Enhancing carbon stock estimation in forests: Integrating multi-data predictors with random forest method

Abstract

Forests are crucial to the global carbon cycle, making accurate measurement of biomass essential for evaluating their carbon capture potential. This study presents a novel approach to estimate and map carbon stocks and sequestration potential in dense forests, by integrating multisensory remote sensing data with often-overlooked abiotic variables such as terrain characteristics, socio-economic factors, and accessibility. By using the LASSO method to analyse predictors and Random Forest regression models, the study achieved a 10 % increase in the coefficient of determination when abiotic variables were included. The optimal satellite data configuration – which combined median of the summer multispectral images with the mean of the November Synthetic Aperture Radar (SAR) images – resulted in normalised root mean square error (nRMSE) and normalised mean absolute error (nMAE) values of 17 % and 14 %, respectively. The green band from Sentinel-2 emerged as the most significant variable, followed by vegetation type or physical predictors such as plot size or population density. Consequently, carbon maps were generated alongside uncertainty maps, providing a clearer assessment of model reliability. Additionally, a Random Forest classification model for land cover achieved an accuracy of 82 %. Therefore, this study highlights the importance of integrating vegetation types and spectral data with physical environmental variables, such as climate data or land register data, to enhance accuracies and robustness in carbon estimation models; thereby providing a more comprehensive understanding of forest carbon stocks and their sequestration potential.