Spatial Disaggregation and Temporal Projection of Building Exposure and Physical Vulnerability using Deep Constrained Clustering and Probabilistic Graph Deep Learning

In the final years of the UN Sendai Framework and the Sustainable Development Goals, significant data gaps in large-scale exposure and vulnerability remain a critical barrier to advancing prospective and probabilistic disaster risk assessment and to measuring progress toward resilience with accountability. Despite rapid advances in information technology and Earth observation, the modelling of building exposure and physical vulnerability has largely remained static, costly, region specific, coarse-grained, overly aggregated, and insufficiently calibrated. To address these limitations, we introduce two methodological innovations that advance the spatial disaggregation and temporal projection of building exposure and physical vulnerability using deep constrained clustering and probabilistic graph-based deep learning. First, we present Deep Conditional Census-Constrained Clustering (DeepC4), a deep learning-based spatial disaggregation framework that integrates local census statistics as cluster-level constraints while jointly modeling multiple conditional relationships among building typology labels through multitask learning from Earth observation data. Second, we present Graph Variational State-Space Model (GraphVSSM), a modular spatiotemporal framework that integrates graph deep learning, state-space modeling, and probabilistic inference to project exposure and vulnerability dynamics over time by combining time-series Earth observation data with prior expert belief systems in a weakly supervised or coarse-to-fine-grained manner. We demonstrate the applicability and scalability of these approaches through case studies in Rwanda, the Philippines, and UN recognized Least Developed Countries. Beyond improving the availability, quality, and granularity of exposure and vulnerability information, our work also highlights unique domain challenges and key considerations regarding the reliability, interpretability, and relevance of deep learning outputs for disaster risk assessment and policymaking.