Intelligent prediction model for earthquake-induced landslide susceptibility based on transfer learning and sampling optimization strategies

Accurate assessment of seismic landslide susceptibility is crucial for disaster prevention and emergency decision-making. Although machine learning methods have been widely applied in this field, they exhibit a strong dependence on large quantities of high-quality samples, resulting in significantly low prediction accuracy of existing studies under data-scarce or cross-regional prediction scenarios, which fail to meet practical application requirements. To address this issue, this study proposes an intelligent prediction model integrating transfer learning and a sampling optimization strategy, aiming to enhance the accuracy and applicability of seismic landslide susceptibility assessment. The model first improves the sample collection method through the sampling optimization strategy to enhance the precision and representativeness of training samples. This not only ensures the accuracy of origin area training but also further strengthens the model's predictive ability in the target area. Subsequently, it incorporates Transfer Component Analysis (TCA) to overcome the differences in environmental characteristics between the origin area and target area, and couples TCA with the LightGBM algorithm to construct the TCA-LightGBM model, realizing the assessment of seismic landslide susceptibility in sample-free areas. Validated through case studies of the Jiuzhaigou and Luding earthquakes, the results demonstrate that the proposed TCA-LightGBM transfer learning method exhibits excellent applicability in seismic landslide susceptibility prediction. After optimization with the TCA algorithm, the model's prediction performance in the target domain is significantly improved, with the AUC value increasing from 0.719 to 0.827, representing an increase of approximately 15.02%. This indicates that TCA technology can effectively alleviate the feature distribution discrepancy between the source domain and target domain, enhancing the model's generalization ability. The method is particularly suitable for scenarios with data scarcity and cross-regional prediction and can provide reliable technical support for the emergency response and risk prevention and control of seismic hazards.