[Purpose] The selection of positive and negative samples is an important factor affecting landslide susceptibility assessment. To address the insufficient representativeness of positive samples and the uncertainty associated with random negative sample selection, this study constructs a landslide susceptibility assessment method oriented toward training sample optimization and validates it in the Lianghekou Reservoir area of the Yalong River.[Methods] First, SBAS-InSAR technology was used to obtain surface deformation information in the reservoir area, and a landslide inventory was constructed by integrating historical landslide records and remote sensing interpretation results. On this basis, training samples were optimized and selected. Positive samples were screened according to the comprehensive similarity between pixels within landslide polygons and typical landslide-conditioning environmental features, so as to retain samples that better represent landslide-prone environments. Negative samples were selected from InSAR low-deformation areas and information-value-model low-susceptibility areas to reduce the risk of incorrectly selecting potentially susceptible areas as negative samples. Then, XGBoost and LR models were constructed under different combinations of positive and negative samples, and their susceptibility assessment results were compared. [Results] A total of 109 potential landslides were identified in the study area, with a total area of 18.9 km², accounting for 2.5% of the study area. Compared with full-pixel sampling within landslide polygons and centroid-point sampling, positive sample screening based on comprehensive similarity reduced sample redundancy while preserving typical landslide-conditioning characteristics, resulting in better model prediction performance. Compared with random sampling outside landslide polygons, selecting negative samples from InSAR low-deformation areas and information-value-model low-susceptibility areas improved model prediction performance. According to the comparative analysis of ROC curves and susceptibility index distribution characteristics, the XGBoost model using comprehensive-similarity-screened positive samples and information-value-model low-susceptibility negative samples performed best, with an AUC value of 0.962. [Conclusion] Representative screening of positive samples within landslide polygons, together with constrained negative sample selection, helps improve training sample quality and model prediction accuracy. The results can provide a reference for landslide disaster prevention and mitigation in reservoir areas and for training sample construction in landslide susceptibility assessment.