Mapping Smallholder Oil Palm Carbon Stocks Using Machine Learning

Abstract

Accurate and spatially explicit estimation of aboveground carbon stocks in smallholder oil palm plantations remains challenging, limiting greenhouse gas accounting and digital carbon traceability in Indonesia. Conventional approaches rely on field-based allometric equations that are often unreliable in peatland and marginal environments, while widely used vegetation indices such as the Normalized Difference Vegetation Index (NDVI) suffer from spectral saturation under dense mature canopies. This study developed a satellite-based machine learning framework for estimating aboveground carbon stocks in smallholder oil palm plantations in Pelalawan Regency, Riau Province, by integrating multitemporal Sentinel-2 imagery, red-edge vegetation indices, Gray Level Co-occurrence Matrix (GLCM) texture features, and GEDI L4A spaceborne LiDAR data within the Google Earth Engine platform. GEDI-derived biomass data were used as pseudo-ground truth to train Random Forest and XGBoost regression models using 4,990 quality-filtered samples, and model performance was evaluated against an independent test dataset and compared with a conventional NDVI-based regression model. Random Forest achieved the highest performance, reducing the Root Mean Square Error (RMSE) by 6.0% relative to the NDVI baseline (16.75 vs. 17.82 tC ha?¹) while attaining an overall classification accuracy of 62.4% (kappa = 0.28). The model estimated a mean aboveground carbon stock of 29.99 tC ha?¹ and a total regional carbon stock of 2.67 MtC across the mapped smallholder plantations. Feature importance analysis identified shortwave infrared reflectance (Band 11) and red-edge GLCM contrast as the most influential predictors, demonstrating that integrating texture and red-edge information effectively improves carbon stock estimation under dense oil palm canopies. The resulting carbon stock maps were deployed through a publicly accessible WebGIS platform, providing a scalable framework for regional carbon monitoring, greenhouse gas inventory development, and digital carbon traceability in fragmented smallholder oil palm landscapes.