Recov-Vision: Linking street view imagery and vision-language models for post-disaster recovery

Building-level occupancy after disasters is critical for triage, inspections, utility re-energization, and equitable resource allocation. Overhead imagery provides rapid coverage but often misses the facade and access cues that signal whether a building is in apparent use. Street-view imagery captures those details, but is sparse and difficult to align with parcels. We present Recov-Vision, a street-level framework that links panoramic video to parcels, rectifies views to facades, and elicits interpretable attributes (e.g., entry blockage, temporary coverings, localized debris). These attributes drive two alternative decision policies over the same evidence: a transparent one-stage count rule and a two-stage design with a separate text-only LLM decision stage. Evaluated across two post-Hurricane Helene surveys, the two policies perform comparably; the two-stage policy attains slightly higher recall while reproducing the ground-truth net recovery (+14 parcels, defined as 20 parcels returning to occupancy minus 6 deteriorating). The non-obvious finding is that residual errors are spatially clustered (Moran’s 𝐼 for the accuracy surface = 0.06, 𝑝 = 0 . 0 3 ), so the framework supports targeted neighborhood audits rather than uniform review. More broadly, decoupling visual evidence extraction from policy-level reasoning yields auditable assessments that agencies can tune for jurisdiction-specific conservatism without retraining the vision component.