Award

Automating initial physical damage assessment (PDA) remains a challenge due to its wide variety of operating domains: sensor modalities, geographic regions, damage modalities, and imaging geometries. Labeling large amounts of training data to cover all of these conditions is not a feasible approach to creating automatic damage assessment tools. In particular, training data is usually completely unavailable for a given geographic domain before the onset of the damage event: few damaged buildings exist in a region before the event, even if there is forewarning of what the relevant region is. With Artificial Intelligence Multimodal Damage Assessment (AI-MDA), Kitware and Etegent will enable efficient creation of damage detectors for new operating domains with minimal human damage sample labeling. We utilize an architecture previously developed on AFRLÆs AI-BDA program and extend it to multiple modalities via a pre-trained remote sensing foundation model. This model can accept as input imagery either Capella SAR (our main focus) or EO data and outputs an information-rich embedding. Small damage detector heads can be trained on these embeddings to achieve effective performance with very few labeled training samples. In addition, we propose low-label and physics-informed methods to quantify model performance, assessment confidence, and operating bounds, which will aid further human review of predicted damage. SAR can be difficult to interpret for non-expert users, complicating the damage assessment process. To simplify results interpretation for users, we extend KitwareÆs RDWATCH software to visualize predicted damage with the help of EO imagery generated from input SAR images, and to provide input/output functionality in common data formats like GeoCOCO.