Source-Model-Guided Target-Group Sampling for Label-Efficient Four-Level Building-Damage Adaptation in Conflict Regions
Year: 2026
Type: article
Abstract: Four-level building-damage assessment from paired satellite imagery requires extensive expert labeling, yet models trained on natural-disaster data transfer poorly to conflict imagery. We propose a source-model-guided framework coupling cross-hazard transfer learning with pre-annotation selection of target-domain image groups. An xBD-trained bi-temporal source classifier initializes conflict-domain fine-tuning and provides building-level predictions for group screening. Predictions are aggregated into a group-level Damage Severity Index (DSI) and predicted-label purity; DSI stratification allocates the budget across predicted severity levels, while purity differentiates groups within strata. In a retrospective label-hiding experiment on a Ukraine dataset, selecting 40 of 120 groups (525 of 1615 labels) reduced labeling by 67.49%. The model achieved 84.29% within-one-grade accuracy, 53.43% exact accuracy, 49.37% Macro-F1, and 0.662 ordinal MAE, close to using all groups. Compared with two uncertainty baselines, it yielded better Macro-F1 and ordinal error estimates at comparable exact accuracy. Together with evaluation in Mariupol, these results suggest that source-model-guided selection can reduce target-domain labeling needs while maintaining competitive performance. However, group selection depends on source-prediction quality, and validation remains limited to regions within one conflict dataset. Future work will investigate more robust screening descriptors and additional conflict datasets.
Source: Applied Sciences
DOI: 10.3390/app162211367
Authors: Na Li(Conceptualization, Data curation, Funding acquisition, Validation, Visualization, Writing – original draft), Meng Ren(Data curation, Software, Validation), Gaijun Zhao(Data curation, Validation), Lilu Zhu(Validation, Writing – review & editing), Zhe Zhang(Methodology, Resources, Supervision, Writing – review & editing)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7235302298
