New Framework Maps Multi-Hazard Risk with Spatial Heterogeneity Awareness

This research introduces a new framework for creating detailed maps of flood and landslide susceptibility and associated risks across large regions, specifically focusing on Kerala, India, and Nepal. A key innovation is its ability to account for the spatial variability of environmental factors that influence these hazards. The framework evaluates two distinct training strategies for Random Forest models: proximity-gated cross-zone training (S1) and ecology-gated zone-constrained training (S2). S1 allows models to learn from geographically nearby areas even across contextual boundaries, while S2 restricts learning and application to within the same ecological zone. Results showed that S1 generally achieved higher accuracy metrics, particularly for flood susceptibility in Nepal, indicating better regional discrimination. Conversely, S2 was more effective at preserving subtle, zone-specific environmental differences in predictor selection and response patterns. Both strategies accurately identified flood-prone lowlands and landslide-prone uplands, but their specific susceptibility and risk classifications varied. The study highlights that integrating both cross-zone and zone-constrained learning can provide a comprehensive understanding of multi-hazard risks, especially when combined with exposure and vulnerability indices.