Forest above-ground biomass estimation using X, C, L, and P band SAR polarimetric observations and different inversion models

Abstract

ABSTRACTSince each frequency senses distinct features of forest structure, it is attractive to understand whether complementarity of multi-frequency can improve the retrieval of forest above ground biomass (AGB). In this study, 15 combinations of X, C, L, and P band SAR observations were applied in the forest AGB estimation through multivariate linear stepwise regression (MLSR), random forest (RF), and deep learning algorithm based on Keras and TensorFlow algorithms to fully explore their potential and capability for forest AGB retrieval. 21 SAR observations were derived from each frequency and worked as SAR observations for forest AGB estimation. According to the retrieval results using three inversion algorithms and 15 SAR observation combinations, MLSR algorithm with combined L and P band SAR observations performed best with R2 = 0.67 and RMSE = 14.51 Mg/ha. Next was RF algorithm with combination of L and P band, the R2 and RMSE are 0.64 and15.20 Mg/ha, respectively. L and P band are more suitable frequencies for estimating forest AGB, only limited improvements can be achieved by combing them with X and C band. Combination of L and P band can obtain comparable AGB estimation accuracy with using combinations from tri- or quad-frequency SAR observations.

Keywords

• Forest AGB
• multi-frequency SAR observables
• MLSR
• RF
• a deep learning algorithm