Intelligent Identification of Key Nutrient Loss Nodes in Topsoil Stockpiling Based on ST-GCN and Multi-Source Heterogeneous Data Fusion

Stripped topsoil from power transmission projects faces continuous nutrient degradation during stockpiling, compromising subsequent vegetation restoration. This study proposes an intelligent method to identify key nutrient loss nodes by fusing Spatio-Temporal Graph Convolutional Networks (ST-GCN) with multi-source heterogeneous data. Indoor simulations monitored physical, chemical, and biological indices of Shandong brown soil, Jiangxi red soil, and Tibetan alpine meadow soil under varying storage conditions. An ST-GCN model predicted spatio-temporal nutrient evolution, while a sliding-window algorithm detected mutation points. Results indicated that alpine meadow soil suffered the most severe decline (39.3% chemical factor loss), followed by brown soil (25.0%) and red soil (17.6%). Strip stacking best preserved physical structure (78.9% retention), while tarpaulin covering maximized biological activity retention (64.0%). Coupling analysis revealed that physical degradation precedes biological response, triggering chemical loss. Under the optimal combination (strip stacking plus tarpaulin), key loss nodes were identified at 30–60 days for alpine meadow soil and red soil (with a delayed peak in the latter), and 60–90 days for brown soil.This method accurately captures differential nutrient attenuation patterns, providing a scientific basis for precise topsoil conservation management.