Remote sensing, model and in-situ data fusion for ...

9 786062 307332

Remote sensing, model and in-situ data fusion for snowpack parameters and related hazards in a climate change perspective

ISBN 978-606-23-0733-2

Ministerul Cercetării şi Inovării

Dr. Gheorghe Stăncălie Coordinator

Remote sensing, model and in-situ data fusion for snowpack parameters and related hazards in a climate change perspective

Editura PRINTECH BUCUREŞTI 2017

Editura PRINTECH Tipar executat la: S.C. ANDOR TIPO S.R.L. – Editura PRINTECH Site: www.andortipo.ro; www.printech.ro Adresa: Str. Tunari nr.11, Sector 2, Bucureşti Tel./Fax: 021.211.37.12; 021.212.49.51 E-mail: [email protected]

Descrierea CIP a Bibliotecii Naţionale a României Remote sensing, model and in-situ data fusion for snowpack parameters and related hazards in a climate change perspective / coordinator: dr. Gheorghe Stăncălie. – Bucureşti : Printech, 2017 Conţine bibliografie ISBN 978-606-23-0733-2 I. Stăncălie, Gheorghe (coord.) 55

Editor: Anișoara Irimescu Cover photo: M-tii Fagaras, Transfagarasan Photographer: Denis Mihailescu

 Copyright 2017 Toate drepturile prezentei ediţii sunt rezervate autorilor. Nicio parte din această lucrare nu poate fi reprodusă, stocată sau transmisă indiferent prin ce formă, fără acordul prealabil scris al autorilor

This book has been realised within the SnowBall project („Remote sensing, model and in-situ data fusion for snowpack parameters and related hazards in a climate change perspective”). The research leading to these results has received funding from European Economic Area (EEA) Financial Mecanism 2009 – 2014, Research within Priority Sectors under the project contract no. 19SEE/2014.

Consortium structure: Coordinator: The National Meteorological Administration, Bucharest, Romania Partner 1: Norwegian Computing Center, Oslo, Norway Partner 2: Technical University of Civil Engineering, Groundwater Engineering Research Center, Bucharest, Romania Partner 3: National Institute of Hidrology and Water Management, Bucharest, Romania Partner 4: West University of Timisoara, Department of Geography, Timişoara, Romania

SnowBall

CONTENT INTRODUCTION...................................................................................

7

CHAPTER 1. Development and operation of a modular automated station for in-situ snow parameters measurements

9

Andrei Diamandi, Cătălin Dumitrache, Oana Nicola, Eduard Luca, Denis Mihăilescu

CHAPTER 2. Single and multi-sensor remote sensing of snow wetness

23

Rune Solberg, Arnt-Børre Salberg, Øivind Due Trier, Øystein Rudjord, Gheorghe Stăncălie, Anişoara Irimescu, Andrei Diamandi, Argentina Nerţan, Simona Catană, Vasile Crăciunescu

CHAPTER 3. Climate change impact on snow-related processes

63

Roxana Bojariu, Ciprian Corbuş, Rodica Mic, Marius Mătreaţă, Vasile Crăciunescu, Narcisa Milian, Alexandru Dumitrescu, Marius-Victor Bîrsan, Sorin-Ionuț Dascălu, Mădălina Gothard, Liliana Velea, Roxana Cică, Cristian Lucian Grecu, Adrian Alin Paşol

CHAPTER 4. Estimation of snowmelt infiltration

91

Dragoş-Ştefan Găitănaru, Roxana-Gabriela Dobre, Radu Gogu

CHAPTER 5. Data fusion methodology for estimating the snow water equivalent, using snow model simulations, ground observations and satellite products

115

Marius Mătreaţă, Simona Mătreaţă, Bogdan Agiu

CHAPTER 6. Avalanche detection in remote sensing images

135

Arnt-Børre Salberg, Florina Ardelean, Marcel Török-Oance

CHAPTER 7. Snow avalanche inventory and hazard assessment in Fagaras Mountains, Southern Carpathians

147

Marcel Török-Oance, Florina Ardelean, Mircea Voiculescu, Arnt-Borre Salberg, Narcisa Milian

CONCLUSIONS.....................................................................................

5

163

SnowBall

6

SnowBall

INTRODUCTION The book presents the main results achieved during the SnowBall project („Remote sensing, model and in-situ data fusion for snowpack parameters and related hazards in a climate change perspective”). The research leading to these results has received funding from European Economic Area (EEA) Financial Mecanism 2009 - 2014 under the project contract no 19SEE/2014. The "Research within Priority Sectors" main goals are to (1) reduce the economic and social disparities, (2) strengthen the bilateral relations through increased research cooperation between scientific communities in Romania, Norway, Iceland and Liechtenstein, and (3) foster long-term cooperation through the building of partnerships between the research entities. The project promoter was the National Meteorological Administration of Romania and the project partners were from Norway: Norsk Regnesentral Stiftelse (Norwegian Computing Center), and from Romania: the Technical University of Civil Engineering Bucharest – Groundwater Engineering Research Center in Bucharest, the National Institute of Hydrology and Water Management in Bucharest, the West University of Timisoara – Faculty of Chemistry, Biology and Geography in Timisoara. The SnowBall project addresses a problem of national interest, namely the accurate and timely knowledge of the seasonal snow distribution and characteristics. The socio-economic impact of snow is significant, ranging from water management and hydropower, to agriculture, transport, tourism, urbanism and emergency situations management. The monitoring of ice and snow is vital for the management of natural resources, extreme events prediction such as snowmelt induced floods, snow avalanches and for analysing the impact of global warming. Thus, it is important to measure the snow state parameters, such as Snow Cover Extent Area (SCE) and Snow Water Equivalent (SWE) which are useful inputs for hydrological models that may assist in the decision making process of water resources management. Furthermore, other parameters like Snow Depth (SD), Snow Density (SDE) and Snow Wetness (SW) are useful inputs for the 7

SnowBall modelling of avalanche hazards. Moreover, knowledge of the snow water content is vital for flash flood forecasting during fast rising air temperature events. Snowmelt during the winter-spring period represents an important opportunity for aquifer recharge as the volume of water released can be many times more than from an individual rain event. In Romania, most of the water used for domestic, industrial, commercial and agricultural purposes is surface water. In the perspective of climate change, surface water resources are becoming more vulnerable and groundwater needs special attention. Considering this aspect, spatial and temporal distribution correlated with a better knowledge of the snowpack parameters may enhance and improve the groundwater resources assessment. Satellite observations are the only efficient means to deliver precise and up-to-date information on snow, glaciers, lake ice and river ice with appropriate spatial coverage. The most important novelty in the remote sensing of snow is due to the launch of the new Sentinel satellites series from 2014, developed for the specific needs of the EU Copernicus programme. The Sentinels are the first operational Earth Observation (EO) satellites in Europe without an entirely meteorological focus. Sentinel-1 (SAR) and Sentinel-3 (moderate resolution optical) are of particular interest for frequent snow monitoring. The satellites open up for new approaches, in particular with respect to multi-sensor and multi-temporal applications. The papers highlight the innovative approaches developed within the SnowBall project: data fusion of satellite products, in-situ observations and modelling monitoring and risk assessment of climate related changes in rapid snowmelt floods, avalanche statistics and snow contribution to aquifers.

8

SnowBall

CHAPTER 1 Development and operation of a modular automated station for in-situ snow parameters measurements Andrei DIAMANDI 1, Cătălin DUMITRACHE, Oana NICOLA, Eduard LUCA, Denis MIHĂILESCU National Meteorological Administration, Bucharest, Romania 1. Introduction Measuring geophysical parameters from space, associated with features and phenomena at the surface of the Earth has a considerable number of advantages over the classical ground based observations in terms of area extent, spatial scales, access to places otherwise difficult – if not impossible to reach, etc. In-situ measurements originate from either automated stations expensive to establish and operate or data collection campaigns – costly and logistically difficult to conduct (especially in remote areas). Remote sensing techniques do not directly measure physical quantities but try to derive them from the electromagnetic radiation emitted/reflected by the objects and captured at the satellite instrument(s). A crucial factor is the availability of accurate ground measurements used to calibrate the remotely sensed data and validate the derived geophysical quantities of interest. That’s why in-situ observations are also called ground-truth measurements. The observation networks operated by national weather services are rather sparse, while calibration/validation of satellite data and products requires a higher spatial density over selected areas, but also beter temporal sampling and new parameters to be measured. Professional snow automatic stations are expensive – they are designed and build to operate in To whom correspondence should be addressed: Andrei Diamandi, National Meteorological Administration, 97 Şos. Bucureşti-Ploieşti, 013686, Bucharest, Romania, e-mail: [email protected] 1

9

SnowBall harsh conditions and to last long while cheap, amateur weather stations, can neither stand difficult winter conditions, nor can they accommodate the type of sensors required for a snow station. The decision to use the existing know how at the National Meteorological Administration of Romania and previous experience to design and build a low cost automated snow stations customised to the needs of the SnowBall project, came in naturally. 2. Station Design In-situ snowpack parameters measurements in the SnowBall project, are serving two purposes: satellite data calibration/validation and validation of satellite derived snow parameters over the test zones. Two sites (Targu Secuiesc and Joseni) have been selected for cal/val data collection while 7 sites in the project test areas have been selected to receive automated snow stations. The design phase started with the collection of the key functional, usability, technical, environmental, support and interaction requirements for both types of stations (cal/val and snow stations). The requirements were further integrated in a Product Requirements Document (PRD), to provide guidelines along the development and production phases. The general requirements are summarised below: • Open source hardware/software; • Power independent (off grid) operation, low power consumption; • Datalogger: software configurable, interfaces for all sensors available; • Data transmission: GPRS; • Measure (physical quantity): snow surface temperature, snow temperature (profile), soil temperature (profile), air temperature, snow depth, snow extent; • Fully automated operation; • Data sampling rate: configurable; • Short/long term data storage; • Environmental: -30 C to + 20 C. Table 1 list the physical quantities to be measured at the snow and cal/val stations and data collection requirements.

10

SnowBall Table 1: List of parameters to be measured, sampling rates and communication; (*) Air temperature measured at weather station is used.

Nr. 1 2 3 4 5 6 7

Parameter

Unit Sampling Rate o Snow surface temperature C Hourly o Snow temperature C Hourly Snow depth m Hourly Snow moisture % Hourly Snow cover extent N/A Hourly o Air temperature C Hourly o Soil temperature C Hourly

Comm

Cal/Val Station Daily Y Daily Y Daily Y Daily Y Hourly Y Daily Y* Daily Y

Snow Station N Y Y Y N Y* N

During the design phase, it quickly became evident that a modular approach where one could combine elements from the same set of hardware and software items to build both types - cal/val and snow stations have the potential of saving time and money while simplifying production, maintenance and deployment. The engineering team decided to start from the proven design of the mobile soil moisture station developed in the ASSIMO project (http://assimo.meteoromania.ro). The same type of data logger and power module (solar panel, battery, charge controller) could be used with some hardware modifications. The software modules for GSM communication, data storage, power management and the snow moisture probe have been reused as well. Therefore, the team could focus on integrating the sensors with the data logger and the mechanical design. The number of probes to be deployed at each cal/val site and the measurement setup led to a split design, i.e. the probes have been split between two dataloggers according to the requirements for each type of measurement. Thus, we have proposed for each cal/val site, two stations instead of one: Type 1 and Type 2. Type 1 would be getting the IR snow surface temperature probe, the soil temperature sensors (6) and one snow depth sensor, while Type 2 would be getting the snow moisture probes, the snow temperature sensors (5) and one snow depth sensor. The snow stations would be equipped with two snow depth, two snow moisture and five snow temperature probes which could share the same data

11

SnowBall logger, power module and mechanical setup. Therefore, there was only one type of snow station to be built. The block schemes of the cal/val stations Type 1 and Type 2 are shown in Figure 1 and Figure 2 and the mechanical setup for both Type 1 and Type 2 is shown in Figure 3. The block scheme and the mechanical setup for the snow station are shown in Figure 4 and Figure 5 respectively.

Figure 1: Cal/Val station Type 1 block scheme.

Figure 2: Cal/Val station Type 2 block scheme. 12

SnowBall

Figure 3: Mechanical set up Cal/Val Station Type 1 (right) and Type 2 (left).

Figure 4: Snow station block scheme.

13

SnowBall

Figure 5: Snow Station mechanical setup.

3. Data conversion and correction The only parameter not measured directly is the snow moisture. There are no known methods of directly measuring the liquid water content of the snowpack, which could also be automated. One of the indirect methods, which could be easily automated, is the measurement of the dielectric permittivity of the snow, which is afterwards converted in volumetric water content (mc/mc) using Topp’s equation (1). 𝑉𝑉𝑉𝑉𝑉𝑉 = 4.3 × 10−6 𝜀𝜀𝑎𝑎3 − 5.5 × 10−4 𝜀𝜀𝑎𝑎2 + 2.92 × 10−2 𝜀𝜀𝑎𝑎 − 5.3 × 10−2 (1)

Frequency Domain Reflectometry has been used with very good results to measure soil water volumetric content. The dielectric probes can be easily calibrated for different types of soils and the calibration coefficients are available but there are no known calibration coefficients for snow. Since a full fledge calibration study was out of the scope of this study, we thougth that the snow moisture data collected in Romania using both a Denoth-meter and a 5TM sensor and Norway could be a good start. With the kind help of the Nowegian team, who volunteered to collect snow moisture data using their Denoth instrument, and our 5TM instrument during the field campaign in Norway, we got a longer series of measurements which would eventually got us the calibration coefficients.

14

SnowBall The temperature correction of the speed of sound with temperature has been applied for the snow depth measurements using the ultrasound probes since the SR50A calculates a distance reading using the speed of sound at 0°C (331.4 m/s). If the temperature compensation formula (2) is not applied, the distance values will not be accurate for temperatures other than 0°C. (2) 4. Prototype construction & testing Three prototypes have been built, one of each Type 1 and Type 2 cal/val stations and one snow station, arround the same hardware datalogger and power module. The software is customised at the sensor level for each type of station with the code for power management and data communication being the same for all types. The main components of the datalogger are shown in Table 2. The prototypes have undergone extensive lab and outdoor testing, helping to fix software bugs and eliminate faulty parts and sensors. The main modifications of the initial design concerned the wiring of the probes to the datalogger. Table 2: Main components of the station.

Item Micro-controller Communications Storage Clock Power

Type Arduino compatible GSM shield SD shield RTC module 1. Solar panel 2. Battery 3. Charge controller

Component Olimexino 328 GSM shield Sparkfun DS 1307 1. XXX 2. YYY 3. ZZZ

Table 3 lists the sensors used for the measurement of each parameter, their type and interface. The snow moisture is measured indirectly, using a capacitance probe and therefore the permittivity data is converted afterwards in % volume of water, using the Topp equation.

15

SnowBall Table 3: Sensors types and interface.

Nr. 1 2 3 4 5 6 7

Parameter Snow surface temperature Snow temperature Snow depth Snow dielectric permitivity Snow cover extent Air temperature Soil temperature

Sensor type IR Digital Ultrasonic Capacitance CCTV Digital Digital

Unit o C o C m F/m N/A o C o C

Sensor MLX90614 DS18B20 SR50A 5TM AVM837 RHT03 DS18B20

Interface I2C I2C SDI-12 SDI-12 Ethernet 1-Wire I2C

5. Deployment The cal/val stations have been deployed at the weather station locations of Targu Secuiesc and Joseni, each site receiving one Type 1 and one Type 2 stations. Additionally, CCTV cameras have been installed on existing masts at 8 meters in Targu Secuiesc and at 10 meters in Joseni, to monitor the snow cover during daylight. The cameras are uploading hourly the still images to the SnowBall server. The seven snow stations have been deployed at the weather stations locations in Sinaia 1500, Predeal, Vf.Omu, Curtea de Arges, Balea Lac, Salvamont Valea Argesului, and Fundata (Figure 6).

Figure 6: Snow (blue) and cal/val (yellow) snow station locations. 16

SnowBall Pictures of the Type 1 and Type 2 cal/val stations deployed at Joseni site are shown in Figure 7, and of the snow stations deployed at Predeal – weather station and Valea Argesului – rescue cottage sites are shown in Figure 8.

Figure 7: Type 1 and Type 2 cal/val station at the Joseni Site.

Figure 8: The snow stations deployed at: Predeal – weather station (left) and Valea Argesului – rescue cottage (right). 17

SnowBall The snow temperature probes, are placed at 5 heights, evenly distributed, starting from the yearly maximum snow depth average. The first of the two snow wetness sensors is mounted at 15 cm from the ground and the second at a distance from the ground equal to half of the yearly maximum snow depth average. The soil temperature sensors are placed at the same depths at both cal/val sites. Example of sensor setup at the Targu Secuiesc cal/val site is given in Table 4. Table 4: Example of sensors setup at the Targu Secuiesc cal/val site.

No. Parameter 1 Snow depth 2 Snow temperature (profile) 3 Snow surface temperature 4 Snow wetness 5 Soil temperature (profile) 6 Snow cover

Sensor/ instrument 2 x ultrasonic 5 x digital thermometer 1 x infra red thermometer 2 x moisture probe 6 x digital thermometer 1 x CCTV camera

Setup At 2m above ground At +10, +20, +30, +40, +50 cm At 2 m height, 90o FOV At +15 and +25 cm At -5, -10, -15, -20, 25 cm Mast, at 8m height

Unit cm °C °C mc/mc °C N/A

6. Operation & data collection Data collection from both call/val and snow stations is fully automated. The stations are uploading to the SnowBall server, the data read from all the sensors during the last 24 hrs, next day, at 00:30. Several Python scripts are checking and pre-processing the data received (e.g. the correction of the sound speed with temperature for the ultrasonic snow depth sensor). It is at this stage that the snow moisture is calculated from the permittivity measured using the Topp equation. Further on, other scripts are concatenating the last 24 hrs data received to the existing timeseries for every sensor and station. Last but not least, daily plots of all parameters are generated for the entire timeseries for the last 7 days. Figures 9 to 12 show plots of some of the parameters measured at the cal/val sites in Joseni and Targu Secuiesc and Figure 13 a picture of the snow cover at the Joseni cal/val site taken with the CCTV camera.

18

SnowBall

Figure 9: Profiles of snow suface temperature and air temperature at 2m recorded by automatic station at Joseni cal/val site.

Figure 10: Profiles of snow surface temperature, air temperature at 2 m and snow depth recorded by automatic station at Joseni cal/val site.

19

SnowBall

Figure 11: Profiles of snow temperature at 10 cm and air temperature recorded by automatic station at cal/val site in Targu Secuiesc.

Figure 12: Profiles of soil temperature at -10 cm and air temperature recorded by automatic station at cal/val site in Targu Secuiesc.

20

SnowBall

Figure 13: Picture of snow cover at the Joseni cal/val site taken with the CCTV camera.

7. Summary and Conclusions Two types of cal/val stations have been designed and built in two exemplaries each and one type of snow automatic station of which seven pieces have been manufactured for the data collection needs of the SnowBall project. The stations have been deployed at the two selected cal/val sites and seven sites on the test zone. The fully automated operation of the stations and pre-processing of the data collected had significantely reduced the time spent on daily routine tasks. The costs (including manpower and deployment) have been reduced by a factor of 5 for a similar performance and accuracy that could be obtained from a professional grade in-situ snow measuring station. At the same time, building the stations in house offered more flexibility in choosing the sensors.

21

SnowBall References SnowBall Project Plan. Campbell SR50A sonic ranger sensor manual https://s.campbellsci.com/documents/cr/manuals/sr50a.pdf. Decagon 5TM soil moisture sensor manual http://manuals.decagon.com/Manuals/13441_5TM_Web.pdf. MLX90614 IR sensor manual https://www.melexis.com/en/product/mlx90614/digital-plug-playinfrared-thermometer-to-can. Olimexino 328 microcontroller manual https://www.olimex.com/Products/Duino/AVR/OLIMEXINO328/resources/OLIMEXINO-328_manual.pdf. DS18B20 temperature sensor https://www.maximintegrated.com/en/products/analog/sensors-andsensor-interface/DS18B20.html.

22

SnowBall

CHAPTER 2 Single and multi-sensor remote sensing of snow wetness Rune SOLBERG 1, Øystein RUDJORD, Arnt-Børre SALBERG, Øivind Due TRIER Norwegian Computing Center, Oslo, Norway Gheorghe STĂNCĂLIE, Anişoara IRIMESCU, Andrei DIAMANDI, Argentina NERŢAN, Simona CATANĂ, Vasile CRĂCIUNESCU National Meteorological Administration, Bucharest, Romania 1. Introduction Snow is a key element of the water cycle. Seasonal snow is characterized by high temporal variability. The variations at daily-to-seasonal time scales are superimposed to long-term trends in the cryosphere, which have been observed during the last decades and are attributed to climate change (Lemke et al., 2007; Serreze et al., 2000). Satellite sensors are the optimum tools for cryosphere monitoring. Accurate observations of snow properties and state are of great interest for hydrology, meteorology and climate change research and applications. The overall objective of the EEA Grants SnowBall project, which carried out the work presented here, is to explore and develop methodology supporting the vision of a future service providing Romanian national authorities with hind-cast and real-time snow and avalanche information from earth observation data. Project work includes development of algorithms and implementation of a prototype snow monitoring system using Sentinel-1 and 3 satellite data for snow surface wetness mapping.

To whom correspondence should be addressed: Rune Solberg, Norwegian Computing Center, Gaustadalléen 23A, P.O. Box 114 Blindem, NO-0314, Oslo, Norway, e-mail: [email protected] 1

23

SnowBall This paper presents the algorithm development and preliminary validation results from the test sites in Romania for the 2015 snow season. The validation was limited to comparison with air temperature as this was what was available for this season. The primary satellite data sources for the algorithms are Sentinel-3 for optical and Sentinel-1 for synthetic aperture radar (SAR). Since Sentinel-3 will only deliver data operationally from early 2017, Terra MODIS data have been used for 2014-2015 and 2015-2016 winter seasons for the algorithm development and validation. The Sentinel-3 data will be used for 2016-2017 winter season for algorithm validation for both Romania and Norway. Sentinel-3A, launched 16 February 2016, has two optical instruments. The Ocean Land Colour Instrument (OLCI) is based on the heritage from ENVISAT's Medium Resolution Imaging Spectrometer (MERIS) instrument. The OLCI operates across 21 wavelength bands from ultraviolet to near-infrared and uses optimized pointing to reduce the effects of sun glint. The swath width is 1270 km, and the spatial resolution 300 m. The Sea Land Surface Temperature Radiometer (SLSTR) is based on the heritage from ENVISAT's Advanced Along-Track Scanning Radiometer (AATSR). The SLSTR uses a dual-viewing technique and operates across nine wavelength bands providing better coverage than AATSR because of a wider swath width (1675 km for the nadir view angle). The sensor has three bands in the visual and near-infrared (555, 659 and 865 nm), three in the mid infrared (1.38, 1.61 and 2.25 μm) and three in thermal infrared (3.74, 10.85 and 12 μm). The spatial resolution is 500 m at visible and infrared wavelengths and 1 km at thermal wavelengths. Sentinel-1 carries a C-band Synthetic Aperture Radar (SAR) building on ESA’s and Canada’s heritage in SAR systems from ERS-1, ERS-2, Envisat and Radarsat. Sentinel-1A was launched 3 April 2014, and Sentinel-1B 25 April 2016. The satellites have repeat cycles of 12 days. The C-band SAR operates in four modes: 1) Strip Map Mode (SM); 2) Interferometric Wide Swath (IW); 3) Extra-Wide Swath Mode (EW); and 4) Wave-Mode (WV). Over land the primary, conflict-free, mode is IW with VV+VH polarizations. The resolution of the IW-mode depends on whether the product is single-look complex (SLC) or ground range detected (GRD) and number of looks. The SLC range-azimuth

24

SnowBall pixel spacing is 2.3 × 17.4 m, whereas for the IW GRD product, the pixel spacing is 10 × 10 m with number of looks equal to 5 × 1. The first validation dataset was collected in winter/spring 2015 in Romania. Precursor algorithms and products for wet snow were previously tested in the Jotunheimen site in Norway by Norwegian Computing Center, and later demonstrated and applied in the whole country of Norway. Jotunheimen is the highest part of the Scandinavian Mountains, and is located in central southern Norway. The SnowBall algorithms are now validated in Jotunheimen, and subsequently in the Bucegi mountain Sinaia region in Romania, part of the Southern Carpathian. The Romanian sites for diagnostic data collection and algorithm cal/val have been selected as the upper sector of the Arges and Ialomita river catchments with altitudes of about 500 to 2500 m. 2. Algorithms 2.1. Optical The ideal approach for retrieval of snow surface wetness based on optical data would be to measure the liquid water contents in the snow, like that proposed and demonstrated by Green et al. (2006). However, this would require an imaging spectrometer with optimally located spectral bands for measuring liquid-water molecular absorption features. Such sensors are currently not available in satellites, only as experimental sensors in aircrafts. Our aim has been to develop an algorithm to be used operationally based on satellite data. Experiments with snow wetness algorithms have confirmed that a combination of snow surface skin temperature and snow grain size, analysed in a time series of observations, can be used to infer wet snow, including giving an early warning of snowmelt start (Solberg et al., 2004). The temperature observations give a good indication of where wet snow potentially could be present, but are in themselves not accurate enough to provide sufficiently strong evidence of wet snow. However, if a rapid increase in the effective snow grain size is observed simultaneously with a snow surface temperature of approximately 0°C, then this is a strong indication of a wet snow surface.

25

SnowBall 2.1.1. The snow surface temperature algorithm

The surface temperature of snow (STS) algorithm is based on an approach proposed by Key et al. (1997). In a comparison study by Amlien and Solberg (2003), this algorithm was identified as one of the best single-view techniques for retrieval of STS for polar atmospheres, and it may be applied with several sensors of moderate resolution, like MODIS, AVHRR, AATSR and OLCI/SLSTR. The absorption of the radiation in the atmosphere depends on the wavelength, and the difference between the brightness temperatures in two channels will therefore yield information about the atmospheric attenuation (Strove et al., 1996; Coll et al., 1994). The split-window technique aims at eliminating the atmospheric effects by utilizing this difference. The surface temperature T is estimated as a weighted sum (or difference) of the brightness temperatures observed. The split-window equation utilizes T11 measured at 11 µm (MODIS band 31, SLSTR band 7) and T12 measured at 12 µm (MODIS band 32, SLSTR band 8): T = b0 + b1T11 + b2T12

(1)

The split-window technique is only sensitive to the effect of the atmospheric water vapour, and not to other atmospheric gases or aerosols. The atmospheric influence on the split-window equation depends on the composition of the atmosphere, and the method must therefore be calibrated for different atmospheres. Coll’s modification to the split-window algorithm was proposed in order to avoid the need of several calibration sets. It is a global non-linear equation for global-scale application. Derivation of regionally optimized linear algorithms has been demonstrated for mid-latitude conditions, but not for colder atmospheres. Coll’s equations are given by: T = T11 + A(T11 − T12 ) + B

(2)

A = b0 + b1 (T11 − T12 )

(3)

b0 = 1.00 b1 = 0.58 B = 0.51

Key’s algorithm (Key et al., 1997) is a modification of the simple splitwindow technique. An additional correction term addresses the variation of

26

SnowBall the view angle θ along a scan line and its effect of the atmospheric path length. The algorithm expresses the surface temperature as: Ts = b0 + b1T11 + b2 (T11 − T12 ) + b3 (T11 − T12 )(sec θ − 1)

(4)

The calibration coefficients depend on the temperature interval and the satellite sensor. 2.1.2. The snow grain size algorithm

For snow grain size (SGS) we have used a normalized grain size index based on work by Dozier (1989) and followed by experiments Fily et al. (1997). MODIS bands 2 and 7 have been used as this index has been shown to be less sensitive to snow impurities. The original algorithm proposed by Dozier (1989) was based on Landsat Thematic Mapper (TM) data. The problem of radiometric terrain effects, the influence of slopes on the reflected light, is minimized by using ratios between two channels as an index for grain size. A number of ratios of the form: Rij =

TM i − TM j TM i + TM j

(5)

were tested. R47 has been selected as the best ratio. Fily (1997) reported that the measured data matches the theoretical curves well. The ratio approach is a simple method. Signals from two channels are sufficient and information about the terrain is not needed (as it is with several other methods). Studies of calculated grain size from the R47 ratio show that the index is well suited for monitoring the changes in grain size due to precipitation and temperature changes. The index increases with increasing temperature and gets a lower value when new snow has fallen. The grain size index for snow for MODIS data lies typically between 0.7 and 1.0. Bare ground of different kinds gives lower index values. 0.7 is not an exact threshold value for snow. Somewhere around 0.7 the index shows that there is probably some snow on the ground. To be sure that the index represents snow grain size, we use a fractional snow cover retrieval algorithm in addition to check that the ground is fully snow covered.

27

SnowBall 2.1.3. The snow surface wetness algorithm

The approach is to infer wet snow from a combination of measurements of STS and SGS in a time series of observations. The Norwegian Computing Center has analysed in-situ and MODIS data for various snow parameters in test sites in Jotunheimen. An example of the temporal development of data from the winter and spring season 2003 is shown in Figure 1 for the two test sites in Jotunheimen. Note the rapid increase of SGS at one site (Valdresflya) when the air temperatures approaches and reaches 0°C level from below. The STS values follow the air temperatures exactly as expected.

Figure 1: STS and SGS retrieved from MODIS data at Valdresflya (VF) and Heimdalshø (HH), Jotunheimen, Norway, in winter/spring 2003 (dotted lines). Temperatures from two weather stations are also shown (solid lines).

The process driving the SGS increase is snowmelt metamorphism. It typically takes place in the spring when snowpack temperatures are close to 0°C. During daytime the air temperature increases to well above 0°C and the snow surface starts to melt. In the late afternoon, evening and nights the temperature falls and the snow surface refreezes. During the melting, smaller grains melt first, and liquid water appears in the upper layer of the snowpack. The bonds between the snow grains are typically destroyed, and the remaining 28

SnowBall snow grains are covered by a layer of liquid water. When the temperature cools again, the liquid water freezes and the snow grain size of the remaining grains increases due to the liquid water added. From this processes, a melting snowpack typically has an aggregation of rounded grains of 1-2 mm (corn snow). Note that snowmelt metamorphism also might take place due to rain. Liquid water from rain percolates downward in the snowpack and refreezes. Based on the types of observation shown above, we developed an approach to infer categories of snow surface wetness (SSW) from a combination of measurements of STS and SGS in a short time series. The temporal behaviour of STS and in particular SGS let us infer the current snowmelt stage the snow surface is within from the multi-temporal observations. We developed a decision-tree approach for classification into the snow wetness categories. The decision boundaries were tuned based on calibration data from test sites in Jotunheimen: in-situ measurements of liquid water in the snow surface and weather station air temperature measurements (Solberg et al., 2004; Solberg et al., 2010). A simplified version of the algorithm applied is expressed below (pixel indexing has been skipped for clarity; MSSW is time-series multi-sensor snow surface wetness): if SGS(today) − SGS(recently) > SGStresh AND −2 1) then MSSW = BARE_GROUND else MSSW = DRY_SNOW

The decision tree further refines MSSW into categories of wet snow. The algorithm also illustrates how bare ground is inferred from temperature observations above 0°C and a rapidly developing negative gradient for SGS (both due to appearance of bare ground patches at the sub-pixel level). Experiments with the snow wetness algorithm have confirmed that the approach of combining STS and SGS, analysed in a time series of observations, can be used to infer wet snow, including giving an early warning of snowmelt start. Air temperature measurements from meteorological stations confirm the maps produced in general. A potential problem sometimes observed is related to clouds. Non-detected clouds or cloud fractions within a pixel will usually decrease the temperature retrieved. One should be aware of this potential problem with partly cloudy SSW maps. 29

SnowBall 2.2. Synthetic aperture radar SAR imaging systems allow for imaging through clouds, and because SAR is an active system, day and night imaging is possible. Due to the nature in which microwaves interact with the surface features the information in the backscattered radar signals can be indicative of moisture content, salinity and physical characteristics (shape, size, orientation). Using SAR to map the snow cover has some limitations. A dry snowpack has minor influence on the SAR signals, and the reflected signals are dominated by the contribution of the snow/ground interface. However, when the snow is wet, the radar signals cannot penetrate the snow, and the backscattered signal is dominated by the contribution from the air/snow interface. The reflected signal is often lower from areas covered with wetsnow, compared to snow-free or dry snow-covered conditions. Wet snow may therefore be mapped in a SAR image by comparing the backscatter coefficient values with corresponding backscatter coefficients from a reference image obtained at snow-free or dry snow conditions (Nagler and Rott, 2000). Hence, the algorithm for mapping wet snow is based on change detection using ratios of wet snow versus snow-free (or dry snow) surfaces. To avoid the need for reference images in the same imaging geometry, we have implemented an alternative method for mapping wet snow areas. This is based on the ‘flattening-gamma’ radiometric terrain correction approach by Small (2011). Without treatment, the hill-slope modulations of the radiometry threaten to overwhelm weaker thematic land-cover induced backscatter differences, and comparison of backscatter from multiple satellites, modes, or tracks loses meaning. The ‘flatting gamma’ SAR methodology suppresses a large part of the brightness variation in the SAR images caused by terrain variation, and may therefore provide a proper treatment to the hill-slope modulations. The ‘flattening-gamma’ products have shown great potential for improving SAR-based mapping of wet snow in mountainous areas, e.g. at time of spring snowmelt. However, the quality of the ‘flatting-gamma’ products depends strongly on the quality of the digital elevation model and the precision of the geocoding.

30

SnowBall 2.3. The wet snow mapping algorithm The steps in the wet snow mapping algorithm may be summarized as follows: • Conversion of the SAR data (digital numbers) to gamma naught. • Multi-looking to reduce speckle noise. The number of looks we apply depends on the desired output resolution. We have applied 6×6 looks (corresponding to a desired pixel spacing of 50 × 50 m). • Conversion to terrain-corrected gamma naught (‘flattening gamma’) backscatter normalization. • Computation of layover and shadow masks. • Geocoding using the range-Doppler algorithm. • Construction of daily mosaic images. • Computation of ratio images, i.e. daily mosaic image versus the reference image. • Thresholding of ratio images to detect wet-snow. If the difference is more than 4 dB, the pixel is classified as wet-snow. • Masking of layover and shadow areas. Currently, the algorithm supports Sentinel-1 GRD and Radarsat-2 SCN/SCW/SLC SAR images. The benefit with this algorithm, compared to the one proposed by Nagler and Rott (2000), is that we only need a single reference image for the area of interest. E.g., if the aim is to perform wet snow mapping for the whole of Romania, then the reference image must cover the whole country, and is constructed by averaging daily mosaic images for the snow-free months. 2.4. Multi-sensor/multi-temporal Whereas the application of optical sensors is limited by cloud cover, the current SAR sensors are limited to the detection of melting snow. Therefore, utilisation of sensor synergy through multi-sensor algorithms is attractive for applications requiring frequent observations (Solberg et al., 2004). The Multi-sensor/multi-temporal Wet Snow (MWS) algorithm we have developed is novel to this project and fuses optical and SAR data to map the wet-snow area. Our idea is to combine multi-temporal observations of optical

31

SnowBall and SAR wet snow in a fusion model to generate improved coverage in space and time. The algorithm we have developed fuses the optical and SAR observations using a Hidden Markov Model (HMM) approach. The snow map includes the thematic snow classes: 5 classes for the results obtained from MODIS and SAR data (dry snow, moist snow, wet snow, very wet snow and soaked snow) and 4 classes for the results obtained from Sentinel-3 and SAR data (dry snow, moist snow, wet snow and very wet snow). The HMM approach based on modelling and assimilation was proposed by Solberg et al. (2008) for retrieval of Fractional snow Cover (FSC). A set of snow states is defined, and for each snow state there is a corresponding reflectance and backscatter model. A similar approach was recently developed in the CryoClim project (www.cryoclim.net), where an accuracy of 93% was obtained for snow extent mapping when validating against synoptic weather stations (Rudjord et al., 2015; Solberg et al., 2015). The basic idea of the approach is to simulate the states that the snow surface goes through during the snow season with a state model. The states are not directly observable, but the remote sensing observations give data describing the snow conditions, which are related to the snow states. Solberg et al. (2008) applied a Hidden Markov Model (HMM) to model the states. There are other state modelling frameworks that could be applied as well. However, HMM is building on statistical theory making it possible to establish a sound probabilistic model derived from observational data (Baum and Petrie, 1966). Note that the HMM solution represents not only a multi-sensor model but also a multi-temporal model. The sequence of states over time is required to follow certain optimisation criteria. Note also that the HMM model is applied per pixel, so each pixel’s history through the snow season is modelled. According to the thematic snow wetness classes applied for the OWS algorithm, five corresponding snow wetness states have been defined in the hidden Markov model (Figure 2). Additionally, there is a state for patchy snow cover and snow-free ground, depletion (< 100% FSC), and a state of temporary snow (thin snow layer making full snow cover for a short period, also named ephemeral snow cover). Allowed transitions between states are shown by arrows. As the model shows, the wet snow classes are “chained” such that the current state might move up to wetter classes or down to drier classes (in terms of liquid water content). 32

SnowBall

Figure 2: Hidden Markov model for snow wetness fusing optical and SAR observations.

In the following paragraphs we describe the basic HMM formalism. In a HMM we observe a system assumed to evolve through a series of different states. Transitions from one state to another happen with certain probabilities. While in a given state the system will produce observables with a certain probability density. We will denote the set of discrete states Q of the internal system by: 𝑄𝑄 = {𝑆𝑆1 , 𝑆𝑆2 , … , 𝑆𝑆𝑣𝑣 }

(6)

where ν is the number of states. Furthermore, the time series of observations, 𝑋𝑋�, will be denoted by: 𝑋𝑋� 𝑇𝑇 = {𝑋𝑋1 , 𝑋𝑋 2 , … , 𝑋𝑋 𝑇𝑇 } (7)

where T is the number of elements of the sequence. The unknown state of the process at time t will be denoted Et, thus Et = Si indicates that the process is in state Si at time t. The states are not directly observable, but are related to observation Xt at times t, (t = 1, 2,…, T) by a probability distribution of measurements: 𝑝𝑝(𝑋𝑋 𝑡𝑡 |𝐸𝐸 𝑡𝑡 = 𝑆𝑆𝑖𝑖 ), 𝑖𝑖 = 1, 2, … , 𝑣𝑣 (8) For a given time period, the model is also described by a set of transition probabilities between each pair of states: 𝑝𝑝�𝐸𝐸 𝑡𝑡 = 𝑆𝑆𝑖𝑖 �𝐸𝐸 𝑡𝑡−1 = 𝑆𝑆𝑗𝑗 �, 𝑖𝑖, 𝑗𝑗 = 1, 2, … , 𝑣𝑣 33

(9)

SnowBall The probabilities of transition between the different states are obviously strongly dependent upon season, thus the process is not stationary and the probabilities of transition are time dependent. The final parameters of the model are the initial conditions defined by the probability of being in a given state at the initial time: 𝑝𝑝(𝐸𝐸1 = 𝑆𝑆𝑖𝑖 ), 𝑖𝑖 = 1, 2, … , 𝑣𝑣

(10)

With HMM, the notion of a class from the classification literature becomes the notion of a model in the HMM formalism. Traditionally, groundcover classification in a temporal sequence of satellite images is the problem of assigning each pixel in the scene to a class based on this pixel’s signal properties (or derived properties). In the HMM case, our aim is to assign each pixel to the model that best explains the observed temporal evolution of the pixel. Solutions to this kind of problem are important in many applications and several algorithms are available. For our problem we have chosen to use the Viterbi algorithm. The Viterbi algorithm is a dynamic-programming algorithm for finding the most likely sequence of hidden states (the Viterbi path) that result in a sequence of the observables. The Viterbi algorithm was proposed by Viterbi (1967) as a decoding algorithm for convolutional codes over noisy digital communication links. The algorithm requires as input the state probability density functions, the transition probabilities between the different states and the initial probability of each state. Let Vt,k be the probability of the most likely state sequence responsible for the first t observations that has k as its final state, then: 𝑉𝑉1,𝑘𝑘 = 𝑝𝑝(𝑋𝑋1 |𝑘𝑘)𝑝𝑝(𝐸𝐸1 = 𝑆𝑆𝑘𝑘 )

𝑉𝑉𝑡𝑡,𝑘𝑘 = 𝑝𝑝(𝑋𝑋 𝑡𝑡 |𝑘𝑘) max𝑖𝑖 �𝑝𝑝�𝐸𝐸 𝑡𝑡 = 𝑆𝑆𝑖𝑖 �𝐸𝐸 𝑡𝑡−1 = 𝑆𝑆𝑗𝑗 �𝑉𝑉𝑡𝑡−1,𝑘𝑘 �

(11) (12)

The Viterbi path can be retrieved by saving back pointers that remember which state i was used in the second equation. The algorithm takes as input the optical wet snow (OWS) map, containing wet snow class probabilities, and a SAR wet snow (SWS) map containing probabilities of wet snow. To train the HMM we take advantage of prior knowledge of local climatology. Per time unit throughout the season (day) we would like to

34

SnowBall establish estimates of the likelihood of each state and the transition probabilities. For Norway, we use a 15-year daily time series (2000-2015) of a 1-km snow surface state product based on a model where data from meteorological stations and numerical weather prediction are combined into a national-wide product provided through the seNorge web portal (Saloranta, 2012). For Romania we estimate climatology from a 10-year time series (20052015) with re-analysis of snow depth and temperature at a 1-km grid resolution (Dumitrescu et al., 2015). Note that we in this first version of the algorithm and product do not take full advantage of the climatology data and rather apply very simplified statistics to establish the probability density functions. 3. Results The algorithm test results for the sites in Romania are presented in the following. The results are here limited to comparison with air temperature for the 2015 winter season, but will be extended with comparison with in-situ snow wetness measurements and the use of Sentinel-3 data when these become available for the 2017 season. 3.1.

2015 snow season

3.1.1. Optical snow wetness

Norway OWS maps were generated from a time series of Terra MODIS data acquired in the period 1 January until 30 June 2015 for the test site in Jotunheimen. Each acquisition was processed independently of cloud cover. A subset of OWS maps was then selected to show the temporal dynamics and development of the snow surface through the season. The chosen maps have moderate to no cloud cover. The acquisition times of the satellite images are listed in Table 1 together with retrieval results and air temperature measurements from the weather stations. The temperature measurements shown are from the morning, the measurement closest to the acquisition time of the satellite image and in the early afternoon. The retrieval results are from the 1 km2 snow map grid cells including the weather stations. The air

35

SnowBall temperatures in the morning and afternoon are included as to give an indication of the temperature gradient before and around the time of the satellite acquisition. Note that all times are given in UTC (which is Norway ‘winter time’, CET, minus 1 hour). The seasonal development of the snow around the weather stations, as shown by the retrieval results, very much follows what is expected for a typical snow season and therefore the local climatology. Valdresflya tends to accumulate very much snow and therefore shows a long snow season. Bitihorn is a mountain peak, so the snow layer is typically shallow and the site is influenced quickly by changing weather. Bygdin is lower than Valdresflya, and usually accumulates much less snow. Eidsbugarden accumulates more snow, and is often influenced by the west-coast weather. Contrary, VinsterenBjørnhølen is dominated by the weather east of the watershed with higher temperatures and less precipitation, and therefore more quickly develops into summer conditions. Table 1: OSW map retrieval results (W) and corresponding air temperature measurements in the morning (08:00), closest to the acquisition time (Ac) and in the afternoon (14:00) for the five weather stations. All times are given in UTC. The retrieval results are shown colour coded as well as with letters (D = Dry, M = Moist, W = Wet, V = Very wet and S = Soaked snow). When there is no OSW retrieval result, other classes are shown (‘+’ = Cloud, ‘-‘ = Partly snow-covered ground and ‘=’ = Bare ground (no snow)). Satellite ac. Date Time 11.03 10:55 08.04 11:20 17.04 11:10 19.04 11:00 20.04 10:10 27.04 10:15 14.05 10:55 15.05 11:35 05.06 11:55 08.06 10:50 13.06 11:05 16:06 10:00 16.06 11:35 20.06 11:15 27.06 11:20

W D D D W W D W D W W W D W S S

Valdresflya 08:00 Ac 14:00 -9.0 -5.9 -4.1 -3.3 -3.4 -1.3 -2,6 -0.9 4.8 0.6 5.6 6.1 8.3 7.7 6.4 0.1 1.9 6.5 1.6 1.2 2.8 0.8 8.4 2.6 0.9 5.0 4.5 3.8 4.4 5.1 10.4 5.1 5.5 2.0 4.2 6.2 2.0 6.6 6.2 14.2 15.6 9.0 10.4 9.6 12.7

W D D D W D S + -

Bitihorn 08:00 Ac -6.9 -3.2 -8.8 -7.8 -5.9 -5.1 -0.2 2.4 4.4 4.8 -3.9 -1.2 -3.4 1.4 -2.5 5.8 -0.9 4.7 -2.7 0.3 -0.1 1.3 -0.8 1.8 -0.8 2.1 7.0 6.1 4.9 15.6

36

14:00 -1.2 -6.0 -4.6 4.1 3.4 1.5 0.6 4.9 4.2 1.5 2.8 2.7 2.7 8.5 10.3

W D D W + = =

Bygdin 08:00 Ac -7.3 -3.8 -1.8 0.2 0.1 1.4 2.9 7.0 4.8 6.6 -5.2 -1.1 2.2 6.8 -1.2 1.4 2.4 6.2 3.8 5.9 4.2 7.1 3.4 5.3 3.4 5.7 7.5 10.2 9.6 14.4

14:00 -3.1 0.2 3.6 8.4 8.9 3.9 4.1 2.4 6.1 6.2 7.2 7.3 7.3 12.2 16.0

SnowBall Satellite ac. Date Time 11.03 10:55 08.04 11:20 17.04 11:10 19.04 11:00 20.04 10:10 27.04 10:15 14.05 10:55 15.05 11:35 05.06 11:55 08.06 10:50 13.06 11:05 16:06 10:00 16.06 11:35 20.06 11:15 27.06 11:20

W D D D D W D W S W -

Eidsbugarden 08:00 Ac 14:00 -9.2 -7.2 -4.9 -2.5 1.4 0.7 -2,7 -0.7 1.5 -1.5 3.1 5.0 4.2 9.8 8.4 -7.6 -3.9 1.9 0.5 2.0 3.8 1.6 9.8 6.0 3.8 8.1 8.8 5.5 6.3 6.5 6.9 7.6 9.6 6.1 6.9 8.9 6.1 8.1 8.9 9.7 13.1 12.1 9.2 18.2 15.7

W D D W S = = = = = =

Vinsteren-Bjørnhølen 08:00 Ac 14:00 -8.6 -5.6 -3.5 -1.5 -0.2 -0.5 0.4 2.2 3.5 3.6 7.9 6,4 -1.0 5.7 5.0 -4.6 -4.2 0.0 0.4 1.7 2.7 -0.9 2.4 3.4 4.1 6.4 7.7 4.9 6.7 7.9 6.7 8.1 9.2 5.4 8.4 11.1 5.4 10.2 11.1 9.9 12.0 15.1 9.3 14.7 16.1

Figure 3 shows as an example the situation on 19 April 2015. The melting is extensive east of the watershed, and has reached higher altitudes. For some of the lowest altitudes the snow is shown as soaked.

Figure 3: Optical snow wetness based on MODIS from 19 April 2015 acquired at 11:00 UTC.

All stations show air temperatures well above 0°C for the time of acquisition. For the station with highest temperature, Vinsteren-Bjørnhølen with 7.9°C, the snow is shown as soaked in the map. Bygdin also shows very 37

SnowBall high temperature with 7.0°C, but here there is already partial snow cover (and therefore no wet snow retrieval). Valdresflya and Bitihorn seem correctly to show wet snow. For Eidsbugarden the snow is shown as dry, even if the air temperature is 3.1°C. It was, however, −1.5°C in the morning. Also, west of the weather station all snow is snow as dry, while there is some snowmelt along the lake coastline towards east. The retrieval result might be explained by that fact that the 1 km2 grid cell around the station is dominated by the snow at somewhat higher altitudes than at the station, which is situated close to the lake shore. Romania The optical snow wetness products have been generated using the time series of Terra MODIS data acquired from 1 January 2015 to 30 June 2015 for the study area represented by the upper parts of the Arges and Ialomita Rivers Catchments. Using MODIS data give the resolution of the OWS retrieval products of 1 km2. In order to present the temporal dynamics and development of the snow conditions a number of optical snow wetness (OWS) products have been selected. Also, the less-to-no cloud cover images have been selected to be presented below (Table 2). Table 2: OSW map retrieval results (W) and corresponding air temperature measurements in the morning (08:00), closest to the acquisition time (Ac) and in the afternoon (14:00) for the four weather stations. All times are given in UTC. The retrieval results are shown colour coded as well as with letters (D = Dry, M = Moist, W = Wet, V = Very wet and S = Soaked snow). When there is no OSW retrieval result, other classes are shown (‘+’ = Cloud, ‘-‘ = Partly snow-covered ground and ‘=’ = Bare ground (no snow)).

38

SnowBall Figure 4 presents an example of the optical snow wetness based on MODIS from 10 April 2015 in the study area. For the OWS analysis only Fundata, Sinaia 1500, Bâlea Lac and Vf. Omu weather stations have been taken into consideration because the other stations are within forested and urban areas (Figure 5). The altitude gradient in the melting season is well covered by the four stations. On the other hand, because Fundata and Sinaia 1500 are situated very close to the forest limit, the neighbour pixels in the east of Fundata and west of Sinaia 1500 had to be used. It means that there might be differences due to the slight variation in altitude for air temperatures close to 0°C.

Figure 4: Optical snow wetness based on MODIS from 10 April 2015 acquired at 09:30 UTC.

Figure 5: Temperature profiles at weather stations on 10 April 2015.

39

SnowBall 3.1.2. SAR snow wetness

Norway In the figure below, white areas correspond to wet snow, green to dry snow or bare ground, blue areas to water bodies, purple to SAR layover or SAR shadow areas, and black to areas outside the area of interest (in this case the sea). On 12 May (Figure 6) we observe a significant amount of wet snow areas Note that the image is acquired in the afternoon (17:02 UTC), ascending pass. All weather stations shows positive temperatures (Figure 7), which correspond with the local-area snow maps (Figure 8).

Figure 6: SAR wet snow map for the test area on May 12, 2015 acquired at 17:02 UTC.

th

Figure 7: Temperature profile at weather stations on May 12 , 2015. The time scale is in local time (UTC + 1h).

40

SnowBall

Eidsbugarde

Bygdin

Vinsteren-

Valdresflya

Bitihorn

th

Figure 8: Wet snow map around the weather stations (red dot) on May 12 , 2015. The SAR image was acquired 17:02 UTC.

Romania The SAR wet snow (SWS) map is showed in Figure 9, for the test site for 12 April 2015. Light blue areas correspond to wet snow, green to dry snow or bare ground, blue areas to water bodies, purple to SAR shadow areas, and dark grey to areas outside the satellite coverage.

Figure 9: SAR wet snow map for the validation area on 12 April 2015. 41

SnowBall On April 12th the dry snow is dominant around the stations, due to negative or close to 0°C temperatures at the image acquisition hour (Figure 10). At the same dates, at Bâlea Lac dry snow dominated the surface around the station due to the fact that this is located on the northern slope. So, even though the temperatures are positive, the snow is dry. All the other stations have temperatures close to or above 0°C, so the retrieval results that there is bare ground is probably correct.

Figure 10: Temperature profiles at weather stations on 12 April 2015.

The hourly air temperatures, around the satellite passage, at weather stations together with SWS results are listed in Table 3. All times are given in UTC and the image acquisition in the afternoon (around 16.30), ascending pass. The hourly air temperatures are used for making the difference between the dry snow and the bare ground, which are included in the same class. Table 3: SWS map retrieval results (Ac) and corresponding air temperature measurements closest to the acquisition time for the weather stations. All times are given in UTC. The results are shown colour coded as well as with letters (W = Wet snow, DB = Dry snow or bare ground). Satellite ac. Date 30.01 23.02 19.03 31.03 12.04 24.04 6.05 18.05

Vf. Omu Ac DB DB DB W DB DB DB DB

16:00 -9.1 -4.0 -14.7 -7.5 -0.4 0.2 6.7 5.2

Bâlea Lac 17:00 -9.3 -4.2 -15.2 -7.9 -1.4 -0.4 6.3 4.2

Ac DB DB DB DB DB DB DB DB

16:00 -5.4 -1.1 -11.8 -2.4 2.6 4.9 11.7 6.3

17:00 -5.3 -1.8 -11.6 -1.1 1.2 4.0 11.7 5.8

42

Sinaia 1500 Ac DB DB DB DB DB DB DB DB

16:00 -2.8 1.0 -6.6 2.0 8.1 8.6 12.2 12.2

17:00 -2.3 0.5 -7.8 1.7 8.0 7.8 12.1 10.2

Fundata Ac DB DB DB DB DB DB DB DB

16:00 -0.5 2.0 -4.8 4.6 11.0 13.4 15.8 15.0

17:00 -0.4 1.9 -5.5 3.6 9.7 11.5 15.0 14.0

SnowBall Satellite ac. Date 30.01 23.02 19.03 31.03 12.04 24.04 6.05 18.05

Predeal Ac DB DB DB DB DB DB DB DB

16:00 1.7 3.9 -3.6 5.6 13.7 15.3 17.4 18.8

Câmpulung 17:00 1.5 2.7 -3.7 4.9 9.9 13.8 16.7 16.8

Ac DB DB DB DB DB DB DB DB

16:00 4.0 6.5 0.9 10.3 17.9 18.7 20.7 21.0

17:00 3.8 6.0 0.0 8.8 16.0 17.3 20.0 20.3

Câmpina Ac DB DB DB DB DB DB DB DB

16:00 2.2 5.9 0.1 13.1 18.5 18.9 20.3 21.7

17:00 2.6 5.8 -0.6 11.4 15.4 17.2 19.6 20.6

Curtea de Argeş Ac DB DB DB DB DB W DB DB

16:00 5.9 7.9 2.5 10.1 18.8 20.7 23.4 22.8

17:00 5.6 7.2 1.2 9.1 17.8 18.4 22.5 21.6

3.1.3. Multi-sensor snow wetness

Norway The multi-sensor snow wetness maps shows that from 23 March most of the area is covered with dry snow. In the maps from 9-11 April, most of the area is still covered with dry snow, but some areas at lower altitudes and along the edge of the snow zone show varying degrees of wet snow. In the map from 21 April most of the area shows varying states of wet snow, with the most severe snowmelt happening in the lower altitudes along the edge of the snow zone. By 9 May, there has been a freezing of the area, as most of the region shows dry snow, with the exception of some regions of the lower altitudes. For 17 May we see the same, but with some areas having entered depletion (partial or no snow cover). Finally, by 6 June (Figure 11), the map shows that most of the remaining snow is in some stage of wetness with the most intensive melt on the edges of the snow-covered area. First, we compare the wetness state at the location of the weather stations with the mean 24 hour air temperature for the same day. This comparison is shown in Table 4. In the cases where the mean air temperature is below 0 and the map indicates wet snow, or vice versa, the potential conflict is marked in red. As the satellite images on which the multi-sensor map is based are normally acquired between approximately 10:00 and 12:00 UTC, the 24h mean air temperature is not always an accurate comparison for snow wetness. The conflicting data are therefore investigated more closely by comparing with hourly air temperature data. From this, we could conclude that there is a good agreement between the multi-sensor snow wetness map and the 24 hour 43

SnowBall mean air temperature measurements. In general, freezing temperatures are shown as dry snow, while higher temperatures tend to be represented as wet snow. The disagreements can mostly be explained when looking at hourly air temperature measurements.

Figure 11: Multi-sensor snow wetness map from Jotunheimen 6 June 2015. Table 4: Comparison of mean 24h air temperature and wetness state for the weather stations. Here, the numbers 0-6 indicate the states dry snow, moist snow, wet snow, very wet snow, soaked snow, depletion and temporary snow, respectively. The apparent deviations are marked in red.

Valdresflya Bygdin Eidsbugarden VinsterenBjørnhølen Bitihorn

Temperature Snow class Temperature Snow class Temperature Snow class Temperature Snow class Temperature Snow class

Mar 23 -0.3

April 9 1.4

April 11 -3.6

April 21 2.7

May 9 -0.7

May 17 -2.1

June 6 2.2

0

0

0

2

0

0

2

-0.2

0.4

-4.7

2.8

-3.0

-0.1

4.1

0

0

2

5

3

5

5

0.5

1.3

-6.2

3.4

-2.5

-0.3

5.2

0

1

3

3

0

2

5

-0.2

-1.3

-4.2

-4.2

-5.6

-0.4

4.3

0

6

5

5

3

5

5

-4.0

-5.5

-3.5

1.5

-8.3

-6.6

1.1

0

0

0

3

0

5

5

44

SnowBall Romania Figure 12 shows a multi-sensor snow wetness maps for 23 February 2015. The temperature has increased from the days before and the snow is transforming into moist, wet and even very wet snow at lower altitudes. Comparison of the wetness state at the location of the weather stations with the mean 24 hour air temperature for the same day is shown Table 5 for eight multi-sensor snow wetness maps.

Figure 12: Multi-sensor snow wetness product for the validation area in Romania, 23.02.2015. Table 5: Comparison of mean 24h air temperature and wetness state for the weather stations. The numbers 0-6 indicate the states dry snow, moist snow, wet snow, very wet snow, soaked snow, depletion and temporary snow, respectively. The uncertainties are marked in red.

Bâlea Lac Fundata Sinaia 1500 Vf. Omu

Temperature Snow class Temperature Snow class Temperature Snow class Temperature Snow class

19 Feb -10.8

23 Feb -2.3

2 Mar -3.2

10 Mar -3.6

25 Mar -0.1

28 Mar -0.5

23 Apr -5.0

25 Apr 3.5

0

2

6

0

1

4

1

3

-6.9

1.1

1.1

0.2

4.5

2.1

3.2

11.0

6

6

6

6

6

6

6

6

-7.9

0.4

1.8

-0.6

3.7

2.1

2.0

7.6

6

6

6

6

6

6

6

6

-14.0

-5.0

-6.2

-7.1

-2.5

-2.7

-6.7

-1.1

6

0

0

0

0

3

2

4

45

SnowBall The expected way of snow surface behaviour is that negative temperatures result in dry snow and positive temperatures results in moist to soaked snow surface. How wet the snow turns depends on the level of positive temperatures as well as the wind speed. But there are cases when the mean air temperature is below 0°C and the map indicates wet snow. A particular case is for pixels with partial snow cover where the snow condition is not evaluated. It is the case of Sinaia 1500 and Fundata stations. In order to have a complete view of the snow evolution, an intercomparative study of the three products OWS, SWS and MWS is needed. The most difficult part is to find SAR data on the same day with moderate or cloudfree optical images. An example of the tree products is shown in Figure 13 for Norway and Figure 14 for Romania. The spatial resolution of the SAR image complements the snow wetness obtained from optical image. The wet snow in SAR image is well correlated with snow wetness classes in optical image. The melting season is visible on MWS map as partial snow-covered pixels. This observation is supported by the positive air temperature at all weather stations between image acquisitions. At high altitudes the negative temperatures during the night favour the snow cover to last longer than at medium and low altitudes where temperature are positive all day long.

Figure 13: Multi-sensor wet snow map for June 2015 (top-left), optical wet snow map for the same day (top-right) and SAR wet snow map for 17 June 2015 (bottom) for Norway.

46

SnowBall

Figure 14: Multi-sensor wet snow map (topleft), optical wet snow map (top-right) and SAR wet snow map (bottom) for 24 April 2015 for Romania.

3.2.

2017 snow season

When this book goes into print, the snow season is still ongoing and the results have so far not been analysed in as much detail as for the 2015 season. This in particular goes for Norway, there the accumulation season still was ongoing when this was written. However, the Sentinel-3 satellite just became available a few months ago, so the following results are in particular included for its novelty. The final scientific results are to be published when the snow season is finished and the results properly analysed. 3.2.1. Optical wet snow

Norway For this snow season we are validating the full product domain covering southern Norway. Weather station data is not available yet, so the comments here are based on what was forecasted and also reported by people in the field. On 26 March 2017 there was a situation of a very warm day in south-east Norway. Air temperatures at altitudes 1000 m.a.s.l. were at maximum in the afternoon around 12°C. The snow map in Figure 15 confirms the high temperatures where present in the Hardangervidda mountain plateau (1100 -

47

SnowBall 1200 m.a.s.l.) in south Norway, orange area representing “wet snow”. At lower altitudes, typically 800-1000 m, the snow is very wet (red).

Figure 15: Optical snow wetness based on MODIS from 26 March 2017 acquired at 13:05 UTC.

Romania The optical snow wetness products have been generated separetly for Terra MODIS and Sentinel-3 SLSTR, so that the results may be compared. For MODIS, the period 1 of November 2016 to 30 March 2017 was used, and for Sentinel-3, 17 November 2016 to 30 March 2017. The seasonal development of the snow around the weather stations follows what is expected for a typical snow season, the altitude profile and therefore the local climatology. The Vf. Omu, Vf. Tarcu and Calimani stations are situated on mountain peaks (above 2000 m.a.s.l.), so the snow depth is usually moderate because the snow is often shattered by wind. On the other hand, the snow season lasts longer due to a longer period of air temperatures below the freezing point. On the other hand, at Bâlea Lac, situated also above 2000 m.a.s.l., but in a glacier caldera, the snow layer tends to accumulate very much and therefore shows the longest snow season of the test site. At the weather stations situated above 1000 m.a.s.l., in various mountain features

48

SnowBall (sunny slopes, peaks of lower heights, the valley floor etc.), the snow conditions are closer to the high mountains, but the snow season is shorter. The snow cover at moderate elevations (from 500 to 1000 m.a.s.l.) may be influenced by the local weather conditions (local winds: foehn, Mediterranean effect, forests etc.) and in some cases the snow cover lasts shorter than at lower elevations (0 - 500 m.a.s.l.). For the OWS analysis, several weather stations located close to forested and urban areas have been taken into consideration. In these cases, pixels around the in-situ measurement points have been considered. It means that there might be differences due to the slight variation of air temperatures. The altitude gradient in the melting season is well covered by the weather stations. The data used for OWS map validation are: OWS retrieval results on Sentinel-3 (W S-3), OWS retrieval results on MODIS (W MODIS), Snow Depth recorded at 6:00 UTC, in cm (SD), Snow Cover Extend - visual observation at 6:00 (SCE) and corresponding air temperature measurements in the morning (08:00), closest to the Sentinel-3 acquisition time (Ac – S-3), to the MODIS acquisition time (Ac – MODIS) and in the afternoon (14:00) for the four weather stations. All times are given in UTC. The retrieval results are shown colour coded as well as with letters (D = Dry, M = Moist, W = Wet, V = Very wet). When there is no OWS retrieval result, other classes are shown (‘+’ = Cloud, ‘-‘ = Partly snow-covered ground and ‘=’ = Bare ground (no snow)). Figures 16 (Sentinel-3) and 17 (MODIS) present the OWS maps for February 4 2017. The snow melting season arrived in the low elevations for Eastern and Southern parts of Romania. The temperatures are positive at almost all stations and all-day long (Figure 18). The only distinction is recorded at Titu weather station and may be explained by the temperature increase between the acquisition hours: from 4.6°C for Sentinel-3 (08:35 UTC) to 5.1°C for MODIS (10:10 UTC), (Table 6).

49

SnowBall

th

Figure 16: OWS map based on Sentinel-3 from 4 of February 2017 acquired at 08:35 UTC.

th

Figure 17: OWS map based on MODIS from 4 of February 2017 acquired at 10:10 UTC.

50

SnowBall

th

Figure 18: Temperature profiles at weather stations on February 4 2017. th

Table 6: Air temperature and snow cover characteristics at February 4 2017. 04.02.2017

Weather station

W S-3

W MODIS

SD (6:00)

SCE (6:00)

8:00

Ac. – S-3

Ac. – MODIS

14:00

Adamclisi Adjud Alexandria Braila Calarasi Cernavoda Constanta Cotnari Fetesti Giurgiu Mangalia Medgidia Negresti Oltenita Ploiesti Roman Slobozia Targoviste Tecuci Titu Urziceni Vaslui Videle

V V +/D V V = = = V D + V W D + +/D = = + M + V D

V + + + V = = = V D + V W D V + = = M W V D

4 3 11 7 6 No data