The study area is located in the frontal zone of the collision between the Indian Plate and the Eurasian Plate, characterized by extremely complex geological structures and frequent seismic activity. Accurate identification of potential strong earthquake source zones in this region and quantification of earthquake occurrence probabilities are crucial for enhancing earthquake prevention and disaster reduction capabilities, as well as for optimizing seismic hazard assessments. Based on the MATLAB radial basis function neural network classification algorithm, and following the principles of seismic tectonic analogy and historical earthquake recurrence, eight geological structural features related to seismic activity were selected as input variables, with the historical maximum magnitude as the output variable, to construct a neural network model. A total of 142 samples were used for model training and testing. By adjusting key parameters such as target error, spread of the radial basis function, and maximum number of neurons, a stable and feasible network model was achieved.This model was then applied to 395 samples to be classified. The output results, combined with manual verification, ultimately formed a zoning scheme for potential strong earthquake sources. Compared with schemes derived from Kriging interpolation and kernel density estimation methods, the proposed zoning scheme shows strong consistency with Seismic Ground Motion Parameter Zonation Map of China (GB18306—2015).This validates the reliability of the method for seismic hazard assessment in complex tectonic regions and its adaptability to small sample datasets. The results provide valuable references for the planning of earthquake prevention and disaster reduction, as well as for seismic risk management.