Rio Freddo hydrographs, but in the case of Bazzoni, the second peak is significantly (~ 40%) lower than the first. Furthermore, the unit peak discharge and runoff ratios of Rio Freddo basin differ by a factor greater than two (see Table 2.2) from the other basins. These differences point out the significance of

17

Figure 2.4. Observed rainfall-runoff time series for 1992 (left) and 1999 (right) event. The flood hydrograph (grey) and the mean areal rainfall (black) is shown for Stancari (a,d), Bazzoni (b,e) and Rio Freddo (c,f). spatial heterogeneity of runoff generation, even within a small size basin, which is possibly attributed to the spatial variability of soil moisture and land surface

18

characteristics (soil type, vegetation cover etc.) (Braud et al. 1999; Vivoni et al. 2007; Sangati et al. 2009; Norbiato et al. 2009).

Figure 5. Flow duration curve derived from 13 years of half hourly averaged data from Posina’s outlet station. The inset magnifies a part of the curve to help display the initial baseflow for 1992 and 1999 event. Note that peak flows for 1992 and 1999 are associated with probability of exceedance equal to 0.003 and < 0.001 percent respectively. The 1999 flood had a different response than 1992 event, which is apparent in the hydrograph signature shown in Figure 2.4. A shorter duration but higher intensity (than 1992) event resulted in a fast rising hydrograph with significantly higher peak discharges (see Table 2.2). The rapid decrease of discharge, immediately after rainfall ceases, indicates that the flood is mainly generated due to fast response mechanisms (i.e surface runoff) (Gaume et al. 2004).

19

Table 2.2. Rainfall and runoff characteristics for the 1992 and 1999 event.

MAP Accum. Total Runoff (mm) Volume (mm)

Runoff Ratio

Peak Discharge (m3/s)

Unit Peak Discharge (m3/s/km2)

1992 1999

1992

1999

1992

1999

1992

1999

1992

1999

Rio Freddo

393

268

86.46

91.12

0.22

0.34

14

56

0.63

2.54

Bazzoni

424

249

212

82.17

0.5

0.33

63

100

1.5

2.38

Stancari

426

254

289.6

101.6

0.68

0.4

196

330

1.69

2.84

Basin

Differences in unit peak discharge and runoff ratios among the basins are not as significant for this event as they were for the 1992 event. This suggests that the spatial heterogeneity of runoff volume is controlled mainly by the rainfall patterns and initial soil moisture conditions rather than the land surface characteristics. Comparison of the hydrologic response of the two events reveals the highly complex character of rainfall-torunoff transformation. The variability in the magnitude of runoff (runoff ratio, unit peak discharge) and hydrograph characteristics (i.e. fast/slow recession limbs) suggests that the relative contribution of surface (fast) and subsurface (slower) processes differs for each event, depending on the rainfall forcing and wetness conditions. These are aspects to be investigated through the hydrologic modeling approach described in the next section.

2.4 Hydrologic Modeling To identify the predominant runoff generation mechanisms and investigate their relative contribution to each flood case, I utilized a hydrologic model. Due to the spatial heterogeneity of the flood response, application of a lumped model would inevitably be

20

unsuccessful as Gaume et al. (2004) has pointed out, thus in this study a distributed physics-based hydrologic model was used.

2.4.1 Model Description The TIN (triangulated irregular network)-based Real-time Integrated Basin Simulator (tRIBS) is the hydrologic model used in this study (Ivanov et al., 2004,a,b; Vivoni et al., 2007).

tRIBS is a continuous fully-distributed model that accounts for the spatial

variability in topography, soils, vegetation and atmospheric forcing.

Infiltration is

simulated in a sloped, heterogeneous and anisotropic soil (Garrote and Bras, 1995). The model simulates the movement of infiltration fronts and their interaction with the groundwater surface, thus emphasizing the dynamic relationship between the vadose and saturated zones (Ivanov et al., 2004a). Depending on the evolution of wetting and top fronts, different states of the soil column can be distinguished in unsaturated, perched and completely saturated zones. As a consequence, runoff generation can occur via different mechanisms that are distinctly identified by the model as infiltration excess, saturation excess, perched return flow and groundwater exfiltration (see Vivoni et al., 2007 for a detailed description of these). In addition to vertical fluxes, tRIBS also accounts for the lateral redistribution of moisture and the losses due to evapotranspiration and groundwater drainage which makes it able to simulate interstorm periods and thus operate in a continuous mode.

21

2.4.2. Model Setup One of the distinct characteristics of tRIBS is that it uses a TIN to represent the complexity of the watershed topographic domain. A TIN is derived from a raster-based digital elevation model (DEM) based on a methodology described in Vivoni et al. (2004). Implementation of TIN offers the advantage of using multiple resolutions, thus reducing significantly the number of computational nodes (relative to the original DEM), while, at the same time, the linear features (e.g. stream network, boundaries etc.) are accurately preserved. In the case of Posina, a high resolution DEM (20 x 20 m2) was used to delineate the watershed, derive the stream network and construct the TIN. The mesh consisted of only 17% of the original DEM nodes, which is sufficient to adequately capture the model response (see Vivoni et al. 2005). This representation allows significant computational savings, making the model suitable for ensemble simulations or even operational real-time applications (e.g., Forman et al., 2008; Mascaro et al., 2009; Vivoni et al., 2009). Besides the representation of the topography, the model accounts also for the spatial distribution of other land surface descriptors. Different soil texture and land use/cover data are mapped to the computational elements.

Each class of soil or land use is

associated with an attribute table (e.g. hydraulic properties, vegetation characteristics), which is used for the computation of the fluxes (i.e. infiltration, evapotranspiration) at each element of the domain. In the case of Posina, a 20 m resolution soil and land use/cover map (Figure 2.2) was used to categorize the basin in different soil and vegetation classes. For the soil texture, the basin was divided in three main types that include sandy loam, silty loam and silty clay loam. The vegetation was categorized in 22

four major classes that included barren land, grasses, broad-leaf forest and conifer forests. A drawback of the use of distributed physics-based models is the requirement of a large set of land-surface parameters, for which in most cases in-situ observations are not available. Since this was also the case for Posina, I had to rely on the literature regarding the properties of the soil and land use/cover classes (Jury et al. 1991; http://ldas.gsfc.nasa.gov/). On the other hand, this is a unique strength of this type of hydrologic models because their application to basins with little or no observations is feasible through physical considerations. The state of moisture in the soil column plays a key role in partitioning runoff into surface and subsurface components (Vivoni et al., 2007), thus successful initialization of the soil moisture profile is important to accurately simulate the hydrologic processes. The unsaturated soil moisture profile in tRIBS is determined from hydrostatic equilibrium using the Brook and Corey (1964) parameterization as:

(2.1)

where !(z) is the soil moisture at depth z, !r and !s are the residual and saturation soil moisture respectively, "b is the air entry bubbling pressure, Nwt the depth to water table and #o the pore-size distribution index (Ivanov et al., 2004a). Thus, if the soil hydraulic properties are determined, the soil moisture profile is derived as a function of the depth to water table. To determine the spatial distribution of the water table position and initialize the model, I followed the methodology described in Vivoni et al. (2007, 2008). The basin was assumed to be fully saturated by defining a depth to water table equal to zero for each element and then the model run for a long time (~330 days) 23

allowing the basin to drain with the evapotranspiration scheme being switched off. This drainage numerical experiment allowed us to construct a groundwater rating curve between flow at the outlet (baseflow) and the mean depth to water table in the basin (at various time steps). On the basis of this groundwater rating curve, I was able to select a realization of the water table position that would produce a baseflow equal to the observed at the basin outlet prior to the flood event, and use that as the initial water table distribution.

2.4.3 Model Calibration and Validation To assess the reliability of hydrologic models, calibration of the model parameters and evaluation of the hydrological estimates against available observations is required. While conceptual models usually need a very long data record (on the order of few years) for successful calibration, in this study the advantage of physically-based models to be calibrated with significantly fewer data is demonstrated. Specifically, it is shown that tRIBS calibrated only for a single flood event and based on runoff observations from a single station (at the basin outlet) can adequately represent the hydrologic response for other storm events and gauging locations.

As mentioned earlier, the drawback of

distributed models is the large number of parameters that need to be determined, which can complicate the calibration procedure. In this case, the calibration procedure was made more efficient by choosing to calibrate only three parameters (out of approximately thirty) for which the flood hydrograph is more sensitive, while relying on values derived from the literature for the rest of the parameters. The parameters that were calibrated are

24

the saturated hydraulic conductivity, the conductivity exponential decay coefficient and the anisotropy ratio (defined as the ratio of horizontal to vertical hydraulic conductivity). Traditionally, hydrologic models are calibrated by comparing the observed and simulated hydrographs at the outlet of the basin. Although a similar approach was followed it should be pointed out that one could take advantage of the distributed nature of the model and include in the calibration the hydrographs of the interior nodes that correspond to the outlets of Rio Freddo and Bazzoni sub-basins, as in Ivanov et al. (2004b). I chose in this study not to include those station observations in order to use them to assess the model’s ability to represent the internal streamflow dynamics. Once the model was setup, an initial estimate for the calibrated parameters was made and the rain gauge-derived rainfall and other meteorological variables were used to force the model and simulate the hydrologic response during the flood events. The length of the simulations was 195 hrs (for 1992 case) and 78 hrs (for 1999 case) and the actual computational time was approximately 40 min and 20 min, respectively. In both cases, simulations started a few hrs prior to the beginning of the storm event (14 hrs for 1992 and 5 hrs for 1999) depending on the availability and gaps in the forcing variables. The SCE (Shuffle Complex Evolution) optimization method (Duan et al., 1992) was utilized to minimize the error between the observed and the simulated hydrograph.

More

specifically, after defining a valid range and an initial value for each of the parameters to be calibrated, SCE routine was inputting tRIBS with different sets of parameters until the mean squared error between the simulated hydrograph and the observed was minimized. Only the flood event of 1999 was used to calibrate the model parameters, while the event of 1992 was used for independent verification.

25

In Figure 2.6, the simulated hydrographs based on the calibrated parameters are presented against the observed hydrographs at the three outlets and for both flood events. Figure 2.6b demonstrates the results of the calibration procedure and all other figures can be considered as an evaluation of the ability of the calibrated model to represent the hydrologic response for a different storm and in interior parts of the basin. A quantitative assessment of the model’s performance is presented in Table 2.3, where calculated error metrics and the Nash-Sutcliffe skill score are shown for all cases (definitions of these error metrics are provided in the Appendix).

Table 2.3. Error metrics, correlation and efficiency score between observed and simulated hydrographs. Note that positive sign means overestimation of the model.

Rel. Error in Peak Discharge

Rel. Error in Total Volume

Nash-Sutcliffe Score

Correlation Coeff.

1992

1999

1992

1999

1992

1999

1992

1999

Rio Freddo

+182%

+77%

+197%

+67%

-5.3

0.3

0.97

0.95

Bazzoni

+90%

+12%

+42%

+47%

0.32

0.47

0.97

0.93

Stancari

+27%

+10%

+1.80%

+19%

0.76

0.74

0.91

0.91

Basin

For the 1999 case, the model exhibits a high sensitivity in the first major rainfall pulse resulting in a first flow peak before the maximum peak, which is less apparent in the observed hydrograph and it is consistent in all sub-basins. However, for the Stancari and Bazzoni stations, the model captures very well the peak discharge (~10% relative error)

26

as well as the recession part of the hydrograph. For the Rio Freddo sub-basin, there is a

Figure 2.6. Simulated hydrographs based on the calibrated parameters (solid) versus observed (dashed) for the 1992 (left) and 1999 (right) flood. Results are shown for the outlet at Stancari (a,b), Bazzoni (c,d) and Rio Freddo (e,f). Note that only the 1999 hydrograph at Stancari (b) was used for the calibration of the model.

27

significant overestimation in peak discharges as well as the total runoff volume (see Table 2.3) which is consistent in both flood cases. A possible explanation for the significant discrepancies in the case of Rio Freddo can be the effect of karstified aquifers in this part of the basin (Norbiato et al. 2008a), which influences the runoff response in a way that the model is not able to capture. For example, existence of karstic structures can create voids that can trap water thus the model’s inability to represent this effect would lead to an overall overestimation of the predicted runoff. Although in all cases the linear correlation (Table 2.3) between observed and simulated hydrographs is very high (>0.9), the Nash-Sutcliffe score is moderate to low (especially for interior basins) due to the high bias in the estimation of the magnitude. Considering that the model was calibrated only based on the single hydrograph of the 1999 event at Stancari, it is worth noting how well it can simulate the hydrograph at the same location for a completely different storm event (Figure 2.6a).

A different

calibration procedure that would include information of the interior hydrographs would probably improve even more the performance of the model. However, even at the current state, the results provide confidence that the model is able to represent the hydrological processes at the basin and, furthermore, demonstrate that distributed hydrological models can achieve good performance with minimal calibration effort.

2.4.4. Runoff Generation Mechanisms By simulating the evolution of the wetting and top fronts, tRIBS partitions runoff into four basic types that include: (i) infiltration excess, (ii) saturation-excess, (iii) perched return flow and (iv) groundwater exfiltration. The first two types constitute the rapid

28

surface runoff response, while the other two correspond to the slower subsurface mechanisms (Vivoni et al., 2007). Note that within the model, infiltration excess type is considered to occur when (a) soil becomes saturated from above due to antecedent rainfall, or (b) rainfall intensity exceeds the soil infiltration capacity (usually referred to as Hortonian runoff) (Ivanov et al., 2004a). Saturation excess runoff occurs only when the soil column becomes fully saturated from a rising water table and rainfall continues to fall on the saturated region. In the case where subsurface water from an element exits onto the surface of another element, it is considered as perched return flow or groundwater exfiltration depending on whether it originated from a perched saturated layer or from the saturated (groundwater) zone. This capability of the model allowed us to identify the dominant runoff generation mechanisms for each flood event and investigate the spatial pattern of each type. Ivanov et al. (2004a) and Vivoni et al. (2007) present a more detailed description of the runoff generation mechanisms simulated in the distributed model. Figure 2.7 presents the contributions of the runoff production mechanisms to the total streamflow at the basin outlet for the different generation types. The first distinct feature of these results is that infiltration excess is the dominant runoff type for both flood cases. The perched return flow can be considered as the second dominant type and the significance of its relative contribution to the total generated runoff differs between the events.

The 1992 event starts with a low contribution of groundwater (dry initial

condition) and, as rainfall accumulates on the basin, there is no runoff production until it reaches ~70 mm (at about 50 hrs, see Figure 2.3). After that point, and while rainfall accumulation continues to rise with higher rates (steeper slope in the accumulation

29

curve), there is a sudden runoff production associated with all four types. This behavior indicates that part of the infiltrated rainfall has reached the water table resulting in an increase of groundwater flow, while in other parts of the basin, perched saturation and fully saturated elements occurred. After approximately 90% of rainfall had fallen (~100 hrs), the streamflow is dominated by the groundwater contribution. The modeling results regarding the evolution of the wetting fronts (not shown) justify the formation of a perched groundwater (expected for low conductivity soils), that resulted in simultaneous surface (infiltration excess) and subsurface (return flow) runoff generation.

Figure 2.7. Time series of simulated runoff components for the 1992 (left) and 1999 (right) flood. The different runoff types include infiltration excess (solid black), saturation excess (solid grey), perched return flow (dashed black) and groundwater exfiltration (dashed grey). During the 1999 flash flood, the basin initial wetness conditions was higher than in the 1992 case, which was translated to an increased initial groundwater flow. The evolution of runoff generation is similar with the 1992 event in the sense that an approximately equal amount of rainfall had to accumulate over the basin before runoff 30

production initiates. In the case of 1999, though, the relative contribution of saturationexcess is higher than that of 1992, but still the major volume of runoff was produced by infiltration excess. Moreover, the peak in saturation excess runoff is earlier, indicating a greater initial saturation in the basin.

Figure 8. Occurrence (percentage of total simulation time) of different runoff generation mechanism as a function of the topographic index ln(Ac/tan$). Besides the temporal evolution of the different types of runoff generation, I also investigated the spatial distribution of runoff occurrence of each mechanism. Figure 2.8 shows the occurrence (defined as the percentage of the total simulation time) of each runoff type, as a function of the topographic index ln(Ac/tan$) of each computational 31

element, where Ac is the upslope contributing area and $ the local terrain slope, following Ivanov et al. (2004b) and Vivoni et al. (2007). Model elements were grouped in unit topographic index bins and the time-average occurrence of a runoff mechanism and standard deviation of each bin are shown. Figure 2.9 shows the spatial distribution of the topographic index over the basin to help visualize the occurrence of different runoff types over the basin area. Clearly, high topographic index areas, corresponding to the streams and the floodplain areas of the basin, are characterized by saturation excess and the groundwater exfiltration contributions. The majority of the basin is characterized by topographic indexes lower than 15 where all runoff types have approximately an equal occurrence.

Figure 9. Spatial map of the calculated topographic index ln(Ac/tan$) over Posina basin.

32

The results indicate that there is high variability in runoff occurrence, for all types, within elements of the same topographic index. As Ivanov et al. (2004b) demonstrated, this variability can be attributed to different soil and vegetation properties of the elements.

Another distinct feature is that for both events the infiltration-excess

mechanism has the lowest occurrence even though the majority of the runoff volume produced during both events was based on this mechanism. A very interesting result apparent in Figure 2.8 is the topographical pattern of each runoff type. For example, infiltration excess occurs mainly at elements with topographic index between 12 and 15, while saturation excess occurs predominantly over elements with index greater than 15. These results are in agreement with previous work by Ivanov et al. (2004b) and Vivoni et al. (2007). On average, these patterns exhibit high similarity between the two storms that, as described earlier in this chapter, exhibit very different characteristics in terms of intensity and spatio-temporal distribution of rainfall. This indicates that the occurrence of different runoff types in the basin is controlled mostly by the topography and seems to have less sensitivity to rainfall forcing. However, the differences in variability within the same index bins can be attributed to differences in the rainfall forcing and initial soil moisture conditions.

2.5 Conclusions This chapter presented an analysis of the rainfall-runoff response during two flood events (an intense flood and a flash flood) that occurred in a complex terrain basin in northeast Italy. In-situ rainfall and discharge observations were used to analyze the hydrologic response for three basin scales and a physically-based, fully distributed

33

hydrologic model was used to simulate runoff and investigate the dominant runoff generation processes of the basin. The main findings from this study are: 1. Comparison of the runoff response during the two flood events shows the complexity of rainfall-to-runoff transformation. While the runoff ratio during 1992 event is significantly higher than 1999, peak discharges show exactly the opposite trend.

The higher intensity and different spatial pattern of rainfall

(concentrated closer to the outlet) during 1999 along with the more wet initial conditions are responsible for the peak runoff differences with the 1992 event as well as the differences in the internal runoff response. 2. Application of tRIBS showed that with only minimal calibration of a few key parameters, the model can successfully represent the runoff response. Relatively small errors of the simulated peak runoff and total volume were achieved for the outlet, while for the interior basins a better calibration scheme that includes the interior hydrographs can potentially improve the results. 3. Simulation results suggest that despite the distinct differences of the two floods, the same runoff mechanisms dominated in both cases. The major runoff volume was produced due to the formation of a perched groundwater table that resulted in simultaneous surface (infiltration excess) and subsurface (return flow) runoff generation. The steep sloped terrain and the moderate-to-low conductivity soils favored the dominance of these mechanisms. 4. Occurrence of each runoff generation mechanism exhibited a specific pattern in relation to the topography. Infiltration excess occurred mainly in the high slope upland areas while saturation excess and groundwater exfiltration occured in the

34

flat regions along the channel network. This pattern is on average consistent for the different events, which indicates that in this basin, occurrence of each runoff type is controlled mainly by topography and the other land surface descriptors. Summarizing, this study attempts to identify and understand the processes that govern the runoff response during two distinctly different flood events in a set of basins in Italy. The results presented here show that observations complemented with modeling work can advance our understanding on the complex nature of runoff generation mechanisms and furthermore enhance the argument that distributed hydrologic models can potentially provide a solution for the prediction in ungauged basins. However, to advance our investigation framework and gain a holistic understanding of the flood generation mechanisms and the interdependencies of runoff controls, more intensive field measurements and associated modeling efforts are required.

35

CHAPTER 3 SENSITIVITY OF A MOUNTAIN BASIN FLASH FLOOD TO INITIAL SOIL MOISTURE AND RAINFALL VARIABILITY

3.1 Introduction The current state of flash flood warning systems is based on the use of models that simulate the hydrologic processes at the watershed scale. Two important variables that control runoff generation and serve as input to models are the initial soil moisture and rainfall distributions. Soil moisture is also a hydrologic state variable that describes the saturation conditions of the land surface, while rainfall is strictly a forcing variable defining the volume of water that falls on the land surface. They are both highly variable in space and it is mainly due to this variability and interaction at the surface that the transformation of rainfall to runoff is a highly nonlinear process. Consequently, accurate knowledge of both variables is required to successfully predict the hydrologic response of a given basin. However, most real world hydrologic modeling applications lack the accurate and adequate observations on both rainfall and initial soil moisture distributions. Rainfall input is usually based on either sparse gauge observations or estimates from remote sensing (e.g., weather radar, satellite platforms), while soil moisture input is provided in most cases using indirect observations as a proxy (e.g., streamflow, depth to water table). In both cases, there are estimation errors due to limitations in the retrieval algorithms and

36

measurement errors, as well as representation errors due to coverage limitations or the coarse resolution and sampling issues of the sensors. Overall, the result is that hydrologic models are typically forced with rainfall and initial soil moisture fields that are associated with high uncertainties which have a subsequent effect on predictions. Assessing and understanding the sensitivity of hydrologic models to these important variables is necessary for the design of hydrologic forecasting strategies and decision making procedures. Evaluating the sensitivity of runoff response to rainfall variability and initial soil moisture conditions is an issue that has been well recognized by the research community. Sensitivity of streamflow generation to rainfall variability was initially addressed nearly four decades ago (Dawdy and Bergmann 1969) and since then, numerous studies have been published on this subject (Beven and Hornberger 1982; Ogden and Julien 1993; Koren et al. 1999; Woods and Sivapalan 1999; Bell and Moore 2000; Segond et al. 2007; Nicótina et al 2008; Sangati and Borga 2009; Saulnier and Lay 2009). Despite the extensive literature on this subject, results do not converge to a unified conclusion and sometimes are even contradictory, as Segond et al. (2007) points out, which highlight the complexity of the problem.

Sensitivity of runoff generation to antecedent wetness

conditions is also a well established subject that has been investigated by several authors (Hino et al. 1988; Loague 1992; Karnieli and Ben-Asher, 1993; Cerdá, 1997; Ceballos and Schnabel, 1998; Fitzjohn et al., 1998; Castillo et al. 2003). Despite the fact that the sensitivity of stream flow to rainfall is linked with antecedent wetness conditions (Singh 1997) most studies so far have focused on the sensitivity effect of each variable and only

37

few have analyzed the combined effect of rainfall and antecedent wetness to runoff generation (Shah et al. 1996; Zehe et al. 2005; Vivoni et al. 2007, Noto et al. 2008). All the aforementioned studies have been based on various hydrologic contexts regarding storm characteristics (movement, duration, and intensity), catchment properties (vegetation, soil), wetness conditions and climatic regimes. The interaction between those factors possibly explains the contradiction of the various results found in the literature. While no generalizations can be made on the issue, still each study has contributed towards the understanding of those complex interactions. In line with this concept, the study presented in this chapter attempts to investigate and explain the sensitivity of flash flood generation to rainfall variability and initial soil moisture conditions, under a specific topographical setup. The study is focused on a major flash flood event that occurred on the upper Tagliamento river basin, in the eastern Italian Alps, which is characterized by complex terrain, high soil moisture capacity and a flashy flow regime (Borga et al. 2007). The influence of rainfall representation on modeling the hydrologic response during flash floods is particularly important due to the high spatiotemporal variability that characterizes storms that induce flash floods (Creutin and Borga, 2003). Regarding the effect of initial soil moisture conditions, it is generally recognized that antecedent wetness is of little importance in determining the magnitude of extreme floods (Wood et al. 1990), but other studies have provided counterexamples of the possible role of antecedent soil moisture when combined with high soil moisture capacity (Borga et al. 2007). Furthermore, Castillo et al. (2003) have shown that runoff sensitivity to soil moisture conditions is closely related to the dominant runoff generation mechanisms. For example, arid basins having infiltration-excess as the dominant runoff

38

mechanism are expected to show much lower sensitivity than basins where subsurface mechanisms dominate. The research objectives addressed in this chapter include: i) the sensitivity of runoff generation to rainfall and initial soil moisture as a function of basin scale, ii) the interdependencies of those sensitivities, and iii) their effect on runoff generation mechanisms. This type of investigation has both practical and scientific merit because it allows us to identify the significance of each variable to flood modeling applications and provide insight on the mechanisms behind the overall effect helping us to understand the complex interactions between rainfall and antecedent wetness. Section 3.2 describes the study area and data. Section 3.3 provides a short description of the flash flood case examined based on the available observations. The outline of the distributed hydrologic model as well as the setup, calibration and validation procedure is provided in section 3.4. The sensitivity results to initial soil moisture conditions and rainfall variability are presented in sections 3.5 to 3.7 and the main conclusions of this study are summarized in section 3.8.

3.2 Study Area and Data The basin considered in this study is the Fella basin located within the Friuli-Venezia Giulia region, northeastern Italy (Figure 3.1). Fella basin is a major left-hand tributary of the Tagliamento River with an area of approximately 623 km2, a mean altitude of 1140 m above sea level and an average annual precipitation of 1920 mm (Borga et al. 2007). Important sub-basins also analyzed in this study include Uque (~24 km2), Pontebba (~165 km2) and Dogna (~329 km2) (see Table 3.1 for details). Fella is a complex terrain basin,

39

part of the eastern Italian Alps, with elevations that range from approximately 300 m a.s.l. (above sea level) close to the outlet to more than 2000 m a.s.l. near the mountain tops. The terrain slope, with an average of 32 degrees, drops to less than 1 degree at the floodplain areas close to the outlet, but reaches values greater than 60 degrees on the mountain slopes. Land is principally occupied by agriculture with significant areas covered by natural vegetation, broad-leaf and conifer forests. The alpine area of Friuli consists mainly of limestone with a spatial sequence of Silurian, Denovian, Triassic, Jurassic and Creataecous formations north to south (Cucchi et al. 2000).

Figure 3.1. Location map of the Fella river basin. The topographic representation is based on 20-m digital elevation model. Numbers 1-4 correspond to the outlets of the basins examined in this study (see Table 3.1).

40

Land cover data used in this study were based on the 100-m Corine Land Cover (CLC) version 9/2007 dataset, provided by the European Environmental Agency. Soil texture information was not available for the Fella basin, thus the land cover map was used to classify and group areas with similar characteristics (e.g., bare rocks with roads) to the same soil class. Subsequently, a runoff curve number map (Borga et al 2007) was used as a proxy to assign initial estimates of the hydraulic properties of each class. The precipitation dataset utilized in this study includes half-hourly radar rainfall fields at 0.5km spatial resolution. Rainfall fields were derived from the reflectivity scans of a Doppler, dual-polarized C-band radar (OSMER radar station) located at Fossalon di Grado, approximately 80 km south of the basin. A number of procedures were applied to the radar data to correct for ground clutter, partial beam blockage and atmospheric attenuation (see Borga et al. 2007 for details). Hydrologic observations include halfhourly data from stream gauges available at the outlet of the basin and the sub-basins 2 and 3 (Figure 3.1). Furthermore, hydraulic modeling combined with pre- and post-flood surveys of the river channel geometry was used to estimate the hydrologic response at other parts of the basin (sub-basin 4, Figure 3.1).

Table 3.1. Characteristics of the Fella basin and its sub-basins. Basin

Elevation (m) mean (std. dev.)

Slope (degrees) mean (std. dev.)

Number

Name

Area (km2)

1

Uque

24

1369 (225)

25 (9)

2

Pontebba

165

1273 (347)

29 (14)

3

Dogna

329

1259 (368)

31 (14)

4

Moggio

623

1171 (426)

32 (14)

41

Other than precipitation, the meteorological forcing was produced with the aid of a numerical weather forecasting system (POSEIDON; Papadopoulos et al., 2002). This system is fully operational since 1999 and is based on a modified version of the Eta/NCEP mesoscale meteorological model (Mesinger et al. 1988; Janjic 1994). The fields of the meteorological variables of interest for the hydrologic model (net radiation, wind speed, air temperature) were provided at 10-km and 1-hr resolution. The same setup has been successfully used and validated for another study in the region of northeastern Italy (see Chapter 2 for more details).

3.3 The 2003 Fella Flash Flood Event The flash flood event of the Fella basin during August 29, 2003 has been characterized as one of the most devastating flash flood event in northeastern Italy since the start of systematic observations in the region (Sangati and Borga 2009). The flood inducing storm started at 10:00 CET (Central European Time), lasted for approximately 12 hrs and resulted in losses of lives and damages close to 1 billion euro (Tropeano et al. 2004) in the area of the upper Tagliamento river.

This event has two distinct

characteristics related to i) the soil moisture conditions prior to the storm event and ii) the spatial distribution of rainfall over the Fella basin. More specifically, the rainfall event occurred at the end of a climatic anomaly of prolonged drought and warm conditions over the Mediterranean (Borga et al. 2007), which resulted in extremely dry soil moisture conditions (baseflow at the outlet was close to zero).

42

Figure 3.2. Left: Time series of basin-averaged half-hourly rainfall accumulation for the Fella basin and sub-basins. Right: Total rainfall accumulation map for Fella basin. Both plots are based on radar-rainfall data during the 2003 flash flood inducing storm. The mesoscale convective system responsible for the flooding, exhibited a characteristic persistence of the convective bands over the northern part of the Fella basin that resulted in very large rainfall accumulations and high spatial variability. The basinaveraged, half-hourly rainfall accumulations exceeded 40 mm (at Uque sub-basin), while the total rainfall accumulation over the 12 hour duration exceeded 400 mm in some parts of the basin (Figure 3.2). The combined effect of dry initial conditions and high spatial variability of rainfall translated to a highly heterogeneous runoff response with runoff ratios (unit peak discharge) ranging from 0.04 (0.43 m3/s/km2) to 0.2 (8.36 m3/s/km2) (Borga et al. 2007) in different parts of the basin. These distinct features make this case suitable to study the effect of initial soil moisture conditions and rainfall variability on the hydrologic response during flash floods.

43

3.4 Hydrologic Modeling The hydrologic response of the Fella river basin during the flash flood event was simulated using a distributed hydrologic model. The distributed nature of the model allowed us to investigate the hydrologic response in several parts of the basin.

A

description of the model, the setup and the calibration/validation exercise that was carried out is provided in the sections below.

3.4.1 Model Description The hydrologic model used in this study is the TIN (triangulated irregular network)based Real-time Integrated Basin Simulator (tRIBS) (Ivanov et al., 2004,a,b; Vivoni et al., 2007). tRIBS is a distributed physics-based model that explicitly accounts for the spatial variability of land surface descriptors (terrain, soil, vegetation), soil moisture and atmospheric forcing. Infiltration is simulated in a sloped, heterogeneous and anisotropic soil (Garrote and Bras, 1995). In addition to vertical fluxes, tRIBS also accounts for the lateral redistribution of moisture and the losses due to evapotranspiration and groundwater drainage which makes it able to simulate interstorm periods and thus operate in a continuous mode. One of the distinct characteristics of tRIBS is the use of a TIN to represent the topography of a complex basin (Vivoni et al. 2004). Use of TINs offers the advantage of using multiple resolutions to represent the topography (Vivoni et al. 2005), depending of the complexity of the terrain, thus reducing significantly the number of computational nodes (relative to the original DEM), while preserving the linear features (e.g., stream network, boundaries etc.) that characterize a river basin. The reader is 44

referred to Ivanov et al. (2004a,b) for additional details on the model physics and capabilities.

3.4.2. Model Setup The hydrologic model setup requires i) the construction of the computational mesh, ii) the parameterization of the soil and land use/cover, and iii) the initialization of the soil moisture conditions. For the Fella basin, a high resolution DEM (20 x 20 m2) (Borga et al. 2007) was used to delineate the watershed, derive the stream network and construct the TIN following procedures described in Vivoni et al. (2004). The Voronoi polygons derived from the TIN constitute the computational elements of the domain. The average area and standard deviation of the derived polygons was equal to 0.0198 km2 and 0.0196 km2, respectively, which corresponds roughly to an equivalent pixel size of 140 m. A comparison (not shown) between the elevation histogram calculated from the original DEM and the TIN ensured that the derived mesh was adequately representing the topography of the basin. tRIBS accounts also for the spatial variability of other land surface descriptors besides topography. Available land surface data maps are used to categorize the basin into different land use/cover and soil texture classes. These classes are mapped to the Voronoi polygons and are associated with an attribute table (e.g., hydraulic properties, vegetation characteristics), which is used for the computation of the fluxes (i.e., infiltration, evapotranspiration, lateral transport) at each element of the domain. The Fella basin was categorized in four major land cover classes that include bare rocks

45

grouped with urban areas, agricultural areas, broad-leaf forests and conifer forests. The soil texture classification was also based on the land cover dataset since no other information was available. Four soil classes were created that distinguished the basin in sandy areas, urban areas, agricultural areas and forested areas. For the parameterization of the soil and land cover variables in the model, I relied on different sources found in the literature (Jury et al. 1991; http://ldas.gsfc.nasa.gov/).

Figure 3.3. Groundwater rating curve for the Fella basin expressed as a relation between baseflow and basin-averaged depth to groundwater table (measured from the surface). The curve was constructed from a simulation-based drainage experiment. The unsaturated soil moisture profile in tRIBS is determined from hydrostatic equilibrium using the Brook and Corey (1964) parameterization (see equation 2.1 and section 2.4.2 of Chapter 2 for more details). Thus, the soil moisture profile is derived as a function of the depth to water table given that the soil hydraulic properties are specified.

46

To determine the spatial distribution of the water table position and initialize the model, I followed the same procedure described in Chapter 2 (section 2.4.2). By using the spatial output of the depth to water table and streamflow at the outlet, I constructed a groundwater rating curve between baseflow and the basin-average depth to water table (Figure 3.3). Based on this groundwater rating curve, a realization of the water table position that would produce a baseflow equal to the observed at the outlet prior to the flood event was selected and used as the initial water table distribution.

3.4.3 Model Calibration and Validation To calibrate the model, I followed the traditional practice of comparing the simulated and observed hydrograph at the outlet. As with all distributed models, tRIBS requires a large number of parameters for each land cover and soil class to be defined, which can make the calibration procedure complicated. Again to be more efficient I chose to calibrate only three parameters (similar to Chapter 2).

The parameters that were

calibrated are the saturated hydraulic conductivity, the conductivity exponential decay coefficient and the anisotropy ratio (defined as the ratio of horizontal to vertical hydraulic conductivity). Calibration was performed using the SCE (Shuffle Complex Evolution) optimization method (Duan et al., 1992), which is used to minimize the mean squared error (MSE) between the observed and simulated hydrograph. As the first step, an initial value and a valid range of each parameter to be calibrated was defined and used as the sampling space for the optimization routine. Each time a new set of parameters was selected based on the SCE algorithm, tRIBS was run (as an intermediate step of the SCE

47

routine) using the new parameters and the resulting output was used to calculate the objective function (MSE) and evaluate the selected set of parameters. This procedure continued until no further reduction of the MSE could be achieved and the final set of parameters was therefore determined. The length of the flash flood simulation was 30 hrs; simulation started 4 hrs before the beginning of the storm and extended 14 hrs after the end of the event in order to fully capture the flood hydrograph (Figure 3.4a). The actual computational time was approximately 15 min for each run and it took about 3 weeks for the optimization routine to converge to the final set of parameters. Unfortunately, in this case the only available data were for the 2003 flash flood event thus the model calibration was based only on a single flood hydrograph. In order to provide independent verification of the model’s performance, I used the available observations at the interior of the Fella basin to compare with the simulation results. Therefore, parameter calibration was performed based on the hydrograph at the basin outlet (Moggio, Figure 3.4a), while for the evaluation of the calibrated model parameters observations at three sub-basins (Uque, Pontebba and Dogna, Figure 3.4b,c,d) were used. Figure 3.4 presents the observed and simulated hydrographs for the basin outlet and the three sub-basins, while Table 3.2 summarizes the results of the comparison. Clearly, the model captures the flood response very well in all cases except for the Dogna basin (Figure 3.4b). Note that data were not available for the complete hydrograph of the Dogna basin, thus only the estimated peak discharge and time to peak were used for the calculation of the comparative statistics in Table 3.2 below.

48

Figure 3.4. Observed versus simulated hydrographs at the outlet of each basin. Note that calibration was based on (a) and validation on (b,c,d).

Table 3.2. Results from the calibration/validation exercise for the 2003 flash flood event. Note that minus denote underestimation of the model. Basin

Relative Error in

Error in Time to Peak

Nash-Sutcliffe score

Number

Name

Peak Runoff

Total Runoff

1

Uque

-1.84%

0.5%

0.5 hr

0.65

2

Pontebba

6.6%

8%

0.5 hr

0.86

3

Dogna

41%

-

1 hr

-

4

Moggio

3.65%

18.6%

-1 hr

0.89

49

Besides the Dogna basin, the relative error (see appendix for definition of error metrics) in peak discharge and total runoff volume ranged between -1.84% to 6.6% and 0.5% to 18.6%, respectively, with the model consistently overestimating the observed response except for the peak discharge in Uque sub-basin that showed a slight underestimation. The errors in time to peak were quite small, with 0.5 hr for the Uque and Pontebba sub-basins and 1 hr for the larger basins of Dogna and Moggio. The NashSutcliffe coefficient ranged from 0.65 (for Uque) to 0.89 (for Moggio), which translates in a high capability of the model to simulate the flood hydrograph. These results provide confidence about the calibration of the model and its ability to represent the hydrologic response of the Fella basin during the flash flood event. In addition, the validation of the model at the three gauging stations suggest the model is adequately capturing the internal processes in the basin.

3.5 Sensitivity to Initial Soil Moisture A numerical experiment was devised to assess the sensitivity of the flash flood response to initial soil moisture conditions. This was achieved by performing a series of simulations for different initial soil moisture conditions and comparing the results with the initially calibrated model run. As mentioned before, the initialization of the moisture profile depends on the depth to the water table, thus the selection of various initial soil moisture conditions was based on the selection of different realizations of the depth to groundwater from the constructed rating curve (Figure 3.3). Each point on the curve represents a realization of the soil moisture field that produces a baseflow (Qb) at the

50

outlet of Fella basin equal to the corresponding y-axis value of the point. The initial soil moisture field used to calibrate the model corresponded to very dry conditions (Qb = 2 m3/s) and for the sensitivity exercise, a series of points of an increased level of saturation (Qb= 4, 6, 8, 10, 20, 120 m3/s) were selected.

Figure 3.5. Left panel: Peak discharge (top) and runoff ratio (bottom) versus basin scale for different initial wetness conditions. Right panel: Difference of peak discharge (top) and runoff ratio (bottom) relative to control simulation (Qb=2) versus basin scale. Figure 3.5 presents the simulated values of peak flow and runoff ratio for different initial soil moisture conditions versus basin area and their difference relative to the

51

control simulation (Qb = 2 m3/s). Peak runoff increases with increasing basin area and the level of saturation, while the runoff ratio increases with the level of saturation but decreases with basin area. For the highest baseflow (Qb= 120 m3/s) where the soil is very wet, the runoff ratio for all basin scales approaches a constant value of 0.65. The relative differences show a tremendous effect of initial soil moisture on both peak flow and runoff ratio. For the whole Fella basin, results show a 100% increase in the flood peak and >200% increase in runoff ratio for the highest level of saturation examined. Figure 3.5 (left panel) shows that the relative difference increases with basin scale (for both peak flow and runoff ratio) for the same level of saturation. For example, for Qb=20 m3/s, the relative difference in peak flow is ~12% for Uque (24km2) and increases to ~30% for Moggio (623km2) which suggest that there is a scaling (relative to basin area) behavior in the effect of initial soil moisture conditions on runoff generation. This behavior becomes more distinct as the level of saturation increases (see for example the slope of increase in differences for Qb=4 and Qb=120 m3/s). The different initial soil moisture fields were selected using the baseflow at the outlet of Fella basin as a proxy for the saturation conditions. One can argue that the difference in the level of saturation is not equal for all sub-basins thus interpretation may not be clear regarding the scale dependence. To further investigate this scale dependence, differences in initial soil moisture (expressed as differences in basin-average depth to water table) and differences in the peak runoff and runoff ratio relative to the control simulation were calculated for each basin (Figure 3.6). Positive increase in average depth to water table means increase in the level of saturation relative to the control. The results in Figure 3.6 show that as relative differences in the initial saturation level increases, the

52

relative differences in peak flow and runoff ratio increases as well and that the rate of this increase (slopes of fitted lines for different basins) differs with basin scale (higher for larger basin scales), which indicates that there is in fact a scale dependence in the sensitivity to initial soil moisture conditions.

For example, one can note that an

approximately equal relative difference in initial soil moisture for the Uque (24 km2) and Moggio (623 km2) basins results in a relative difference in peak flow of about 35% and 100%, respectively.

Note that the fitted lines were based on a least-squares linear

regression and were used only for visualization purposes (they are not used to suggest a linear dependence).

Figure 3.6. Relative difference in peak discharge (left) and runoff ratio (right) versus relative difference in average depth to water table (used to express the difference in saturation level) for each basin. Note that fitted lines were based on least-squares linear regression and aim only in aiding visualization of the scale effect (different slopes) on the sensitivity to initial soil moisture conditions. Surface soil moisture state is a key factor in partitioning runoff into surface and subsurface components (Vivoni et al. 2007) and determining the dominant runoff

53

generation mechanism. Thus, different initial wetness conditions are expected to result in different dominant runoff generation mechanisms. To investigate the effect of different initial soil moisture conditions on the runoff generation processes, I used the capability of the hydrologic model to differentiate the runoff generation associated with the different mechanisms. tRIBS partitions runoff into four basic types that include: (i) infiltration excess, (ii) saturation-excess, (iii) perched return flow and (iv) groundwater exfiltration (Ivanov et al. 2004a). The first two types constitute the rapid surface runoff response, while the other two correspond to the slower subsurface mechanisms (Vivoni et al. 2007).

Table 3.3. Runoff contribution of each generation mechanism for different initial soil moisture conditions. Percentage of Total Runoff Volume (%) 3

Qb (m /s)

Infiltration Excess

Saturation Excess

Perched Return Flow

Groundwater Exfiltration

2

62

22

5.4

10.6

4

54

26

6.3

13.7

6

49

28

6.8

16.2

8

45

30

7.2

17.8

10

42

31

7.5

19.5

20

33

34

7.8

25.2

120

9

39

5

47

The contribution of each mechanism to the total runoff volume is presented in Table 3.3 for the different initial wetness conditions. For all cases, except for Qb = 120 m3/s, the runoff production is dominated by the surface mechanisms. Infiltration excess is the dominant mechanism for the control (Qb=2 m3/s) simulation, and as the level of saturation increases there is a transition from infiltration- to saturation-excess with a

54

simultaneous significant increase of the groundwater contribution. For Qb = 120 m3/s, it appears that the soil is close to saturation and that most of the surface runoff was produced due to saturation-excess, while the groundwater contribution is 47% of the overall runoff volume. The sensitivity results to initial soil moisture conditions found in this study, somehow contradict the findings by Castillo et al. (2003) who concluded that the hydrologic response after high intensity low frequency storms is independent of initial soil water content.

However, the same authors mention that the sensitivity depends on the

predominant runoff generation mechanisms and as I showed in the results, the transition from surface to subsurface runoff contribution for increasing saturation level can have a tremendous effect in the runoff generation. Thus for this case it was shown that even for heavy precipitation events, the effect of initial water content can be very significant and a possible explanation may relate to the high soil moisture capacity of the basin (Borga et al. 2007).

3.6 Sensitivity to Rainfall Variability The influence of rainfall variability on the hydrologic simulation of the 2003 Fella flash flood case has been previously investigated by Sangati and Borga (2009). While I generally followed their methodology for the experimental setup, this work differs in the following aspects. First, a hydrologic model of different complexity was used, thus the work presented here could be potentially used as a reference to examine how sensitivity to rainfall variability varies with model complexity. Second, the sensitivity to rainfall

55

variability was examined as a function of antecedent wetness, and third, the effect on the runoff generation mechanisms was investigated. The setup of the sensitivity experiment included the aggregation of the original rainfall fields (0.5 km resolution) to the resolution of 1, 2, 4, 8 and 16 km that were then used to force the hydrologic model. Aggregation of the rainfall fields results in: i) error in basin-averaged rainfall volume and ii) smoothing of the spatial variability which translates to the distortion of the rainfall gradients over the basin and relative to the channel network (Sangati and Borga 2009). Table 3.4 summarizes the total mean areal rainfall amounts for all basins and rainfall resolutions examined. As it is shown, the rainfall aggregation results in reduction of the total basin-averaged rainfall that becomes more significant (for all basins) at coarser resolutions.

Table 3.4. Total basin-averaged rainfall for different aggregation levels. Basin

Area (km2)

Total mean areal rainfall (mm) using resolution 0.5 km

1 km

2 km

4 km

8 km

16 km

Number

Name

1

Uque

24

286.5

287

285

279

285

192

2

Pontebba

165

247

247

246.42

244

239

211

3

Dogna

329

237

237

236.57

235

232

224

4

Moggio

623

189.8

189

188.78

187

183

170

Furthermore, Figure 3.7 demonstrates the distortion of the rainfall field as the aggregation level and subsequently the variability smoothing increases. It is interesting to note how the high rainfall areas are shifting with aggregation level (especially above 4 km resolution) and that the precipitation structure is completely lost at 16 km resolution. The significance of those two effects on the hydrologic response during the flash flood is

56

evaluated by comparing the simulated hydrographs for each rainfall resolution against the control (0.5 km resolution).

Figure 3.7. Hourly radar-rainfall accumulation map at the original spatial resolution (0.5 km) and at 1, 2, 4, 8, 16 km resolutions after aggregation. Data correspond to 17:00 UTC during the flash flood event. Figure 3.8 shows the effect of rainfall aggregation on peak runoff and total runoff volume for the different basin scales. For both metrics and for all basins, the relative difference is below 10% for rainfall resolution up to 4 km, which means that the rainfall field is adequately captured for resolutions up to 4 km. For coarser resolutions (" 8 km),

57

the differences increase significantly (> 25%) and the underestimation in peak discharge exceeds 50% in some cases (Uque and Moggio).

Figure 3.8. Left panel: Peak discharge (top) and total runoff volume (bottom) versus basin scale for different rainfall aggregation resolutions. Right panel: Difference of peak discharge (top) and total runoff volume (bottom) relative to control simulation (0.5km resolution) versus basin scale.

Those results reflect the combined effect of error in rainfall volume and variability smoothing due to aggregation. In order to decompose the two error sources and examine their effect independently, a series of simulations were performed for two scenarios. The

58

first scenario used as rainfall forcing fields the original resolution fields (0.5 km) that were biased based on the calculated basin-averaged bias between original and aggregated fields. Thus, the original rainfall gradients were preserved but the bias resulting from the different aggregation levels (for 4, 8, 16 km) was introduced in order to assess the bias effect alone. Note that the basin-averaged bias was calculated and applied to the original fields at the same time-step (hourly) that was used to force the hydrologic model. The second scenario used as rainfall forcing fields the different resolution fields (4, 8, 16 km), but adjusted them so the basin-averaged rainfall was unbiased. Thus, in this case the basin-averaged rainfall volume was kept constant, but the rainfall gradients were smoothed out in order to assess the effect of rainfall variability alone.

Again, the

adjustment of the rainfall fields for the bias was done at each time-step. As it is shown in Figure 3.9, the effect of bias is dominant for the smallest scale basin and as the basin scale increases the effect of bias and resolution becomes equally important.

For example, for the 16 km resolution, the relative difference in peak

discharge due to bias and resolution effect is approximately 53% and 0.5%, respectively, for the Uque sub-basin, while for the whole Fella basin, it is 30% and 23%. The general pattern of these results suggests that the effect of rainfall volume error decreases with basin area, while the effect of resolution increases with scale. This is expectable behavior since the bias in rainfall decreases with basin scale (see Table 3.4) and rainfall coefficient of variation (not shown) increases. Sangati and Borga (2009) showed that the error in peak discharge and runoff volume increases as the ratio of rainfall resolution to basin area increases. Thus, the largest errors should be reported for the Uque sub-basin and 16 km resolution. It was shown that this is true when you are assessing the effect of bias alone

59

but when you are evaluating the overall effect (bias and variability smoothing), the larger basins can show errors that are approximately equal to the smaller basins even for the coarsest rainfall resolution (see Figure 3.8).

This is explained due to the interplay

between bias and resolution effect with basin scale as mentioned above.

Figure 3.9. Effect of bias in basin-averaged rainfall volume due to aggregation (left) and effect of variability smoothing due to coarsening of resolution (right) expressed as relative differences of peak discharge (top) and runoff volume (bottom) for different basin scales. The smoothing of rainfall variability and the simultaneous decrease of rainfall intensity due to averaging has an effect on the runoff generation mechanisms. Similar to

60

the previous section, this was examined by comparing the relative contribution of each runoff generation process to the total runoff production (Table 3.5). As the resolution becomes coarser (reduced intensity and variability), the contribution of saturation-excess and subsurface mechanisms becomes more significant, but the infiltration-excess remains the dominant mechanism even at the coarsest resolution examined (16 km). Segond et al. (2007) showed, on the basis of numerical experiments in a midsized catchment (1400 km2), that runoff production mechanisms are the dominant source of variability in runoff response.

Table 3.5. Same as Table 3.3 but for different rainfall resolutions.

Rainfall Resolution (km)

Percentage of Total Runoff Volume (%) Infiltration Excess

Saturation Excess

Perched Return Flow

Groundwater Exfiltration

0.5

62

22

5.4

10.6

1

61.4

22.3

5.5

10.8

2

61

22.5

5.6

10.9

4

59.2

23.5

5.8

11.5

8

51.7

26.7

7.2

14.4

16

40

33

8.6

18.4

Thus, these findings have an important practical implication and one should keep in mind that forcing a physically-based model with coarse resolution rainfall fields (i.e. satellite-rainfall) could result in altering the runoff generation processes with a subsequent effect in the simulated hydrologic response.

61

3.7 Combined Sensitivity to Initial Soil Moisture and Rainfall Variability The last part of this sensitivity analysis includes the combined effect of antecedent wetness and rainfall variability. In this case, the simulations were performed for the rainfall aggregation levels used previously and for three different initial conditions, dry (Qb = 2), moderate (Qb = 20) and wet (Qb = 120 m3/s), in order to examine the dependence of the effect of rainfall aggregation to initial soil moisture conditions.

Figure 3.10. Relative difference of peak discharge for different aggregation levels and different initial wetness conditions for each basin examined. Note that white, grey and black symbols correspond to initial wetness conditions of Qb =2, 20 and 120 respectively. 62

Figures 3.10 and 3.11 show the combined effect on peak discharge and total runoff volume, respectively, for each basin. The main finding is that for decreasing wetness the effect of rainfall aggregation increases. Namely, the drier the basin the more significant is the impact of rainfall aggregation.

For resolutions # 4km, the effect of rainfall

aggregation is small and the dependence to initial soil wetness is not shown clearly, but for coarser resolutions (" 8km) the dependence is distinctly shown for all basin scales. This result is in agreement with the work by Shah et al. (1996) who related the role of rainfall heterogeneities to antecedent catchment conditions and found a significant change in runoff only for low initial water content.

Figure 3.11. Same as Fig. 3.10 but for total runoff volume. 63

A point to note is that while the relative differences (for 16 km) in total rainfall for Moggio (623 km2) and Uque (24 km2) are approximately 10% and 33%, respectively, their corresponding difference in peak discharge and runoff volume are very close for the dry initial condition case. This can be explained by the previous results where it was shown that for the case of Uque the difference in peak runoff and volume is dominated by the bias in rainfall, while for Moggio the smoothing of spatial variability had a significant effect as well. As the wetness increases, the difference between the Uque and Moggio runoff generation errors increases, which suggest that the effect of rainfall bias becomes more significant than the spatial variability for increased wetness conditions.

3.8 Conclusions This chapter presented a sensitivity analysis which investigates the effect of initial soil moisture conditions and rainfall variability on runoff generation during a major flash flood event in a mountainous basin in northeastern Italy. A series of different simulation scenarios were carried out using a distributed hydrologic model and the principal conclusions of this study include: 1. Runoff generation showed high sensitivity to initial soil moisture conditions. The peak discharge and runoff ratio for Fella basin increased by 100% and >200%, respectively, for the very wet relative to dry initial conditions.

Furthermore, sensitivity of runoff to initial soil moisture

conditions exhibits scale dependence with the sensitivity increasing with basin scale.

64

2. Modeling results suggest that infiltration-excess was the dominant runoff generation mechanism during the 2003 flash flood. For increasing initial wetness, or for coarser rainfall resolution, the contribution from saturation excess and groundwater becomes significant. 3. Aggregation of rainfall fields at resolution > 8 km resulted in a significant decrease of peak discharge (>30%) and total runoff volume (>25%) for all basin scales. The effect of bias in rainfall volume had a significant effect for all basins examined, while the effect of variability smoothing was important only for the larger scale basins where the rainfall gradients were stronger. 4. Rainfall aggregation effect and initial wetness conditions effect have an apparent dependence.

Results suggest that coarsening of rainfall

resolution has a greater impact on runoff generation for drier soil conditions. 5. Rainfall intensity and variability smoothing due to aggregation results in an increase of saturation-excess and subsurface runoff generation but infiltration excess remains the dominant generation mechanism.

I acknowledge that the results presented here are at a certain degree model dependent as is the case for all simulation-based experiments. Also, the findings are based only on the specific area and single event thus the trends and dependencies shown are related to the basin characteristics (e.g., high soil moisture capacity) and storm structure and one should not generalize the conclusions of this study for other flash flood cases. However,

65

the results point out the significance of the role of initial soil moisture and rainfall variability in runoff generation during flash flood events.

Better initialization of

hydrologic models and more accurate estimation of rainfall (in terms of magnitude and variability) can lead to significant improvement of current flash flood warning systems.

66

CHAPTER 4 UNDERSTANDING THE SCALE RELATIONSHIPS OF UNCERTAINTY PROPAGATION OF SATELLITE RAINFALL THROUGH A DISTRIBUTED HYDROLOGIC MODEL

4.1 Introduction Precipitation is one of the most important components of the hydrological cycle and the driving force for one of the most devastating natural hazards, i.e., floods. Model simulation of the hydrologic processes (e.g. the generation of runoff) at watershed scale is the basis for nowcasting floods and providing information that is essential for the preservation of property and human lives. The primary input to a hydrological model is precipitation. Consequently, the accuracy of the flood prediction is tied to the accuracy of the precipitation estimation. However, estimating rainfall rates at high accuracy and continuously in space and time is an extremely difficult task due to the limitations of current sensor technologies, both in terms of resolution and spatio-temporal coverage, as well as due to uncertainty in the inversion techniques. Traditionally, rain gauges have been used to measure surface rainfall rates. Gauges are considered as the most accurate sensors for measurements over a limited area (nearly a point), but their small coverage (especially over complex terrain and tropical regions) limits the adequacy in representing the spatial structure of highly variable rainfall fields over large spatial scales. Weather radars on the other hand, have advanced precipitation

67

monitoring due to the spatially distributed information these systems can provide, and have created significant potential on the use of radar observations for flood related applications (Tilford, 1987; Garrote, 1992; Pessoa et al., 1993 and Borga et al., 2000, among others). Radar-rainfall estimates also suffer from a number of uncertainties associated with issues in radar calibration, variability in the reflectivity-to-rainfall relationship, vertical reflectivity profile, and atmospheric/rain-path attenuation (more details on the issue can be found in Krajewski and Smith 2002). Furthermore, beam blockage effects due to complex terrain constrain the applicability of radar observations in mountainous areas (mostly prone to flooding), while the establishment of a network of sensors that could resolve the coverage issue is rarely a viable solution due to the high cost of deploying such systems. In recent times, there has been significant development in space-based precipitation estimation that has now opened up new horizons in hydrological applications at global scale. Satellite sensors offer unique advantages comparing to gauges and weather radars because they provide (i) global coverage and (ii) observations in regions where in situ data are inexistent or sparse.

Due to this uniqueness, the use of satellite data for

hydrologic applications has gained growing interest. Guetter et al. (1996) conducted numerical experiments using synthetic satellite-rainfall data forcing a rainfall-runoff model to estimate soil water and streamflow at three large-scale basins (> 2000 km2) in the United States. Tsintikidis et al. (1999) examined the potential use of visible and infrared images to derive mean areal rainfall estimates to force a hydrologic model in the Blue Nile region. Wilk et al. (2006) used passive microwave datasets to derive estimates of the water balance over the Okavango River basin. Su et al. (2008) evaluated the use of

68

TRMM (Tropical Rainfall Measuring Mission) Multisatellite Precipitation Analysis product (3B42) for streamflow simulations in the La Plata Basin. In a similar manner, Collischon et al. (2008) used the 3B42 dataset to estimate daily streamflow in the Amazon basin. Although these studies revealed the potential on the use of satellite-based rainfall estimates for hydrologic applications, they also reported deficiencies that differ in significance depending on the use. The two main sources of those deficiencies are i) the error structure of the satellite-rainfall estimates and ii) the rainfall error propagation through the hydrologic model. For the first source, several studies have been reported that deal with the assessment and the characterization of the retrieval error for a number of global satellite rainfall products (see McCollum et al. 2002, Gebremichael and Krajewski 2004, Ebert et al 2007 and Dinku et al. 2007, among others). Although these studies provide useful information on satellite-rainfall uncertainties, as Hossain and Anagnostou (2006b) pointed out, they focus on the accumulation of rainfall over large spatiotemporal scales as opposed to the flux, involving error statistics that are more relevant to large scale meteorological phenomena. Hence, many of these studies do not provide insight on the smaller-scale surface hydrologic processes such as floods and flash floods, particularly over complex terrain (see for example, Griffith et al., 1978; Arkin and Meisner, 1987; Huffman et al., 2001; Steiner et al., 2003, among others). Hossain and Anagnostou (2006a) and Bellerby and Sun (2005) are some of the examples of recent effort to address this issue of characterizing the satellite error structure at scales relevant to flood processes.

69

On the second error source, the propagation of error in rainfall through hydrologic model is a subject that has long been identified as a critical issue. However, most studies so far have involved the propagation of radar-rainfall error (see for example Borga et al. 2000; Borga 2002; Sharif et al 2004; Vivoni et al. 2007) and only few have investigated the satellite-rainfall error propagation in hydrologic simulations (Hossain and Anagnostou 2004, 2005; Hong et al. 2006, 2007; Nijssen and Lettenmaier, 2004). Evaluating the error propagation of satellite-rainfall through the prism of surface hydrology is a very challenging task because it relates too many factors, which include, among others: i) specifications of the satellite-rainfall product, ii) scale of the basin, iii) spatio-temporal scale of the hydrologic variable of interest, iv) the level of complexity and physical processes represented by the hydrologic model used and v) regional characteristics. As we now stand at the doorstep of a global scale precipitation mission, named

Global

Precipitation

Measurement

(GPM,

http://gpm.gsfc.nasa.gov/),

a

comprehensive investigation/evaluation of the use of current satellite products in smallscale hydrologic applications appears mandatory and can serve as a valuable reference to the mission’s designers as well as highlight its usefulness to society. This chapter, aims in evaluating the scale characteristics of satellite-rainfall error propagation through a distributed hydrologic model, emphasizing on flood simulations over complex terrain. Because the surface hydrologic processes leading to surface runoff generation are controlled by the complexity of the terrain, the use of distributed and physically-based model is a necessary, but often, absent ingredient in literature on error propagation. The general framework of this study follows the work by Hossain and Anagnostou (2004) and Hong et al. (2006), in the sense that an error model is used for

70

generation of ensembles of satellite-rainfall fields, which subsequently propagate through a hydrologic model in order to evaluate the runoff error in a probabilistic manner. The most notable novelty of this work is the higher level of complexity in terms of both the satellite-rainfall error model and hydrologic model used as well as the fact that two satellite products with contrasting space-time scales are examined. Furthermore, the scale dependence of the error propagation is addressed and investigated by comparing the results for a number of basins ranging from 100 km2 to 1200 km2. Results in this study are based on a single flash flood event that took place in a mountainous basin in Northeast Italy. Although it is recognized that results can be driven by the specific characteristic of the analyzed event, hopefully this study will provide a proof-of-concept to trigger further studies involving a wide range of systematic investigations involving hydrologic models of varying complexity, storm cases of different spatio-temporal rainfall variability, basins of varying characteristics and satellite products of varying resolutions and error characteristics. In section 4.2 the study area and the datasets utilized in this study are presented. The methodology for the generation of satellite-rainfall ensembles along with an error analysis is described in section 4.3 and the resulted simulated hydrographs from the propagation of those ensembles is discussed and analyzed in section 4.4. The main results of the error propagation analysis are described in section 4.5 and overall conclusions are summarized in section 4.6.

71

4.2 Study Area and Data The basin (Bacchiglione basin) considered in this study is located in the Veneto region, which is part of the northeastern Italian Alps (Figure 4.1).

An area of

approximately 1200 km2 drains to its outlet just upstream the Montegaldella city with a drainage direction that has a north-south alignment. The variability of the terrain is very distinct, with the upper part (first 30km from the northern boundary of the basin) being highly irregular with elevation that ranges from 200 m to greater than 2000 m, and the mid-lower part being quite flat (elevation around 100m and slopes lower than 0.5 degrees). The vegetation follows in a sense the elevation pattern and differentiates between forested (broad-leaved and conifer) areas in the higher elevation region (northern basin, which is part of the eastern Italian Alps) and the lower elevation (midlower basin) that is predominately covered by croplands. The high precipitation amounts in the area (> 1000 mm annually) along with the very steep irregular terrain (slopes greater than 40 degrees in the highlands) make the region prone to the generation of floods and thus suitable for hydrologic investigations. Several sources of data are utilized in this study. The precipitation data included both in situ measurements (rain gauges) and remote sensing based (radar and satellite) retrievals. The rain gauges, located in the region (see Figure 4.1), provided half-hourly rainfall accumulations that were used for (i) bias adjustment of the radar-rainfall fields (Borga et al. 2000) and (ii) calibration of the satellite-rainfall error model (Hossain et al. 2009). Distributed rainfall maps are obtained from a C-band Doppler weather radar located at Mt. Grande, approximately 10 km southeast of the basin’s outlet, and were available at 1km spatial and 1hr temporal resolution. The algorithm used to generate 72

rainfall estimates from those radar measurements and performance evaluations are described in Borga et al. (2000).

Figure 4.1. Map showing the locations of the Posina and Bacchiglione basins in the northeastern Italian Alps. Note that the thin (4km) and thick (25km) grid shown, provide a visual comparison between the spatial resolution of the satellite products used and the basin scales.

Two different satellite products were used, the TRMM 3B42 version 6 (Huffman et al. 2007) and a dataset obtained from the calibration of high-resolution global IR data from available passive microwave satellite rainfall estimates on the basis of algorithm described in Kidd et al. (2003), hereafter named KIDD. Table 4.1 summarizes the native space-time resolution of each precipitation dataset used in this study. The hydrologic data include stream gauge observations that were available at half hourly scale but only for one sub-basin (called Posina, see Figure 4.1).

73

Table 4.1. Nominal spatial and temporal resolution of the precipitation data used.

Data Type

Remote Sensing In situ

Product Name

Resolution Temporal (hr)

Spatial

3B42

3

0.25x0.25 deg.

KIDD

0.5

~ 4x4 km

Radar

1

1 km

Gauge

0.5

point

Furthermore, soil and land use/cover maps were used to derive the soil and vegetation properties required for the hydrologic model setup.

4.3 Satellite-Rainfall Ensembles For the generation of the satellite-rainfall ensembles I used the satellite-rainfall error model developed by Hossain and Anagnostou (2006a) (hereafter named SREM2D from Satellite Rainfall Error Model-2 Dimensional). SREM2D uses stochastic space-time formulations to characterize the multi-dimensional error structure of satellite retrievals and combines that with input “reference” rain fields (representing the “true” surface rainfall process) of higher accuracy and resolution, in order to simulate probable realizations of satellite-like rainfall estimates (for more details on the error model see Hossain and Anagnostou 2006a).

One of the model prerequisites is the regional

calibration for every satellite product, meaning that for the simulation of each satellite product a different set of parameters needs to be obtained. In this case, the model parameters were calibrated for the study region using six months (June – November, 2002) of gauge and satellite data for the 3B42 and KIDD satellite-rainfall products. The calibration of SREM2D parameters and verification of the predicted variability for the

74

region and satellite products used in this study is described in Hossain et al. (2009). The focus of this chapter is on the analysis of the generated ensembles from the SREM2D error model and their error propagation. The high resolution (1km-1hr) radar-rainfall fields were used as the “reference” to generate realizations for i) 3B42 product at its nominal scale (see Table 4.1), ii) KIDD product at high resolution (4km-1hr, hereafter KIDD-4km) and iii) KIDD product aggregated at coarser space-time resolution (25km-3hr, hereafter KIDD-25km). The reason for aggregating the KIDD product was to compare the error propagation characteristics of (i) the two satellite products at the same resolution (3B42 vs KIDD25km) and (ii) the two resolutions for the same product (KIDD-4km vs KIDD-25km). A total of 100 realizations were generated for each satellite product but due to computational limitations sub-samples of those realizations were used to force the hydrologic model.

This was done in the following procedure.

Ensembles of each

satellite product were ranked based on their overall rainfall bias (compared to the reference field) and realizations were selected starting at the 5th percentile with step increment of 5 percentiles (i.e. 5th, 10th, 15th etc.). Thus, a total of 20 realizations (see Figure 4.2 and 4.3) plus the average of all 100 realizations from each set was used for the error propagation experiment. The results presented in this study are focused on a major flood event that occurred in the study area during October 1996 (started around Oct. 15 at 15:00 CET). The rainfall event that caused the flooding lasted for more than 60 hours and resulted in mean areal rainfall accumulation (based on radar estimates) of 200 mm for the Bacchiglione basin (~1200 km2) and approximately 350 mm (see Figures 4.2 and 4.3) for the mountainous

75

Figure 4.2. Top: Mean areal rainfall accumulation curves for Bacchiglione basin calculated from radar (black) and SREM2D ensembles for 3B42 (left), KIDD-25km (middle) and KIDD-4km (right). Bottom: Bias (left), relative RMSE (middle) and Nash-Sutcliffe score (right) between MAP time series derived from SREM2D ensembles and the reference rainfall (radar). 76

Figure 4.3. Same as Figure 4.2 but for Posina basin.

77

sub-basin of Posina (~116 km2).

In Figures 4.2 and 4.3, I present the mean areal

precipitation (MAP) derived from reference (radar) and from the SREM2D ensembles for the Bacchiglione and Posina basins respectively. MAP was calculated over a 96-hr window that extended few hours before and after the flood-induced storm. Mean areal precipitation for each basin was based on the arithmetic average of all pixels (radar or satellite) that fully or partially overlapped the basin’s area.Due to the high resolution of radar fields (1km) and the relatively large area of the basin’s examined (>100 km, see Table 4.2), areal weighting was not applied for the calculation of the “reference” MAP but only for the satellite-rainfall MAP (for all products). The generated ensembles are compared with the reference rainfall and the results are presented as box plots of bias, relative root mean squared error (rel. RMSE) and Nash-Sutcliffe (N-S) score (see appendix for definition of those metrics).

Table 4.2. List of size and topography slope information of the basins used in this study.

Basin

Area (km2)

1

Slope (degrees) Mean

Std. Dev.

108

5.07

4.8

2

148

12.45

8.85

3

244

11.02

7.37

4

269

9.59

9.03

5

398

13.55

9.27

6

584

15.24

9.89

7

627

14.87

9.83

Posina

116

20.61

9.23

Bacchiglione

1200

10.57

9.82

78

Several points are noted from Figures 4.2 and 4.3. First, all satellite realizations underestimate the total reference rainfall (bias30%) and reduced ability to characterize the “true” process (N-S < 0.5).

4.4 Hydrologic Simulations The hydrologic model used in this study (hereafter named as tRIBS out of triangulated irregular network Real-time Integrated Basin Simulator) is a fully distributed model that can simulate multiple storm events and account for the moisture losses during interstorm periods (see Ivanov et al. 2004a for more details). tRIBS has been applied successfully to several flood related studies (e.g. Vivoni et al. 2006a,b) that demonstrate the ability of the model to represent the hydrologic response during extreme events. One of the major advantages of using tRIBS is its ability to represent the complex terrain with high accuracy while being computationally very efficient (reduced number of computational nodes) by leveraging the triangulated irregular networks scheme (Vivoni et al. 2004, 2005). This attribute is very important when ensemble simulations are considered (e.g., Forman et al., 2008; Mascaro et al., 2009). In this case study I used the model to simulate a single storm-induced flood event (the October 1996 flood). The distributed nature of the model allows retrieving the hydrograph response for several interior nodes of the basin thus providing the ability to compare the error propagation for different scales of drainage area. The error propagation was evaluated for a number of basin scales that ranged between 100 and 1200 km2 (see Figure 4.4 and Table 4.2). The successful setup of this type of model requires detailed information regarding the land descriptors (e.g. land use/cover, soil type etc.) of the simulation domain.

The

Bacchiglione basin was categorized into four major land use/cover classes that include barren land, grass, broad-leaf forests and conifer forests and into three soil classes based on the general pattern of land use/cover. For the parameterization of the soil and land

80

cover variables in the model, different sources found in the literature were used (Jury et al. 1991; http://ldas.gsfc.nasa.gov/).

Figure 4.4. Stream network of Bacchiglione basin and the locations (black dots) of the basin outlets analyzed in this study. Note that the numeric IDs correspond to the IDs presented in Table 4.2. To ensure that the model could describe properly the rainfall-runoff transformation processes of the flood event used in this simulation exercise, a minimal calibration similar with the procedure described in Chapter 2 and Chapter 3, was carried out. The model was forced with the reference (radar) rainfall input and the SCE (shuffle complex evolution) optimization method (Duan et al., 1992) was used in order to minimize the mean squared error between the observed and the simulated hydrograph.

81

The calibration was performed on the flood event used in this study and as mentioned earlier the only available hydrograph measurements for the basin were at the Posina outlet.

Thus, instead of using the model’s output at the outlet of the domain

(Bacchiglione basin), I used the interior simulations that corresponded to the Posina subbasin and compared those to the available observations. Note that the total simulation time was 160 hrs and all the runoff quantities presented herein were calculated over that time window.

Table 4.3. Calibrated parameters of saturated hydraulic conductivity, conductivity decay coefficient and anisotropy ratio for the three soil classes of Bacchglione basin.

Soil Class

Sat. Hydr. Conduct. (mm/hr)

Conduct. Decay Coefficient (x 10-4)

Anisotropy Ratio

A

28

7.75

620

B

22.5

6.56

144

C

29

4.96

183

Table 4.3 shows the final calibrated values of the three parameters and Figure 4.5 presents a comparison of the simulated (based on calibrated parameters) and observed hydrographs. While the simulated hydrograph appears to be more sensitive to rainfall variations than the observed, the general response is realistic and associated with a relative error of ~20% for the peak discharge, which is low relative to the range of runoff simulation errors associated with the herein error propagation experiment. A point to note is that the main objective of the calibration exercise presented here was to ensure that the transformation of rainfall to runoff for the specific event would be as realistic as possible since that would have an effect on the results of the error propagation analysis.

82

Thus, I do not claim that the calibration exercise presented here provides a model suitable for general flood prediction of storm cases in this basin.

Figure 4.5. Top: Observed (black) and simulated (blue) hydrographs for the Posina basin during the Oct. 1996 flood event. Bottom: Mean areal precipitation over Posina basin based on radar-rainfall data. The calibrated hydrologic model was forced with the generated SREM2D ensembles from each satellite rainfall product and the simulated hydrographs for all 9 basins (Table 4.2) were analyzed. Figures 4.6 and 4.7 show the corresponding hydrographs for the rainfall ensembles presented in Figures 4.2 and 4.3, respectively. As expected, the conclusions derived from the comparison between the SREM2D-derived hydrographs

83

and the reference is in agreement with the comparison of the rainfall fields. Again, 3B42 and KIDD-4km behave similarly and significantly better than the KIDD-25km for the large-scale (Bacchiglione) basin, while for the smaller-scale basin (Posina) the hydrologic simulations based on the high-resolution KIDD-4km product outperforms all other products. An interesting point to note is that for the larger basin (Bacchiglione) the 3B42 exhibits higher variability (larger whisker lengths of the boxplots) than the KIDD4km in bias and N-S score, while for the smaller basin (Posina) this is reversed. However, this difference is not that apparent in the variability of rainfall statistics (Figure 4.3), which indicates that rainfall-to-runoff transformation can magnify (or attenuate) the variability in rainfall retrieval error.

The spread of the simulated hydrographs

corresponds to the uncertainty in hydrologic simulations due to uncertainty in rainfall, thus based on the above findings the propagation of rainfall uncertainty depends significantly on basin scale and hydrologic response (i.e. dominant runoff generation mechanisms). This is discussed next.

4.6 Analysis of Error Propagation To investigate the propagation of error in satellite-rainfall through the rainfall-runoff transformation, the error in rainfall versus error in runoff is presented in Figure 4.8 based on the metrics of relative RMSE and relative error (see appendix). A very distinct feature of the results is that the propagation of error exhibits a linear behavior in terms of its relative term. This linearity appears stronger for the case of Bacchiglione basin (1200 km2). Especially for the scatter plot of relative errors in total rainfall versus relative errors in total runoff, the points are aligned very close to the 1-1 line, which indicates that

84

Figure 4.6. Top: Simulated hydrographs based on radar (black) and SREM2D (grey) rainfall ensembles for 3B42 (left), KIDD-25km (middle) and KIDD-4km (right), for Bacchiglione basin. Bottom: Bias (left), rel. RMSE (middle) and N-S score (right) between reference (radar) and SREM2D hydrographs. 85

Figure 4.7. Same as Figure 4.6 but for Posina basin.

86

the relative error in rainfall translates to an equal relative error in runoff. A point to note from Figure 4.8 is that the performance of each satellite product manifests in distinctclusters of the rainfall-runoff error domain (those clusters are separated by different colors associated with different satellite products). This effect is much more profound in the case of Posina, where the points representing the high-resolution satellite product (KIDD-4km; blue color) cluster in a distinct (from the other products) region on the figure that is associated with lower relative error and higher damping effect on the propagation of relative RMSE. For the Bacchiglione basin, the two clusters of KIDD4km and 3B42 mix in the same domain since they perform equally as mentioned before. This strengthens the argument made in the previous section that for the smaller scale basins high-resolution products are critical to moderate the retrieval error propagation in runoff. The propagation of rainfall error to peak runoff is also presented in Figure 4.8. In the case of peak runoff, the propagation has a different effect comparing to the total runoff. Most of the ensembles, for all products and both basins, show magnification of the relative error. For Posina basin, the KIDD-4km product shows high variability with the relative error in the peak discharge ranging between -20% and 50% while the rainfall error is between 20% and 40%. This can be attributed to the highly nonlinear rainfall-torunoff transformation and to the increased spatial variability of the high-resolution product, which, I speculate, can have a significant effect in the runoff production (especially for the highly complex terrain of Posina). Similar analysis was carried out for all basins presented in Table 4.2, to further investigate the dependency of error propagation with basin scale. Results are presented in Figure 4.9 in terms of the ratio of the error metric (relative error and relative RMSE) in runoff over the corresponding error metric in rainfall versus basin scale. Ratios equal to

87

Figure 4.8. Error propagation metrics: Top panel shows the relative error in total runoff versus relative error in total rainfall for Bacchiglione (left) and Posina (right) basin. Middle panel shows relative RMSE in discharge versus relative RMSE in rainfall, and bottom panel shows relative error in peak runoff versus relative error in total rainfall respectively. Errors were calculated between SREM2D ensembles and the reference (radar). Note that blue black and red triangle corresponds to the ensemble average of KIDD 4km, KIDD 25km and 3B42 respectively.

88

one indicate that statistics of the error in rainfall would translate to an equal statistical measure of the error in runoff, while ratios lower (higher) than one would indicate that the error dampens (magnifies) through the rainfall-runoff transformation process. An increase of ratio with catchment area up to 600 km2 and an approximate plateau for larger basins is consistent for all ratios presented in Figure 4.9. These results clearly states the scale dependence of the propagation of error and indicates that for smaller scale basins (< 600 km2), the damping effect of the error is greater than in larger basins. Moreover the variability among products (scatter of solid circles) as well as the variability within each product (length of error bars) is higher for smaller scale basins ( 1.2 mm/hr

0.32

0.70

0.77

0.69

2.6 for rain> 3 mm/hr

PERSIANN 0.08

0.85

0.41

0.85

4.9 for rain> 0 mm/hr

Bias

Correl. Coeff.

3B42

0.22

CMORPH

Although these results are based on a single event, and do not provide statistical significance, they provide nevertheless an example of how satellite estimates perform over complex terrain for an extreme storm case that caused a flash flood and point out the severe underestimation of all products. These findings are in agreement with more

100

extensive evaluation studies of satellite precipitation products that have shown a) underestimation of PERSIANN and CMORPH (particularly for heavy precipitation events) over a mountainous region in Ethiopia (Bitew and Gebremichael, 2009) and b) that PERSIANN significantly underestimate precipitation in high elevation areas (Hirpa et al., 2010). On the other hand, the high correlation between radar and satellite rainfall estimates and the consistent underestimation of all satellite products suggest that a bias adjustment of the satellite fields could potentially improve the representation of watershed mean areal rainfall estimates. In order to derive the mean field bias of each dataset, radar pixels were averaged to match the spatial resolution and orientation of each satellite data grid. This was done for the whole radar domain (120km radius, Figure 5.1) and for each radar field (at both hourly and three hourly scales). The distribution of all rainy pixels from satellite and the corresponding averaged radar fields were compared by calculating and plotting the quantiles of their rainfall distribution (Figure 5.3). As it is shown in Figure 5.3, quantiles between radar and all satellite products exhibit a strong linear dependence with correlation coefficients > 0.9 for all products. A distinct feature from Figure 5.3 is that there is an obvious shift in the slope from the 1-1 line (slope= 1) which is considered to be an estimate of the bias of each satellite product. This slope (or adjustment factor) was calculated based on linear regression (Figure 5.3) between radar and satellite quantiles and the values are summarized in Table 5.2. Note that for PERSIANN the slope was calculated based on the entire spectrum of rainfall values while for the case of CMORPH and 3B42, the slope was calculated for satellite-rainfall values > 3 mm/hr and

101

1.2 mm/hr respectively because there was no apparent bias with the radar below these thresholds.

Figure 5.3. Radar-rainfall quantiles vs satellite-rainfall quantiles for CMORPH (left), PERSIANN (middle) and 3B42 (right). Solid line corresponds to the slope =1 line and dash line corresponds to the linear regression applied.

Figure 5.4. Same as Figure 5.2 but adjusted based on the mean-field bias factor derived from quantile-quantile plots.

102

The calculated slopes (Table 5.2) were used as a multiplicative factor to adjust the original satellite datasets. Obviously, the adjustment for each product was made based on the same threshold that was used to derive the slope. The adjusted basin-averaged rainfall and the comparative statistics are presented in Figure 5.4 and Table 5.2.Correlation was preserved and the total rainfall bias was improved for all products. PERSIANN and 3B42 gave similar bias (0.41 and 0.45 respectively), while CMORPH gave significantly better bias (0.77).

However, underestimation of rainfall remains

significant for all products thus the mean field bias representing larger domain averages could not explain completely the observed bias of basin-averaged precipitation over Fella basin. The “unexplained” part of the discrepancies can possibly be attributed to i) resolution effects ii) algorithmic deficiencies and iii) orographic enhancement of precipitation.

5.4. Hydrologic Simulations The hydrologic response of the Fella river basin during the flash flood event was simulated using the TIN (triangulated irregular network)-based Real-time Integrated Basin Simulator (tRIBS) model (Ivanov et al., 2004,a,b; Vivoni et al., 2007). tRIBS is a distributed physics-based model that explicitly accounts for the spatial variability of land surface descriptors (terrain, soil, vegetation), soil moisture and atmospheric forcing. A detailed description regarding the characteristics, setup, calibration and validation of the model for the Fella basin is provided in section 3.4 of Chapter 3. The satellite-rainfall datasets (CMORPH, PERSIANN and 3B42) before and after the adjustment were used to force the hydrologic model and the resulted hydrographs at the outlet of the basin were compared against the “reference” hydrograph that was based on radar-rainfall input (Figure 5.5). As it is shown, the error in rainfall has a tremendous impact on the 103

simulated hydrographs for all cases. The bias in total runoff, the relative error in peak discharge and error in time to peak are summarized in Table 5.3 for all cases. Before the bias adjustment of the satellite rainfall, there is no apparent peak in the hydrographs (relative error !97% for all products) and even after the adjustment, only the CMORPHbased simulation produces a clear signal of a flood hydrograph, which however underestimates peak discharge with a relative error of 85%.

Figure 5.5. Observed and simulated hydrographs at the outlet of Fella basin based on radar (reference) and satellite rainfall input before and after the bias adjustment.

104

Bias in rainfall (see Table 5.2) magnifies as it propagates through the nonlinear transformation of rainfall to runoff and results in bias of total runoff that exceeds an order of magnitude for all products (Table 5.3).

Table 5.3. Error metrics based on the comparison of radar simulated and satellite simulated hydrographs before and after the mean field bias adjustment. Note that the simulated hydrographs were based on the radar-calibrated parameters. Before Adjustment Data

After Adjustment

Bias

Rel. Error in Peak Disharge

Error in Time to Peak

+9

0.12

-95%

+5.5

-97%

+7.5

0.26

-85%

+4

-98%

+13

0.13

-94%

+5

Bias

Rel. Error in Peak Disharge

Error in Time to Peak

3B42

0.07

-98%

CMORPH

0.09

PERSIANN

0.05

Furthermore, in all cases hydrographs show a significant delay in time to peak (!4 hrs). Previous work on satellite-based hydrologic simulations of streamflow have shown that better results can be obtained when hydrologic models are calibrated based on satellite-rainfall input (Hughes 2006; Artan et al., 2007).

To investigate whether

recalibration could potentially improve the results, the hydrologic model was calibrated for each satellite product and for both original (before adjustment) and adjusted fields. Figure 5.6 presents the simulated hydrographs for all cases and Table 5.4 summarizes the calculated error metrics. For the case of using the original satellite input, CMORPH improves dramatically the results with only 5% underestimation in peak relative error and outperforms all other products. 3B42 reduced the relative error to 46% but PERSIANN continues to severely underestimate (-81%). Although the timing of the simulated peak discharge improved for all cases, it continues to have a delay of the order of 2 hrs for 3B42 and CMORPH and 3.5 hrs for PERSIANN (Table 5.4). After bias adjustment and

105

recalibration all products perform extremely well, with most surprising the case of PERSIANN, which besides the improvement in the error metrics results in a hydrograph that captures extremely well the rising and falling limp of the reference hydrograph.

Table 5.4. Same as Table 5.3 but for parameters calibrated separately for each satellite product before and after the adjustment. Before Adjustment Data

After Adjustment

Bias

Rel. Error in Peak Disharge

Error in Time to Peak

+2

1.44

-2.6%

+0.5

-5%

+2

1.24

-3.6%

+2.5

-81%

+3.5

1.06

3.0%

+1

Bias

Rel. Error in Peak Disharge

Error in Time to Peak

3B42

0.82

-46%

CMORPH

1.25

PERSIANN

0.27

Figure 5.6. Simulated hydrographs based on individual calibration of radar input and each satellite product before (left) and after (right) bias adjustment.

106

Although a first look on these results gives a very optimistic message about the use of satellite products, on a second thought it is puzzling how recalibration managed to match so closely the reference hydrograph when the rainfall input is underestimated by approximately 30-60 %. That by itself should tell a lot about the calibration and the parameter values used in hydrologic models. A close look at the calculated bias in total runoff for the case of adjusted products (Table 5.3) shows an overestimation (bias > 1) which means that while less rainfall is used as input, more runoff is produced than in the reference case. This is unrealistic and the explanation for this relies on the recalibrated parameter values (Table 5.5).

As it is shown, an impermeable soil (hydraulic

conductivity close to zero) was practically created in order to increase runoff production and match the reference hydrograph. Essentially the bias in rainfall was compensated in the calibration parameters thus soil hydraulic conductivity was forced to become zero (in almost all cases). The parameters lose completely their physical meaning and even if the simulated hydrographs appear satisfactory in this case, the simulation results are not expected to be consistent under different conditions.

Table 5.5. Values for saturated hydraulic conductivity obtained after individual calibration for each product. Note that the values shown correspond to the dominant soil class that occupies over 80% of the basin area.

Data

Sat. hydraulic conduct. (mm/hr) Before Adjustment

After Adjustment

3B42

0.00

0.02

CMORPH

0.00

5.76

PERSIANN

0.00

0.30

Radar

19.00

107

Figure 5.7. Same as Figure 5.6 (right) but for fully saturated initial soil moisture conditions. To demonstrate this argument the simulations were repeated with same calibrated parameters of the adjusted satellite-rainfall input but for different initial soil moisture conditions. Specifically, the hydrologic response was investigated under the scenario that initial wetness conditions over the whole basin are fully saturated. This was done within the model by setting the level of groundwater table equal to the soil surface. The results (Figure 5.7) suggest exactly what was speculated. While in Figure 5.5 the hydrographs have a similar behavior, once the initial conditions changed, the non-realistic parameterization could not consistently represent the radar-derived hydrologic response. These findings do not intent to counter the argument that recalibration based on satellite input can improve satellite-based simulation but point out the fact that recalibration should be done under physical considerations.

108

5.5. Conclusions This study investigated the potential use of high resolution satellite precipitation products for simulating a major flash flood event over a complex terrain basin in Northeastern Italy. Three satellite products, 3B42, CMORPH and PERSIANN were compared against radar-rainfall observations for the storm event that caused the flooding. Results showed that satellite precipitation estimates suffer from large bias but were able to represent the temporal variability of basin-averaged precipitation. The mean field bias derived from the comparison of the whole radar domain was not able to explain the observed bias over the study basin. This indicates that other sources related to i) spatial variability in the retrieval algorithm uncertainty, ii) the orographic enhancement of precipitation not accounted in satellite estimates and iii) resolution effects, introduce variability in the error not captured by the mean-field bias. The hydrologic simulations forced with satellite-rainfall input could not capture the basin’s hydrologic response during the flash flood event. Even after adjusting satellite rainfall fields for the mean-field bias, the systematic error in rainfall remained significant and was magnified severely in terms of bias in runoff parameters (volume, peak runoff, time to peak). Recalibration of model parameters using precipitation derived from each satellite product showed that can account for this bias and improve the modeling results. However, this should be treated with caution because it was also shown that to match the reference hydrograph, parameter values were forced to take unrealistic values thus losing their physical meaning and consequently the ability to represent the hydrological processes in the basin.

109

CHAPTER 6 CONCLUSIONS AND FUTURE RESEARCH

The work presented in this dissertation aims at advancing our understanding regarding the i) evolution and controls of runoff response and mechanisms during flash floods and ii) issues on the potential use of satellite precipitation observations for flood prediction. A hydrologic investigation/validation framework that incorporated in-situ and remote sensing observations complemented with modeling tools was established in order to address these complex issues.

The main findings and principal conclusions are

summarized below.

6.1 Conclusions Summary Chapter 2 involved the application of a physically-based distributed hydrologic model for the simulation of two contrasting flood events (an intense flood and a flash flood) that occurred over a small-size (116 km2) mountainous basin in Northeastern Italian Alps. Results showed that with only minimal calibration of a few key parameters, the model could successfully represent the runoff response at the outlet of the basin. Comparison of the runoff observations during both events revealed the complexity of rainfall-to-runoff transformation attributed to the differences in initial soil moisture conditions and spatiotemporal distribution of rainfall.

However, simulation results suggested that

despite the distinct differences of the two floods, the same runoff mechanisms dominated in both cases. Furthermore, it was shown that the occurrence of each runoff generation

110

mechanism exhibited a specific pattern in relation to the topography, which was consistent in both flood cases. These findings indicate that in the examined basin, the occurrence of each runoff generation mechanism is controlled mainly by topography and the other land surface descriptors. In Chapter 3, the effect of initial soil moisture conditions and rainfall variability on runoff generation is examined for a major flash flood event that occurred over a mediumsize basin (623 km2) in Northeastern Italian Alps. Results revealed that despite the general belief, runoff response during the flash flood showed high sensitivity to initial soil moisture conditions. Furthermore the sensitivity exhibited scale dependence with the sensitivity increasing with basin size. Aggregation of rainfall fields to scales greater than 8 km had a significant impact on the simulation of both peak discharge and total runoff. The effect of i) bias in rainfall volume and ii) rainfall variability smoothing due to aggregation was examined separately and the results showed that bias in rainfall volume had a significant effect for all basin scales examined, while the effect of variability smoothing was important only for the larger scale basins (> 300 km2) where the rainfall gradients were stronger. The combined effect of rainfall aggregation and initial soil moisture conditions indicated that coarsening of rainfall resolution had a greater impact on runoff generation for drier soil conditions. In terms of the dominant runoff generation mechanisms, the intensity and variability smoothing due to aggregation of rainfall fields resulted in an increase of subsurface runoff generation although in all cases the contribution from surface runoff processes was the dominant. A probabilistic evaluation of the propagation of satellite-rainfall error through a distributed hydrologic model was investigated in Chapter 4. The analysis was based on

111

two satellite products at different resolutions and retrieval accuracies and a number of basins that ranged in scale (100-1200 km2). Results showed that mean areal precipitation is consistently underestimated by satellite realizations. Bias in satellite-rainfall varied significantly depending on the product and the basin size, which clearly indicates that the performance of a given product relates to both its resolution and the scale of application. Furthermore, the ability of satellite fields to represent the variability of rainfall was low, which is a critical limitation for flash flood simulations, since as it was shown in the results of Chapter 3, error in representation of the distribution of rainfall in space and time can alter significantly the hydrologic response of the basin. Based on the analysis of the simulated hydrographs, the highest resolution rainfall estimates outperformed the coarser resolution estimates in all cases, which demonstrate the importance of highresolution satellite precipitation products for flood related applications. Arguably, the propagation of error from rainfall-to-runoff exhibits a linear behavior and a definitive dependence to basin scale.

Results revealed that the ability of dampening the

precipitation error reduces as the basin scale increases and approaches a plateau for basins larger than 600 km2. The work presented in Chapter 5 extended the simulation-based exercise of Chapter 4 using actual high resolution satellite precipitation products associated with the major flash flood event that was studied in Chapter 3. Three satellite products with different resolution and algorithmic characteristics were used to force a hydrologic model. Results showed that all satellite precipitation estimates suffered from large bias but in contrast with the results of Chapter 4, they were able to represent the temporal variability of basin-averaged precipitation for the flash flood examined. The mean field bias derived

112

from the comparison with the reference radar data could not explain the observed bias in the satellite products over the study basin. Results indicate that possibly other sources related to i) spatial variability in the retrieval algorithm uncertainty, ii) the orographic enhancement of precipitation not accounted in satellite estimates and iii) resolution effects, introduce variability in the error not captured by a mean-field bias parameter. Due to the severe underestimation of rainfall amounts, even after adjusting for the mean field bias, the hydrologic simulations forced with satellite-rainfall input could not capture the basin’s hydrologic response during the flash flood event. The systematic error in rainfall over the study basin was magnified severely in terms of bias in runoff parameters (volume, peak runoff, time to peak).

Recalibration of model parameters using

precipitation derived from each satellite product showed that can account for this bias and improve the modeling results. However, this should be treated with caution because it was also shown that to match the reference hydrograph, model parameters were given unrealistic values thus losing their physical meaning and consequently the ability to represent the hydrological processes in the basin under varying conditions.

6.2 Future Research Directions One of the key issues of this thesis was the application of a physically-based distributed hydrologic model for simulating flash floods. Results showed that this type of models can successfully represent the hydrologic response during flash flood event and thus provide a promising tool to help us improve our understanding of the physical mechanisms that govern runoff generation. Future assessment of the robustness of DHM

113

for operational use, should involve continuous mode simulations in order to investigate the ability of these models to represent soil moisture dynamics during inter-storm periods. Soil moisture was shown in Chapter 3 to be very important for accurately simulating flash floods. While present work was focused on few distinct flash flood cases and over a specific region; future steps to further advance this investigation should involve other major flash flood events over different regions. Dependencies of the runoff generation mechanisms (dominance and spatial occurrence) should be examined relative to climatic conditions (humid vs arid ) and land surface characteristics (soil, vegetation cover, etc.) to gain a better understanding of the role of each factor on the hydrologic processes during a flash flood. As an example, a currently available database from an EU-funded project on flash floods (named HYDRATE, http://www.hydrate.tesaf.unipd.it/ ) provides an excellent opportunity for researchers to address the issues highlighted above for a range of basin scales and climatic regions in Europe. The other main component of this dissertation was the evaluation of the potential use of satellite precipitation estimates for simulating flash floods. As stated previously, the findings presented in this dissertation can be considered as a proof-of-concept regarding the use of satellite-rainfall for complex terrain flood simulations. While the conclusions derived can depend to a certain degree on the i) models used ii) storm morphology and iii) basin characteristics, they highlight some key factors that impede the direct integration of satellite QPE with hydrologic models for flood related applications. One definite message from this work is that there is a need for substantial improvement in the satellite rainfall retrievals. This can be done either on the retrieval algorithm level or through better characterization/correction of systematic errors in satellite QPE.

A

114

potential direction could be the dynamic adjustment (gauge or radar-based bias adjustment) of satellite-rainfall estimates at high temporal scales (i.e. 3 hourly) rather than monthly that is currently done in certain products (e.g. NASA’s 3B42). Perhaps the idea of using satellite precipitation estimates for monitoring flash floods is a stretch for the current capabilities of satellite QPE and perhaps the vision of developing a “Global Flood Warning System” based on satellite observations sounds like total fiction but what is research if not to explore the unexplored and imagine the unimaginable.

115

APPENDIX

The error metrics presented in this dissertation are defined as follows:

where XRef (t) and XS (t) corresponds to the reference and simulated variable of interest (e.g. rainfall, discharge etc) respectively, at time step t. T is the total number of time steps in the time series analyzed. Note that in the case of peak runoff, bias is defined as the ratio of maximum values of XRef and XS respectively. Tpo and Tps correspond to the reference and simulated time to peak respectively. Note that throughout the results presented in this thesis the term “reference” can relate to either observations or radarbased simulations.

116

BIBLIOGRAPHY

Arkin, P.A., and B.N. Meisner, 1987: The relationship between large-scale convective rainfall and cold cloud cover Western Hemisphere during 1982-1984. Monthly Weather Review, 115, 51-74. Artan, G., H. Gadain, J.L. Smith, K. Asante, C.J. Bandaragoda and J.P. Verdin, 2007: Adequacy of satellite derived rainfall data for streamflow modeling. Nat. Hazards. 43, 167-185. Barret, E.C., and D.W. Martin, 1981: The use of satellite data in rainfall monitoring. Academic Press, London. Bell, V.A., and R.J. Moore, 2000: The sensitivity of catchment runoff models to rainfall data at different spatial scales. Hydrology ans Earth System Science, 4(4), 653-667. Bellerby, T.J., and J. Sun, 2005: Probabilistic and Ensemble Representations of the Uncertainty in an IR/MW Satellite Precipitation Product. J. Hydrometeor., 6, 10321044. Beven, K.J., and G.M. Hornberger, 1982: Assessing the effect of spatial pattern of precipitation in modeling stream flow hydrographs. Water Resources Bulletin, 18(5), 823-829. Bitew, M.M., and M. Gebremichael, 2009: Evaluation Through Independent Measurements: Complex Terrain and Humid Tropical Region in Ethiopia. In: Satellite Applications for Surface Hydrology, eds. Gebremichael and Hossain, Springer-Verlag: NY. Borga, M., 2002: Accuracy of radar rainfall estimates for streamflow simulation. J. Hydrol., 267, 26-39. Borga, M., E.N. Anagnostou, and E. Frank, 2000: On the use of real-time radar rainfall estimates for flood prediction in mountainous basins, J. Geophys. Res. -Atmos., 105(D2), 2269-2280. Borga, M, P. Boscolo, F. Zanon, and M. Sangati, 2007: Hydrometeorological Analysis of the 29 August 2003 Flash Flood in the Eastern Italian Alps. J. Hydrometeor., 8(5), 1049-1067. Borneman, R., 1988: Satellite rainfall estimating program of the NOAA/NESDIS Satellite Analysis Branch. Natl. Wea. Dig., 13(2), 7-15. Borrell, V. E., D. Dartus, and M. Alquier, 2007: Forecasting flash floods in ungauged basins with satellite data. Proceeding of the PUB Kick-off meeting held in Brasilia, 20-22 November 2002. IAHS Publ. 309. Braud, I., P.C. Fernández, and F. Bouraoui, 1999: Study of the rainfall–runoff process in the Andes region using a continuous distributed model. J. Hydrology, 216, 155–171.

117

Brooks, R.H., and A.T. Corey, 1964: Hydraulic properties of porous media. Hydrol. Pap. 3, Colo. State Univ., Fort Collins. Carpenter, T. M., J. A. Sperfslage, K.P. Georgakakos, T. Sweeney, and D.L. Fread, 1999: National threshold runoff estimation utilizing GIS in support of operational flash flood warning systems, J. Hydrol., 224, 21–44. Castillo, V.M., A. Gomez-Plaza, and M. Martinez-Mena, 2003: The role of antecedent soil water content in the runoff response of semiarid catchments: a simulation approach. J. Hydrol., 284, 114-130. Ceballos, A., and S. Schnabel, 1998: Hydrological behavior of a small catchment in the dehesa landuse system (Extremadura, SW Spain), J. Hydrol., 210, 146-160. Cerdá, A., 1997: Seasonal changes of the infiltration rates in a Mediterranean scrubland on limestone. J. Hydrol., 198, 209-225. Collischonn, B., W. Collischonn , and C.E.M. Tucci, 2008: Daily hydrological modeling in the Amazon basin using TRMM rainfall estimates. J. Hydrol., 360, 207-216. Creutin, J.D., and M. Borga, 2003: Radar hydrology modifies the monitoring of flashflood hazard. Hydrol. Process. 17, 1453-1456. Cucchi, F., S. Piano, E. Marinetti, G. Massari, S. Oberti, and L. Zini, 2000: Studies for the realization of the hydrological map of Friuli-Venezia Giulia. Ipogea, 3, 57-71. Dawdy, D.R., and J.M. Bergmann, 1969: Effect of rainfall variability on streamflow simulations. Water Resour. Res., 5, 958-966. Delrieu, G., and Coauthors, 2005. The Catastrophic Flash-Flood Event of 8-9 September 2002 in the Gard Region, France: A First Case Study for the Cèvennes-Vivarais Mediterranean Hydrometeorological Observatory. J. Hydrometeor., 6, 34-52. Desilets, SLE, T.P.A. Ferrè, and B. Ekwurzel, 2008: Flash flood dynamics and composition in a semiarid mountain watershed. Water Resour. Res., 44. doi: 10.1029/2007WR006159 Dinku, T., P. Ceccato, E. Grover-Kopec, M. Lemma, S.J. Connor, and C.F. Ropelewski, 2007: Validation of satellite rainfall products over East Africa's complex topography. International Journal of Remote Sensing, 28:7, 1503-1526. Duan, Q., S. Sorooshian, and V.K. Gupta, 1992: Effective and Efficient Global Optimization for Conceptual Rainfall-runoff Models. Water Resour. Res., 28(4), 1015-1031. Ebert, E.E., J.E. Janowiak, and C. Kidd, 2007: Comparison of near-real-time precipitation estimates from satellite observations and numerical models. Bull. Amer. Meteor. Soc., 88, 47–64. Fitzjohn, C., J.L. Ternan, and A.G. Williams, 1998: Soil moisture variability in a semiarid gully catchment: implications for runoff and erosion control. Catena, 32, 55-70.

118

Forman, B.A., E.R. Vivoni, and S.A. Margulis, 2008: Evaluation of Ensemble-based Distributed Hydrologic Model Response with Disaggregated Precipitation Products. Water Resour. Res. 44, doi:10.1029/2008WR006983. Garrote, L. 1992: Real-time modeling of river basin response using radar-generated rainfall maps and a distributed hydrologic database. MS thesis. Department of Civil and Environmental Engineering, Massachusetts Institute of Technology, pp 501. Garrote, L., and R.L. Bras, 1995: A Distributed Model for Real-time Flood Forecasting using Digital Elevation Models. J. Hydrol., 167. 279-306. Gaume, E., M. Livet, and M. Desbordes, 2003: Study of the hydrological processes during the Avène River extraordinary flood (south of France): 6–7 October 1997. Phys. Chem. Earth, 28, 263–267. Gaume, E., M. Livet, M. Desbordes, and J.P. Villeneuve, 2004: Hydrological analysis of the River Aude flash-flood on 12 and 13 November 1999. J. Hydrol., 286, 135–154. Gaume, E., and Coauthors, 2009. A compilation of data on European flash floods. J. Hydrol., 367, 70-78. doi: 10.1016/j.jhydrol.2008.12.028. Gebremichael, M. and W.F. Krajewski, 2004: Characterization of the temporal sampling error in space-time-averaged rainfall estimates from satellites. J. Geophys. Res., 109(D11), doi: D11110 0.1029/2004JD004509. Giannoni, F., J.A. Smith, Y. Zhang, and G. Roth, 2003: Hydrologic modeling of extreme floods using radar rainfall estimates. Adv. Water Resour., 26, 195-203. Giertz, S., and B. Diekkruger, 2003: Analysis of the hydrological processes in a small headwater catchment in Benin (West Africa). Phys. Chem. Earth, 28, 1333–1341. Goodrich, D.C., T.O. Keefer, C.L. Unkrich, M.H. Nichols, H.B. Osborn, J.J. Stone, and J.R. Smith, 2008. Long-term precipitation database, Walnut Gulch Experimental Watershed, Arizona, United States. Water Resour. Res. 44, doi:10.1029/2006WR005782. Griffith, C.G., W.L. Woodley, and P.G. Grube, 1978: Rain estimation from geosynchronous satellite imagery-visible and infrared studies. Mon. Wea. Rev., 106, 1153-1171. Grimes, D. I. F., and M. Diop, 2003: Satellite-based rainfall estimation for river flow forecasting in Africa. I: Rainfall estimates and hydrological forecasts. Hydrol. Sci. J., 48, 567–584. Gruntfest, E., and J. Handmer, 2001: Dealing with flash floods: Contemporary issues and future possibilities. In Coping With Flash Floods (NATO Science Partnership Subseries: 2), Gruntfest E, Handmer JW (eds). Kluwer Acad., Boston, Mass; 3-10. Guetter, A.K., K.P. Georgakakos, and A.A. Tsonis, 1996: Hydrologic applications of satellite data: 2. Flow simulation and soil water estimates. J. Geophys. Res., 101, 26527-26538.

119

Hino, M., Y. Odaka, K. Nadaoka, and A. Sato, 1988: Effect of initial soil moisture content on the vertical infiltration process – a guide to the problem of runoff-ratio and loss. J. Hydrol., 102, 267-284. Hirpa, F.A., M. Gebremichael, and T. Hopson, 2009: Evaluation of High Resolution Satellite Precipitation Products over Very Complex Terrain in Ethiopia. J. Appl. Meteor. Climatol., (in press). Hong, Y., K.-L Hsu, S. Sorooshian, and X. Gao, 2004: Precipitation Estimation from Remotely Sensed Imagery Using an Artificial Neural Network Cloud Classification System. J. Appl. Meteor., 43, 1834-1852. Hong, Y., K.-L. Hsu, H. Moradkhani, and S. Sorooshian, 2006: Uncertainty quantification of satellite precipitation estimation and Monte Carlo assessment of the error propagation into hydrologic response. Water Resour. Res., 42, doi:10.1029/2005WR004398 Hong, Y., R.F. Adler, F. Hossain, S. Curtis and G.J. Huffman, 2007: A first approach to global runoff simulation using satellite rainfall estimation. Water Resour. Res., 43, doi:10.1029/2006WR005739 Hossain, F., and E. N. Anagnostou, 2004: Assessment of current passive microwave and infrared based satellite rainfall remote sensing for flood prediction. J. Geophys. Res. Atmos., 109, doi: 10.1029/2003JD003986. Hossain, F., and E. N. Anagnostou, 2005: Numerical investigation of the impact of uncertainties in satellite rainfall estimation and land surface model parameters on simulation of soil moisture. Adv. Water Res., 28, 1336-1350. Hossain, F., and E.N. Anagnostou, 2006a: A two-dimensional satellite rainfall error model. IEEE Trans. Geosci. Remote Sens., 44(4), 1511-1522. Hossain, F., and E.N. Anagnostou, 2006b: Assessment of a Multidimensional Satellite Rainfall Error Model for Ensemble Generation of Satellite Rainfall Data. IEEE Trans. Geosci. Remote Sens. Letters, 3, 419-423. Hossain, F., and D. P. Lettenmaier, 2006: Flood prediction in the future: Recognizing hydrologic issues in anticipation of the Global Precipitation Measurement mission, Water Resour. Res., 42, doi:10.1029/2006WR005202. Hossain, F., L. Tang, E.N. Anagnostou and E.I. Nikolopoulos, 2009: A Practical Guide to a Space-Time Stochastic Error Model for Simulation of High Resolution Satellite Rainfall Data, In: Satellite Applications for Surface Hydrology, eds. Gebremichael and Hossain, Springer-Verlag: NY. Huet, P., X. Martin, J.L. Prime, P. Foin, C. Laurain, P. Cannard, 2003 : Retour d’expèriences des crues de septembre 2002 dans les dèpartements du Gard, de l’Hèrault, du Vaucluse, des bouches du Rhône, de l’Ardèche et de la Drôme. In Inspection gènèrale de l’Environnement. Paris, France. 124. Huffman, G.J., R.F. Adler, M.M. Morrissey, D.T. Bolvin, S. Curtis, R. Joyce, B. McGavock, and J. Susskind, 2001: Global Precipitation at One-Degree Daily Resolution from Multisatellite Observations. J. Hydrometeor., 2, 36–50.

120

Huffman, G.J., and Coauthors, 2007: The TRMM Multi-satellite Precipitation Analysis: Quasi-Global, Multi-Year, Combined-Sensor Precipitation Estimates at Fine Scale. J. Hydrometeor., 8, 38-55 . Hughes, D.A. 2006: Comparison of satellite rainfall data with observations from gauging station networks. J. Hydrol., 327(3-4), 399-410. Ivanov, V.Y., E.R. Vivoni, R.L. Bras, and D. Entekhabi, 2004a: Cathcment hydrologic response with a fully distributed triangulated irregular network model. Water Resour. Res., 40(11),doi: 10.1029/2004WR003218. Ivanov, V.Y., E.R. Vivoni, R.L. Bras, and D. Entekhabi, 2004b: Preserving highresolution surface and rainfall data in operational-scale basin hydrology: a fullydistributed physically-based approach. J. Hydrol., 298(1-4), 80-111. Janjic, Z.I., 1994. The step-mountain Eta coordinate model: Further developments of the convection, viscous sublayer and turbulence closure schemes. Mon. Wea. Rev., 122, 927-945. Joyce, R. J., J. E. Janowiak, P. A. Arkin, and P. Xie, 2004: CMORPH: A method that produces global precipitation estimates from passive microwave and infrared data at high spatial and temporal resolution. J. Hydromet., 5, 487-503. Jury, W.A., G.R. Wilforrd, and G.H. Walter, 1991: Soil Physics. John Wiley & Sons: New York. Karnieli, A., and J. Ben-Asher, 1993. A daily runoff simulation in semiarid watersheds based on soil water deficit. J. Hydrol., 149, 9-25. Keefer, T.O., M.S. Moran, and G.B. Paige, 2008 : Long-term meteorological and soil hydrology database, Walnut Gulch Experimental Watershed, Arizona, United States. Water Resour. Res., 44, doi:10.1029/2006WR005702. Kidd, C., D.R. Kniveton, M.C. Todd, and T.J. Bellerby, 2003: Satellite Rainfall Estimation Using Combined Passive Microwave and Infrared Algorithms. J. Hydromet., 4, 1088-1104. Koren, V.I., B.D. Finnerty, J.C. Schaake, M.B. Smith, D.-J. Seo, and Q.-Y. Duan, 1999: Scale dependencies of hydrologic models to spatial variability of precipitation. J. Hydrol., 217, 285-302. Krajewski, W.F., and J.A. Smith, 2002: Radar Hydrology: rainfall estimation. Adv. Water Resour., 25, 1387-1394. Kuligowski, R.J., 1997: An overview of National Weather Service quantitative precipitation estimates (QPE). Techniques Development Laboratory Office Note 974, 27 pp. La Barbera, P., L. Lanza, and F. Siccardi, 1993: Flash flood forecasting based on multisensor information. Proceedings of the Yokohama Symposium on Extreme Hydrological Events: Precipitation, Floods and Droughts. IAHS Publ. no. 213. Loague, K., 1992: Soil water content at R-5. Part 2. Impact of antecedent conditions on rainfall-runoff simulations. J. Hydrol., 139, 253-261. 121

Martina, M. L. V., E. Todini, and A. Libralon, 2005: A Bayesian decision approach to rainfall thresholds based flood warning. Hydrol. Earth Syst. Sci. Discuss., 2, 2663– 2706. Martina, M.L.V., and D. Entekhabi, 2006: Identification of runoff generation spatial distribution using conventional hydrologic gauge time series. Water Resour. Res., 42, doi:10.1029/2005WR004783. Mascaro, G., E.R. Vivoni, and R. Deidda, 2009: Implications of Ensemble Quantitative Precipitation Forecast Errors on Distributed Streamflow Response. J. Hydrometeor., 11(1), 69-86. McCollum J.R., W.F. Krajewski, R.R. Ferraro, and M.B. Ba, 2002: Evaluation of biases of satellite rainfall estimation algorithms over the continental United States. J. Appl. Meteor., 41(11), 1065-1080. Mesinger, F., Z.I. Janjic, S. Nickovic, D. Gavrilov, and D.G. Deaven, 1988 : The steepmountain coordinate: Model description and performance for cases of Alpine lee cyclogenesis and for a case of an Appalachian redevelopment. Mon. Wea. Rew., 116, 1493-1518. Nicótina, L., E.A. Celegon, A. Rinaldo, and M. Marani, 2008 : On the impact of rainfall patterns on the hydrologic response. Water Resour. Res., 44, doi:10.1029/2007WR006654. Nijssen, B., and D.P. Lettenmaier, 2004: Effect of precipitation sampling error on simulated hydrological fluxes and states: Anticipating the Global Precipitation Measurement satellites. J. Geophys. Res.-Atmos., 109(D2), Art. No. D02103, 0.1029/2003JD003497 Norbiato, D., M. Borga, S.D. Esposti, E. Gaume, and S. Anquetin, 2008a: Flash flood warning based on rainfall thresholds and soil moisture conditions: An assessment for gauged and ungauged basins. J. Hydrol., 362, 274-290. Norbiato, D., 2008b: Regional Analysis of Flooding and Flash Flooding. Ph.D. thesis, Department of Land and Agroforest Environment, University of Padova, Padova, Italy, pp 42. Norbiato, D., M. Borga, R. Merz, G. Bloschl, and A. Carton, 2009: Controls on event runoff coefficients in the eastern Italian Alps. J. Hydrol., 375, 312-325. Noto, L.V., V.Y. Ivanov, R.L. Bras, and E.R. Vivoni, 2008: Effects of initialization on the response of a fully-distributed hydrologic model. J. Hydrol., 352 (1-2), 107-125. Ogden, F.L., and P.Y. Julien, 1993 : Runoff sensitivity to temporal and spatial rainfall variability at runoff plane and small basin scales. Water Resour. Res., 29(8), 25892597. Ogden, F.L., H.O. Sharif, S.U.S. Senarath, J.A. Smith, M.L. Baeck, and J.R. Richardson, 2000: Hydrologic analysis of the Fort Collins, Colorado flash flood of 1997. J. Hydrol., 228, 82–100.

122

Papadopoulos, A., P. Katsafados, G. Kallos, and S. Nickovic, 2002: The weather forecasting system for POSEIDON-An overview. Global Atmosphere and Ocean System, 8 (2-3), 219-237. Pessoa, M. L., R.L. Bras and E.R. Williams, 1993: Use of Weather Radar for Flood Forecasting in the Sieve River Basin: A Sensitivity Analysis. J. Appl. Meteor., 32, 462-475. Petty, G. and W.F. Krajewski, 1996: Satellite estimation of precipitation over land, Hydrol. Sci. J., 41(4), 433-451. Pielke, J.R.A., M.W. Downton, J.Z.B Miller, 2002: Flood Damage in the United States, 1926 – 2000: A Reanalysis of National Weather Service Estimates, UCAR, Boulder, Colo. Sangati, M., and M. Borga, 2009: Influence of rainfall spatial resolution on flash flood modelling. Natural Hazards and Earth System Sciences, 9, 575-584. Sangati, M., M. Borga, D. Rabuffetti, and R. Bechini, 2009: Influence of rainfall and soil properties spatial aggregation on extreme flash flood response modelling: An evaluation based on the Sesia river basin, North Western Italy. Adv. Water Resour., 32(7), 1090-1106. Sapiano, M. R. P., and P. A. Arkin, 2009: An intercomparison and validation of high resolution satellite precipitation estimates with 3-hourly gauge data. J. Hydrometeor., 10(1), 149-166. Saulnier, G.-M., and M. Le Lay, 2009 : Sensitivity of flash-flood simulations on the volume, the intensity, and the localization of rainfall in the Cévennes-Vivarais region (France). Water Resour. Res., 45, doi: 10.1029/2008WR006906. Scofield, R. A., 1987: The NESDIS operational convective precipitation estimation technique. Mon. Wea. Rev., 115, 1773–1792. Scofield, R. A., and V. J. Oliver, 1977: A scheme for estimating convective rainfall from satellite imagery. NOAA Tech. Memo. NESS 86, pp 47. Scofield, R.A., and R.J. Kuligowski, 2003: Status and Outlook of Operational Satellite Precipitation Algorithms for Extreme-Precipitation Events. Weather and Forecasting, 18, 1037–1051. Segond, M.-L., H.S. Wheater, and C. Onof, 2007: The significance of spatial rainfall representation for flood runoff estimation: A numerical evaluation based on the Lee catchment, UK. J. Hydrol., 347, 116-131. Shah, S., P. O’Connel, and J. Hosking, 1996: Modeling the effects of spatial variability in rainfall on catchment response. 2. Experiments with distributed and lumped models. J. Hydrol., 175, 89-111. Sharif, H.O., F.L. Ogden, W.F. Krajewski, and M. Xue, 2004: Statistical analysis of radar-rainfall error propagation. J. Hydrometeor., 5(1), 199-212. Singh, V.P., 1997: Effect of spatial and temporal variability in rainfall and watershed characteristics on stream flow hydrograph. Hydrol. Process., 11, 1649-1669. 123

Smith, E. A., and coauthors, 2007: International Global Precipitation Measurement (GPM) program and mission: An overview. In: Measuring precipitation from space – EURAINSAT and the future. V. Levizzani, P. Bauer, and F. J. Turk, Eds., Springer, 611-653. Steiner, M., T.L. Bell, Y. Zhang, and E.F. Wood, 2003: Comparison of two methods for estimating the sampling-related uncertainty of satellite rainfall averages based on a large radar dataset. J. Climate, 16, 3759-3778. Stone, J.J., M.H. Nichols, D.C. Goodrich, and J. Buono, 2008: Long-term runoff database, Walnut Gulch Experimental Watershed, Arizona, United States. Water Resour. Res., 44. doi:10.1029/2006WR005733. Su, F., Y. Hong, and D.P. Lettenmaier, 2008: Evaluation of TRMM Multisatellite Precipitation Analysis (TMPA) and its Utility in Hydrologic Prediction in the La Plata Basin. J. Hydrometeor., 9, 622-640. Tilford, K.A., 1987: Real-time flood forecasting using low intensity resolution radar rainfall data. M.Sc. thesis, Department of Civil Engineering, University of Birmingham, England, pp 67. Tropeano, D., L. Turconi, and S. Sanna, 2004: Debris flow triggered by the 29 August 2003 cloudburst in Val Canale, eastern Italian Alps. Proceedings International Symposium INTERPRAEVENT 2004, Riva del Garda, Italy, 121-132. Tsintikidis, D., K.P. Georgakakos, G.A. Artan, and A.A. Tsonis, 1999: A feasibility study on mean areal rainfall estimation and hydrologic response in the Blue Nile region using METEOSAT images. J. Hydrol., 221, 97-116. Vivoni, E.R., V.Y. Ivanov, R.L. Bras, and D. Entekhabi. 2004: Generation of triangulated irregular networks based on hydrological similarity. ASCE J. Hydrol. Eng., 9, 288– 303. Vivoni, E.R., V.Y. Ivanov, R.L. Bras, and D. Entekhabi, 2005: On the effects of triangulated terrain resolution on distributed hydrologic model response. Hydrol. Process., 19(11), 2101-2122. Vivoni, E.R., D. Entekhabi, R.L. Bras, V.Y. Ivanov, M.P.Van Horne, C. Grassotti, and R.N. Hoffman, 2006a: Extending the Predictability of Hydrometeorological Flood Events using Radar Rainfall Nowcasting. J. Hydrometeor., 7 (4), 660-677. Vivoni, E.R., R.S. Bowman, R.L. Wyckoff, R.T. Jakubowski, and K.E. Richards, 2006b: Analysis of a Monsoon Flood Event in an Ephemeral Tributary and Its Downstream Hydrologic Effects. Water Resour. Res., 42 (3), doi: 10.1029/2005WR004036. Vivoni, E.R., D. Entekhabi, and R.N. Hoffman, 2007: Error propagation of Radar Rainfall Nowcasting Fields through a Fully-Distributed Flood Forecasting Model. J. Appl. Meteor. Climatol., 46(6), 932-940. Vivoni, E.R.,F. Di Benedetto, S. Grimaldi, and E.A.B Eltahir, 2008: Hypsometric Control on Surface and Subsurface Runoff. Water Resour. Res., 44, doi:10.1029/2008WR006931.

124

Vivoni, E.R., K. Tai, D.J. Gochis, 2009. Effects of Initial Soil Moisture on Rainfall Generation and Subsequent Hydrologic Response during the North American Monsoon. J. Hydrometeor., 10(3), 644-664. Wilk, J., D. Kniveton, L. Andersson, R. Layberry, M.C. Todd, D. Hughes, S. Ringrose, and C. Vanderpost, 2006: Estimating rainfall and water balance over the Okavango River Basin For hydrological applications. J. Hydrol., 331, 18-29. Wood, E.F., M. Sivapalan, and K. Beven, 1990: Similarity and scale in catchment storm response. Review of Geophysics, 28, 1-18. Woods, R.A., and M. Sivapalan, 1999: A synthesis of space-time variability in storm response: Rainfall, runoff generation and routing. Water Resour. Res.,35, 2469-2485. Zehe, E., R. Becker, A. Bárdossy, and E. Plate, 2005: Uncertainty of simulated catchment runoff response in the presence of threshold processes: Role of initial soil moisture and precipitation. J. Hydrol., 315, 183-202. Zhang, Y., J.A. Smith, and M.L. Baeck, 2001: The hydrology and hydrometeorology of extreme floods in the Great Plains of Eastern Nebraska. Adv. Water Resour. 24: 10371049.

125