Optimizing ecological security patterns in western Jilin, China: comparative ecological source identification and explainable machine-learning resistance modeling

Agricultural expansion and landscape fragmentation are altering ecological connectivity in agro-pastoral transition zones. Ecological security patterns can support landscape conservation and restoration, but their reliability is constrained by uncertainty in ecological source identification and resistance surface construction. This study identified suitable methods for constructing ecological security patterns in ecologically fragile agricultural landscapes. We compared three ecological source identification approaches and five machine learning models for resistance surface construction, and assessed their implications for ecological network optimization in western Jilin, China. We assessed five ecosystem services and compared ordered weighted averaging, morphological spatial pattern analysis, and ecosystem service significance assessment for ecological source identification. We compared five machine learning models to construct the ecological resistance surface. Based on the selected sources and resistance surface, we used minimum cumulative resistance and circuit theory to identify corridors, pinch points, and barriers. Ecosystem services showed spatial heterogeneity, with most services decreasing from east to west. Scenario 4 of ordered weighted averaging (OWA4) achieved the highest conservation efficiency for biodiversity, Ramsar wetlands, and nature reserves, whereas morphological spatial pattern analysis (MSPA) provided the greatest coverage of these elements. Given its higher conservation efficiency and more balanced representation of multiple ecosystem services, OWA4 was selected for subsequent connectivity planning. RF and XGBoost showed the strongest overall performance; RF achieved the higher AUC, whereas XGBoost provided a more balanced representation of resistance factors. Both were retained for resistance surface construction. The final ecological security pattern included 23 ecological sources covering 3719.62 km 2 , 30 corridors, 243 pinch points, and 31 barrier areas. Urban areas, croplands, roads, and saline-alkali lands were major constraints on ecological flows. Integrating comparative source identification, machine learning-based resistance modeling, and circuit theory reduced uncertainty in ecological security pattern construction. The resulting “two corridors, two belts, and three zones” framework can guide wetland restoration, corridor protection, and landscape planning in agricultural-wetland landscapes.