New satellite constellations will soon be deployed with locational revisit rates of 40 captures/day resulting in an image acquisition volume that quickly exceeds capacity for manual image analysis. This inspires an automated solution for analyzing spatio-temporal imagery that can augment analyst workflow and enhance geospatial investigation of locations over time. Building on System & Technology Researchâs (STRâs) previously developed spatio-temporal software and using STRâs access to over 20 petabytes of commercial satellite imagery, we propose the Spatio-Temporal Analysis & Recognition System (STARS) solution. STARS incorporates integrated uncertainty characterization and temporal evidence aggregation for reliable object detection and disambiguation, a mature data conditioning pipeline, robust few-shot learning approach, and extensive object detection experience with diverse satellite datasets. This solution is designed to identify, detect, and disambiguate rare objects across spatio-temporal sequences of images. Our proposed system leverages feature enhancement and reweighting to learn optimal features for few-shot object detection. Detections can be leveraged by our proposed Bayesian non-parametric probability estimation method to provide meaningful and quantifiable confidence estimates and be tolerant to variations in collection resolution, orientation, and scene illumina
