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

1. Problem Statement

Following large-scale disasters (earthquakes, hurricanes, wildfires, etc.), humanitarian responders require building-level damage maps within hours to direct search-and-rescue teams and allocate resources. Current machine learning approaches for damage assessment face significant operational barriers:

• Data Scarcity: High-performing models typically require matched pre- and post-event imagery and training sets drawn from similar past events. In real-world responses, pre-event imagery is often years old or missing, and new disasters differ significantly from historical training data (domain shift).
• Labeling Bottlenecks: Collecting sufficient labeled data for new events is impossible within the critical first-day window.
• Operational Constraints: Existing solutions rely on expert human interpretation or volunteer tracing, which are difficult to scale to cover every building in a large city rapidly.

The core problem is generating accurate, building-level damage maps quickly using only post-disaster imagery and minimal local labeling, without requiring a dedicated machine learning engineering team.

2. Methodology

HASTE (High-speed Assessment and Satellite Tracking for Emergencies) is a no-code web platform that addresses these gaps by training small, event-specific models from a handful of analyst labels. It offers two distinct methods that share a unified interface but trade off between labeling effort and computational turnaround.

Method 1: Per-Scene Segmentation and Footprint Join

• Workflow: The analyst draws polygons over the post-disaster scene to label classes (e.g., background, intact building, damaged building).
• Training: A fresh U-Net model (ResNeXt-50 encoder) is trained specifically for that single scene using cross-entropy loss. To handle sparse labels, the system clusters them into tiles and ignores unlabeled background pixels. An optional "constraint class" allows analysts to mark large areas as "known undamaged" to penalize false positives.
• Inference: The model runs over the full scene in overlapping tiles. Pixel predictions are aggregated per building footprint to calculate a "damage fraction" (ratio of damaged pixels to valid pixels) at 0, 10, and 20-meter buffers to account for misalignment.
• Turnaround: A full iteration (label, train, review) takes approximately 1–2 hours.

Method 2: Footprint Embedding and In-Browser Logistic Regression

• Workflow: This method skips GPU training. It uses a pretrained vision model to extract features for every building footprint in the scene.
• Embedding: Features are pooled from the backbone (defaulting to MOSAIKS or DINOv2 variants) over each footprint. The pooling strategy includes concatenating per-channel statistics (min, max, mean, std) of the tokens.
• Classification: As the analyst clicks and labels a few buildings, a logistic regression classifier is fitted entirely in the browser using the frozen feature vectors. The model standardizes features and minimizes regularized logistic loss via gradient descent.
• Interactivity: The classifier re-trains and re-scores the visible scene in seconds as new labels are added. It supports multi-class "one-versus-rest" scenarios and runs efficiently on WebGPU or CPU.
• Turnaround: Iteration happens in seconds, allowing analysts to refine the model rapidly without waiting for GPU jobs.

Platform Architecture

HASTE is built as a React single-page application communicating with a REST API and a tile server. Long-running tasks (training, embedding, inference) are offloaded to a queue system managed by a shared core library (hastegeo). Data (imagery, footprints, models) resides in cloud storage, while metadata is stored in a NoSQL database. The platform integrates with existing building footprint datasets (e.g., Microsoft, Overture) and supports various imagery sources (optical and radar).

3. Key Contributions

• Deployment of a No-Code Platform: HASTE operationalizes rapid damage assessment, allowing non-ML analysts to generate per-building damage maps from post-disaster imagery alone.
• Dual-Method Approach: The paper details two complementary methods: a pixel-level segmentation approach (Method 1) and a footprint-embedding approach (Method 2), analyzing their trade-offs in labeling effort versus turnaround time.
• Label Efficiency Experiments: Preliminary experiments on the xBD dataset demonstrate that the footprint-embedding method (Method 2) achieves high accuracy with significantly fewer labels than fully supervised baselines.
• Operational Validation: The paper documents the platform's use in over 30 real-world disaster responses since 2023, delivering results to partners (e.g., Red Cross, UN agencies) within hours to days.

4. Results

Experimental Findings (xBD Dataset)

Experiments focused on Method 2 using post-disaster imagery only:

• Label Efficiency: The embedding-based approach with a DINOv2 backbone achieved a macro ROC-AUC of 0.84 using only 1% of the training labels (approx. a few dozen buildings per disaster).
• Comparison to Supervised Baselines: At 1% labeling, the embedding method (0.84 AUC) outperformed a fully supervised ResNet-50 fine-tuned on the same 1% of data (0.77 AUC). At 5% labeling, the embedding method matched the performance of a ResNet-50 trained on 100% of the labels. At 10% labeling, it surpassed the fully supervised baseline (0.91 vs. 0.88).
• Backbone Performance: Self-supervised backbones (DINOv2, DINOv3) significantly outperformed the random-feature MOSAIKS baseline. Backbone size mattered less than the quality of pretraining.

Operational Performance

• Turkey Earthquakes (2023): Mapped four cities, estimating 3,800 damaged buildings and 160,000 affected people.
• Rolling Fork Tornado (2023): Achieved 0.86 precision and 0.80 recall against field ground truth in under two hours per scene.
• Maui Wildfire (2023): Delivered an assessment of 1,700 damaged buildings by 1 PM on the day imagery became available, with validation showing ~97% accuracy (99% recall, 96% precision).
• Hurricane Melissa (Jamaica, 2025): Identified widespread damage in Black River and Montego Bay. Validation samples showed high recall (86–96%) but lower precision (71–82%), attributed to the conservative nature of the footprint join (rubble outside outlines is not credited).
• External Evaluation: An independent study on the Turkey damage product found high specificity (94%) but low sensitivity (43%), confirming the system is conservative and rarely flags undamaged buildings as damaged, though it may miss some damaged structures.

5. Significance and Claims

The paper positions HASTE as a critical bridge between advanced machine learning research and immediate humanitarian needs. Its significance lies in:

• Overcoming Domain Shift: By training small models on a single event with local labels, HASTE avoids the accuracy degradation seen when applying globally trained models to new, unseen disaster scenarios.
• Democratizing AI: The no-code interface allows humanitarian partners to run sophisticated damage assessment workflows without maintaining a dedicated ML team.
• Speed vs. Accuracy Trade-off: The platform provides a practical framework for responders to choose between a slower, pixel-accurate segmentation method (Method 1) and a faster, iterative embedding method (Method 2) based on their immediate time constraints.
• Real-World Impact: The system has moved beyond theoretical benchmarks to support over 30 real-world responses, delivering actionable data to decision-makers within the critical "golden hours" of a disaster.

The authors conclude that while the methods are simple by design, they effectively address the specific constraints of disaster response. Future work focuses on integrating vision-language models for automated labeling, active learning to optimize label selection, and expanding assessment to infrastructure like roads and bridges.