Abstract: [Objective] In order to address the issue that non-landslide samples selected using the buffer zone sampling strategy in landslide susceptibility evaluation may still harbor landslide risks, as well as the problem that most evaluation models lack thorough feature extraction, resulting in low evaluation accuracy, [Methods] This paper proposes a non-landslide sample optimization method based on Convolutional Auto-Encoder (CAE), built upon the buffer sampling strategy. The method optimizes the non-landslide samples by learning the fea-tures of landslide samples and utilizing reconstruction errors. For the evaluation model, the Convolutional Block Attention Module (CBAM) is integrated into the Residual Network (ResNet) to construct the Res-Net-CBAM landslide susceptibility model, capturing deeper, more complex, and more representative features. The experiment uses Wanzhou District in the Three Gorges Reservoir as the study area, selecting 12 influencing factors, such as elevation. Four models—SVM, DNN, CNN, and ResNet-CBAM—are employed to compare and analyze the evaluation accuracy and results of buffer zone sampling versus CAE-based optimized sam-pling.[Results] The results show that, under the same evaluation model, the CAE-based optimization sampling strategy for non-landslide samples yields higher reliability and accuracy. Under the same sampling strategy, the ResNet-CBAM model outperforms the other models in terms of accuracy, precision, recall, F1 score, and AUC. The evaluation results are similar across models, with high and very high susceptibility areas predominantly located in regions with low vegetation cover and frequent human activities, such as along the Yangtze River. Moreover, the ResNet-CBAM model with CAE-based optimized sampling demonstrates superior prediction results and is more suitable for landslide susceptibility evaluation in this area. [Conclusion] The non-landslide sample optimization strategy based on CAE and the ResNet-CBAM evaluation model effectively improve the accuracy of landslide susceptibility evaluation. The research method and experimental results provide valuable theoretical support and scientific guidance for landslide prevention and control in Wanzhou District.