FMLD: A Vertical Federated Learning Framework for Privacy‐Preserving Multimodal Landslide Detection

Abstract Landslides are among the most severe global geohazards posing a significant threat to human life and infrastructure. To support landslide detection and prediction, various geohazard monitoring approaches have been developed, such as optical remote sensing imagery, light detection and ranging, and ground‐based sensors, generating vast volumes of landslide‐related data. However, these data often involve privacy and security concerns for data owners. Institutions such as private companies, national space agencies, and geological survey departments are often reluctant to share geohazard data, which hinders the development of machine learning‐based technologies for landslide assessment. To address this challenge, this study proposes a privacy‐preserving framework named Federated Multimodal Landslide Detection (FMLD), based on vertical federated learning, enabling secure and collaborative model training across multiple institutions. FMLD integrates complementary multimodal data from different organizations, including optical imagery, digital elevation models, and hillshade maps, allowing the model to exploit the strengths of each modality while keeping raw data private. Extensive experiments conducted in three study areas show that FMLD achieves comparable landslide detection accuracy with its centralized counterpart. The proposed method effectively protects data privacy and security, thereby enhancing geological data sharing. This study demonstrates a practical pathway toward secure, collaborative, and knowledge‐complementary artificial intelligence applications in geoscience.