Abstract: In landslide susceptibility assessment, optimizing both sample selection and model configuration is critical for prediction quality. Existing studies often focus on improving only a single aspect—either sample selection or the evaluation model—which may not ensure the overall reliability of the results. Taking the Ili River Valley landslide group as a case study, this paper proposes a multi-strategy optimization approach that integrates three aspects: landslide/non-landslide sample screening, evaluation factor system construction, and hyper-parameter tuning. Using the widely applied random forest(RF)and back-propagation neural network(BPNN)as base models, we combined Bayesian optimization(BO)and differential evolution(DE)to compare the performance of two unoptimized models(RF, BPNN)and four optimized models(DE-RF, BO-RF, DE-BPNN, BO-BPNN). The comparison was based on susceptibility zoning maps, zonal statistics, and model accuracy metrics. The results show that the optimized models achieve improved prediction accuracy. The RF model generally outperforms the BPNN model in landslide susceptibility prediction, and the DE algorithm delivers stronger optimization effects than BO. Based on comprehensive evaluation metrics, the DE-RF model produces more regular and accurate predictions. The proportions of extremely low, low, moderate, and high susceptibility zones are 39.14%, 22.62%, 21.54%, and 16.70%, respectively. The corresponding evaluation accuracies reach 85.4%, 82.4%, 89.9%, and an AUC of 0.933. Field investigations confirm that the actual landslide distribution aligns well with the predicted susceptibility levels, validating the effectiveness of the proposed multi-strategy optimization method and providing a reference for landslide prediction and mitigation in the region.