DisasterVision Data Integrity and a Siamese Baseline for Building Damage Assessment
DisasterVision investigates data integrity and software readiness for satellite-based building damage assessment using paired pre- and post-disaster imagery from the xBD dataset. The study analyzes 2,799 image pairs across ten disaster events through deterministic inventory construction, strict polygon validation, geographic overlap screening, and annotation forensics. Validation accepted 2,760 image pairs and rejected 39, while forensic examination identified 44 invalid building polygons. All 5,598 dated images were matched with nominal geographic transformations, with no cross-role overlap or boundary contact identified at four tested expansion margins. Simulated whole-pair exclusion would remove 6,057 source post-disaster labels, compared with 43 under UID-linked exclusion. An implemented Siamese ResNet-18 segmentation baseline completed CPU software smoke testing. The work establishes a reproducible foundation for further damage-assessment experiments, while independent geographic verification, full model evaluation, and proposed MobileViT and RAFT comparisons remain future work.
Authors
- Ritesh Kumar Chandel
- Saara
Institutions
- Guru Gobind Singh Indraprastha University (IN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24375576
- Primary Topic
- Remote-Sensing Image Classification
- Type
- preprint