The Dark Matter of Intelligence: A Remote Sensing Foundational Self-Supervised Model

Overhead imagery acquisition, procurement, labeling, and analysis is a costly task from an expertise, labor, and monetary perspective normally needed to train supervised models, creating massive barriers to entry. Self-supervised learning (SSL), described as “the dark matter of intelligence,” has an innate ability to generalize to downstream tasks without the need for any labeled data, thus lowering the barrier to entry. NT Concepts aims to build a foundational overhead imagery computer vision model able to perform numerous downstream tasks such as object detection, segmentation, classification, etc. The model will leverage state-of-the-art deep learning techniques and large-scale remote sensing datasets from various sources. The models’ generalizable intelligence enables faster and more accurate analysis of overhead imagery, critical in the persistent race with the United States’ near peers. Strategically limiting reliance on expensive, well-curated datasets used today in supervised models allows the United States to compete not with manpower, but with higher-order algorithms that are more performant, more generalizable, and more robust than traditional supervised models. The proposed foundational model will advance the state-of-the-art in overhead imagery research and practice, and contribute national security advancements, and economic development due to its broad range of use and commercialization prospects spanning all business sectors.