High-accuracy Machine Learning Models to Estimate above Ground Biomass over Tropical Closed Evergreen Forest Areas from Satellite Data

Abstract

Quantifying the amount of biomass stored in forested areas has been traditionally done with manual field measurements, which is costly, time consuming and doesn't scale well over large areas. This paper investigates the possibility to estimate the amount of Above Ground Biomass (AGB) using machine learning models with publicly available satellite data, where a large-scale training dataset has been created from a detailed biomass mapping project in the Democratic Republic of Congo (DRC). Several model architectures including the current state-of-the-art tree-based models were tested along with deep neural network (DNN) ones. It was found that DNN models provide slight improvement in accuracy, whilst can potentially be used for further fine-tuning with smaller local dataset for usage elsewhere outside of the DRC.