Battle-Damage via Electro-Optical Automated Recognition (BEAR)

Satellite imagery and its analysis play a vital role in physical damage assessment (PDA) and battle damage assessment (BDA) by enabling accurate detection and characterization of damage for targets of interest. Timely PDA and BDA insight for analysts is critical to ensure prompt report preparation and analysis, delivering confidence in subsequent operational plans. However, manually analyzing extensive regions in satellite imagery is time-intensive and demands specialized expertise.áNGA and the COCOMs have a need for automated change detection and damage assessment utilizing commercial remote-sensing data to accelerate analyst workflows and provide time-critical intelligence for Human Assistance and Disaster Relief (HADR) and the warfighter. BlackSkyÆs constellation can enhance the monitoring and promptness of PDA and BDA by leveraging its high-revisit collections to catch ongoing damage indicators, such as smoke and fire, and provide post-event damage assessment, such as infrastructure damage. BlackSkyÆs high-resolution and very high-resolution electro-optical sensor capabilities are suitable for both target validation and damage characterization insights. Our proposed solution performs PDA and BDA using a single image while providing enhanced confidence and robust target identification when image pairs and feature polygons (OSM, BlackSkyÆs object detections, etc.) are available. The proposed solution builds upon industry-leading designs from xView2 challenge winners and utilizes the publicly available xBD dataset for natural disaster damage assessmentùadapting its architectural decisions for real-world BDA implementation. Additionally, we plan to curate a new benchmark BDA dataset using BlackSkyÆs existing imagery catalog and new tasking collections where annotations are promptly derived using novel state-of-the-art textual and visual retrieval methods. Moreover, our proposed solution focuses on human interpretability by automatically accompanying detections with similar scenarios from BlackSkyÆs curated damage dataset. Our Machine Learning Operations (MLOps) platform efficiently scales machine learning training runs, seamlessly tracks and compares model performance, and promptly deploys models. These features will enable us to thoroughly define and characterize the operating bounds of our proposed solution by comparing its performance against critical metadata features such as geographic area, sensor characteristics, time of day, land type, target object, and more