A radiative transfer model-based method for the estimation of grassland aboveground biomass

International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag

A radiative transfer model-based method for the estimation of grassland aboveground biomass Xingwen Quan a , Binbin He a,∗ , Marta Yebra b,c , Changming Yin a , Zhanmang Liao a , Xueting Zhang a , Xing Li a a

School of Resources and Environment, University of Electronic Science and Technology of China, Chengdu, Sichuan 611731, China Fenner School of Environment and Society, The Australian National University, ACT, Canberra, Australia c Bushfire & Natural Hazards Cooperative Research Centre, Melbourne, Australia b

a r t i c l e

i n f o

Article history: Received 25 April 2016 Received in revised form 14 August 2016 Accepted 3 October 2016 Keywords: Grassland aboveground biomass Landsat 8 OLI product Leaf area index PROSAILH Ill-posed inversion problem

a b s t r a c t This paper presents a novel method to derive grassland aboveground biomass (AGB) based on the PROSAILH (PROSPECT + SAILH) radiative transfer model (RTM). Two variables, leaf area index (LAI, m2 m−2 , defined as a one-side leaf area per unit of horizontal ground area) and dry matter content (DMC, gcm−2 , defined as the dry matter per leaf area), were retrieved using PROSAILH and reflectance data from Landsat 8 OLI product. The result of LAI × DMC was regarded as the estimated grassland AGB according to their definitions. The well-known ill-posed inversion problem when inverting PROSAILH was alleviated using ecological criteria to constrain the simulation scenario and therefore the number of simulated spectra. A case study of the presented method was applied to a plateau grassland in China to estimate its AGB. The results were compared to those obtained using an exponential regression, a partial least squares regression (PLSR) and an artificial neural networks (ANN). The RTM-based method offered higher accuracy (R2 = 0.64 and RMSE = 42.67 gm−2 ) than the exponential regression (R2 = 0.48 and RMSE = 41.65 gm−2 ) and the ANN (R2 = 0.43 and RMSE = 46.26 gm−2 ). However, the proposed method offered similar performance than PLSR as presented better determination coefficient than PLSR (R2 = 0.55) but higher RMSE (RMSE = 37.79 gm−2 ). Although it is still necessary to test these methodologies in other areas, the RTMbased method offers greater robustness and reproducibility to estimate grassland AGB at large scale without the need to collect field measurements and therefore is considered the most promising methodology. © 2016 Elsevier B.V. All rights reserved.

1. Introduction Aboveground biomass (AGB) determines biosphere–atmosphere interactions, and it is key to our understanding of the terrestrial carbon balance (Anaya et al., 2009; Cartus et al., 2012; He et al., 2015; Houghton, 2005; Liu et al., 2015; Su et al., 2016). A spatial and temporal assessment of AGB at different stages can be used to determine which processes drive changes in the global carbon cycle and can help land managers to develop strategies for climate change mitigation (Yan et al., 2015). Although field surveys provide the most accurate method for obtaining grassland AGB, they are too time-consuming and costly over large areas (Paul et al., 2013; Xie et al., 2009). Remote sensing can provide a uniquely effective and efficient means of achieving

∗ Corresponding author. E-mail address: [email protected] (B. He). http://dx.doi.org/10.1016/j.jag.2016.10.002 0303-2434/© 2016 Elsevier B.V. All rights reserved.

this end due to its high temporal and spatial resolution image of large landscape observation (Barrachina et al., 2015; Chen et al., 2015; Fassnacht et al., 2014; Lu, 2006). Different methods exist to estimate AGB using different remote sensing data: (i) light detection and ranging (LiDAR) data (Chen, 2015; Tsui et al., 2012), (ii) synthetic aperture radar (SAR) data (Baghdadi et al., 2015; Englhart et al., 2011; Liu et al., 2015), (iii) optical satellite data (Cho et al., 2007; Cho and Skidmore, 2009; Ramoelo et al., 2015; Doraiswamy et al., 2005; Liu et al., 2010; Tian et al., 2012), and (iv) multi-sensor data (Clevers and van Leeuwen, 1996; Koch, 2010; Li et al., 2015; Su et al., 2016; Zhang et al., 2014). LiDAR is an active sensing technology which uses a laser (light amplification by stimulated emission of radiation) to transmit a light pulse towards a target and a receiver to measure the backscattered or reflected light from that target (Cho et al., 2012; Lefsky et al., 2005). The LiDAR data have shown great potential for the retrieval of vegetation biophysical parameters that are largely related to AGB, such as vegetation height, volume and structure

160

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

(Lefsky et al., 2005; Zolkos et al., 2013). However, LiDAR data are generally used for forest and shrub AGB mapping (Boudreau et al., 2008; Cho et al., 2012; Lefsky et al., 2005; Zolkos et al., 2013) and as far as we know there are not studies focused on grassland AGB. Moreover, it is not suitable in applications for large regions because wall-to-wall coverage of large areas with LiDAR is impractical (Cartus et al., 2012; Liu et al., 2015). SAR data also showed great potential for AGB mapping by using its intensity information, polarized information and phase information, such as the development of polarimetric SAR (PolSAR), SAR interferometry (InSAR) and polarimetric SAR interferometry (PolInSAR) (Cartus et al., 2012; Cloude and Papathanassiou, 1998; Liu et al., 2015; Solberg et al., 2013). Svoray and Shoshany (2002) derived biomass in a semi-arid grassland region by modifying the water-cloud model. Moreau and Le Toan (2003) utilized SAR data as a function of plant biomass at homogeneous areas. However, the use of SAR data remains challenging due to its high costs and lack generality for the application of the empirically based equations to large scale. Moreover, the SAR data are hypersensitive to the underlying surface in grasslands, which will sufficiently influence the accuracy of AGB estimations (Hajj et al., 2014). With regard to the use of optical satellite data to estimate AGB, the primary methods are empirical approaches, including vegetation indices regression (Anderson et al., 1993; Zheng et al., 2004, 2007), partial least squares regression (PLSR) (Cho et al., 2007), artificial neural networks (ANN) (Xie et al., 2009) and machine learning algorithms (Clevers et al., 2007). The vegetation indices regression method was to establish the relationship (such as linear fitting, exponential regression, polynomial regression, etc.) between vegetation indices and vegetation variables of interest, and with this relationship, the target variables can be derived. PLSR is a method for relating two data matrices, X and Y, by a linear multivariate model, but goes beyond traditional regression in that it models also the structure of X and Y. This method derives its usefulness from its ability to analyze data with many, noisy, collinear, and even incomplete variables in both X and Y (Wold et al., 2001). The ANN and machine learning algorithms need a training database consisting of canopy reflectance spectra together with the corresponding vegetation parameters, and their performance largely relies on the training database and the training process itself (Houborg et al., 2009). Generally, these methods are empirical and indirect (Koch, 2010) as they relate AGB with other directly retrieved vegetation parameters, such as height and crown closure. Consequently, these methods are limited to a certain region and time and, being indirect, they may introduce extra uncertainties in the estimation of AGB. Another method for the estimation of AGB is the modeldata assimilation approach which incorporates field and multiple remote sensing data into dynamic mechanistic models, such as the CERES-Wheat (Godwin et al., 1989) and World Food Studies (WOFOST) model (Diepen et al., 1989; Ma et al., 2013a, 2013b). This approach has increasingly been used for crop growth monitoring and AGB or yield prediction, with considerable success (He et al., 2015; Huang et al., 2016, 2015a). However, these dynamic mechanistic models are hard to parameterize due to the large requirement of input parameters, and the iterative optimizing process is time-consuming, especially for the four-dimensional variational algorithm (4D-Var) (Quan et al., 2015b; Talagrand and Courtier, 1987). Optical satellite data and radiative transfer model (RTM) inversion techniques have been widely used to retrieve vegetation biophysical and biochemical variables, such as leaf area index (LAI) (Houborg et al., 2007; Quan et al., 2014), canopy water content (Quan et al., 2015a), canopy or leaf chlorophyll content (Darvishzadeh et al., 2008b; Yin et al., 2016) and fuel moisture content (Quan et al., 2015b; Yebra and Chuvieco, 2009b; Yebra et al.,

2013). RTM inversion techniques have proven to be a promising way to retrieve bio-physical and bio-chemical variables because compared with the empirical methods, RTM are more universal as they are based on physical laws that provide explicit relations between canopy properties and spectra (Houborg et al., 2009, 2007; Quan et al., 2015a; Yebra et al., 2013, 2008). Thus, these RTMbased approaches have the advantage of reproducibility. However, to date, no study has explored the use of RTM inversion techniques for the estimation of grassland AGB. In this paper, a novel method based on the PROSAILH (PROSPECT and SAILH) RTM (hereafter referred as RTM-based method) was explored to estimate grassland AGB from two model parameters: LAI (m2 m−2 , defined as a one-side leaf area per unit of horizontal ground area) and dry matter content (DMC, gcm−2 , defined as the dry matter per leaf area). To test the performance of this method vs. the traditional empirical methods, the exponential regression, PLSR and ANN were also implemented. A case study of the proposed methods was applied to a plateau grassland in China to estimate its AGB, and the results were validated using the field measurements. 2. Materials and methods 2.1. Study area and data 2.1.1. Study area and sampling The study area is the Qinghai Lake watershed, located in Qinghai Province, China (36◦ 15 –38◦ 20 N, 97◦ 50 –101◦ 22 E) (Fig. 1). This watershed is a closed inland basin surrounded by mountains, with an area of approximately 29,600 km2 . The watershed ranges in elevation from 3194 m to 5174 m with annual mean temperatures between −1.10 ◦ C and 0.80 ◦ C. The annual precipitation is between 324.50 mm and 412.80 mm, and the majority of the precipitation falls during the period from May to September. Due to its unique geographic location, geomorphic features, climate conditions and saline-alkali soil, the Qinghai Lake watershed forms a complex habitat with diverse grass species. The field surveys were carried out in late July 2014 and early August 2015 in collaboration with the Qinghai Ecosystem Remote Sensing Monitoring Centre (http://www.qherc.org/). The sampling plots were selected based on the image of 1:100,0000 grassland cover types in Qinghai province, China. A total of 135 30 × 30 m plots were sampled. A GPS was used to locate their geographical positions. For each plot, LAI was obtained from fish-eye photographs and the CAN-EYE V6.3.8 analysis software. Three pictures were taken in the diagonal direction of each plot. In each plot, 3 subplots (0.5 × 0.5 m) were randomly selected to destructively sample the aboveground grass by removing all grass to the ground level. The collected samples were transported to the laboratory, oven-dried for 24 h at 105 ◦ C (Matthews, 2010) and weighed (dry weight). 2.1.2. Satellite data and pre-processing Landsat 8 Operational Land Imager (OLI) products acquired within the field survey periods were used as the source of reflectance data to carry out the RTM inversion. The Landsat 8 OLI sensor images the Earth every 16 days at a pixel size of 30 m × 30 m (same size as the field plots). The data were downloaded from the United States Geological Survey (USGS) (http://glovis.usgs.gov/). A total of five Landsat-8 OLI scenes per date were needed to completely cover the Qinghai Lake watershed. Only images covered by less than 70% cloud were selected and downloaded. The images were atmospherically corrected using the FLAASH tool in the ENVI (version 5.2) image processing software (Matthew et al., 2000). FLAASH is structured based on the MODTRAN (MODerate resolution atmospheric TRANsmission) atmospheric RTM, which can be

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

161

Fig. 1. Map showing the location of the study area and the sampling plots.

used to convert the digital number (DN) values of these images into surface reflectance after being transformed to the top of atmosphere radiance. A mid-latitude summer atmospheric model, rural aerosol type with a visibility of 40 km and an average water content of 1.00 gcm−2 were specified to run the atmospheric correction model. Cloud and cloud shadow in each of Landsat-8 OLI images were detected, and were masked using the FMASK algorithm (Zhu and Woodcock, 2012). Vegetation indices have exhibited the advantage of a high sensitivity to vegetation and can reduce the influences from the atmosphere, topography and soil background (Houborg et al., 2007; Liao et al., 2015) and therefore their usages were explored in this study. The normalized difference infrared index (NDII) (Hardisky et al., 1983) and normalized difference tillage index (NDTI) (Van Deventer et al., 1997) were used for their sensitivities to water content and dry matter (Wang et al., 2013; Yebra et al., 2013). The NDII and NDTI are defined as, NDII =

NIR − SWIR1 NIR + SWIR1

(1)

NDTI =

SWIR1 − SWIR2 SWIR1 + SWIR2

(2)

where near infrared (␳NIR ) band and shortwave infrared band (␳SWIR1 ) and ␳SWIR2 were extracted from band 5 (845–885 nm), band 6 (1560–1660 nm) and band 7 (2100–2300 nm) of atmospherically corrected Landsat 8 OLI product. 2.2. RTM-based method A schematic diagram of the RTM-based grassland AGB retrieval scheme is given in Fig. 2. The methodology can be divided into two main parts: the forward modeling (section 2.2.1 and 2.2.2) and backward inversion (section 2.2.3). The forward modeling is to set up a look up table (LUT) containing free parameters and corresponding simulated spectra based on the PROSAILH RTM. The backward inversion is to estimate the AGB based on the observed and simulated spectra through a merit function. 2.2.1. Parameterization of PROSAILH and forward modeling The PROSAILH model combines the PROSPECT (Jacquemoud and Baret, 1990) leaf optical properties model and the SAILH canopy reflectance model (Verhoef, 1984; Kuusk, 1991). PROSPECT calculates the leaf reflectance and transmittance ranging from 400 nm

to 2500 nm with a spectral resolution of 1 nm. This model includes six parameters: leaf structure parameter, N (unit-less); chlorophyll a + b content, Cab (␮gcm−2 ); equivalent water thickness, EWT (gcm−2 ); DMC (gcm−2 ); carotenoid content, Car (␮gcm−2 ); and leaf brown pigment, Cbp (unit-less). SAILH is a four-stream RTM and is used for simulating the spectra of the homogenous canopy from 400 nm to 2500 nm, also with a spectral resolution of 1 nm. The SAILH model requires ten parameters: two leaf inclination distribution function (LIDF) parameters, LIDFa and LIDFb (unit-less); soil factor, psoil (unit-less); LAI (m2 m−2 ); hot spot factor, hspot (unit-less); the Sun zenith angle, tts (◦ ); observer zenith angle, tto (◦ ); relative azimuth angle, psi (◦ ); leaf hemispheric reflectance, LR (nm); and leaf transmittance, LT (nm). LR and LT can be simulated by the PROSPECT model. The PROSAILH model requires the determination of 14 parameters that define the leaf and canopy optical properties (Table 1). The tts, tto and psi were obtained from the Landsat 8 OLI metadata files. LIDFa and LIDFb have a different value depending on the LIDF of the canopy (planophile, erectophile, plagiophile, extremophile, spherical, and uniform). In this study, the LIDF was set as spherical based on field surveys and therefore LIDFa = −0.35 and LIDFb = −0.15 (default values for spherical LIDF in the SAILH model). The Cbp was set to 0 because the grass was in the growing stage during the two sampling campaigns. Car and Cab were set as the model default values (Car = 8 ␮gcm−2 and Cab = 40 ␮gcm−2 ) since they are mainly sensitive to visible wavebands with ignorable effect on our target variables (AGB = LAI × DMC). The hspot was parameterized as a function of LAI according to He et al. (2013) and Houborg et al. (2009). LAI was ranged from 0.1 to 8 based on the field measurement and the information from the MOD15A2 product (Myneni et al., 2002) of the study area. The PROSAILH model integrates the soil spectrum, and the parameter psoil, ranging from 0, representing wet soil, to 1, representing dry soil, to characterize the condition of the soil. Based on the prior information from the field survey, psoil was set as a free parameter ranging from 0.2 to 1.0. The N was set to 2 for a monocotyledon leaf type (Houborg et al., 2007). The EWT (gcm−2 ), ranged from 0.001 to 0.02 gcm−2 , while the DMC from 0.002 to 0.01 gcm−2 based on the prior information from the field surveys. The biomass (i.e., LAI × DMC gm−2 ) was also constrained to the range observed in the field surveys (20–400 gm−2 ) during the forward simulation process to avoid unrealistically high or low values in the simulated combinations. Therefore, the LAI,

162

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

Fig. 2. Flowchart showing the RTM-based method for the estimation of grassland AGB.

Table 1 Parameterization of the PROSAILH RTM. Models SAILH

PROSPECT

Parameters

Symbol

Range or fixed value

Prior information

tts tto psi LAI hspot psoil

/ 0 / 0.1–8 0.5/LAI 0.2–1.0

Product metadata Product metadata Product metadata Field survey & MCD15A2 (Houborg et al., 2009) Field survey

Two leaf inclination distribution function (LIDF) parameters

LIDFaLIDFb

−0.35

Leaf structure parameter Chlorophyll a + b content Leaf equivalent water thickness Dry matter content Brown pigment Carotenoid content

N Cab EWT DMC Cbp Car

2 40 0.001–0.02 0.002–0.01 0 8

Sun zenith angle View zenith angle Relative azimuth angle Leaf area index Hot spot factor Soil factor

psoil, DMC and EWT were set as free parameters used to forward model the spectra using the PROSAILH RTM. 2.2.2. Filtering out of unrealistic spectra using ecological criteria The estimation of vegetation variables using RTM is generally hampered by the ill-posed inversion problem (Houborg et al., 2009; Yebra and Chuvieco, 2009b), meaning that the different combinations of RTM input parameters may correspond to similar spectra, whereas some of the combinations are unrealistic, resulting in a large uncertainty if these unrealistic combinations were used. The use of statistical relationship between input parameters, also called ecological criteria (Jurdao et al., 2013), has the effect on improving the accuracy level in the retrieval process since the soil and model parameters are not independent but correlated for specific species (Yebra and Chuvieco 2009a,b; Jurdao et al., 2013; Feret et al., 2011; Quan et al., 2015a). Ignoring such information leads to the simulation of unrealistic combinations that may never occur in the target study area, aggravating the ill-posed inversion problem (Yebra and Chuvieco, 2009a,b). Following these studies, a positive linear correlation was found between EWT and DMC with a

Units ◦

( ) (◦ ) (◦ ) m2 m−2

␮gcm−2 gcm−2 gcm−2 ␮gcm−2

−0.15

Field survey Houborg et al., 2007 Model default Field survey Field survey Field survey Model default

coefficient of determination (R2 ) of 0.46 and root mean square error (RMSE) of 0.0017 gcm−2 (Fig. 3). In order to avoid these spectra that were not likely to occur, an empirical relationship between EWT and DMC (Eq. (3)) was used as a filter criterion to remove the EWT vs. DMC combinations that do not meet this relationship (Jurdao et al., 2013; Yebra and Chuvieco, 2009b; Yebra et al., 2008). Eq. (3) was derived from the fitted curve based on measured EWT and DMC as showed in Fig. 3. RMSE > |0.381EWT + 0.0025 − DMC|

(3)

2.2.3. Backward inversion The LUT inversion methodology is a commonly used optimal method for solving the inversion of an RTM. It is a global search algorithm that avoids the local minima problem, and this method is simple (Darvishzadeh et al., 2008a; Schlerf and Atzberger, 2006). For the inversion process, a merit function is needed to identify the parameter combinations that yield the best fit between measured

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

163

for each pixel corresponding to the Landsat 8 OLI product were derived using the trained ANN and the Landsat 8 EVI values. The LAIe is definitively not unbiased due to, among other reasons, the uncertainty from these satellite products and this simple method. Therefore, rather than the use of fixed LAIe value as that used in previous studies (Yebra et al., 2008; Yebra and Chuvieco et al., 2009b; Zarco-Tejada et al., 2003), a LAI range (0.8 × LAIe , 1.2 × LAIe ) was used to filter the unrealistic combinations of retrieved free parameters (LAI, psoil, EWT and DMC) whose retrieved LAI values were placed out of the range. Finally, the mean value of LAI × DMC extracted from rest combinations of retrieved free parameters was regarded as the estimated grassland AGB. 2.3. Empirical methods

Fig. 3. Relationship between measured DMC and EWT and the fitted linear curve.

and LUT spectra. The spectral angle (SA) method (Jurdao et al., 2013; Kruse et al., 1993) was used as the merit function, SA

→

→





v , w = cos

−1

→

→

v ×w

→



→

 v  × w

(4)

where v and w are the observed (atmospherically corrected Landsat 8 OLI products) and the simulated (LUT) spectra, respectively. Both of them are considered as an m-dimensional feature vector, with m being the number of spectral channels (m = 9, seven reflectance bands, NDII and NDTI). To enhance the consistency of the retrieved AGB, 30 best matched SA values in the LUT was regarded as the combinations of retrieved free parameters (LAI, psoil, EWT and DMC) (Darvishzadeh et al., 2008a). However, the unrealistic combinations of retrieved free parameters may also exist due to the ill-posed inversion problem. Following several studies which obtained improvements in water content and fuel moisture content estimation when LAI was fixed to a known value extracted from the MOD15A2 product (Yebra et al., 2008; Yebra and Chuvieco et al., 2009b3), the MOD15A2 was also used to filter these unrealistic combinations. The MOD15A2 product is a global 8-day composited LAI dataset, and its accuracy in grassland has been evaluated by Fensholt et al. (2004), Hill et al. (2006), Tian et al. (2002) and Wang et al. (2004). However, the spatial resolution of MOD15A2 (1 km) and the Landsat 8 OLI product (30 m) mismatch. Generally, as for a specific grassland and growth stage, the LAI extracted from MOD15A2 product has a high correlation with the enhanced vegetation index (EVI) extracted from the MODIS vegetation indices product (MOD13A2) (Quan et al., 2016). Therefore, a simple method was used to derive the LAI in the spatial resolution of 30 m based on the MOD15A2, MOD13A2 and Landsat 8 OLI products. Firstly, the EVI and LAI values for the study area and in the same sampling periods (late July 2014 and early August 2015) were extracted from MOD13A2 and MOD15A2 product. These products were downloaded from the United States Geological Survey (USGS) (http://www.usgs.gov). Secondly, the matched LAI and EVI combinations (spatial resolution: 1 km) for the study area with vegetation covered (defined by EVI > 0.1) were set as the inputs to train an ANN. Thirdly, the EVI values with spatial resolution of 30 m were calculated for the Landsat 8 OLI product based on Eq. (5) as, EVI = 2.5 ×

NIR − red NIR + 6 × red − 7.5 × blue + 1

(5)

where ␳NIR ␳red ␳blue are the reflectance band 5, band 4 and band 2 from the atmospherically corrected Landsat 8 OLI products. Finally, the LAI values with the spatial resolution of 30 m (LAIe )

Three empirical methods were also used for the estimation of grassland AGB, with the purpose of comparing its performance versus the RTM-based method. These methods were exponential regression, PLSR and ANN. The models were regressed/calibrated using a 60% of the 135 measured grassland ABG (randomly selected) and corresponding satellite observation, while the rest 40% was used for validation. 2.3.1. Exponential regression In the studies of Zheng et al. (2007) and Zheng et al. (2004), they found that the AGB has a high correlation with the normalized difference vegetation index (NDVI, Eq. (6)), and they used this vegetation index for the estimation of AGB from satellite images. Following their work, the NDVI derived from atmospherically corrected Landsat 8 OLI products was used to estimate the AGB. An exponential equation was used to set up the statistical relationship between measured AGB and NDVI (Eq. (7)): NDVI =

NIR − red NIR + red

AGB = a × eb×NDVI

(6) (7)

where red and NIR were extracted from atmospherically corrected Landsat 8 OLI product band 4 (630–680 nm) and band 5 (845–885 nm); a and b are the regression coefficients. Eq. (7) was fitted using the curve fitting toolbox in Matlab software (2015a). 2.3.2. PLSR PLSR is a multivariate statistical technique to analyze data with many, noisy, collinear, and even incomplete variables (Wold et al., 2001). This technique reduces a large number of measured collinear spectral variables to a few non-correlated latent variables or factors while maximizing co-variability to the variable(s) of interest (Cho et al., 2007; Darvishzadeh et al., 2008a,b; Geladi and Kowalski, 1986; Hansen and Schjoerring, 2003). As in multiple regression, the aim of PLSR is to build a linear model (Cho et al., 2007; Darvishzadeh et al., 2008a,b) as, Y = Xˇ + E where Y is the vector of the response variable (AGB in this study), X is the matrix of the predictor (spectral reflectance and vegetation indices), ␤ is the matrix of coefficients, and E is the matrix of residuals. In principle, PLSR is closely related to principal component regression, but principal component regression performs the decomposition on the spectral data alone (Geladi and Kowalski, 1986). PLSR uses the response variable information during the decomposition process and performs the decomposition on both the spectral and the response simultaneously (Cho et al., 2007; Darvishzadeh et al., 2008a,b; Geladi and Kowalski, 1986). In most cases, this procedure can reduce the number of spectral variables to a few independent variables (Hansen and Schjoerring, 2003). For

164

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

the details of this method and its applications, please refers to (Cho et al., 2007; Darvishzadeh et al., 2008a,b; Ehsani et al., 1999; Feret et al., 2011; Geladi and Kowalski, 1986; Schmidtlein and Sassin, 2004). The selection factors or latent variables is vital for a (multiple) linear regression. PLSR may be over-fitting if too many factors are used. To avoid this problem, the leave-one-out cross-validation method following recommendations of Cho et al. (2007) was used. A specific factor will be only used if reduces the root mean square error of cross-validation (RMSECV, Eq. (8)) by >2% (Cho et al., 2007; Darvishzadeh et al., 2008a,b; Kooistra et al., 2004).



RMSECV =

n i=1

yˆ i − yi n

2

(8)

where yˆ i and yi were the predicted AGB from each cross-validation phase and the measured AGB, respectively, and n the number of samples. Once the optimum number of PLS factors were determined, a definitive model was developed. Fig. 4. Result of estimated grassland AGB using the RTM-based method.

2.3.3. ANN The ANN offers a powerful method for analyzing complex relationships among vegetation variables without making assumptions about the data, and therefore is capable of handling non-normality, nonlinearity and collinearity in a system (Haykin, 2001; Ingram et al., 2005). The ANN is generally composed of three parts: an input layer, one output layer and one or more hidden layers (Jensen et al., 1999). The ANN predicts output data from patterns learned from a set of input training data. By comparing the current output layer to the desired output response, the difference between the two layers can be obtained and used to adjust weights within the network, with the purpose to achieve a set of weights that produce results that closely resemble the target outputs. This adaptive learning process is repeated until the difference between predicted and training values drop below a predetermined threshold of user-defined accuracy (Ingram et al., 2005; Jensen et al., 1999). Once constructed and after patterns in the data have been learned, the ANN can be used to estimate the target variables from another independent dataset with observations (Baret et al., 1995; Hongliang and Shunlin, 2003; Smith, 1993). The training process is usually computationally intensive. Because some satellite bands are closely related and to avoid the over-fitting problem, only those bands and derived vegetation indices that have the largest information content (Landsat bands 1–7, NDII and NDTI in this study) were used in the training iteration. To facilitate the comparison of methods, the ANN was trained using the same measurements and corresponding spectral data that were used in the PLSR. This work was accomplished by the Matlab ANN toolbox (2015a). 3. Results Fig. 4 shows the retrieval results of grassland AGB using the RTM-based method with R2 = 0.64 and RMSE = 42.67 gm−2 . However, some of the AGB values were underestimated. This is probably because the soil spectra incorporated in the PROSAILH model do not conform to the actual saline-alkali soil property in the study area. As for the exponent regression, the 60% measured AGB and corresponding NDVI was used for the fitting. Fig. 5 shows the result of fitted curve with R2 = 0.43 and RMSE = 36.42 gm−2 . Fig. 6 (a) shows the accuracy level of the exponential regression in estimating the AGB (using the rest 40% measurements). The accuracy level was poor (R2 = 0.48 and RMSE = 41.65 gm−2 ) compared with the RTMbased method. It can be also found that the estimated AGB gradually get saturate when AGB greater than 175 gm−2 , due to the saturation problem of NDVI in detecting dense vegetation (Fig. 5) while

this problem was not found in the RTM-based method. Therefore, the exponential regression method based on NDVI is not suitable for the estimation AGB of dense vegetation. As for the PLSR, band 5 (NIR) and NDII were selected in the PLSR using the leave-one-out cross-validation method to avoid over-fitting problem, and the regressed equation based on the two factors is, AGB = 38.14 + 152.6 × NIR + 274.98 × NDII

(9) (R2

Fig. 6(b) shows the accuracy level of regressed = 0.61 and RMSE = 30.12 gm−2 ) and validated AGB (R2 = 0.55 and RMSE = 37.79 gm−2 ) based on the PLSR. The accuracy level is slightly better than or similar to the RTM-based method and is better than the exponential regression method. However, the saturation problem can be also found. As is shown in Fig. 6(b), the estimation AGB generally gets saturated when AGB greater than 250 gm−2 . Compared with the exponential regression method, this problem is alleviated. The same inputs for PLSR were used to train the ANN and to estimate AGB using this trained ANN to avoid the over-fitting problem. Fig. 6(c) shows the accuracy level of estimated AGB using this method with R2 = 0.43 and RMSE = 46.26 gm−2 . The accuracy level is the worst compared with the RTM-based method, exponential regression and PLSR. However, the saturation problem did not occur in this case, even for the high AGB values (>250 gm−2 ). 4. Discussions In this study, a RTM-based method was proposed to estimate grassland AGB and its performance was compared with three traditional empirical methods: the exponential regression, PLSR and ANN. The accuracy level of retrieved AGB using the RTMbased method (R2 = 0.64 and RMSE = 42.67 gm−2 ) was superior to the exponential regression (R2 = 0.48 and RMSE = 41.65 gm−2 ) and ANN method (R2 = 0.43 and RMSE = 46.26 gm−2 ) while with slightly better or similar accuracy than PLSR (R2 = 0.55 and RMSE = 37.79 gm−2 ). However, the RTM sets up the explicit relations between the canopy properties and the spectra and therefore the RTM-based method can be applied in other areas without the prior use of regression/calibration data as the PLSR. Moreover, the saturation problem was generally occurred in the exponential regression and PLSR (Cho et al., 2007; Darvishzadeh et al., 2008a,b; Feret et al., 2011). This problem was not found in the RTM-based method and ANN. However, the regression/training data were not required in the RTM-based method while the per-

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

165

Fig. 5. Exponent regression between grassland AGB and NDVI using the measured AGB and observed NDVI values.

Fig. 6. The regression and validation results for the exponent regression (a), PLSR (b) and ANN (c). The R2 and RMSE between measured and estimated AGB are also shown.

formance of ANN largely relies on a sufficient training database (Hongliang and Shunlin, 2003). Therefore, the RTM-based method is more replicable than the empirical methods (Houborg et al., 2009, 2007; Quan et al., 2015a,b; Yebra et al., 2013). In our previous studies, both the empirical method and data assimilation technique were used for the estimation of mixed grassland and shubland AGB (Xing et al., 2014) and grassland AGB (He et al., 2015). For the empirical method, a modified water cloud

model (Attema and Ulaby, 1978) was used to estimate the AGB from ASAR data and Landsat TM products derived LAI values. The estimated AGB showed that the predicted AGB highly correlated with the measured AGB (R2 = 0.8007 and RMSE = 280.8 gm−2 ) (Xing et al., 2014). For the data assimilation technique, the 4DVar algorithm was applied to optimize the input parameters of a soil–water–atmosphere–plant (SWAP) model (Huang et al., 2015b; Supit et al., 1994). With the optimized input parameters, a good

166

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

result was also obtained (R2 = 0.76 and RMSE = 54.25 gm−2 ) (He et al., 2015). Compared with the estimated grassland AGB in the two previous studies, the accuracy of retrieved AGB using the RTMbased method in this study appears slightly lower R2 (R2 = 0.64) but a lower RMSE (RMSE = 42.67 gm−2 ). However, compared with the empirical method, the RTM-based method in this study does not need the regression data and therefore it has the advantage of reproducibility. This method is also more feasible than the data assimilation technique because the PROSAILH RTM requires fewer parameters than the SWAP model. Therefore, the RTM-based method is considered the most promising methodology. In Fig. 4, the estimated AGB was generally underestimated, especially for the low AGB values. This may be because the PROSAILH RTM was built without the consideration of stem properties. The PROSPECT model is built for the simulation of the reflectance and transmittance spectra of leaves (Jacquemoud and Baret, 1990), and the SAILH model was built for simulating the reflectance for the canopy while ignoring the effect from stems (Verhoef, 1984). Therefore, the PROSAILH model may be effective for estimating foliage biomass, but it may be not that feasible for the estimation of stem biomass, indicating the further improvement of this RTM. Romero et al. (2012) found the use of a spectral index, defined as (␳2305 –␳1495 )/(␳2305 + ␳1495 ), performed well in the estimation of DMC. Wang et al. (2011b) also found that the normalized dry matter index (NDMI), defined as (␳1649 −␳1722 )/(␳1649 + ␳1722 ), can be used for the estimation of DMC from the Leaf Optical Properties Experiment (LOPEX) dataset. In a subsequent work, Wang et al. (2011a) extended this study to the canopy level for the estimation of foliar biomass from laboratory-measured spectral reflectance, and a linear correlation was found between NDMI and foliar biomass. These studies indicated that the grassland AGB can be estimated based on these spectral indices due to their high sensitivities to DMC. However, for most satellite sensors to date, ␳1495 and ␳1722 were not available. As an alternative, the Landsat 8 OLI products were used in this study to estimate AGB, and the results showed that the grassland AGB can be estimated from this satellite product.

5. Conclusion In this study, we proposed a novel method to estimate grassland AGB based on the PROSAILH RTM. To alleviate the ill-posed inversion problem, the ecological criteria was used to filter the unrealistic combination of EWT vs. DMC. The constrained LAI ranges derived from MOD15A2, MOD13A2 and Landsat 8 OLI products were used to filter the retrieved combinations (LAI, psoil, EWT and DMC) whose retrieved LAI values were placed out of the constrained LAI range. To test the performance of the proposed method, three traditional empirical methods: the exponential regression, PLSR and ANN methods were also used. The results showed that the RTM-based method was more accurate than the exponent regression and ANN while with slightly better or similar accuracy level than PLSR. However, the RTM-based method can be applied in other areas without the prior use of regression data, and therefore it is more reproductive than the empirical methods. Additionally, the exponential regression and PLSR were generally challenged by the saturation problem while this problem did not occur in the RTM-based method. Future work will focus on further improving the presented method for the accurate and robust estimation of grassland AGB in large-scale areas based on other RTMs such as ACRM (Kuusk, 2001), DART (Gastellu-Etchegorry et al., 2004), FLIGHT (North, 1996), etc.

Acknowledgment This work was supported by the National Natural Science Foundation of China (Contract No. 41471293 & 41671361), the Fundamental Research Fund for the Central Universities (Contract No. ZYGX2012Z005) and the National High-Tech Research and Development Program of China (Contract 2013AA12A302). The authors wish to thank the Qinghai Ecosystem Remote Sensing Monitoring Centre for providing part of the field measurement dataset used in this study; Yu Zhu, Dasong Xu and Fangwei Man for their assistance during the field campaigns; and the China scholarship council (CSC) for its support of Xingwen Quan during his scholarship at Australian National University. References Anaya, J.A., Chuvieco, E., Palacios-Orueta, A., 2009. Aboveground biomass assessment in Colombia: a remote sensing approach. For. Ecol. Manage. 257, 1237–1246. Anderson, G.L., Hanson, J.D., Haas, R.H., 1993. Evaluating landsat thematic mapper derived vegetation indices for estimating above-ground biomass on semiarid rangelands. Remote Sens. Environ. 45, 165–175. Attema, E.P.W., Ulaby, F.T., 1978. Vegetation modeled as a water cloud. Radio Sci. 13, 357–364. Baghdadi, N., Le Maire, G., Bailly, J.S., Ose, K., Nouvellon, Y., Zribi, M., Lemos, C., Hakamada, R., 2015. Evaluation of ALOS/PALSAR L-Band data for the estimation of Eucalyptus plantations aboveground biomass in Brazil. IEEE J. Sel. Topics Earth Obs. Remote Sens. (JSTARS) 8, 3802–3811. Baret, F., Clevers, J.G.P.W., Steven, M.D., 1995. The robustness of canopy gap fraction estimates from red and near-infrared reflectances: a comparison of approaches. Remote Sens. Environ. 54, 141–151. Barrachina, M., Cristóbal, J., Tulla, A.F., 2015. Estimating above-ground biomass on mountain meadows and pastures through remote sensing. Int. J. Appl. Earth Obs. 38, 184–192. Boudreau, J., Nelson, R.F., Margolis, H.A., Beaudoin, A., Guindon, L., Kimes, D.S., 2008. Regional aboveground forest biomass using airborne and spaceborne LiDAR in Québec. Remote Sens. Environ. 112, 3876–3890. Cartus, O., Santoro, M., Kellndorfer, J., 2012. Mapping forest aboveground biomass in the Northeastern United States with ALOS PALSAR dual-polarization L-band. Remote Sens. Environ. 124, 466–478. Chen, Q., Vaglio Laurin, G., Valentini, R., 2015. Uncertainty of remotely sensed aboveground biomass over an African tropical forest: propagating errors from trees to plots to pixels. Remote Sens. Environ. 160, 134–143. Chen, Q., 2015. Modeling aboveground tree woody biomass using national-scale allometric methods and airborne LiDAR. ISPRS J. Photogramm. Remote Sens. 106, 95–106. Cho, M.A., Skidmore, A.K., 2009. Hyperspectral predictors for monitoring biomass production in Mediterranean mountain grasslands Majella National Park, Italy. Int. J. Remote Sens. 30, 499–515. Cho, M.A., Skidmore, A., Corsi, F., van Wieren, S.E., Sobhan, I., 2007. Estimation of green grass/herb biomass from airborne hyperspectral imagery using spectral indices and partial least squares regression. Int. J. Appl. Earth Obs. 9, 414–424. Cho, M.A., Mathieu, R., Asner, G.P., Naidoo, L., van Aardt, J., Ramoelo, A., Debba, P., Wessels, K., Main, R., Smit, I.P.J., Erasmus, B., 2012. Mapping tree species composition in South African savannas using an integrated airborne spectral and LiDAR system. Remote Sens. Environ. 125, 214–226. Clevers, J.G.P.W., van Leeuwen, H.J.C., 1996. Combined use of optical and microwave remote sensing data for crop growth monitoring. Remote Sens. Environ. 56, 42–51. Clevers, J., Van der Heijden, G., Verzakov, S., Schaepman, M., 2007. Estimating grassland biomass using SVM band shaving of hyperspectral data. Photogramm. Eng. Remote Sens. 73, 1141–1148. Cloude, S.R., Papathanassiou, K.P., 1998. Polarimetric SAR interferometry. IEEE Trans. Geosci. Remote Sens. 36, 1551–1565. Darvishzadeh, R., Skidmore, A., Schlerf, M., Atzberger, C., 2008a. Inversion of a radiative transfer model for estimating vegetation LAI and chlorophyll in a heterogeneous grassland. Remote Sens. Environ. 112, 2592–2604. Darvishzadeh, R., Skidmore, A., Schlerf, M., Atzberger, C., Corsi, F., Cho, M., 2008b. LAI and chlorophyll estimation for a heterogeneous grassland using hyperspectral measurements. ISPRS J. Photogramm. Remote Sens. 63, 409–426. Diepen, C.v., Wolf, J., Keulen, H.v., Rappoldt, C., 1989. WOFOST: a simulation model of crop production. Soil Use Manage. 5, 16–24. Doraiswamy, P.C., Sinclair, T.R., Hollinger, S., Akhmedov, B., Stern, A., Prueger, J., 2005. Application of MODIS derived parameters for regional crop yield assessment. Remote Sens. Environ. 97, 192–202. Ehsani, M.R., Upadhyaya, S.K., Slaughter, D., Shafii, S., Pelletier, M., 1999. A NIR technique for rapid determination of soil mineral nitrogen. Precis. Agric. 1, 219–236. Englhart, S., Keuck, V., Siegert, F., 2011. Aboveground biomass retrieval in tropical forests — The potential of combined X- and L-band SAR data use. Remote Sens. Environ. 115, 1260–1271.

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168 Fassnacht, F.E., Hartig, F., Latifi, H., Berger, C., Hernández, J., Corvalán, P., Koch, B., 2014. Importance of sample size, data type and prediction method for remote sensing-based estimations of aboveground forest biomass. Remote Sens. Environ. 154, 102–114. Fensholt, R., Sandholt, I., Rasmussen, M.S., 2004. Evaluation of MODIS LAI, fAPAR and the relation between fAPAR and NDVI in a semi-arid environment using in situ measurements. Remote Sens. Environ. 91, 490–507. Feret, J.B., Francois, C., Gitelson, A., Asner, G.P., Barry, K.M., Panigada, C., Richardson, A.D., Jacquemoud, S., 2011. Optimizing spectral indices and chemometric analysis of leaf chemical properties using radiative transfer modeling. Remote Sens. Environ. 115, 2742–2750. Gastellu-Etchegorry, J.P., Martin, E., Gascon, F., 2004. DART: a 3D model for simulating satellite images and studying surface radiation budget. Int. J. Remote Sens. 25, 73–96. Geladi, P., Kowalski, B.R., 1986. Partial least-squares regression: a tutorial. Anal. Chim. Acta 185, 1–17. Godwin, D., Ritchie, J., Singh, U., Hunt, L., 1989. A Users Guide to CERES Wheat, Vol. v2 10. Hajj, M.E., Baghdadi, N., Belaud, G., Zribi, M., Cheviron, B., Courault, D., Hagolle, O., Charron, F., 2014. Irrigated grassland monitoring using a time series of terraSAR-X and COSMO-skyMed X-Band SAR Data. Remote Sens. 6, 10002–10032. Hansen, P.M., Schjoerring, J.K., 2003. Reflectance measurement of canopy biomass and nitrogen status in wheat crops using normalized difference vegetation indices and partial least squares regression. Remote Sens. Environ. 86, 542–553. Hardisky, M., Klemas, V., Smart, R.M., 1983. The influence of soil salinity, growth form, and leaf moisture on the spectral radiance of Spartina alterniflora canopies. Photogramm. Eng. Remote Sens. 49, 77–83. Haykin, S.S., 2001. Neural Networks: a Comprehensive Foundation. Tsinghua University Press. He, B., Quan, X., Xing, M., 2013. Retrieval of leaf area index in alpine wetlands using a two-layer canopy reflectance model. Int. J. Appl. Earth Obs. 21, 78–91. He, B., Li, X., Quan, X., Qiu, S., 2015. Estimating the aboveground dry biomass of grass by assimilation of retrieved LAI into a crop growth model. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 8, 550–561. Hill, M.J., Senarath, U., Lee, A., Zeppel, M., Nightingale, J.M., Williams, R.J., McVicar, T.R., 2006. Assessment of the MODIS LAI product for Australian ecosystems. Remote Sens. Environ. 101, 495–518. Hongliang, F., Shunlin, L., 2003. Retrieving leaf area index with a neural network method: simulation and validation. IEEE Trans. Geosci. Remote Sens. 41, 2052–2062. Houborg, R., Soegaard, H., Boegh, E., 2007. Combining vegetation index and model inversion methods for the extraction of key vegetation biophysical parameters using Terra and Aqua MODIS reflectance data. Remote Sens. Environ. 106, 39–58. Houborg, R., Anderson, M., Daughtry, C., 2009. Utility of an image-based canopy reflectance modeling tool for remote estimation of LAI and leaf chlorophyll content at the field scale. Remote Sens. Environ. 113, 259–274. Houghton, R.A., 2005. Aboveground forest biomass and the global carbon balance. Global Change Biol. 11, 945–958. Huang, J., Tian, L., Liang, S., Ma, H., Becker-Reshef, I., Huang, Y., Su, W., Zhang, X., Zhu, D., Wu, W., 2015a. Improving winter wheat yield estimation by assimilation of the leaf area index from Landsat TM and MODIS data into the WOFOST model. Agr. For. Meteorol. 204, 106–121. Huang, J., Ma, H., Su, W., Zhang, X., Huang, Y., Fan, J., Wu, W., 2015b. Jointly assimilating MODIS LAI and ET products into the SWAP Model for winter wheat yield estimation. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 8, 4060–4071. Huang, J., Sedano, F., Huang, Y., Ma, H., Li, X., Liang, S., Tian, L., Zhang, X., Fan, J., Wu, W., 2016. Assimilating a synthetic Kalman filter leaf area index series into the WOFOST model to improve regional winter wheat yield estimation. Agr. For. Meteorol. 216, 188–202. Ingram, J.C., Dawson, T.P., Whittaker, R.J., 2005. Mapping tropical forest structure in southeastern Madagascar using remote sensing and artificial neural networks. Remote Sens. Environ. 94, 491–507. Jacquemoud, S., Baret, F., 1990. PROSPECT: a model of leaf optical properties spectra. Remote Sens. Environ. 34, 75–91. Jensen, J.R., Qiu, F., Ji, M., 1999. Predictive modelling of coniferous forest age using statistical and artificial neural network approaches applied to remote sensor data. Int. J. Remote Sens. 20, 2805–2822. Jurdao, S., Yebra, M., Guerschman, J.P., Chuvieco, E., 2013. Regional estimation of woodland moisture content by inverting Radiative Transfer Models. Remote Sens. Environ. 132, 59–70. Koch, B., 2010. Status and future of laser scanning: synthetic aperture radar and hyperspectral remote sensing data for forest biomass assessment. ISPRS J. Photogramm. 65, 581–590. Kooistra, L., Salas, E.A.L., Clevers, J.G.P.W., Wehrens, R., Leuven, R.S.E.W., Nienhuis, P.H., Buydens, L.M.C., 2004. Exploring field vegetation reflectance as an indicator of soil contamination in river floodplains. Environ. Pollut. 127, 281–290. Kruse, F.A., Lefkoff, A.B., Boardman, J.W., Heidebrecht, K.B., Shapiro, A.T., Barloon, P.J., Goetz, A.F.H., 1993. The spectral image processing system (SIPS)—interactive visualization and analysis of imaging spectrometer data. Remote Sens. Environ. 44, 145–163.

167

Kuusk, A., 1991. The hot spot effect in plant canopy reflectance. In: Myneni, R., Ross, J. (Eds.), Photon-Vegetation Interactions. Springer Berlin Heidelberg, pp. 139–159. Kuusk, A., 2001. A two-layer canopy reflectance model. J. Quant. Spectrosc. Radiat. Transfer 71, 1–9. Lefsky, M.A., Hudak, A.T., Cohen, W.B., Acker, S., 2005. Geographic variability in lidar predictions of forest stand structure in the Pacific Northwest. Remote Sens. Environ. 95, 532–548. Li, W., Niu, Z., Liang, X., Li, Z., Huang, N., Gao, S., Wang, C., Muhammad, S., 2015. Geostatistical modeling using LiDAR-derived prior knowledge with SPOT-6 data to estimate temperate forest canopy cover and above-ground biomass via stratified random sampling. Int. J. Appl. Earth Obs. 41, 88–98. Liao, Z., He, B., Quan, X., 2015. Modified enhanced vegetation index for reducing topographic effects. APPRES 9, 096068–096068. Liu, J., Pattey, E., Miller, J.R., McNairn, H., Smith, A., Hu, B., 2010. Estimating crop stresses, aboveground dry biomass and yield of corn using multi-temporal optical data combined with a radiation use efficiency model. Remote Sens. Environ. 114, 1167–1177. Liu, Y.Y., van Dijk, A.I.J.M., de Jeu, R.A.M., Canadell, J.G., McCabe, M.F., Evans, J.P., Wang, G., 2015. Recent reversal in loss of global terrestrial biomass. Nat. Clim. Change 5, 470–474. Lu, D., 2006. The potential and challenge of remote sensing-based biomass estimation. Int. J. Remote Sens. 27, 1297–1328. Ma, G., Huang, J., Wu, W., Fan, J., Zou, J., Wu, S., 2013a. Assimilation of MODIS-LAI into the WOFOST model for forecasting regional winter wheat yield. Math. Comput. Modell. 58, 634–643. Ma, H., Huang, J., Zhu, D., Liu, J., Su, W., Zhang, C., Fan, J., 2013b. Estimating regional winter wheat yield by assimilation of time series of HJ-1 CCD NDVI into WOFOST–ACRM model with Ensemble Kalman Filter. Math. Comput. Modell. 58, 759–770. Matthew, M.W., Adler-Golden, S.M., Berk, A., Richtsmeier, S.C., Levine, R.Y., Bernstein, L.S., Acharya, P.K., Anderson, G.P., Felde, G.W., Hoke, M.L., 2000. Status of atmospheric correction using a MODTRAN4-based algorithm, AeroSense 2000. Int. Soc. Opt. Photonics, 199–207. Matthews, S., 2010. Effect of drying temperature on fuel moisture content measurements. Int. J. Wildland Fire 19, 800–802. Moreau, S., Le Toan, T., 2003. Biomass quantification of Andean wetland forages using ERS satellite SAR data for optimizing livestock management. Remote Sens. Environ. 84, 477–492. Myneni, R.B., Hoffman, S., Knyazikhin, Y., Privette, J.L., Glassy, J., Tian, Y., Wang, Y., Song, X., Zhang, Y., Smith, G.R., Lotsch, A., Friedl, M., Morisette, J.T., Votava, P., Nemani, R.R., Running, S.W., 2002. Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data. Remote Sens. Environ. 83, 214–231. North, P., 1996. Three-dimensional forest light interaction model using a Monte Carlo method. IEEE Trans. Geosci. Remote Sens. 34, 946–956. Paul, K.I., Roxburgh, S.H., England, J.R., Ritson, P., Hobbs, T., Brooksbank, K., John Raison, R., Larmour, J.S., Murphy, S., Norris, J., Neumann, C., Lewis, T., Jonson, J., Carter, J.L., McArthur, G., Barton, C., Rose, B., 2013. Development and testing of allometric equations for estimating above-ground biomass of mixed-species environmental plantings. For. Ecol. Manage. 310, 483–494. Quan, X., He, B., Wang, Y., Tang, Z., Li, X., 2014. An extended fourier approach to improve the retrieved leaf area index (LAI) in a time series from an alpine wetland. Remote Sens. 6, 1171–1190. Quan, X.W., He, B.B., Li, X., 2015a. A bayesian network-based method to alleviate the ill-posed inverse problem: a case study on leaf area index and canopy water content retrieval. IEEE Trans. Geosci. Remote Sens. 53, 6507–6517. Quan, X., He, B., Li, X., Tang, Z., 2015b. Estimation of grassland live fuel moisture content from ratio of canopy water content and foliage dry biomass. IEEE Geosci. Remote Sens. Lett. 12, 1903–1907. Quan, X., He, B., Li, X., Liao, Z., 2016. Retrieval of grassland live fuel moisture content by parameterizing radiative transfer model with interval estimated LAI. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 9, 910–920. Ramoelo, A., Cho, M.A., Mathieu, R., Madonsela, S., van de Kerchove, R., Kaszta, Z., Wolff, E., 2015. Monitoring grass nutrients and biomass as indicators of rangeland quality and quantity using random forest modelling and WorldView-2 data. Int. J. Appl. Earth Obs. 43, 43–54. Romero, A., Aguado, I., Yebra, M., 2012. Estimation of dry matter content in leaves using normalized indexes and PROSPECT model inversion. Int. J. Remote Sens. 33, 396–414. Schlerf, M., Atzberger, C., 2006. Inversion of a forest reflectance model to estimate structural canopy variables from hyperspectral remote sensing data. Remote Sens. Environ. 100, 281–294. Schmidtlein, S., Sassin, J., 2004. Mapping of continuous floristic gradients in grasslands using hyperspectral imagery. Remote Sens. Environ. 92, 126–138. Smith, J.A., 1993. LAI inversion using a back-propagation neural network trained with a multiple scattering model. IEEE Trans. Geosci. Remote Sens. 31, 1102–1106. Solberg, S., Astrup, R., Breidenbach, J., Nilsen, B., Weydahl, D., 2013. Monitoring spruce volume and biomass with InSAR data from TanDEM-X. Remote Sens. Environ. 139, 60–67. Su, Y., Guo, Q., Xue, B., Hu, T., Alvarez, O., Tao, S., Fang, J., 2016. Spatial distribution of forest aboveground biomass in China: estimation through combination of spaceborne LiDAR, optical imagery, and forest inventory data. Remote Sens. Environ. 173, 187–199.

168

X. Quan et al. / International Journal of Applied Earth Observation and Geoinformation 54 (2017) 159–168

Supit, I., Hooijer, A., Van Diepen, C., 1994. System Description of the Wofost 6.0 Crop Simulation Model Implemented in CGMS. Joint research centre, European commission. Svoray, T., Shoshany, M., 2002. SAR-based estimation of areal aboveground biomass (AAB) of herbaceous vegetation in the semi-arid zone: a modification of the water-cloud model. Int. J. Remote Sens. 23, 4089–4100. Talagrand, O., Courtier, P., 1987. Variational assimilation of meteorological observations with the adjoint vorticity equation. I: Theory. Q. J. R. Meteorolog. Soc. 113, 1311–1328. Tian, Y., Woodcock, C.E., Wang, Y., Privette, J.L., Shabanov, N.V., Zhou, L., Zhang, Y., Buermann, W., Dong, J., Veikkanen, B., Häme, T., Andersson, K., Ozdogan, M., Knyazikhin, Y., Myneni, R.B., 2002. Multiscale analysis and validation of the MODIS LAI product: i Uncertainty assessment. Remote Sens. Environ. 83, 414–430. Tian, X., Su, Z., Chen, E., Li, Z., van der Tol, C., Guo, J., He, Q., 2012. Estimation of forest above-ground biomass using multi-parameter remote sensing data over a cold and arid area. Int. J. Appl. Earth Obs. 14, 160–168. Tsui, O.W., Coops, N.C., Wulder, M.A., Marshall, P.L., McCardle, A., 2012. Using multi-frequency radar and discrete-return LiDAR measurements to estimate above-ground biomass and biomass components in a coastal temperate forest. ISPRS J. Photogramm. 69, 121–133. Van Deventer, A., Ward, A., Gowda, P., Lyon, J., 1997. Using Thematic Mapper data to identify contrasting soil plains and tillage practices. Photogramm. Eng. Remote Sens. 63, 87–93. Verhoef, W., 1984. Light scattering by leaf layers with application to canopy reflectance modeling: the SAIL model. Remote Sens. Environ. 16, 125–141. Wang, Y., Woodcock, C.E., Buermann, W., Stenberg, P., Voipio, P., Smolander, H., Häme, T., Tian, Y., Hu, J., Knyazikhin, Y., Myneni, R.B., 2004. Evaluation of the MODIS LAI algorithm at a coniferous forest site in Finland. Remote Sens. Environ. 91, 114–127. Wang, L., Hunt, J.E.R., Qu, J.J., Hao, X., Daughtry, C.S.T., 2011a. Towards estimation of canopy foliar biomass with spectral reflectance measurements. Remote Sens. Environ. 115, 836–840. Wang, L.L., Qu, J.J., Hao, X.J., Hunt, E.R., 2011b. Estimating dry matter content from spectral reflectance for green leaves of different species. Int. J. Remote Sens. 32, 7097–7109. Wang, L.L., Hunt, E.R., Qu, J.J., Hao, X.J., Daughtry, C.S.T., 2013. Remote sensing of fuel moisture content from ratios of narrow-band vegetation water and dry-matter indices. Remote Sens. Environ. 129, 103–110. Wold, S., Sjöström, M., Eriksson, L., 2001. PLS-regression: a basic tool of chemometrics. Chemom. Intell. Lab. Syst. 58, 109–130.

Xie, Y., Sha, Z., Yu, M., Bai, Y., Zhang, L., 2009. A comparison of two models with Landsat data for estimating above ground grassland biomass in Inner Mongolia, China. Ecol. Modell. 220, 1810–1818. Xing, M., Quan, X., Li, X., He, B., 2014. An extended approach for biomass estimation in a mixed vegetation area using ASAR and TM data. Photogramm. Eng. Remote Sens. 80, 429–438. Yan, F., Wu, B., Wang, Y., 2015. Estimating spatiotemporal patterns of aboveground biomass using Landsat TM and MODIS images in the Mu Us Sandy Land. China Agr. For. Meteorol. 200, 119–128. Yebra, M., Chuvieco, E., 2009a. Generation of a species-specific look-up table for fuel moisture content assessment. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2, 21–26. Yebra, M., Chuvieco, E., 2009b. Linking ecological information and radiative transfer models to estimate fuel moisture content in the Mediterranean region of Spain: solving the ill-posed inverse problem. Remote Sens. Environ. 113, 2403–2411. Yebra, M., Chuvieco, E., Riano, D., 2008. Estimation of live fuel moisture content from MODIS images for fire risk assessment. Agr. For. Meteorol. 148, 523–536. Yebra, M., Dennison, P.E., Chuvieco, E., Riano, D., Zylstra, P., Hunt, E.R., Danson, F.M., Qi, Y., Jurdao, S., 2013. A global review of remote sensing of live fuel moisture content for fire danger assessment: moving towards operational products. Remote Sens. Environ. 136, 455–468. Yin, C., He, B., Quan, X., Liao, Z., 2016. Chlorophyll content estimation in arid grasslands from Landsat-8 OLI data. Int. J. Remote Sens. 37, 615–632. Zarco-Tejada, P.J., Rueda, C.A., Ustin, S.L., 2003. Water content estimation in vegetation with MODIS reflectance data and model inversion methods. Remote Sens. Environ. 85, 109–124. Zhang, G., Ganguly, S., Nemani, R.R., White, M.A., Milesi, C., Hashimoto, H., Wang, W., Saatchi, S., Yu, Y., Myneni, R.B., 2014. Estimation of forest aboveground biomass in California using canopy height and leaf area index estimated from satellite data. Remote Sens. Environ. 151, 44–56. Zheng, D., Rademacher, J., Chen, J., Crow, T., Bresee, M., Le Moine, J., Ryu, S.-R., 2004. Estimating aboveground biomass using Landsat 7 ETM+ data across a managed landscape in northern Wisconsin, USA. Remote Sens. Environ. 93, 402–411. Zheng, G., Chen, J.M., Tian, Q.J., Ju, W.M., Xia, X.Q., 2007. Combining remote sensing imagery and forest age inventory for biomass mapping. J. Environ. Manage. 85 (3), 616–623. Zhu, Z., Woodcock, C.E., 2012. Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sens. Environ. 118, 83–94. Zolkos, S., Goetz, S., Dubayah, R., 2013. A meta-analysis of terrestrial aboveground biomass estimation using lidar remote sensing. Remote Sens. Environ. 128, 289–298.