A Novel Data Mining Approach for Building-Level Storm Damage Labels from Municipal Permits

Predicting building-level disaster damage is bottlenecked by a lack of parcel-scale ground truth. We introduce an empirical data mining pipeline that extracts physical structural damage labels from municipal building permits using keyword surge detection and LLM forensics. Across Hurricane Harvey (Houston, TX) and Hurricane Ian (Lee County, FL), we generate 173k parcel-level damage labels (55k and 117k, respectively). Extracted flood labels align strongly with 3-meter inundation rasters (80\% vs. 15\% baseline) and NFIP claims (Harvey: $r = 0.70$; Ian: $\rho = 0.68$), while wind labels match private insurance claims ($r = 0.72$). This establishes an open, scalable method for curating parcel-level damage labels.