Evaluating machine learning approaches for aboveground biomass prediction in fragmented high-elevated forests using multi-sensor satellite data

Abstract

Accurate aboveground biomass (AGB) estimations over large areas are essential for assessing carbon stocks and forest resources. This study evaluated machine learning approaches for AGB modeling in Pakistan's mountainous region of Diamir district using freely available Sentinel-1 and Sentinel-2 data and 171 field-measured AGB training points. Random Forest, Gradient Tree Boosting, CatBoost, LightGBM, and XGBoost algorithms were implemented and optimized. Models were developed using individual and combined datasets. Sentinel-2 optical data outperformed Sentinel-1 radar data, but the fusion of both sensors achieved the highest accuracy (R2 > 0.7, RMSE = 105.64 Mg/ha, MAE = 85.34 Mg/ha). Tree canopy height was the most informative predictor for this data, besides terrain variables and radar textures. The machine learning models significantly improved AGB estimates compared to traditional regression techniques, and gradient boosters outperformed Random Forest. This research demonstrates the potential of multi-sensor remote sensing data and advanced algorithms for forest biomass mapping in complex terrain, with modeling accuracies reaching root mean squared errors below 90 Mg/ha. The framework provides an effective solution for monitoring biomass using freely available satellite data. Further refinements include integrating higher-resolution optical data and additional field samples for better validation. This study contributes to remote sensing capabilities for assessing vegetation carbon stocks and dynamics.