Regional heterogeneity-aware domain adversarial training for trans-regional landslide susceptibility assessment

Landslide susceptibility assessment is fundamental to geological hazard prevention and mitigation. Although deep learning has recently emerged as a powerful tool in this field, its performance remains highly dependent on sufficient landslide inventories, which are often unavailable in data-scarce regions. Moreover, significant regional heterogeneity may cause domain shifts between sample-rich and sample-scarce areas, limiting the transferability of conventional models. To address these challenges, this study proposes a regional heterogeneity-aware domain adversarial training method for trans-regional landslide susceptibility assessment. The core novelty lies in the development of a domain-label representation that quantitatively characterizes inter-regional environmental discrepancies using landslide conditioning factors and embeds this heterogeneity information into the adversarial learning process. The proposed method leverages Longshan County as a sample-rich source domain and seven adjacent counties in northwestern Hunan, China, as data-scarce target domains to extract transferable domain-invariant features. Experimental results show that the proposed method achieves the highest average accuracy of 79.09% across the seven target domains, markedly outperforming five benchmark models, whose average accuracies range from 56.65% to 61.33%. In addition, ablation results confirm that incorporating the heterogeneity-aware domain-label representation improves feature alignment and enhances the spatial generalization of landslide susceptibility prediction. These findings demonstrate that the proposed method provides an effective solution for robust trans-regional landslide susceptibility assessment under limited target-domain samples.