Abstract This study focuses on the clustered landslide event triggered by intense rainfall on 16 June 2024 in the Fujian–Guangdong–Jiangxi border region, aiming to develop an efficient deep learning model for high‐accuracy landslide susceptibility mapping. Based on the mapped landslide distribution and insights from field investigations, we constructed a data set integrating 13 controlling factors, including terrain, geomorphology, hydrology, geology, ecology, and human activities. Addressing limitations of conventional neural networks, such as constrained receptive fields, weak multi‐scale generalization, boundary degradation, and inadequate semantic characterization, we enhance ConvNeXt‐Tiny with feature pyramids and skip connections to improve small‐scale feature delineation. We further introduce a Lite‐Transformer to learn multi‐head self‐attention representations of high‐level semantics and to capture global structural relationships among the controlling factors. The resulting CNXT‐Ti‐LT (ConvNeXt‐Tiny with Lite‐Transformer) model was evaluated against widely used deep‐learning and machine‐learning baselines. The results show that CNXT‐Ti‐LT achieves clear improvements on nearly all evaluation metrics, while maintaining a favorable balance between accuracy and robustness. It more accurately captures feature associated with highly susceptible slopes, highlighting its strong potential for practical applications. Meanwhile, our findings suggest that the dynamic coupling between the physical mechanisms of rainfall‐induced clustered landslides in southeastern China and the spatiotemporal heterogeneity and uncertainty of rainfall remains insufficiently understood. Future work will advance both event‐process characterization and mechanism‐oriented modeling to improve interpretability and transferability across spatiotemporal scales and different regional conditions.
Luo et al. (2026) studied this question.