Model-based Estimation of Forest Growing Stock Volume with ALOS PALSAR Backscatter and Polarimetric Parameter
Ling, F.¹; Chen, E.^2
1Fuzhou University, CHINA; ²Chinese Academy of Forestry, CHINA

The water cloud model (WCM) (Attema and Ulaby, 1978) is a simple semi-empirical radar backscattering model for vegetation and has been adapted to estimate AGB or growing stock volume (GSV), which can be converted to forest aboveground biomass (AGB), from synthetic aperture radar (SAR) backscatter (Pulliainen et al., 1994; Askne et al., 1995) and interferometric coherence (Askne et al., 1997). These adapted models have been successfully applied to boreal forests (Pulliainen et al., 2003, Santoro et al., 2002, Drezet et al., 2007); in these applications, the models were trained with in-situ AGB or GSV measurements from forest inventory. However, forest inventory datasets are rarely available for regional or global mapping, and the methods depending on the forest inventory cannot take into account of spatial variations of environmental conditions, which affect the accuracy of estimation (Askne and Santoro, 2009).

Several methods were developed for GSV estimation from SAR data independent of forest inventory datasets. In Wagner et al. (2003), the unknown parameters of an empirical model were derived from the coherence histogram of each ERS scene. By stepwise regression, Sun et al. (2011) showed the potential of the combined use of lidar samples and radar imagery for forest biomass mapping with the Laser Vegetation Imaging Sensor data and the Advanced Land Observing Satellite (ALOS) Phased Array type L-band Synthetic Aperture Radar (PALSAR) data. To estimate GSV from ERS-1/2 tandem coherence, Askne and Santoro (2009) demonstrated an automatic approach to train the interferometric water cloud model (IWCM) (Askne et al., 1997) using multi-temporal data. In this approach, the areas of open ground and dense forest were identified with temporally consistent low and high coherence separately; the model parameters were then inferred from the pixel values of open ground and dense forest. This training approach, however, need multi-temporal coherence images, making it useful for only selected test sites. Santoro et al. (2011) proposed the BIOMASAR algorithm to estimate GSV in boreal forest using hyper-spectral series of Envisat ASAR ScanSAR backscatter measurements and achieved satisfactory results at 1 km pixel size. In the BIOMASAR algorithm, MODIS Vegetation Continuous Fields product (MODIS VCF) (Hansen et al., 2003) was used to identify the open ground and dense forest for the estimation of the parameters of WCM. Cartus et al. (2011) also used MODIS VCF to train the IWCM to estimate GSV from ERS-1/2 tandem coherence and generated a GSV map for Northeast China at 50 m pixel size. Although MODIS VCF was successfully used in these two applications, its course resolution (500 m) prevents it from accurate estimation of the model parameters when targeting at GSV map with higher resolutions. To map forest AGB in the Northeast United States with ALOS PALSAR dual-polarization backscatter, Cartus et al. (2012) used a national map of canopy density with a pixel size of 30 m to train the WCM with gaps (Askne et al., 1995) for automation; however, such kind of canopy density map with high resolution is usually unavailable for other regions.

In order for any GSV estimation from SAR data to be independent of forest inventory datasets, extra information is needed either from multi-temporal SAR measurements (Askne and Santoro, 2009) or from optical data (Sun et al., 2011; Santoro et al., 2011; Cartus et al., 2011; Cartus et al., 2012). The phase information in ALOS PALSAR Fine Beam Dual (FBD) mode data is well preserved and can be exploited with SAR polarimetry to describe the scattering of the earth targets. In this study, we used the WCM with gaps (Askne et al., 1995) to estimate forest GSV with ALOS PALSAR FBD data at single acquisition for the temperate forest in a test site in Northeast China. In our method, the model was trained with the aid of Shannon entropy from the H/alpha decomposition of the C2 covariance matrix, making our method independent of forest inventory datasets.