Ecologically Constrained Multi-Task Deep Learning for Interpretable Joint Estimation of Mountain Forest Biomass and Growing Stock Volume

Reliable quantification of forest aboveground biomass (AGB) and growing stock volume (GSV) underpins carbon stock assessment and ecosystem monitoring. However, the complex mountainous terrain makes it difficult to estimate forest parameters, and research on the simultaneous estimation of forest AGB and GSV remains limited. To this end, this study proposed an eco-constrained explainable multi-task deep learning framework (Eco-AMT-MLP) that integrated Sentinel-2 data obtained from the Google Earth Engine platform with topographic, soil, and climatic factors to jointly estimate forest AGB and GSV while incorporating an ecological constraint linking the two prediction tasks in the complex mountainous forests of northwestern Yunnan Province. The model’s performance was evaluated against Random Forest (RF), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and K-Nearest Neighbor Regression (KNN), and ablation experiments were conducted to assess the contribution of the ecological constraint module. The Eco-AMT-MLP model demonstrated the best overall predictive performance among the compared models, achieving an R2 of 0.64 ± 0.08 and an RMSE of 132.99 ± 19.96 m3 ha−1 for GSV prediction, and an R2 of 0.58 ± 0.07 and an RMSE of 87.37 ± 13.10 Mg ha−1 for AGB prediction. Compared with the MLP model, Eco-AMT-MLP increased the R2 values for GSV and AGB predictions by 0.04 and 0.11, respectively, while reducing the corresponding RMSE values by 7.46 m3 ha−1 and 10.12 Mg ha−1, respectively. However, the model shows only limited improvement in AGB estimation accuracy, indicating that its enhancement of AGB estimation is less pronounced than its improvement of GSV estimation. SHAP analysis further identified environmental covariates and Sentinel-2 spectral features as important predictors of forest AGB and GSV. By integrating joint estimation of forest AGB and GSV with ecological constraints and interpretable predictor analysis, this framework provides a methodological reference for regional forest carbon-stock assessment, forest resource inventories, and ecosystem monitoring in complex mountainous regions.