Learning to map building damage from a few confirmed examples

Fewer labels, not just fewer trainable parameters

Pretrained geospatial models offer a starting point for interpreting satellite imagery. But adapting them to a particular disaster still requires examples of what damage looks like. Updating only a small portion of a model’s parameters can reduce the scale of adaptation without necessarily reducing its appetite for labeled data.

In an arXiv preprint posted on November 5, 2026, Hengwei Zhao and colleagues propose Positive-Unlabeled GFM Adaptation, or PUMA. Its training data combine confirmed damaged buildings with a larger collection of buildings whose damage status is unknown. It does not require labeled undamaged examples for damage-classifier adaptation.

That distinction matters. Treating every unlabeled building as undamaged introduces incorrect training signals whenever the unknown group contains damaged structures.

Two mechanisms that correct each other

PUMA examines building-centered patches combining imagery from before and after a disaster. In its default configuration, it adapts a RemoteCLIP-pretrained model using deep visual prompt tuning, while keeping the pretrained model’s parameters frozen.

The framework tackles two connected problems: learning visual features that distinguish damage, and assigning provisional labels to buildings without annotations.

Its first component, NeWER, refines the model’s internal image representations. Incorrect provisional labels can make buildings from different classes look artificially similar to the model, or encourage methods to separate buildings that belong together. NeWER uses a non-contrastive architecture that avoids explicitly pushing presumed different-class pairs apart. Its proposed training losses are designed to reduce the influence of incorrectly matched pairs relative to better-aligned ones.

The second component, LaTTE, updates provisional labels using representative embeddings for each class and an automatically adjusted confidence threshold. Confident predictions can temporarily provide undamaged-building supervision. Crucially, those decisions are reversible: a building can return to a softer provisional label if confidence falls.

The approach does not require a separate estimate of the proportion of damaged buildings. Better features support better provisional labels, which then support further feature refinement. The authors provide a mathematical interpretation of this interaction under simplifying assumptions, rather than a guarantee that every training update improves the result.

What the benchmark tests found

The researchers tested damage classification on three events from the xBD dataset—Hurricane Florence, the Joplin tornado and the Santa Rosa wildfires—and on a dataset of buildings affected by Japan’s 2011 Tohoku tsunami.

Adaptation used 300 labeled damaged buildings per event, together with 5,000 unlabeled buildings for each xBD event or 4,000 for the tsunami dataset. Results were averaged across three runs.

PUMA variants achieved the highest reported F1 scores among the compared methods for the tsunami, hurricane and wildfire events. F1 balances precision, or how often damage predictions are correct, with recall, or how much actual damage is detected.

The default square-root-loss variant reached a tsunami F1 of 93.71%. The logarithmic-loss variant reached 93.72% for the hurricane and 95.79% for the wildfire. PUMA was competitive but not the best-performing method on the tornado event.

A retrospective wildfire test

The authors also evaluated image-covered subareas affected by the 2025 Eaton and Palisades fires in Los Angeles County. With 50 labeled damaged buildings and 500 unlabeled buildings for adaptation in each area, PUMA achieved F1 scores of 92.81% in Eaton and 93.89% in Palisades.

Those results substantially exceeded the baseline that treated unlabeled buildings as undamaged. Comparisons with richer supervision were more mixed: in Palisades, a method additionally given labeled undamaged examples performed slightly better, while labeling the entire training pool produced a higher score still.

These were retrospective evaluations against reference annotations refined by two expert teams, not evidence that the system had guided an emergency response.

What remains unresolved

The damage labels simplify a more complicated reality. In the xBD experiments, major damage and destruction formed the damaged class, while minor damage was grouped with no damage. The tsunami dataset instead concerns whether buildings were washed away. The system therefore does not provide a detailed, uniform assessment of structural condition across hazards.

Finding buildings is another constraint. When the researchers substituted automatically extracted footprints, object-level damage F1 scores were about 80–82% across the three xBD events—below the corresponding main classification results.

PUMA offers promising preprint evidence for reducing damage-label requirements during model adaptation. It does not eliminate the need for suitable before-and-after imagery, building localization, evaluation labels or further validation before operational use.