Abstract: In the Bazhong area, Sichuan Province, China, the predominant lithology red bed rock mass exhibits unique geological structures and lithological compositions that heighten its susceptibility to landslides during rainfall. Assessing landslide susceptibility in this region is an effective measure for mitigating disaster risks and minimizing economic losses. This study, focuses on Bazhong city as the research area, incorporates eleven conditioning factors including elevation, relief, and annual average rainfall to develop a geographically weighted regression - random forest (GWR-RF) coupling model. This model optimize the negative sampling strategy by comparing it against traditional random sampling across the entire area. The results indicate the following: (1) Random sampling from the entire area leads to significant disparities in susceptibility assessments, accompanied by a relatively diminished accuracy, rendering it unsuitable for slope unit-based assessments. (2) The coupled GWR-RF model demonstrates spatial variations in landslide susceptibility, predominantly distributing in the Enyang, Bazhou, Pingchang counties, and the central - southern region of Nanjiang county. By performing random sampling within areas with the lowest proportion of landslide-prone slope units, this model effectively optimizing the negative sample sampling strategy for slope units, thereby enhancing the reliability and robustness of the predictive model.