Mangrove extent and above-ground biomass mapping in Sarawak, Malaysia using Sentinel data and machine learning

Mangrove forests are critical blue carbon ecosystems yet remain underrepresented in national carbon accounting frameworks due to persistent data gaps at fine spatial scales. This study presents a regionally focused, operationally reproducible remote sensing and machine learning framework to map mangrove extent and estimate above-ground biomass (AGB) across western and central Sarawak, Malaysia. Open-access Sentinel-1 C-band SAR and Sentinel-2 multispectral data were fused with extensive field inventory data from 245 plots to develop classification and regression models using the eXtreme Gradient Boosting (XGBoost) algorithm. While L-band SAR offers known advantages for high-biomass estimation, such datasets were not freely or consistently available for the full temporal and spatial coverage required; the chosen C-band–optical fusion maximises accessibility, repeatability, and integration into regional monitoring workflows. The mangrove extent map, generated at 10 m resolution, achieved an overall accuracy of 88.8%, a Cohen's Kappa of 0.776, and an AUC of 0.960. The predicted mangrove area totalled 1078.7 km2, with 83.8% exhibiting very low classification uncertainty. For AGB, model performance was quantified using held-out predictions to avoid optimistic bias; out-of-fold cross-validated predictions across all plots (n = 245) yielded R2 = 0.72, RMSE = 28.59 Mg ha−1, MAE = 20.80 Mg ha−1, and Pearson's r = 0.85. Monte Carlo simulations (n = 100) confirmed the model's stability, yielding RMSE = 26.25 ± 1.33 Mg ha−1 and R2 = 0.783 ± 0.023 (evaluated on held-out predictions within each iteration). Species-level AGB analysis revealed the highest mean biomass in Rhizophora mucronata (102.1 Mg ha−1) and the lowest in Bruguiera parviflora (74.5 Mg ha−1). SHAP interpretation identified canopy height, NDVI, VH backscatter, and elevation as the most influential predictors. This accessible, fused, and interpretable modelling framework provides a scalable solution for national carbon inventories, including annual emission-factor refinement and spatial change-detection grids , while supporting targeted mangrove conservation and restoration under REDD+ and other climate initiatives.

Text 126123.pdf - Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (9MB) Official URL or Download Paper: https://linkinghub.elsevier.com/retrieve/pii/S2352...

Actions (login required)

View Item