Method for Identifying Spatial Heterogeneity of Natural Disasters from the Perspective of Geographic Causal Reasoning

Abstract

To overcome the shortcomings of traditional correlation and static spatial statistics in disaster causal analysis, a geographic causal inference mechanism is constructed, combining causal discovery, individualized effect estimation, and spatially constrained clustering methods to identify regional differences in disaster risk. The computational process involves a structural causal model and watershed topology weights. The DML algorithm collaborates with the causal forest algorithm to extract CATEs (category-experienced exponents), and the consistency of GNN causal regularization with panel models is tested. Counterfactual simulation techniques are used to establish the ΔRisk–ΔCost Pareto frontier, establishing guiding principles for policy zoning and resource allocation. Regional empirical testing based on EM-DAT (2015–2024) data reveals that this technique significantly reduces CATE errors in regions with high exposure and strong spillover effects, enhancing the stability of causal consistency indicators and policy rankings. This study proposes specific technical implementation paths for regional management, budget improvement, and online risk control in the context of extreme climate change.