Hybrid Feature Selection and Ensemble Learning for Aboveground Carbon Mapping in Oil Palm Plantations Using Multi-Source Satellite Data

Oil palm plantations play an important role in agricultural production and carbon storage in tropical regions. The accurate estimation of aboveground carbon (AGC) is essential for sustainable plantation management, climate change mitigation, and carbon monitoring. Although field measurements provide reliable estimates, they are often time-consuming, labor-intensive, and costly, particularly over large plantation areas. Recent advances in remote sensing and machine learning offer efficient alternatives for AGC estimation using satellite imagery. In this study, we developed a machine learning framework for AGC estimation in oil palm plantations using Sentinel-2 multispectral imagery and Sentinel-1 synthetic aperture radar (SAR) data. Field measurements were integrated with spectral variables, vegetation indices, and SAR-derived parameters extracted from satellite data. A hybrid feature selection approach combining Pearson correlation, mutual information and mRMR was used to identify the most relevant variables. Six machine learning algorithms were evaluated, including Linear Regression, Random Forest, XGBoost, Gradient Boosting, LightGBM, and Extra Trees. Because the 160 observations comprise sixteen 10 m × 10 m grid cells nested within ten 40 m × 40 m field plots, model performance was assessed with leave-one-plot-out cross-validation: all sixteen cells of a plot were held out together, and the hybrid feature selection was repeated inside every fold using only that fold’s training plots. Performance was measured on pooled out-of-fold predictions using R², root mean squared error (RMSE), and average absolute relative error (AARE%). Under this spatially independent design the combined Sentinel-1 + Sentinel-2 dataset gave the highest accuracy (R² = 0.7950, RMSE = 4.14 t C ha⁻¹, AARE = 33.53%), followed by Sentinel-1 alone (R² = 0.7631, RMSE = 4.46 t C ha⁻¹) and Sentinel-2 alone (R² = 0.6108, RMSE = 5.71 t C ha⁻¹). Linear Regression and Extra Trees were the most robust models, whereas the boosted ensembles did not generalize to unseen plots. Repeating the evaluation with an ungrouped random split of the same data inflated R² by up to 0.70, showing that a large part of the accuracy obtainable under that design reflects within-plot spatial autocorrelation rather than predictive skill. These findings indicate that optical-SAR imagery combined with machine learning can provide useful AGC estimates in oil palm plantations, and that spatially independent validation is essential for reporting them honestly. The proposed framework can be used to support plantation-scale carbon mapping, monitoring, and carbon stock assessment, subject to further calibration and independent validation across additional plantations.