Estimation of Forest Aboveground Biomass Using Multitemporal Quad-Polarimetric PALSAR-2 SAR Data by Model-Free Decomposition Approach in Planted Forest

Remote sensing technology has emerged as a promising approach for indirectly mapping forest aboveground biomass (AGB) over large areas with limited ground measurements. Polarimetric synthetic aperture radar (PolSAR) images exhibit significant potential in accurately mapping AGB in mountainous forests due to their excellent penetration capability. However, the accuracy and reliability of mapping forest AGB are often constrained by the quality of features and uncertain atmospheric conditions. Therefore, it is crucial to obtain relevant features closely associated with forest AGB. In this study, to overcome the disadvantages of common polarization decomposition methods, model-free decomposition methods were firstly applied to map forest AGB by extracting decomposition features from L-band quad-polarimetric PALSAR-2 images. Additionally, various data combination methods used to integrate the multi-temporal quad-polarimetric SAR images were also applied to investigate the combined effects of the datasets from different acquisition dates and evaluate their capability in reducing errors caused by uncertain atmospheric conditions. Ultimately, a forward feature selection method was employed along with four regression models to effectively map forest AGB. The results demonstrated that the model-free decomposition method exhibits less sensitivity towards changes in window size compared to other methods. Furthermore, there is a significant improvement observed in the sensitivity between forest AGB and features extracted through model-free decompositions compared to common polarization decomposition methods; some features also demonstrated superior capabilities in capturing forest structure information related to AGB. Moreover, given the model-based or model-free decomposition methods, using the combined datasets from multi-temporal SAR images led to a substantial increase in determination coefficient (R2) and a great decrease of relative root mean square error (rRMSE) of mapping forest AGB for each regression method than using the individual images. Given a dataset, the model-free decomposition methods offered a significant improvement of the accuracy of mapping AGB compared with the model-based decomposition approaches. The optimal result was achieved through the combination of six images and using the model-free decomposition method and random forest regression, with the rRMSE value of 22.95%. Furthermore, this study also implied that the integration of multi-temporal images partially alleviates saturation effect.