Parsimonious machine learning for the global mapping of aboveground biomass potential

Abstract

Advances in computational power and methods, and the widespread availability of remote sensing data have driven the development of machine learning models for estimating global carbon storage. Current models often rely on dozens of predictor variables to estimate aboveground biomass density (AGBD), resulting in accurate but complex models that are challenging to interpret from a biological and ecological standpoint. Yet, it remains unclear whether such model complexity is essential to achieving accurate predictions. This manuscript investigates the potential to create a simpler, yet accurate, global AGBD model. Our approach leverages only climate-based predictors, using a systematic predictor selection process to determine the optimal subset of variables that maximize model accuracy. Surprisingly, we found that a minimal model trained with only four bioclimatic variables outperformed more complex models. When compared to a state-of-the-art complex model and ground-based data, our model achieved comparable accuracy using only four predictors, far fewer than the 186 predictors used in the complex model. In conclusion, we present a lightweight, interpretable climate-based model for AGBD estimation, with the additional advantage of being adaptable for projecting AGBD under future climate scenarios.