Overview
- Field: Computer Vision
- Authors: Caleb Robinson, Anthony Ortiz, Simone Fobi Nsutezo, Cameron Birge, Meygha Machado
- Published: 2026-07-13
- arXiv: 2607.11838
- Paper: https://arxiv.org/abs/2607.11838
Summary
When a large disaster strikes, responders need a map of which buildings are damaged within hours. Models that perform well on public benchmarks assume matched before-and-after imagery and a training set drawn from similar past events—neither of which is usually available for a new disaster on its first day.
HASTE (High-speed Assessment and Satellite Tracking for Emergencies) is a no-code web platform that lets analysts who are not machine learning engineers produce per-building damage maps from post-disaster satellite imagery. HASTE implements two methods that share one interface:
1. Single-scene semantic segmentation: The user labels polygons over the post-disaster scene, a small semantic segmentation model is trained on that single scene, run over the whole image, and the per-pixel output is joined to existing building footprints. 2. Foundation-model embeddings + logistic regression: A pretrained vision model embeds each building footprint; the user labels a small number of buildings, logistic regression is fitted in the browser, and the rest of the scene is scored within seconds.
Results
Preliminary experiments on xBD show that embeddings from a foundation model, pooled over building footprints, can distinguish damaged from undamaged buildings using only post-disaster imagery—matching a fully supervised ResNet-50 baseline while requiring only one-twentieth of the labels.
Real-World Impact
HASTE and its predecessors have supported more than 30 real disaster responses since 2023, covering earthquakes, hurricanes, cyclones, floods, wildfires, and tornadoes, delivering results to humanitarian partners within hours to days of imagery availability.