DOI: 10.11648/j.acm.20261505.16 ISSN: 2328-5613

Estimating and Mapping Forest Carbon Sequestration in the Mau Forest Complex Using Sentinel-2 Remote Sensing and Machine Learning

Fred Onyango, Charity Wamwea, Amos Langat
Show PDF Cite
Accurate forest-carbon estimates are essential for climate mitigation, restoration planning and monitoring, reporting and verification, yet conventional field inventories are expensive and spatially discontinuous. This study evaluated whether freely available Sentinel-2 imagery combined with statistical and machine-learning models can estimate and map forest carbon across the Mau Forest Complex, Kenya. Existing geo-referenced forest inventory data comprising 3,907 measured trees were linked to seven Sentinel-2 vegetation indices: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Moisture Index (NDMI), Normalized Burn Ratio (NBR), red-edge Chlorophyll Index (CIre) and Red-Edge Normalized Difference Vegetation Index (RENDVI). Above-ground biomass was estimated using a pan-tropical allometric equation; below-ground biomass, carbon stock and CO2 equivalent were subsequently derived using IPCC factors. Ordinary linear regression, a linear mixed-effects model, Random Forest and Extreme Gradient Boosting (XGBoost) were compared using a 70:30 training-test split, with five-fold cross-validation for machine-learning hyperparameter tuning. All vegetation indices were positively associated with above-ground biomass (p<0.001), with CIre (Pearson r=0.530), RENDVI (r=0.516) and NDMI (r=0.513) among the strongest correlates. On the independent test set, XGBoost performed best (R 2 =0.473, RMSE=0.424, MAE=0.327), followed by Random Forest (R 2 =0.468, RMSE=0.426, MAE=0.329), the mixed-effects model (R 2 =0.454, RMSE=0.431, MAE=0.333) and linear regression (R 2 =0.436, RMSE=0.438, MAE=0.340). XGBoost reduced RMSE by 3.3% relative to the linear baseline. Residual spatial autocorrelation was negligible (Moran’s I=0.059). Spatial prediction identified higher carbon densities in western and south-western portions of the Complex and lower values in more disturbed eastern blocks. The study contributes a reproducible, low-cost benchmark of linear, mixed-effects and ensemble models for spatial forest-carbon monitoring in a heterogeneous African montane forest; the modest gain from nonlinear models shows that optical indices alone limit accuracy, so future improvements will likely require structural information from radar or LiDAR.

More from our Archive