Abstract

not-yet-known not-yet-known not-yet-known unknown Projecting how climate change will reshape global patterns of forest carbon storage is paramount to investigate the future of the global carbon cycle. Advances in machine learning and the proliferation of remote-sensing missions have stimulated the development of statistical models for the estimation of carbon storage on a global scale. However, existing models present two major limitations. First, they are not suitable for forecasting because they are trained on variables that cannot be projected. Second, they are based on large numbers of predictors, which hinders the interpretability and explainability of their results from a biological and ecological perspective. In this manuscript, we present a machine learning model for the global mapping of potential Aboveground Biomass Density that overcomes the above mentioned limitations. On the one hand, our model is based exclusively on bioclimatic variables and is therefore suitable for forecasting and interpretable from a biological perspective. On the other hand, we calibrated our model with a state-of-the-art predictor selection process to identify the optimal subset of predictors for the maximization of model accuracy. We found that a minimal model based solely on four bioclimatic variables performed the best. Furthermore, our model reproduces field measurements from experimental forest plots equally well than a state-of-the-art model trained on $190$ predictors. In conclusion, we provide a simple and computationally cheap model that can be coupled with climate projections to investigate how climate change will reshape carbon storage on a global scale.