Terrain-informed indicators for gully erosion in the sensitive ecological zones of loess plateau via flow algorithms and deep learning

Sensitive ecological zones, Loess Plateau, China We develop a terrain-informed workflow that fuses an improved flow-direction scheme with a lightweight semantic-segmentation network (ICGNet) to map runoff pathways and gully channels from high-resolution digital elevation models (DEMs). The framework quantifies three indicators central to erosion risk—upslope slope length, open-channel extent, and gully density—using an iGD8 formulation of specific contributing area and multi-scale feature fusion to preserve hydrologic continuity. Requiring only DEMs, the approach enables consistent, region-wide mapping, rapid scenario testing, and transfer across small catchments with heterogeneous slope types while retaining physical interpretability and computational efficiency. ICGNet delineated open channels from DEM-derived features with high predictive performance, achieving a mean pixel accuracy of 92.56 % and a mean Intersection-over-Union of 89.66 %, and outperforming the benchmark segmentation networks. The optimized slope-length model reproduced runoff pathways and channel-related topographic structure with high accuracy (overall accuracy 96.74 %, F1-score 92.77 %, MSE 0.0981, RMSE 0.3063 and MAE 0.2133) and stable inference time (43.48 s). Predicted gully density across seven representative slope types closely matched field observations (RMSE 0.0144), revealing terrain-driven convergence zones and erosion hotspots characteristic of the Loess Plateau. Error analysis based on PBIAS, RMSE and MAE showed that the proposed framework substantially reduced bias and overall error in slope-length and gully-density estimates compared with conventional flow-direction schemes. The resulting indicators support the prioritisation of sub-catchments and the placement of check structures, vegetative buffer strips and channel-protection works, and provide process-consistent diagnostics for evaluating land-use and climate-change scenarios in erosion-sensitive landscapes. ● Terrain-informed framework derives DEM-based soil-erosion indicators. ● Refined flow-direction and slope-length methods capture topographic controls. ● Lightweight deep learning maps open channels accurately in complex terrain. ● Modeled gully density aligns with observations across heterogeneous slopes. ● Scalable, region-ready diagnostics support targeted erosion mitigation.