Efficient disaster damage prediction method using building point data and LSTM: a case of flood disaster

Accurate information on the location and use of individual buildings is essential for estimating impacts from disasters. However, even in developed countries, such data remains scarce, forcing reliance on aggregated statistics that obscure building-level impacts. We therefore propose a method for efficiently constructing point data on business facilities with industrial attributes for disaster analysis. We developed a multimodal industrial classification model within a Long Short-Term Memory (LSTM) framework. This model integrates business names from telephone directories with spatial context -business establishment statistics and land use zoning to probabilistically assign primary and secondary business types. As a result, an accuracy of approximately 83%–88% was achieved in industrial classification. The multimodal classification model contributed an average improvement of 13.0% in business establishment statistics and 5.4% in land use zoning for manufacturing predictions versus the non-multimodal case. The results of applying the damage and restoration functions from the manual to the prepared building data indicate variations ranging from 0% to 236% compared to a 500ⅿ grid-based damage method. The difference is significant compared to the accuracy of the building estimates, suggesting that it is desirable to change to building-based estimates.