Global-scale biomass estimation based on machine learning and deep learning methods

Abstract

Modeling accurate aboveground biomass (AGB) is a critical aspect of remote sensing research. Besides selecting the appropriate model from significant inputs, integrating optical and microwave products can improve biomass estimation accuracy. Three models were used to estimate biomass to accomplish these goals: Convolutional Neural Network (CNN), Random Forest (RF), and Support Vector Machine for Regression (SVR). Two feature selection techniques, Support Vector Machine for Regression Feature Selection (SVRFS) and Random Forest Feature Selection (RFFS), were used to select the most significant variables to include in the AGB estimation models. Grid search with cross-validation was employed to optimize the hyperparameters for each algorithm. This research aimed to reduce the effects of Radio Frequency Interference (RFI), where RFI filtering outperformed the remaining three models (R2 values were 0.79, 0.80, and 0.82 without reducing RFI effects and were 0.85, 0.87, and 0.88 with reducing RFI effects for SVR, RF and CNN models respectively). Two feature selection techniques jointly selected the variables tree height, vegetation optical depth (VOD), and normalized difference vegetation index (NDVI). The highest Root-Mean-Square deviation (RMSE) values were obtained when CNN and RF models were used in conjunction with the SVRFS method (31.22 Mg/ha and 31.28 Mg/ha for CNN and RF models using the SVRFS model, respectively). This study investigated machine learning and deep learning algorithms’ capability to improve global-scale AGB estimation.