Abstract: [Objective]The current research on the susceptibility of debris flow disasters has yet to address the limitations of geographic location relationships and spatial dependence. [Methods]This article constructs a debris flow dataset for Gansu Province with 10,198 sample points and proposes a susceptibility assessment method based on LA-GraphCAN. Initially, a nearest neighbor graph is built using KNN based on the cprojection coordinates of sample points. Secondly, GCN is used to efficiently aggregate local neighborhood information and extract key geographic and environmental features. Additionally, GAT is introduced to add a dynamic attention mechanism, enhancing the representation of features. Then ,validate the effectiveness of the proposed method, conduct comparative analyses from different perspectives, and finally, evaluate the susceptibility of debris flows in Gansu Province. [Results]The results indicate that LA-GraphCAN achieves accuracy, precision, recall, and F1 scores of 0.9441, 0.9287, 0.9375, and 0.9331, respectively, outperforming mainstream machine learning models such as Random Forests and CNN. Based on the evaluation of LA-GraphCAN, the number of historical debris flow disaster points in the highly susceptible areas of Gansu Province is 4055, accounting for 95% of the historical debris flow occurrences in Gansu Province, which is consistent with the distribution of historical disasters. [Conclusion]Both the performance evaluation and the susceptibility assessment results for Gansu Province indicate that the LA-GraphCAN method, which considers the spatial dependencies of debris flow disasters, yields superior results and is well-suited for debris flow susceptibility research.