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

Abstract: When a large disaster strikes, responders need a map of which buildings are damaged within hours. The models that do well on public benchmarks assume matched before-and-after imagery and a training set drawn from similar past events, and neither is usually available for a new disaster in its first day. We present HASTE (High-speed Assessment and Satellite Tracking for Emergencies), 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. The first requires the user to label polygons over the post-disaster scene, trains a small semantic segmentation model on that single scene, runs it over the whole image, and joins the per-pixel output to existing building footprints. The second embeds every footprint with a pretrained vision model, requires the user to label a handful of buildings, and fits a logistic regression in the browser that scores the rest of the scene in seconds. We describe the platform, both methods, and the engineering that supports them. We also report preliminary experiments on xBD showing that foundation-model embeddings pooled over footprints separate damaged from intact buildings using post-disaster imagery alone, matching a fully supervised ResNet-50 baseline with a twentieth of its labels. HASTE and its predecessors have supported more than thirty real-world disaster responses since 2023, spanning earthquakes, hurricanes, cyclones, floods, wildfires, and tornadoes, delivering results to humanitarian partners within hours to days of imagery becoming available. We close with the directions we think are most promising, including vision-language assessment, active learning, and damage models for roads and other infrastructure. HASTE is open source at https://github.com/microsoft/haste.

• The paper introduces HASTE, a no-code platform that rapidly produces per-building damage maps from post-disaster satellite imagery using minimal local labeling.
• The paper details two methods—a per-scene segmentation using a U-Net with ResNeXt-50 and an embedding-based logistic regression—achieving ROC-AUC scores up to 0.92 with limited labels.
• The paper demonstrates real-world effectiveness with high accuracy (e.g., 97% accuracy in Maui wildfire) and rapid turnaround, proving its value in emergency response.

HASTE: Rapid, No-Code Post-Disaster Building Damage Assessment from Satellite Imagery

Motivation and Context

Post-disaster building damage mapping via satellite or aerial imagery is a critical task for emergency response and humanitarian operations. However, benchmark-winning ML systems in this domain assume access to matched pre/post-event imagery and event-specific training data—assumptions rarely met in real-world, time-critical scenarios. The HASTE (High-speed Assessment and Satellite Tracking for Emergencies) platform addresses these operational and methodological deficits by enabling domain analysts, not just ML experts, to produce per-building, georeferenced damage maps rapidly from only post-event satellite acquisitions and minimal local labeling effort.

System Design and Methodology

HASTE is a no-code, web-based platform with cloud-integrated back-end, enabling deployment at scale. It offers two alternative methods for building-level damage assessment:

Method 1: Per-Scene Semantic Segmentation

This approach treats the post-disaster image as an independent segmentation task:

• Interactive Labeling: Analysts annotate polygons over the scene, specifying classes such as background, intact, and damaged building.
• Model Training: A U-Net with a ResNeXt-50 encoder is instantiated per event and trained only on the labeled regions, sidestepping domain shift and sensor-forgetting endemic to globally trained models.
• Inference and Output: The per-pixel damage probability is aggregated within (and around) each building footprint, yielding a fractional, per-footprint damage score. Multiple buffer radii are used to mitigate footprint-image misregistration and account for dispersed debris.
• Constraint Regularization: Large, negative supervision polygons (e.g., "all buildings here are intact") are utilized to regularize learning and suppress false positives.

Method 2: Embedding-Based, In-Browser Logistic Regression

This method skips GPU retraining entirely by leveraging frozen, pretrained vision model embeddings:

• Feature Extraction: Each building footprint is encoded by pooling patch-token embeddings (mean, min, max, std) from a pretrained backbone (default: MOSAIKS random features; also supports DINOv2/v3 foundation models) applied to the post-disaster image.
• Interactive Classification: As the analyst labels a handful of buildings (as few as three across two classes), HASTE fits an L2-regularized logistic regression in-browser (WebGPU-accelerated if available) that instantly scores all footprints, yielding real-time feedback and iterative refinement.
• Matching and Throughput: At 1024 features, the system scores ~100,000 buildings/sec even on commodity hardware, making it practical for city-scale mapping without back-end roundtrips.

These approaches trade off annotation effort and computational resources: Method 1 enables pixel-level inference at the cost of GPU jobs and polygon annotation, while Method 2 enables minimal-label, instant feedback but only labels buildings with known footprints.

Experiments and Performance

Label Efficiency and Comparative Results

The preliminary xBD dataset experiments collapse the damage scale to binary and measure macro ROC-AUC as a function of labeled sample percentage:

• Label Efficiency: With only 1% of event-specific building labels, DINOv2/3 backbones achieve 0.84 ROC-AUC; MOSAIKS trails by 0.1-0.14. As label budget increases to 10-50%, DINOv3 reaches 0.91–0.92 ROC-AUC.
• Supervised Baseline Comparison: A fully supervised ResNet-50 fine-tuned end-to-end (using all available labels) attains 0.88; embeddings plus LR match or exceed this using only 5–10% of labels. Thus, HASTE achieves superior label efficiency and matching accuracy without specialized end-to-end re-training.
• Operational Metrics: In real events, time from image availability to actionable output is hours, not days. E.g., for the 2023 Maui wildfire, HASTE delivered ~1,700 validated damage estimates (97% accuracy, 99% recall, 96% precision) within 4 hours.

Validation in Real-World Operations

Over 30 international disaster responses—including earthquakes, hurricanes, wildfires, and cyclones—have used HASTE-derived products, disseminated to major humanitarian agencies (Red Cross, WFP, UNDP, OCHA). Notably, cloud obstruction, variable footprint quality, and label ambiguities are explicitly handled via "unknown" flags and population estimates with robust confidence intervals.

Generalization, Limitations, and Uncertainty

• Domain Shift: By focusing on per-event retraining/embedding adaptation, HASTE avoids cross-event performance collapse observed in traditional ML approaches.
• Annotation Uncertainty: HASTE validation is typically against expert visual interpretation; field-surveyed ground-truth comparisons (as in Rolling Fork tornado) show high specificity and variable sensitivity.
• Methodological Constraints: Damage not visible from above, misaligned/missing footprints, and out-of-footprint debris remain sources of unaccounted loss. Method 2 can only label buildings present in the (often outdated) footprint dataset.

Future Directions

• Vision-Language Extensions: Integrating vision-LLMs (e.g., GeoVision Labeler) for zero-shot geospatial classification and annotation could reduce or eliminate the initial human labeling step and provide interpretable, prompt-based assessment.
• Active Learning: Incorporating active label suggestion—based on model uncertainty/diversity in embedding space—could further reduce label requirements and drive higher accuracy with less analyst time.
• Domain-Specific Pretraining: Pretraining on large, multi-event disaster domains or performing scene-specific token adaptation for segmentation or embedding pooling could enhance generalization across diverse sensor types and event phenotypes.
• Beyond Buildings: The platform's modularity supports extension to objects such as roads, bridges, or vehicles, which is essential for comprehensive infrastructure impact assessments.
• Robustness Improvements: Automatic registration between imagery and footprints and better handling of clouds/haze would improve reliability especially in rapid response settings with varying data quality.

Conclusion

HASTE operationalizes rapid, event-adapted building damage mapping from satellite imagery, minimizing the technical bottlenecks of traditional ML pipelines. By leveraging foundation visual features and efficient analyst-in-the-loop loops, it enables label- and compute-efficient deployment under field conditions where time, imagery, and labels are all scarce. Its quantitative results demonstrate both high accuracy and unmatched speed, and its practical deployments provide actionable, uncertainty-aware outputs to disaster response bodies worldwide. Future work will further generalize the platform, deepen automation, and broaden the suite of assessable infrastructure targets.

Reference: "HASTE: A Platform for Rapid Post-Disaster Building Damage Assessment" (2607.11838)