HASTE: A Platform for Rapid Post-Disaster Building Damage Assessment

In the immediate aftermath of a disaster, humanitarian responders require rapid, accurate maps of building damage to prioritize search-and-rescue and aid delivery. Traditional machine learning models for this task typically require matched pre- and post-disaster imagery and large, labeled training sets from similar historical events. In real-world emergencies, these prerequisites are rarely met, leaving responders to rely on slow, manual interpretation or volunteer-based mapping efforts.

HASTE (High-speed Assessment and Satellite Tracking for Emergencies) provides a no-code interface that enables non-engineers to generate damage maps using two distinct, user-driven methods. The first method involves training a U-Net semantic segmentation model on a per-scene basis using polygon labels drawn by the analyst. The second method uses a footprint-embedding approach: it extracts features from building footprints using a pretrained vision model and fits a logistic regression classifier directly in the web browser. This second method allows for near-instantaneous retraining and re-scoring as the analyst adds more labels, facilitating a highly iterative human-in-the-loop workflow.

Preliminary experiments on the xBD dataset demonstrate that the footprint-embedding method is highly label-efficient. Using only 5% of available labels, the platform achieves performance comparable to a fully supervised ResNet-50 baseline trained on 100% of the data. HASTE has been successfully deployed in over thirty real-world disaster responses since 2023, including earthquakes, hurricanes, and wildfires, consistently delivering actionable insights to humanitarian partners within hours of imagery availability.

By shifting the paradigm from globally trained, static models to locally trained, interactive tools, HASTE bridges the gap between sophisticated machine learning research and the urgent, unpredictable needs of field responders. The platform's ability to function without pre-event imagery and its low barrier to entry make it a critical tool for scaling damage assessment in resource-constrained environments.