<p><?tk 4?>In the process of regional landslide susceptibility prediction, the traditional GA-BP model will be disturbed by the landslide and non-landslide samples spatial proximity when selecting the training samples, and this disturbance has a significant impact on the prediction results. To address this issue, this study introduces the Self-Organizing Map (SOM) model, utilizing its unsupervised clustering capabilities to mitigate bias in landslide susceptibility predictions associated with the GA-BP model. The method is validated and applied to a representative red-bed landslide area, namely Shunqing District, Nanchong City, which serves as the study area. The results show that: (1) The model performance was evaluated using the Area Under the Curve (AUC) and F1 score. The SOM–GA–BP model achieved an AUC of 0.894 and an F1-score of 0.885. Due to the clustering effectiveness of the SOM model, the SOM–GA–BP model demonstrated superior predictive capability. (2) The prediction is based on landslide frequency ratios of each factor at the raster level. The proportion of extreme high and high susceptibility zones in the SOM-GA-BP model reaches 91.04%, compared to 81.79% in the GA-BP model, indicating a 9.25% reduction in prediction bias. (3) Landslide susceptibility in Shunqing’s red beds was mapped via an SOM–GA–BP model. The extreme high and high susceptibility zones were distributed in the hilly areas formed by tectonic erosion and on steep slopes adjacent to rivers; the moderate, low and extreme low susceptibility zones were located on gentle slopes at valley margins and occur in flat, geologically stable geomorphic regions.</p>