Fusion of ALOS Palsar and Landsat ETM data for land cover classification and biomass modeling using non-linear methods

This work demonstrates the utility of reduced resolution ALOS PALSAR data for biomass mapping and land cover classification over the tropical forests of Indonesia. This study is important because we processed the ALOS PALSAR mosaic, which is made freely available within K&C initiatives project and will be updated regularly. We first used 38 sample plots collected on the ground during dry season in September 2004, to develop a tree diameter (dbh)-biomass model. The HH, HV, HV/HH and HH-HV backscatters of ALOS PALSAR data allowed the empirical estimation of forest above ground biomass (AGB). Each band of PALSAR data was separately used to estimate the biomass, and we found HV band resulted in better correlation with the AGB compared to other SAR bands. Validation of the prediction results was carried out by comparing the biomass estimates with those predicted from an existing allometric equation. Optical data are sensitive to the physical properties of the reflectors whereas SAR data are more influenced by the geometric properties of the scatterers. Therefore, the second part of this study concerned the integration of mosaic SAR textures and ETM data for land cover classification. The classification was conducted using ETM data and variations of ETM, SAR bands, and SAR textures calculated using GLC Matrix. The image classifications were carried out using a Machine Learning based classifier, so-called Support Vector Machine (SVM), and a conventional Maximum Likelihood method. An ensemble of neural networks method using Kalman filter and scaled conjugate gradient algorithm was applied. The classification accuracy was assessed using confusion matrices and Kappa statistics. We show that the introduction of SAR textures significantly enhanced the classification accuracies. This study showed that the joint processing of SAR and multispectral data increased the accuracies of biomass estimation and landuse classifications. The efficiency of the method at medium spatial r- esolutions allows its application of global datasets.