Leveraging SAR and Optical Remote Sensing for Enhanced Biomass Estimation in the Amazon with Random Forest and XGBoost Models

Abstract

This study addresses the challenge of estimating above-ground biomass (AGB) in the Amazon rainforest by developing a reference geographical database, which provides the ground truth, and comparing the relative importance of using Synthetic Aperture Radar (SAR) and optical remote sensing data to automatically infer AGB. In the experiments reported in this article, we assessed how those two remote sensing data sources impact the accuracy of AGB estimates produced by regression models built with Random Forest (RF) and Extreme Gradient Boosting (XGBoost). The research involved compiling a comprehensive database from many available forest inventories, integrating parcel- and tree-level data to enable precise biomass estimation. The methodology included setting up a spatial data analysis environment, standardizing data, and implementing an experimental protocol with feature selection and leave-one-out cross-validation. The results demonstrate that both kinds of data, i.e., SAR and optical, and their combination can be used for estimating AGB, providing valuable insights for forest management and climate change mitigation efforts. The reference database is available upon request to the corresponding authors.