Capabilities for the Detection or Classification of Physical Damage using Remote Sensing

This STTR Phase I project proposes REACT (Rapid Event Assessment & Confidence Triage), a scalable, interpretable, and deployment-ready framework for automated physical damage assessment from remote sensing data. REACT introduces three core innovations: (1) self-supervised generative pretraining using masked autoencoders on vision transformers to enable cross-domain generalization from unlabeled imagery; (2) a hierarchical Siamese transformer architecture with learned differencing modules and adaptive token merging for efficient, fine-grained change detection; and (3) a confidence quantification module that delivers calibrated uncertainty estimates aligned with ICD 203. The system outputs structured damage classifications consistent with CJCSI 3162.02 and confidence-labeled geospatial data in formats such as GeoJSON, KML, and NAS/GCES for direct integration into mission workflows. REACT is designed to address key operational challenges in battle damage assessment (BDA), disaster relief, and infrastructure triage, especially under degraded imagery, unfamiliar sensors, and limited labeled data. Phase I will demonstrate the feasibility of REACT through performance benchmarking, cross-domain generalization, robustness to degraded inputs, and task-specific uncertainty calibration. By tightly coupling explainability, efficiency, and analytic rigor, REACT establishes a foundation for rapid, trusted, and standards-compliant damage assessment in both defense and humanitarian scenarios.á