Geohazard Susceptibility Assessment of Riverside Highway Zones under Multiple Feature Spaces Adaptation Network

Riverside highway zones are prone to high risks of geohazards such as landslides due to proximity to water bodies, steep terrain, and frequent anthropogenic activities. However, current single-source domain transfer learning methods face limitations in geohazard susceptibility prediction when significant discrepancies exist between source and target domains in hydrogeological conditions (e.g., river density, rainfall intensity) and engineering disturbances (e.g., highway construction), often leading to negative transfer issues and reduced model generalizability. This study proposes a multi-source domain transfer learning framework based on a multiple feature spaces adaptation network (MFSAN). Focusing on three riverside highway zones in Fujian Province, China, nine environmental factors (including highway density and river density as core hydrogeological features) were extracted to construct a landslide spatial database. The susceptibility models from Anxi County (source domain 1) and Dehua County (source domain 2) were transferred to Youxi County (target domain) with unlabeled samples for cross-regional landslide susceptibility evaluation. Comparative analyses were conducted against non-transferable learning models (NTL) and single-source domain adaptive models (domain adaptive neural network, DANN). The results demonstrate: (1) The MFSAN model achieved a cross-regional prediction accuracy of 0.851, outperforming single-source transfer models with improvements of 3.61% in accuracy, 1.91% in AUC, and 9.64% in overall assessment metric (OA). (2) Historical landslide validation revealed that 79.2% of landslides occurred within high-to-extreme susceptibility zones predicted by MFSAN, the highest among all models. (3) MFSAN exhibited superior capability in capturing hydrogeological coupling effects unique to riverside environments. For instance, the concentration of hazard-prone sites within 3 km of highways (70%-83%) was accurately reflected in predictions. The MFSAN framework effectively integrates spatial features and disaster development patterns from multiple source domains, comprehensively capturing regional heterogeneity and providing an optimized solution for cross-regional geohazard susceptibility prediction. This approach demonstrates enhanced generalization capability and practical value for mitigating landslide risks in complex engineering environments.