Predicting Tropical Carbon Stock using Multi-Layer Perceptron: A Multi-Sensor Fusion Approach

Abstract

Accurately mapping Aboveground Biomass (AGB) in tropical regions remains a significant challenge due to the saturation of common remote-sensing sources, such as Sentinel-2 optical imagery and Sentinel-1 C-band radar, in high-biomass forests. This study aims to address these limitations by proposing a data-fusion approach that incorporates L-band radar from ALOS PALSAR to predict carbon stock across the complex landscape of Yogyakarta, Indonesia.We combined GEDI L4A AGBD "ground-truth" data (2020) with features extracted from Sentinel-2, Sentinel-1 texture, ALOS PALSAR L-band backscatter, and SRTM topography. Four machine learning models—Multiple Linear Regression (MLR), Random Forest (RF), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP)—were trained and evaluated. The results demonstrate that models relying solely on conventional sensors (Sentinel-1 and Sentinel-2) performed poorly, explaining only ~18% of the AGBD variation. Integrating L-band and topographic variables more than doubled model performance. The MLP model achieved the highest accuracy, yielding an R² of 0.3389 and an RMSE of 74.26 t/ha.Although the accuracy is moderate, it realistically reflects the inherent noise in GEDI L4A estimates and the region’s highly fragmented forest structure. Ultimately, this research confirms that fusing L-band radar is essential for improving tropical AGB mapping.