Topological Data Analysis for Automated Annotation of EO/SAR Datasets

In recent years, it has become increasingly important to conduct Geospatial Intelligence (GEOINT) operation via commercial and government persistent sensor systems, which have produced a copious amount of data relevant to the National Geospatial-Intelligence Agency (NGA). As the supply of data expands, it is necessary to employ automated analytics to exploit the data efficiently. We cannot rely on sheer human power to analyze data; instead, results must be automated and human analysts cued for final judgment. While there has been successful empirical application of AI algorithms on complex image domain problems (e.g., target/facial recognition), the mathematical basis has not kept pace, nor has there been a similar success in more difficult sensor modalities such as Synthetic Aperture Radar (SAR). Once trained, the current state of the art in AI is fragile on new data and methods are often ad hoc, not relying on technical/empirical rigor. For NGA, these issues are a significant impediment to ensuring reliability, explainability, and trustworthiness of neural networks, thus failing to achieve the trusted autonomy required to deliver on NGA's AAA Strategy. In high stakes environments of DoD applications, this situation is intolerable as the United States races for AI dominance. In response to this solicitation, Arete has developed a unique mathematical approach to dramatically reduce the human time required in AI systems based on recent advances in topological data analysis. Our method is also robust to noisy data with widely different views. During Phase I, we will perform a feasibility study, which will lead to the Homological Object Recognizer and Nomenclature Extraction Toolkit (HORNET) operating on SAR in Phase II.