Review of Landslide Identification Based on Multisource Data Fusion

Landslide detection is widely recognized as a fundamental prerequisite for geological hazard monitoring, early warning, and risk mitigation. However, traditional detection approaches are constrained by inherent limitations, including low efficiency, heavy reliance on expert experience, and limited spatial coverage. Rapid advances in remote sensing, geoscience surveying, and deep learning have made multisource data-driven intelligent landslide detection a frontier research direction. This paper systematically reviews the characteristics, application scenarios, and complementary or conflicting interactions of multisource data. A multisource data integration framework tailored for landslide detection is proposed, with fusion strategies classified into data-level, feature-level, and decision-level according to the integration hierarchy. The advantages, limitations, and applicable conditions of each strategy are summarized. Following the developmental trajectory of deep learning in detection tasks, representative deep learning architectures, including convolutional neural networks, encoder–decoder models, and transformers, are comprehensively reviewed with respect to their structural characteristics and fusion application frameworks. The results indicate that the complementary properties of multisource heterogeneous data in terms of physical sensitivity, spatiotemporal resolution, and geometric structure, when combined with the powerful feature representation capabilities of deep learning models, substantially improve the accuracy, robustness, and cross-regional generalization performance of landslide detection. Nevertheless, challenges remain regarding engineering applicability and model robustness. Future research should prioritize the following directions: enhancing cross-source data integration through unified semantic frameworks; meeting real-time engineering demands via lightweight models and efficient deployment; strengthening generalization and transfer learning through cross-modal pretraining with large artificial intelligence models; and developing more reliable and robust landslide detection frameworks by incorporating model interpretability and uncertainty assessment.