Retrieval of Forest Canopy Parameters by Inversion of ... - IEEE Xplore

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013

715

Retrieval of Forest Canopy Parameters by Inversion of the PROFLAIR Leaf-Canopy Reflectance Model Using the LUT Approach Khalid Omari, H. Peter White, Karl Staenz, and Douglas J. King

Abstract—The potential of simulating broad leaf forest canopy spectral reflectance using a canopy-leaf PROFLAIR (PROSPECT + FLAIR) model was investigated in this study. The model was inverted with hyperspectral Hyperion data using a look up table (LUT) approach to retrieve canopy leaf area index (LAI), leaf chlorophyll content and canopy integrated chlorophyll content . The LUT was populated by simulating the model in forward mode using a space of realization generated based on the specific distribution of the input parameters and based on a priori information from the field. The estimated variables were then compared to ground measurements collected in the field. The results showed the ability of the PROFLAIR model to realistically simulate canopy spectral reflectance. When compared to ground measurements, the model showed a reasonable performance to retrieve canopy LAI with an RMSE of 0.47 and leaf chlorophyll content with an RMSE of 4.46 g/cm . Index Terms—Environmental imaging, vegetation mapping.

remediation,

hyperspectral

I. INTRODUCTION

A

S hyperspectral technologies continue to develop towards regular and operational Earth Observation missions, there exists significant interest in developing models which can extract quantitative estimation of canopy biophysical properties such as leaf area index (LAI), fraction of absorbed photosynthetically active radiation (fAPAR), albedo and canopy biochemical properties. These vegetation parameters are primary inputs to models of land surface processes [37]. The accuracy and precision of their estimation is critical to successful quantification of the energy exchange between the atmosphere and terrestrial vegetation [21]. Fortunately, the increased availability of a variety of remote sensing satellite data types contributes effectively in deriving such parameters on a regular basis. Moreover, Manuscript received August 30, 2012; revised December 18, 2012; accepted January 04, 2013. Date of publication January 30, 2013; date of current version May 13, 2013. K. Omari was with the Ottawa-Carleton Geoscience Centre, Department of Earth Sciences, University of Ottawa, ON, Canada. He is now with the Canada Centre for Remote Sensing, Natural Resources Canada, Ottawa, ON, Canada. H. P. White is with the Canada Centre for Remote Sensing, Natural Resources Canada, Ottawa, ON, Canada (corresponding author, e-mail: [email protected]). K. Staenz is with the Alberta Terrestrial Imaging Centre (ATIC)/Department of Geography, University of Lethbridge, Lethbridge, AB, Canada. D. J. King is with the Department of Geography and Environmental Studies, Carleton University, Ottawa, ON, Canada. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JSTARS.2013.2240264

the spatial, spectral and directional dimensions of such data help in improving the accuracy of the estimated variables. There is a large variety of methods that have been developed to extract the biochemistry and biophysics of vegetation canopies from remote sensing data, particularly from hyperspectral imagery. Most studies have focused on the development of semi-empirical and empirical approaches in estimating these parameters [5], [7], [9], [16], [28], [38]. These approaches are mainly based on the derived relationships between these parameters and some particular combinations of spectral features (vegetation indices). Although, these approaches are computationally efficient and can handle large amounts of data, predictive estimates often cannot be inferred from these methods outside the domain in which the model data were collected. They are highly site and sensor specific and, therefore, less suitable for applications in large areas [14]. In contrast, great effort has been devoted to the development of radiative transfer models, geometric optical models, and computer simulation models [15], [33], [51]. These physically based approaches often estimate biophysical and biochemical parameters by describing in detail the interactions between the electromagnetic radiation and canopy elements [15], [43]. The accuracy of these models is generally related to the degree of detail of the description of the medium structure. A precise model tends to describe the way the electromagnetic radiation is scattered or absorbed by the different atmosphere-canopy-soil elements. Unfortunately, when the model precision increases its complexity increases as well. Its invertibility, therefore, becomes difficult to impossible (linearly) and other inversion strategies have to be developed [15]. Directly inverting a physically based canopy radiative transfer model can be mathematically and computationally challenging. With non-linear models the use of optimization inversion approaches are required. Theses methods rely on minimizing the cost function expressing the distance between observed and modeled reflectance values. The definition of suitable cost functions is critical to the success of the inversion process. These functions account for the uncertainties associated with both the model and the measurements. Model uncertainties contain, in part, the assumptions made on the structure of the canopy and the way the electromagnetic radiation interacts with the canopy elements [6]. The uncertainties associated with the measurements include the errors associated with the processing of remote sensing data (radiometric correction, atmospheric correction, etc). In practice, the use of the cost function is an iterative process based on searching the

1939-1404/$31.00 © 2013 British Crown Copyright

716

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013

combination of model inputs that can lead to the closest match between estimated and measured values. However, this process is often time consuming and sensitive to the initial guesses of the solution [34]. In addition, cost functions may trap in local minima because of the noisy nature of the observed data. The introduction of the priori knowledge information such as means and standard deviations of observed data as weights in the cost function may overcome the local minima issue [6]. The priori knowledge can also used to quantify the impact of each observed variable on the model retrieval. This method known as the contribution index (CI) accounts for both the uncertainty of the observed variable and its sensitivity to model parameters [53]. Common procedures of nonlinear models inversion are based on optimization approaches such as simplex or quasi-Newton methods [42], [49]. They are widely used for their simplicity and minimum constraints demand on the input parameters. However, these iterative approaches can converge to local minima rather than the desired solution. To avoid becoming trapped inside local minima, the initial guess of input parameters should be close to the real solution. Other techniques based on pre-computed values stored in databases such as look-up-tables (LUTs) and Neural Networks (NN) are the most popular optimisation inversion approaches used for retrieval of canopy biophysical variables. The NN approach is based on training Artificial Neural Networks (ANNs) according to databases previously generated [1], [46]. These relationships (networks) between reflectance values and canopy biophysical and biochemistry variables are then adapted to the actual situations. The NN method was previously applied to biophysical parameter retrieval from optical and microwave remote sensing data [2], [46]. The LUT approach, on the other hand, consists of running a radiative transfer model in forward mode to simulate canopy bidirectional reflectance factors (BRFs) from a wide range of canopy biophysical and biochemical variables. The resulting reflectance values are stored in a LUT. The measured reflectance values are then compared to the generated LUT using a minimizing function [6], [10], [11], [20], [21], [45]. The major advantage of the LUT approach is that it is not restricted to a given type of target cover [35]. Because of the ill-posed nature of the inverse problem [6], a unique inversion solution is not always achievable, that is, several canopy biophysical variables may yield similar spectral signatures [45]. Combal et al. [6] proposed the use of a priori information as an approach to solve such ill-posed problems. Ancillary information such as soil type, land-cover type, and canopy density can be used to constrain the model inversion for estimation canopy characteristics. The aim of this study is to evaluate the potential of the combined leaf optical Properties Spectra model PROSPECT [19] and the Four-Scale Linear Model for AnIsotropic Reflectance model (FLAIR) [31], [50], referred here as the PROFLAIR model, for the interpretation of hyperspectral data and the estimation of surface biophysical properties. FLAIR was initially designed to describe the angular behaviour of canopy reflectance [50]. It uses the physical description of canopy structure as outlined by the Four-Scale model [4]. The FLAIR model was advanced using a more physically robust multiple

scattering scheme and the bidirectional behaviour of the new model has been tested against multi-angular airborne data [31] acquired during the BOReal Ecosystem-Atmosphere Study (BOREAS) [36]. We are using the FLAIR model instead of other known radiative transfer models such as SAIL (Scattering from Arbitrarily Inclined Leaves) because of its physical basis in describing the canopy architecture [4], [50]. With PROFLAIR, leaf spectral properties generated by PROSPECT are used, in part, as FLAIR input, the use of this model to further investigate the biochemistry of the canopy such as Chlorophyll content. The following questions are addressed: • Can the coupled model PROFLAIR reproduce realistically measured canopy spectra using a set of input parameters measured in the field? • How well does the PROFLAIR model perform in retrieving biophysical characteristics of broad leaf forest canopies using the LUT inversion approach? PROFLAIR was tested using Hyperion data acquired over a forested area in the vicinity of a heavy metal acid tailings site. The LUT inversion approach was then used to retrieve LAI and canopy chlorophyll content. These two parameters are the focus of this study since they are key indicators toward ecosystem assessment [3]. The extracted parameters were evaluated against in-situ measurements acquired in the field. Assuming a reasonable performance of the modelling approach, PROFLAIR was used to verify the hypothesis that there is a gradient in canopy structural and chemical damage in the study site related to its adjacency to the abandoned mine tailings site. In Section II, the data set is described, and the model is tested in forward mode against field-measured data. Section III describes the adopted inversion approach and discusses the results of leaf chlorophyll content, LAI and canopy integrated chlorophyll content retrievals. II. DATA SETS AND METHODS A. Data Sets 1) In Situ Measurements: Ten study plots ( Fig. 1) were defined in a forest area adjacent to the abandoned Kam Kotia mine site near Timmins, Ontario, Canada. The area is relatively flat with mixed boreal forest species dominated by mature trembling aspen ( Populus tremuloides), a few pockets of co-dominant balsam poplar ( Populus balsamifera) and jack pine ( Pinus banksiana). The understory is composed by young black spruce ( Picea mariana), white spruce ( Picea glauca) and balsam fir ( Abies balsamea). Each of the ten plots was 60 m 60 m in size. They were selected along a previously defined transect within the forest [27] that has been demonstrated to exhibit a gradient of damage with increasing distance from the tailings edge [8], [47]. The study is limited to surrounding Aspen dominated forest in transect area because the sensitivity of of Aspen to airborne pollutants and heavy metal emissions present in the site [27]. The number of plot is limited to ten as the study area between the two creeks that starts from the mine tailings and follow the drainage direction covers only about 680 m. Six of the sites is a continuity to previous field campaigns [26], [27] and additional four sites were established along the transect. The

OMARI et al.: RETRIEVAL OF FOREST CANOPY PARAMETERS BY INVERSION OF THE PROFLAIR LEAF-CANOPY REFLECTANCE MODEL

717

Fig. 2. Outline of the PROFLAIR model.

Fig. 1. The study site plots shown in white on a 2004 air-photo. The drainage direction is in a north-northeast direction from the tailings source (yellow line).

plots are oriented in the same direction as the EO-1 Hyperion flight path. The plot size was chosen to minimize the effect of any misregistration that might be related to errors in geometric correction, and to contain the influence of all significant components of the radiative transfer theory described in the original FLAIR model as mentioned in Omari et al., [31]. Still, this minimization is relative since about 90 m 90 m would ensure that 1 full Hyperion pixel (30 m 30 m) is covered. However, 90 m 90 m requires further resources and, more important, the heterogeneous nature of forests in general and our study site in particular make it very difficult to find homogenous plots with such size. During a field campaign carried out on August 14, 2005, the dominant Trembling Aspen trees were sampled from each plot. A total of 16 leaf samples per plot (one to three samples per tree) were collected from the top of the stand for laboratory analysis. Leaf samples were immediately stored in plastic bags kept in cold environment for spectral measurements (reflectance and transmittance). Leaf samples were transported in frozen condition with dry ice to the laboratory for biochemical analysis (chlorophyll concentration and water content). The mean reflectance of plot leaf samples (both canopy foliage and understorey) was taken as the average of leaf sample reflectances measured using a LI-COR 1800 Integrating sphere coupled by a 200 Am diameter single mode fiber to an ASD field spectrometer with 1 nm sampling interval between 350 and 2500 nm. LAI and clumping index values of each plot were extracted from digital hemispherical photos (DHPs) taken in the upward direction

in each plot. Downward DHPs were also taken to extract the vegetation cover (f-cover) of the understory. The processing of DHPs was performed using CAN-EYE software [44]. 2) Hyperspectral Image Acquisition and Data Pre-Processing: A Hyperion image was acquired over the study site on August 23, 2005. The Hyperion sensor is on board NASA’s Earth Observing 1 (EO-1) satellite [32]. This technology demonstrator is the first spaceborne hyperspectral instrument to acquire visible and near-infrared (VNIR), and shortwave infrared (SWIR) spectral data. This sensor nominally provides 242 spectral bands, covering a range from 360 to 2600 nm at approximately 10 nm spectral resolution and 30 m ground sampling distance. Some of the bands, particularly those at the lower and upper ends of the Hyperion’s wavelength range, exhibit a poor signal-to-noise ratio [41]. As a result, about 192 of the 242 bands are useful. Hyperion consists of two distinct spectrometers, one with 70 bands covering the VNIR and the other one with 172 bands in the SWIR with some overlap in 1000 nm region. Processing of the hyperspectral data was carried out using the Imaging Spectrometer Data Analysis System (ISDAS), a software package developed at the Canada Centre for Remote Sensing [39]. The procedure for pre-processing Hyperion data is described in Khurshid et al. [23] and further detailed in Hitchcock and White [17]. There are four major corrections performed in order to derive reflectance at the canopy level. First is the detection and correction of the ‘smile-frown’ condition, which refers to a systematic across-track wavelength shift from the centre wavelength occurring in each image plane. In ISDAS, the detection of the band centre wavelength and its bandwidth are based on using atmospheric spectral absorption features for at-sensor radiance spectra [30]. Second, atmospheric correction is performed on the data using the radiative transfer model, MODTRAN 4.2, to retrieve at-surface reflectance from at-sensor radiance using a LUT approach [40]. This technique provides the extra benefit of deriving atmospheric water vapour on a per pixel basis and applying that information during atmospheric correction. Third, a de-striping technique was performed to remove systematic noise such as non-linear detector response. This is done by replacing the affected columns and pixels using the spatial and spectral information of adjacent pixels [41]. Finally, in order to improve the quality of the hyperspectral data, noise is reduced using an average smoothing procedure as outlined in Khurshid et al. [23].

718

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013

Fig. 3. An example showing the effects of leaf biochemical properties on canopy spectral reflectance simulated by the PROFLAIR model: (a) chlorophyll content using , and , (b) dry matter using , and , (c) using , and and (d) leaf structure (N) using , equivalent water thickness and . In all simulations LAI is set to 3 and view, illumination and phase angles are set to 5 , 40 and 0 , respectively.

B. PROFLAIR Model: Forward Mode In this study, the combined leaf-canopy model PROFLAIR is used to derive chlorophyll content at the canopy scale from hyperspectral data. PROSPECT is one of the most common leaf radiative transfer models used to investigate the effects of leaf biochemical and structural properties on leaf hemispherical reflectance and transmittance spectra [19]. An improved (1 nm resolution) and recalibrated version of PROSPECT, used in this study, has been validated and shown to achieve good accuracy in the VNIR and SWIR wavelengths (400–2500 nm) [12], [24]. The output spectrum is a function of leaf chlorophyll content , leaf water content , dry matter content and the structure parameter representing leaf anatomy. PROSPECT was combined with various canopy models in numerous studies to extract canopy and leaf biophysical and biochemical properties [18]. The FLAIR model was initially intended to describe the angular behaviour of canopy reflectance [31], [50]. It uses the physical description of canopy structure previously reported in the Four-Scale model [4]. To fully exploit all spectral information from hyperspectral data, a multiple scattering scheme was introduced recently into the FLAIR model [31]. This scheme is based on the decomposition of the multiple scattering radiation field into two parts. The first part deals with multiple scattering within the canopy considering a competently black soil and the second

part deals with multiple scattering between the canopy and the background. The bidirectional behaviour of the new model has been tested against multi-angular airborne data [31] using the BOReal Ecosystem-Atmosphere Study (BOREAS) campaign of 1994 [36]. Fig. 2 is a schematic representation of the coupled model. As can be seen, leaf spectral properties generated by PROSPECT using a set of leaf biochemical constituents are used, in part, as FLAIR input parameters to generate canopy spectral reflectance. Fig. 3 shows the effects of leaf biochemical constituents on the canopy spectral reflectance simulated with the PROFLAIR model. From this figure, the effect of each constituent is observable in specific wavelength ranges. For example, in Fig. 3(a), reflectance in the wavelength range between 500–720 nm responds to pigments in foliage and is lower for healthier leaves and higher for stressed ones. Moreover, the wavelengths 550 nm and 700 to 710 nm are the most adequate for foliage pigment monitoring [13]. These simulations may also help to pinpoint appropriate wavelengths for retrieval of other canopy parameters when inverting the model as described in Section III. The validation of PROFLAIR in forward mode was performed by running the model on a set of chlorophyll, water content, ground background spectral reflectance, structural canopy parameters, and illumination/view geometry data collected in the field. The derived spectral canopy BRFs were

OMARI et al.: RETRIEVAL OF FOREST CANOPY PARAMETERS BY INVERSION OF THE PROFLAIR LEAF-CANOPY REFLECTANCE MODEL

719

Fig. 5. Flow-chart representing the inversion strategy. Fig. 4. Example of simulated spectra generated with the PROFLAIR model with the incorporation of multiple scattering scheme (with MS) and without incorporating multiple scattering scheme (without MS) compared to Hyperion data, background reflectivity measured as a composite of 40% of black soil reflectance and 60% mean understorey samples reflectance and finally the foliage reflectance considered as the mean of leaf samples reflectance of the same plot.

then compared to Hyperion data acquired over the study site. To evaluate the performance of the coupled model, a set of data collected in the field was used. Fig. 4 presents the simulated spectral canopy reflectance produced with the PROFLAIR model and the corresponding Hyperion data. In this figure, the Hyperion spectral reflectance is an average of 3-by-3 pixels ( 90 m by 90 m) to reduce any uncertainties caused when image registration is performed. Also, a 3-by-3 average smoothing window was applied to Hyperion spectra to reduce residual radiometric noise. On the other hand, as previously mentioned, the collected ground measurement data were an average of 16 samples collected in a 60 m by 60 m area. This figure shows generally a good agreement between simulated and measured canopy spectral reflectances over the range of 500–1300 nm. It also demonstrates the benefit of incorporating the multiple scattering scheme of Omari et al. (2008a and b) for this spectral region (comparison between the PROFLAIR with and without the incorporation of multiple scattering component). The contribution of the multiple scattering component exceeds 50% of the total reflected radiation in the near infrared region. Fig. 4 also shows that estimated canopy-level reflectance is lower than the mean reflectance of plot leaf samples and the background reflectivity. This is expected as canopy-level reflectance includes shadows as well as vegetation. III. INVERSION STRATEGY The LUT was generated by simulating the top of canopy (TOC) reflectance using the view/illumination geometry of the Hyperion acquisition from a wide range of PROFLAIR input parameters using the protocol of Fig. 5. Some prior information on the range of variation of all the canopy variables and the type of soil is incorporated in order to obtain reliable and stable solutions [6]. In a first step, sampling the space of canopy realization was achieved by generating each model input parameter within a specific range with the assumption that these variables follow normal distributions ( Table I). A routine was designed and used to draw each variable values by providing its mean and standard deviation and the desired number of samples with the

normal distribution assumption ( Table I). Field data was used to determine mean and range (based on the standard deviation) of chlorophyll content and LAI. Two parameters are the focus of this study since they are key variables to assess the physiological health of vegetation as part of an ecosystem assessment [3]. Defaults values were considered for the other PROSPECT input parameters [12]. The spectral reflectance regions used in this study are not sensitive to changes by variations in theses parameters. The PROFLAIR model was then used to generate the corresponding reflectance and transmittance LUT. The generated 1 nm PROSPECT leaf reflectances and transmittances were convolved to the Hyperion spectral characteristics assuming that the Spectral Response Function (SRF) of Hyperion follows a Gaussian distribution [32]. Previous studies highlighted the effect of the background spectral properties on biophysical parameter retrieval (for example, see [29]). A proper background spectral reflectance measured in the field will enhance the retrieval. Moreover, it is difficult to know the exact spectral contribution of the background within each pixel when mapping these parameters. Meroni et al., [29] suggest that the accuracy of the retrieval is not dramatically compromised if standard background spectra are used in the inversion process. To limit the impact of the background on retrieval, three understorey vegetation samples that occur most frequently in each plot were collected, and an average background spectral reflectance per site was determined. The background reflectance was generated by combining a standard black soil reflectance and the average of the vegetation understorey reflectance . The selection of the type of bare soil was based on observation of the most frequently occurring to be typical of the forest area under investigation. The fraction of each component is considered by measuring the understorey f-cover from digital hemispherical photos taken in downward direction. Therefore, the background reflectance can be expressed as follows [1]:

(1) The illumination and view angles were derived from Hyperion acquisition characteristics as well as the date, time and location of the image acquisition. A total of 200 000 canopy realizations have been generated. Only selected spectral bands were

720

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013

TABLE I DISTRIBUTION OF THE PROFLAIR INPUT VARIABLES USED TO SIMULATE THE SPACE CANOPY REALIZATION USED TO POPULATE THE LOOK UP TABLE (LUT). HERE, IT IS ASSUMED THAT INPUT PARAMETERS FOLLOW NORMAL DISTRIBUTIONS. MEAN AND STANDARD DEVIATION VALUES ARE BASED ON PRIORI INFORMATION EXTRACTED FROM MEASUREMENTS

used according to Weiss et al., [45]. In the case of Hyperion, optimal spectral sampling is crucial to biophysical parameter estimation. Using the complete Hyperion spectra increases the noise and decreases retrieval performance. An approach similar to Weiss et al. [45] was followed by testing different band combinations with the root mean square error (RMSE) computed as the distance between measured and estimated canopy integrated chlorophyll content . The best combination corresponding to the minimum RMSE included bands are centered at 498.87, 630.32, 691.37, 711.72, 793.13, 844.00, and 915.23 nm. The inversion approach is then a process of finding the “closest” simulated reflectance to the measured ones. In this case, the corresponding set of input variables represents the solution of the problem. The (cost function) can be written as follows [45]: (2) and are the canopy reflectance acquired by where Hyperion and simulated using PROFLAIR (found in the LUT), respectively. The cost function is weighted by the number of selected wavelengths . The possible solutions considered were those corresponding to the median of the 10 lowest (as suggested in [6]). A second cost function to remove residual perturbations in the data as suggested by Combal et al. [6] was evaluated but was not found to improve the output of the results. The means and standard deviations of measured points were used, in this study, to determine the range of each parameter when populating the LUT. IV. RESULTS AND DISCUSSION Comparison between leaf chlorophyll content measured in the field and chlorophyll content estimated from the model is presented in Fig. 6. The horizontal errors bars represent one standard deviation for measured chlorophyll of all samples and the vertical bars represent the confidence range of 10%. The selection of this value was based on the agreement to the standard deviation of measurements. The results show that the LUT enables retrieval of chlorophyll content with an RMSE of 4.46 for the ten plots. This figure also shows that there is no trend in under or overestimation of chlorophyll as most of the points are approximately equally distributed above and under

Fig. 6. Comparison between measured and estimated chlorophyll content . Each measured value is an average of 16 samples taken from the top of the trees in a 60 by 60 m plot. The estimated values are the output of inverting the PROFLAIR model using Hyperion data. Each Hyperion spectra is an average of 3 by 3 pixels. The uncertainty bars indicate within canopy point variability, which is lower than inter-site variability. The low (estimated and measured) refers to the closest plot to the tailings and therefore the most affected plot by the deposition of particles transported by the blowing winds from the tailings.

the 1:1 line. A linear regression yielded (dashed line) a poor coefficient of determination . Pigment estimate uncertainties can be caused in part by the heterogeneity within the site. As mentioned previously, the sampled trees were limited to trembling aspen. This species is dominant in the study site with 50 to 80% of the cover. Other species including poplar and jack pine trees were not sampled. The degree of heterogeneity changes between plots as in most natural boreal forests. The understorey in this plot was thick and composed mostly of alder ( Alnus), raspberry ( Rubus) and jack pine trees 2 to 5 m in height. The impact of plot heterogeneity increases when taking into account that estimated chlorophyll is an average over 90 m by 90 m (3 3 average smoothing window applied on Hyperion image) compared to the sampled plot of 60 m by 60 m. Other sources of uncertainties propagate during the pre-processing steps of Hyperion data. These uncertainties have been documented during the atmospheric and sensor artefacts corrections [23], [30], [41]. The sampling scheme may also contribute to these estimate uncertainties. As an example, a high standard deviation in chlorophyll measurement is shown within each plot

OMARI et al.: RETRIEVAL OF FOREST CANOPY PARAMETERS BY INVERSION OF THE PROFLAIR LEAF-CANOPY REFLECTANCE MODEL

Fig. 7. Comparison between measured and estimated leaf area index (LAI). Each measured value was extracted from digital hemispherical photos of a 60 by 60 m plot. The estimated values are the output of inverting the PROFLAIR model using Hyperion data. Each Hyperion spectra is an average of 3 by 3 pixels.

and even between samples taken from the same tree. The temporal shift of 9 days between the date of acquisition of the Hyperion data and the date of field measurements has a small effect since both dates were during the peak LAI period several weeks before senescence. It also has been reported that inverting PROSPECT to retrieve the biochemical parameters with good accuracy is not always achievable due to large within plot variability [25]. To evaluate the performance of LAI retrieval, modelled LAI values were compared to those measured in the field as illustrated in Fig. 7. This Figure shows good agreement, with an RMSE of 0.47 and an . Fig. 7 also shows a slight overestimation of LAI values. The uncertainties may be explained by the underestimation of LAI because of foliage clumping or/and the possibility that the model is taking the understorey into account as well. Field measurements, such as acquiring and processing digital hemispherical photos, may also have contributed to these errors even though comprehensive procedures to optimize acquiring and processing the data were followed. Canopy integrated chlorophyll content can be estimated as . Fig. 8 illustrates the comparison between measured and simulated canopy integrated chlorophyll content showing an RMSE of 24.44 and an of 0.60. In general, these results show the potential of using the PROFLAIR model in estimating broad leaf forest canopy biophysical characteristics. The accuracies of the chlorophyll and LAI estimates fall within the validation range of other physically based models tested on broadband multispectral and hyperspectral data [1], [22], [29], [52]. The ten plots ( Fig. 1) were established along a transect within the forest with an increasing distance from the tailings and following the drainage direction [26]. Fig. 9 is an attempt to sense a potential gradient of damage along the transect by estimating the canopy integrated chlorophyll content in the field plots. This Figure shows a slight increase of canopy integrated chlorophyll with an increasing distance from the tailings pond and following the drainage direction. The lack of change of leaf chlorophyll

721

Fig. 8. Comparison between measured and estimated canopy integrated by inverting PROFLAIR with LUT chlorophyll content approach.

Fig. 9. Estimated canopy integrated chlorophyll content alongside the defined transect.

content along the transect suggest that the trend shown in Fig. 9 is mainly due to a structural change (LAI) and that the leaf chlorophyll content information contributed very little to this trend. Also, it was noted during this study that overstorey LAI and canopy clumping was more impacted at sites closer to the tailings site. Chlorophyll analysis on sampled leaves suggest generally that the physiological status of the plots varied from healthy to very healthy suggesting that the damage at this site is structural rather than physiological. This confirms previous results in the same site using high spatial resolution multispectral imagery (Lèvesque and King, 1998 and 2003; [8]), suggesting that the damage is structural rather than physiological. Their finding implies that the leaf reflectance is too varied within and between crowns to see a subtle gradient. They also found that aspen trees are resistant to metal and pH stress, even though the soil analyses showed low pH and high concentration of Al, Mg and Mn. Many of these studies have demonstrated that within and between crown shadows and texture can provide valuable indications of the stress at this site. High stress results in a more open canopy structure near the tailings, with the forest floor cover becoming dominated with understory vegetation such as alders and raspberry. These produce a reflectance signal that mixes heavily with the overstorey and falsely gives the impression, in quantitative analysis, that the forest is healthier.

722

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 6, NO. 2, APRIL 2013

V. CONCLUSION In this study, the potential of the PROFLAIR model was investigated to retrieve biochemical and biophysical characteristics of broadleaf forest canopy from hyperspectral Hyperion imagery. This linked model showed its ability to simulate realistic forest canopy spectra if properly parameterized. The simulated spectra by the PROFLAIR model were similar in magnitude and detail to those extracted from Hyperion in the VNIR and SWIR regions used in this study. The look up table (LUT) approach was employed to invert the model; it was found that this approach is flexible and easy to implement. The LUT was populated by running the model for a space of canopy realization generated by randomly involving each input model variable. Some a priori information about the data and the canopy structure were used to optimize the size of the LUT, and a cost function was used to retrieve leaf area index (LAI) and chlorophyll . The canopy integrated chlorophyll content was content then deduced for selected wavelengths. The results showed an agreement between estimated and measured LAI with a RMSE of 0.47 and an of 0.58. This is an acceptable agreement when compared to other models’ performance. The results also inand a of 0.25 between esdicate a RMSE of 4.46 timated and measured chlorophyll content. As previously discussed, uncertainties are influenced in part by atmospheric correction process and within plot variability. In spite of that, further improvement of the PROFLAIR model, such as enhancing the hot spot function and the anisotropy of the background, is needed. Also, an effort has to be made in improving field data collection, particularly the design of the sampling scheme, number of plots and number of sampled trees per plot. In addition, enhancing the pre-processing techniques of hyperspectral data is critical for better retrievals. The growing interest placed in developing hyperspectral technologies and initiating hyperspectral missions in several organizations will make high quality and fully exploitable data available. This will certainly enhance the quantitative estimation of vegetation biochemistry and biophysics. ACKNOWLEDGMENT The authors would also like to thank Richard Fernandes, Marie Weiss and Nadia Rochdi for their productive discussions and contributions, field crew members for their help in acquiring field data and the anonymous reviewers for their constructive comments and assistance in editing resulting in completion of this paper. REFERENCES [1] C. Bacour, F. Baret, D. Béal, M. Weiss, and K. Pavageau, “Neural network estimation of LAI, fAPAR, fCover and , from top of canopy MERIS reflectance data: Principles and validation,” Remote Sens. Environ., vol. 105, no. 4, pp. 313–325, 2006. [2] F. Baret, J. Clevers, and M. Steven, “The robustness of canopy gap fraction estimates from red and near infrared reflectances: A comparison of approaches,” Remote Sens. Environ., vol. 54, pp. 141–151, 1995. [3] G. A. Carter, “Ratios of leaf reflectances in narrow wavebands as indicators of plant stress,” Int. J. Remote Sens., vol. 15, pp. 697–703, 1994. [4] J. M. Chen and S. Leblanc, “A 4-scale bidirectional reflection model based on canopy architecture,” IEEE Trans. Geosci. Remote Sens., vol. 35, pp. 1316–1337, 1997. [5] J. G. P. W. Clevers and L. Kooistra, “Using hyperspectral remote sensing data for retrieving canopy cholorophyll and nitrogen content,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 5, no. 2, pp. 574–583, 2012.

[6] B. Combal, F. Baret, M. Weiss, A. Trubuil, D. Macé, A. Pragnère, R. Myneni, Y. Knyazikhin, and L. Wang, “Retrieval of canopy biophysical variables from bidirectional reflectance using prior information to solve the ill-posed inverse problem,” Remote Sens. Environ., vol. 84, pp. 1–15, 2002. [7] N. C. Coops, M. L. Smith, M. E. Martin, and S. Ollinger, “Prediction of eucalypt foliage nitrogen content from satellite-derived hyperspectral data,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 6, pp. 1338–1346, 2003. [8] P. Cosmopoulos and D. J. King, “Temporal analysis of forest structural condition at an acid mine site using multispectral digital camera imagery,” Int. J. Remote Sens., vol. 25, no. 12, pp. 2259–2275, 2004. [9] P. J. Curran, J. A. Kupiec, and G. M. Smith, “Remote sensing the biochemical composition of slash pine canopy,” IEEE Trans. Geosci. Remote Sens., vol. 35, no. 2, pp. 415–420, 1997. [10] R. Darvishzadeh, A. A. Matkan, and A. D. Ahangar, “Inversion of radiative transfer model for estimation of rice canopy chlorophyll content using lookup-table approach,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 5, no. 4, pp. 1222–1230, 2012. [11] W. A. Dorigo, “Improving the robustness of cotton status characterisation by radiative transfer model inversion of multi-angular CHRIS/ PROBA data,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 5, no. 1, pp. 18–29, 2012. [12] J. B. Féret, C. François, G. P. Asner, A. A. Gitelson, R. E. Martin, L. P. R. Bidel, S. L. Ustin, G. le Maire, and S. Jacquemoud, “PROSPECT-4 and 5: Advances in the leaf optical properties model separating photosynthetic pigments,” Remote Sens. Environ., vol. 112, pp. 3030–3043, 2008. [13] A. A. Gitelson, U. Gritz, and M. N. Merzlyak, “Relationships between leaf chlorophyll content and spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant leaves,” J. Plant Physiol., vol. 160, pp. 271–282, 2003. [14] N. Gobron, B. Pinty, M. M. Verstraete, and Y. Govaerts, “A semi-discrete model for the scattering of light by vegetation,” J. Geophys. Res., vol. 102, pp. 9431–9446, 1997. [15] N. S. Goel, “Models of vegetation canopy reflectance and their use in estimation of biophysical parameters from reflectance data,” Remote Sens. Rev., vol. 4, pp. 1–212, 1988. [16] P. Gong, R. Pu, G. S. Biging, and M. R. Larrieu, “Estimation of forest leaf area index using vegetation indices derived from hyperion hyperspectral data,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 6, pp. 1355–1362, 2003. [17] R. Hitchcock and H. P. White, Processing EO-1 Hyperion Data Using ISDAS, Geomatics, Canada, Technical Note 2, 2007, 30 pages. [18] S. Jacquemoud, C. Bacour, H. Poilve, and J. P. Frangi, “Comparison of four radiative transfer models to simulate plant canopies reflectance direct and inverse mode,” Remote Sens. Environ., vol. 74, pp. 471–481, 2000. [19] S. Jacquemoud and F. F. Baret, “PROSPECT: A model of leaf optical properties spectra,” Remote Sens. Environ., vol. 34, pp. 75–91, 1990. [20] D. Kimes, Y. Knyazikhin, J. L. Privette, A. A. Abuelgasim, and F. Gao, “Inversion methods for physically-based models,” Remote Sens. Rev., vol. 18, pp. 381–439, 2000. [21] Y. Knyazikhin, J. V. Martonchik, R. B. Myneni, D. J. Diner, and S. W. Running, “Synergistic algorithm for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from MODIS and MISR data,” J. Geophys. Res., vol. 103, no. D24, pp. 32257–32276, 1998. [22] B. Koetz, F. Baret, H. Poilvé, and J. Hill, “Use of coupled canopy structure dynamic and radiative transfer models to estimate biophysical canopy characteristics,” Remote Sens. Environ., vol. 95, no. 1, pp. 115–124, 2005. [23] S. Khurshid, K. Staenz, L. Sun, R. A. Neville, H. P. White, A. Bannari, C. M. Champagne, and R. Hitchcock, “Preprocessing of EO-1 hyperion data,” Can. J. Remote Sens., vol. 32, no. 3, pp. 84–97, 2006. [24] G. Le Maire, C. François, K. Soudani, D. Berveiller, J.-Y. Pntailler, N. Breda, H. Genet, H. Davi, and E. Dufrêne, “Calibration and validation of hyperspectral indices for the estimation of broadleaved forest leaf chlorophyll content, leaf mass per area, leaf area index and leaf canopy biomass,” Remote Sens. Environ., vol. 112, pp. 3846–3864, 2008. [25] G. Le Maire, C. Francois, and E. Dufrene, “Towards universal broad leaf chlorophyll indices using PROSPECT simulated database and Hyperspectral reflectance measurements,” Remote Sens. Environ., vol. 89, pp. 1–28, 2004. [26] J. Lévesque and D. J. King, “Spatial analysis of radiometric fractions from high-resolution multispectral imagery for modeling individual tree crown and forest canopy structure and health,” Remote Sens. Environ., vol. 84, no. 4, pp. 589–602, 2003.

OMARI et al.: RETRIEVAL OF FOREST CANOPY PARAMETERS BY INVERSION OF THE PROFLAIR LEAF-CANOPY REFLECTANCE MODEL

[27] J. Lévesque and D. J. King, “Airborne digital camera image semivariance for evaluation of forest structural damage at an acid mine site,” Remote Sens. Environ., vol. 68, pp. 112–124, 1999. [28] M. E. Martin and J. D. Aber, “High spectral resolution remote sensing of forest canopy lignin, nitrogen, and ecosystem processes,” Ecol. Applicat., vol. 7, no. 2, pp. 431–443, 1997. [29] M. Meroni, R. Colombo, and C. Panigada, “Inversion of a radiative transfer model with hyperspectral observations for LAI mapping in poplar plantations,” Remote Sens. Environ., vol. 92, pp. 195–206, 2004. [30] R. A. Neville, L. Sun, and K. Staenz, “Spectral calibration of imaging spectrometers by atmosphere absorption matching,” Can. J. Remote Sens., vol. 34, pp. S29–S42, 2008, Suppl. 1. [31] K. Omari, H. P. White, and K. Staenz, “An enhanced description of multiple scattering within the FLAIR model using the photon re-collision probability approach,” IEEE Trans. Geosci. Remote Sens., vol. 47, no. 8, pp. 2931–2941, 2009. [32] J. S. Pearlman, P. S. Barry, C. C. Segal, J. Shepanski, D. Beiso, and S. L. Carman, “Hyperion, a space-based imaging spectrometer,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 6, pp. 1160–1173, 2003. [33] B. Pinty, J.-L. Widlowski, N. Taberner, N. Gobron, M. M. Verstraete, M. Disney, J.-P. Gastellu-Etchegorry, L. Jiang, A. Kuusk, P. Lewis, X. Li, W. Ni-Meister, T. Nilson, P. R. J. North, W. Qin, L. Su, R. Thomson, W. Verhoef, H. Wang, J. Wang, G. Yan, and H. Zhang, “Radiation Transfer Model Intercomparison (RAMI) exercise: Results from the second phase,” J. Geophys. Res., vol. 109, no. D06210, 2004, doi 10.1029/2003JD004252. [34] B. Pinty and M. M. Verstraete, “Extracting information on surface properties from bidirectional reflectance measurements,” J. Geophys. Res., vol. 96, no. D2, pp. 2865–2874, 1991. [35] G. Robert, “A review of the application of BRDF models to infer land cover parameters at regional and global scales,” Progr. Phys. Geogr., vol. 25, no. 4, pp. 483–511, 2001. [36] P. Sellers, F. Hall, H. Margolis, B. Kelly, D. Baldocchi, G. Hartog den, J. Cihlar, M. G. Ryan, B. Goodison, P. Crill, K. J. Ranson, D. Lettenmaier, and D. E. Wickland, “The Boreal Ecosystem-Atmosphere Study (BOREAS): An overview and early results from the 1994 field year,” Bull. Amer. Meteorolog. Soc., vol. 76, no. 9, pp. 1549–1577, 1995. [37] P. J. Sellers, “Canopy reflectance, photosynthesis and transpiration,” Int. J. Remote Sens., vol. 6, no. 8, pp. 1335–1372, 1985. [38] M. L. Smith, M. E. Martin, L. Plourde, and S. V. Ollinger, “Analysis of hyperspectral data forest canopy nitrogen concentration: Comparison between an airborne (AVIRIS) and a spaceborne (Hyperion) sensor,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 6, pp. 1332–1337, 2003. [39] K. Staenz, T. Szeredi, and J. Schwarz, “ISDAS-A system for processing/analyzing hyperspectral data,” Can. J. Remote Sens., vol. 42, no. 2, pp. 99–113, 1998. [40] K. Staenz and D. Williams, “Retrieval of surface reflectance from hyperspectral data using a look-up table approach,” Can. J. Remote Sens., vol. 23, no. 4, pp. 354–368, 1997. [41] L. Sun, R. A. Neville, K. Staenz, and H. P. White, “Automatic destriping of hyperion imagery based on spectral moment matching,” Can. J. Remote Sens., vol. 34, pp. S68–S81, 2008, Suppl. 1. [42] A. Tarantola, Inverse Problem Theory. Methods for Data Fitting and Model Parameter Estimation. Amsterdam, The Netherlands: Elsevier Science, 1987. [43] W. Verhoef, “Light scattering by leaf layers with application to canopy reflectance modeling: The SAIL model,” Remote Sens. Environ., vol. 16, pp. 125–141, 1984. [44] M. Weiss, CAN-EYE User Guide. Toulouse, France: NOV-3075-NT1260, NOVELTIS, 2002. [45] M. Weiss, F. Baret, R. B. Myneni, A. Pragnere, and Y. Knyazikhin, “Investigation of a model inversion technique to estimate canopy biophysical variables from spectral and directional reflectance data,” Agronomie, vol. 20, pp. 3–22, 2000. [46] M. Weiss and F. Baret, “Evaluation of canopy biophysical variable retrieval performances from the accumulation of large swath satellite data,” Remote Sens. Environ., vol. 70, pp. 293–306, 1999. [47] H. P. White and A. Abuelgasim, “Sustainable management and rehabilitation of mine sites for decision support—Remote sensing innovations and applications,” Geomatics Canada, Bulletin 1, 2012, in press. [48] H. P. White, H. Dickinson, A. Trebble, and T. Alfoldi, “Presenting a web based evaluation tool to interactively explore the relationship between observed bi-directional reflectance and canopy characteristics,” in Proc. 26th Can. Symp. Remote Sens., Wolfville, Nova Scotia, Jun. 14–17, 2005, pp. 451–455.

723

[49] H. P. White, L. Sun, K. Staenz, R. Fernandes, and C. M. Champagne, “Determining the contribution of shaded elements of a canopy to remotely sensed hyperspectral signatures,” in Proc. 1st Int. Symp. Recent Advances in Quantitative Remote Sensing, Valencia, Spain, 2002a, pp. 159–166. [50] H. P. White, J. R. Miller, and J. M. Chen, “Four-scale linear model for anisotropic reflectance (FLAIR) for plant canopies. I: Model description and partial validation,” IEEE Trans. Geosci. Remote Sens., vol. 39, no. 5, pp. 1072–1083, 2001. [51] J.-L. Widlowski, M. Taberner, B. Pinty, V. Bruniquel-Pinel, M. Disney, R. Fernandes, J.-P. Gastellu-Etchegorry, N. Gobron, A. Kuusk, T. Lavergne, S. Leblanc, P. E. Lewis, E. Martin, M. Mõttus, P. R. J. North, W. Qin, M. Robustelli, N. Rochdi, R. Ruiloba, C. Soler, R. Thompson, W. Verhoef, M. M. Verstraete, and D. Xie, “The third RAdiation transfer Model Intercomparison (RAMI) exercise: Documenting progress in canopy reflectance modelling,” J. Geophys. Res., vol. 112, 2007, doi 10.1029/2006JD007821. [52] P. J. Zarco-Tejada, J. R. Miller, T. L. Noland, G. H. Mohammed, and P. H. Sampson, “Scaling-up and model inversion methods with narrowband optical indices for chlorophyll content estimation in closed forest canopies with hyperspectral data,” IEEE Trans. Geosci. Remote Sens., vol. 39, no. 7, pp. 1491–1506, 2001. [53] K. Zhang, B. Hu, J. Wang, E. Pattey, and A. Smith, “ Improving the retrieval of the biophysical parameters of vegetation canopies using the contribution index,” Can. J. Remote Sens., vol. 37, no. 6, pp. 643–652, 2011. Khalid Omari received the B.Sc. degree in physics from the University of KadiAyyad, Marrakech, Morocco, in 1996, and is working toward the Ph.D. from the Ottawa-Carleton Geoscience Center, Department of Earth Sciences, University of Ottawa, Ottawa, ON, Canada. His Ph.D. dissertation focused on radiative transfer modeling in vegetation canopies and the extraction of biophysical and biochemical parameters of vegetation canopies from hyperspectral data. H. Peter White (M’02) received a B.Sc. degree in physics and astronomy from Saint Mary’s University, Halifax, NS, Canada, in 1988, and M.Sc. and Ph.D. degrees in physics and astronomy from York University, Toronto, ON, Canada, in 1994 and 1999, respectively. He is currently a Research Scientist with the Canadian Centre for Remote Sensing, Natural Resources Canada, Ottawa, ON, Canada. His remote sensing interests involve using hyperspectral remote sensing for modelling optical bidirectional reflectance of canopies, environmental monitoring, and geological exploration. Karl Staenz received the M.Sc. and Ph.D. degrees in geography from the University of Zurich, Zurich, Switzerland, in 1976 and 1978, respectively. After 20 years as a Research Scientist with the Canada Centre for Remote Sensing, Ottawa, ON, Canada, he joined the University of Lethbridge, Lethbridge, AB, Canada, where he is currently a Professor for remote sensing and the Science Director of the Alberta Terrestrial Imaging Center. His research over the years has focused on hyperspectral remote sensing, particularly in the development of methodologies and their application to natural resources and environment inventorying and monitoring.

Douglas J. King is presently Professor and Chair of the Dept. of Geography and Environmental Studies at Carleton University, Ottawa, ON. His research focuses on airborne and high resolution satellite (optical and radar) remote sensing, with additional larger area studies using moderate resolution imagery such as SPOT and Landsat. This is pursued to develop remote sensing and geo-spatial methods for modelling, mapping and monitoring forests, wetlands, and other critical habitat areas such as shorelines at landscape and patch scales.