Dominant Expression of SAR Backscatter in Predicting Aboveground Biomass: Integrating Multi-Sensor Data and Machine Learning in Sikkim Himalaya

Accurate assessment of aboveground biomass (AGB) is crucial for understanding carbon budgets, climate change impacts, and evaluating forest responses to environmental shifts. In this study, AGB was estimated in Sikkim State of India by leveraging the capabilities of machine learning (ML) and integrating multi-sensor satellite data. Specifically, the random forest (RF) and categorical boosting algorithm (CatBoost) models were utilised. Field estimated AGB ranges from 1.99 to 530.02 Mg/ha with an average of 252.58 Mg/ha, utilised for model prediction and validation. The RF model slightly outperformed the CatBoost model, with a coefficient of determination (R²) of 0.71 and root mean square error (RMSE) of 72.98 Mg/ha, compared to the CatBoost model’s R² of 0.67 and RMSE of 80.69 Mg/ha, The former showed a greater capacity to combat overfitting. Synthetic aperture radar variables have emerged as significant predictors because of their contribution to the structural properties of plants. This study acknowledges the limitations and challenges due to data availability, especially for ground truth measurements, which pose constraints on the accuracy and representativeness of AGB estimates. Uncertainties associated with AGB estimation, such as variations in vegetation structure and species composition, also affected model performance. Despite these limitations, this study emphasises the significance of multi-sensor data integration and ML models in AGB estimation and highlights their potential applications in forest management and climate change mitigation efforts in the Himalayan mountainous region.