Hybrid Frequency Ratio–CNN Framework for Landslide Susceptibility Mapping: A Case Study of South Bostanlyk, Uzbekistan

This study presents landslide susceptibility mapping (LSM) using frequency ratio (FR) and convolutional neural network (CNN) models in the South Bostanlyk area of Uzbekistan, a region that is highly susceptible to landslides due to steep mountainous terrain, complex geology, and active seismicity.The study area features carbonate bedrock overlain by loess and colluvial deposits, which exacerbate slope instability, particularly during seasonal rainfall and snowmelt events.This study aimed to systematically delineate areas susceptible to future landslides through an integrated analysis of geomorphological, geology, soil, normalized difference vegetation index (NDVI), land cover, and historical landslide inventory datasets.FR analysis revealed that landslide occurrence correlates strongly with valley depth, relative slope position, terrain ruggedness index, channel network distance, and land cover, while XGBoost feature importance confirmed valley depth, and relative slope position as the top two dominant factors.To complement this statistical approach, a CNN model was applied to the spatial datasets, resulting in improved predictive performance.Both the CNN and FR models achieved predictive accuracies exceeding 75%, confirming their robustness for regional-scale landslide susceptibility assessment.The resulting susceptibility maps offer practical insights for disaster risk management, infrastructure development, and climate change adaptation, highlighting the effectiveness of integrated modeling frameworks for landslide risk mitigation in rapidly developing mountainous regions such as the South Bostanlyk area.