Deep learning-based ecological resistance surface prediction for watershed-scale ecological security pattern optimization in the Fenhe River Watershed, China

Constructing a robust ecological security pattern (ESP) is fundamental for territorial ecological restoration and spatial optimization. Conventional ecological resistance surface (ERS) construction relies on subjective weighting and linear aggregation, while machine learning methods often inadequately capture spatial neighborhood relationships and multi-scale context. We developed a Weighted Residual–Convolutional Block Attention Module U-Net (WRCA-UNet) and integrated it with ecological source (ES) identification, ERS prediction, ecological corridor (EC) and node extraction, and 2054 multi-scenario simulation for the Fenhe River Watershed using multi-source spatiotemporal data from 2000 to 2024. WRCA-UNet outperformed Extreme Gradient Boosting (XGBoost) in predicting the resistance-reference label, reducing mean absolute error (MAE) by 16.31% and increasing R2 to 0.86. Important ES increased from 8117.71 to 9457.14 km2 and then declined to 7764.26 km2, while EC number and spatial configuration varied markedly over time. A “one ring, two cores, three belts, and four zones” optimization framework was identified. By 2054, potential ES declined under natural development, farmland protection, and economic development scenarios, but remained near the 2024 baseline under ecological protection. The framework provides data-driven spatial support for relative proxy ERS prediction, ecological restoration, and watershed-scale spatial optimization.