基于随机森林与K最近邻模型的澜沧江中游GX水电站库区滑坡易发性评价OA

Landslides represent one of the most significant geological hazards threatening infrastructure safety and human settlements in mountainous regions worldwide.In the context of large-scale hydropower development in tecton-ically active and topographically complex alpine environments,accurate landslide susceptibility assessment is essential for risk mitigation and sustainable engineering planning.This study is focused on the reservoir area of the GX hydro-power station located in the middle reach of the Lancang River,a region characterized by extreme relief,active fault-ing,and diverse lithological assemblages within the southeastern margin of the Tibetan Plateau.Covering an area of ap-proximately 980 square kilometers,the study zone exhibits pronounced spatial heterogeneity in landslide distribution,with a notable concentration along the left bank of the Lancang River valley.To address the challenges posed by such complex terrain,the aim of this research is set to evaluate and compare the performance of two machine learning mod-els,Random Forest(RF)and K Nearest Neighbor(KNN),in mapping landslide susceptibility and identifying domi-nant controlling factors.Based on remote sensing imagery from Google Earth Pro and field validation,a comprehen-sive inventory of 980 landslides is established.To ensure balanced model training,an equal number of non-landslide samples are randomly generated outside a 100-meter buffer around all identified landslides.Eleven environmental fac-tors are initially considered,including elevation,slope,aspect,plan curvature,profile curvature,roughness,relative re-lief,lithology,distance to faults,normalized difference vegetation index(NDVI),and distance to rivers.After assess-ing multicollinearity using Pearson correlation coefficients,roughness is excluded due to its high correlation with slope,resulting in a final set of 10 independent predictors.All spatial analyses and modeling are implemented within a GIS environment using a 30-meter resolution NASADEM dataset.The RF and KNN models are trained on 70 percent of the samples and validated on the remaining 30 percent.Model performance is evaluated using the Area Under the Re-ceiver Operating Characteristic Curve(AUC).The RF model achieved an AUC of 0.776 9,slightly outperforming the KNN model with an AUC of 0.766 4,indicating superior discriminative capability in this geologically intricate set-ting.Spatially,both models have identified high susceptibility zones primarily along valley flanks.However,the RF derived map exhibited more coherent,moderately distributed high risk areas with strong spatial continuity,whereas the KNN result showed noticeable overgeneralization and fragmentation of high susceptibility patches.Factor importance analysis revealed that the RF model effectively integrated multidimensional controls,with slope,curvature-related met-rics,NDVI,relative relief,and lithology emerging as the top six contributors.In contrast,the KNN model placed great-er emphasis on topographic proximity metrics such as relative relief,distance to rivers,and elevation,while assigning minimal weight to lithology and vegetation,reflecting its reliance on local spatial similarity rather than complex factor interactions.Using the natural breaks classification method on RF predicted probabilities,the study area is divided into five susceptibility levels.Statistical validation confirmed that the majority of actual landslides fell within the high and very high susceptibility zones,while non landslide points are predominantly located in low risk areas,demonstrating the model's reliability.Notably,the pronounced asymmetry in landslide density between the left and right banks corre-lates with differences in lithology,where weak clastic sedimentary rocks dominate the left bank,and proximity to major fault systems.A systematic comparison of RF and KNN in a high-relief is provided in this study,tectonically active reservoir setting,highlighting the advantages of RF in capturing nonlinear relationships and producing spatially coher-ent landslide susceptibility patterns.The results further clarify the roles of lithology and fault proximity in controlling the spatial heterogeneity and left-right bank asymmetry of landslide occurrence in the GX reservoir area,offering practical support for landslide risk management during hydropower construction and operation.