SBIR-STTR Award

To support geospatial intelligence (GEOINT) through exploitation and analysis of synthetic aperture radar (SAR) imagery and address the challenges created by task-specific classification methodologies where feature extraction is mission-limited, ARiA will utilize deep learning (DL) expertise to enable automated unsupervised feature extraction (AUFE). Building upon our development of unsupervised DL for classification, segmentation, and augmentation of remote-sensing imagery ARiA will develop and demonstrate the feasibility of the Feature Learning, Encoding, and Extraction Toolkit (FLEET), a feature-extraction software for use with SAR that: (1) intelligently learns robust encodings from unlabeled data, thereby (2) reduces manual feature engineering, and (3) provides a framework for feature visualization and analysis (FVA). Fleet will facilitate a future National Geospatial-Intelligence Agency (NGA) capability for high performance GEOINT to support mission critical tasks.

To support geospatial intelligence (GEOINT) through exploitation and analysis of synthetic aperture radar (SAR) imagery, ARiA will enable segmentation and analysis of SAR imagery by building on prior work with unsupervised and semisupervised deep learning for semantic segmentation and terrain-sensitive automated target recognition (ATR) to develop and enhance FLEET, developed in Phase I as an unsupervised deep-learning framework for segmentation of geospatial imagery. In Phase II, ARiA will extend FLEET as an active-learning and few-shot learning software system where users can (1) refine and retrain underlying deep-learning models on-the-fly while analyzing SAR imagery and (2) enhance generalization and robustness of underlying deep-learning models through transfer learning and domain adaptation. ARiA will demonstrate that FLEET will be performant in novel domains such as when the user deploys a pretrained FLEET model in a new geographical region with minimal label requirements—i.e., few-shot annotations or pixels for segmenting new imagery. Data labeling requirements will be further reduced by integration of additional sources of information from other modalities such as EO and mapping services.