VODCA2AGB - A novel approach for the estimation of global AGB stocks based on vegetation optical depth data and random forest regression

Abstract

Aboveground biomass (AGB) plays an important role in the Earth’s carbon cycle, both as a source and sink of carbon, and therefore its monitoring is critical in the fight against climate change. In the past, several AGB maps on global and regional scales have been created to provide an estimate of stocks at a certain point in time. However, most of these maps are static and thus do not allow for monitoring the change in biomass.As an alternative to current approaches, vegetation optical depth (VOD) could be used to monitor global AGB dynamically. VOD is a measure of the attenuation of microwave signals traveling through the vegetation canopy. It has a linear relationship with vegetation water content, which in turn is related to biomass and vegetation water status. Previous studies have already proven the feasibility of AGB estimation based on L-band VOD (L-VOD) using nonlinear univariate parametric regression approaches. However, due to the high complexity of the global AGB-VOD relationship, this approach might be better suited for regional AGB estimation.To overcome this problem, a novel approach for global AGB estimation relying on multivariate non-parametric random forest regression (RF) was developed and tested. For this purpose, RF models were trained on an annual reference AGB product, and predictions were compared with a hold-out test set as well as an independent AGB map. The RF models rely on a set of different feature combinations, including data from VOD, leaf area index (LAI), land cover (LC) and sun induced fluorescence (SIF). The prediction accuracy was assessed via R^2 scores, with results ranging from 0.79 to 0.96 %. The largest increase in model performance was observed when land cover information was used in combination with VOD. Prediction accuracy was the lowest over dense and high vegetation, such as broadleaf evergreen forests, and conversely, the best results were achieved for sparse and low vegetation, like crops and shrubs, which is in line with previous studies. Overall, the thesis showed promising results for the future application of AGB estimation based on VOD and random forest regression.