300 damaged buildings are enough: PUMA adapts a geospatial foundation model to map post-disaster damage

Building damage mapping can be done when only a small number of damaged buildings are annotated and every other building is unlabeled, without large labels for both damaged and undamaged classes. Earlier parameter-efficient fine-tuning methods mostly train under full supervision, implicitly assuming reliable annotations at scale; after a disaster such annotations are costly and too slow to obtain. On the xBD hurricane, tornado and wildfire events and the ABCD tsunami data, fine-tuning uses only 300 damaged buildings plus 4,000 to 5,000 unlabeled buildings, and PUMA reaches the highest overall accuracy and F1 among the compared methods on tsunami, hurricane and wildfire, while on tornado it is competitive with the best NcPU.

The method splits damage mapping into two steps: locate each building in the imagery, then decide whether that building is damaged. Damage classification is written as positive-unlabeled learning, with a few confirmed damaged positives and many samples of unknown status, avoiding complete annotation of the undamaged class. Localization uses existing building footprints when available and otherwise an RPN plus Segment Anything Model pipeline on pre-disaster imagery; with automatically extracted footprints, pixel-level F1 is 84.11%, 87.96% and 84.90% for hurricane, tornado and wildfire.

Two modules alternate: one separates damaged and undamaged buildings in feature space, the other uses that separation to progressively correct pseudo-labels on unlabeled buildings. Under weak supervision, direct representation learning introduces cross-class noisy pairs that wrongly pull different-class samples together and make the decision boundary hard to learn. Non-contrastive representation refinement uses only same-class pairs and two losses that suppress the gradient of noisy pairs; latent-state label disambiguation adds undamaged supervision through a self-adaptive threshold and reverts a pseudo-label to unlabeled when confidence drops, so no class prior needs estimating.

The demonstrated scope is binary building damage mapping: the xBD hurricane, tornado and wildfire events and the ABCD tsunami data, plus the Eaton and Palisades subareas of the 2025 Los Angeles wildfires; metrics are computed at object level and reported as mean and standard deviation over three repeats. In the real case, the Eaton subarea runs with only 50 damaged-building labels, and a separate setting uses 300 damaged buildings plus 4,000 unlabeled buildings. The authors propose extending the framework to more complex damage assessment scenarios and broader weakly supervised remote sensing applications.

The theoretical analysis shows the lower bound is not monotonically tightened at every step but becomes asymptotically tight as within-class representations align. Ablations show latent-state label disambiguation alone gives high recall but low precision on the hurricane event, while adding the self-adaptive threshold raises precision but sharply lowers recall on the tsunami event; only the combination gives a more balanced precision and recall. Full fine-tuning degrades markedly and varies more under weak supervision. The real-case reference annotations were cross-checked by two expert teams, and buildings whose status remained ambiguous after review were excluded.