High-density forest AGB estimation in tropical forest integrated with PolInSAR multidimensional features and optimized machine learning algorithms

Abstract

Accurately estimating the above-ground biomass (AGB) of high-density tropical rainforests is a challenging issue. In this study, airborne multi-baseline PolInSAR data were used to estimate the tropical rainforest AGB in Gabon, Africa. The most suitable baseline combination of the PolInSAR data was initially determined through baseline selection, and the PolInSAR parameters related to forest height were obtained based on the forest canopy height estimation theory and microwave penetration depth theory. The height parameter, baseline parameter, and observation geometry parameter were then used as independent variables to construct the AGB regression model. Support vector regression (SVR) was chosen as the AGB estimation model, and the global best particle swarm algorithm (GLB-PSO) was used to optimize the SVR model’s parameters. The results show that the RFECV variable selection method is superior to the Pearson method. The GLB-PSO algorithm can also further improve the saturation point of the SVR model—the estimation results show that the saturation point of AGB estimation of PolInSAR multidimensional features combined with the SVR machine learning algorithm is up to 500 Mg/ha, while this saturation point can be increased to 650 Mg/ha when using the GLB-PSO-SVR algorithm.