Advancing Landslide Susceptibility Mapping across Heterogeneous Regions with Deep Learning-Based Domain Adaptation

Data-driven approaches such as machine learning (ML) have been increasingly used in landslide susceptibility assessments. However, their performance is often hindered by the scarcity of landslide inventory and the limited generalizability of ML algorithms. This research evaluates the generalization capability of ML models and introduces a deep learning-based domain adaptation framework to improve landslide susceptibility mapping across heterogeneous ecoregions. The approach involves training the neural network (NN) model on a source ecoregion and retraining it with data from the target ecoregion. Utilizing the Statewide Landslide Information Database for Oregon (SLIDO) and employing spatial cross-validation, this study benchmarks NN models without domain adaptation against the proposed framework in the diverse ecoregions of the Pacific Northwest, United States. The results demonstrate that the generalization capabilities of NN models trained in one ecoregion tend to decline when applied to different ecoregions, and this decline varies significantly from one ecoregion to another. Nevertheless, models trained on data from source ecoregions can reduce the data demand for new ecoregions, thereby addressing data scarcity challenges. However, the effectiveness of these pre-trained models heavily depends on the similarity of training data between source and target ecoregions. If the ecoregions are significantly different, using transfer learning may not achieve the same level of performance as ecoregion-specific models. This study underscores the importance of leveraging knowledge from existing landslide susceptibility models to develop robust models in regions with scarce data and identifies key considerations for selecting suitable existing models for successful adaptation.