Deep learning applications on satellite imagery datasets for nuclear nonproliferation and counter-proliferation

Abstract

This study examined the applicability of deep-learning techniques for extracting artificial structures from high-resolution satellite imagery to support verification processes in nuclear nonproliferation and counter-proliferation efforts. This examination relied on a tailored dataset and an open-source dataset. The tailored dataset was curated using satellite images of well-known nuclear complexes and was further refined to enhance domain relevance. Furthermore, using the attention U-Net model, optimal values of parameters such as batch size were determined to enhance performance. The model was then tested on satellite images of nuclear facilities from various sources, demonstrating effective performance even when applied to distinct and complex environments. To assess the robustness of the model, accuracy evaluations were conducted using both pixel-based and object-based tests. This dual evaluation approach provided a comprehensive analysis of the model, highlighting its practical utility for real-world verification tasks, particularly those related to nuclear activities. Although some false positives were detected, the proposed approach enabled the successful extraction of the majority of structures of interest. This achievement is anticipated to substantially reduce the interpretational workload for analysts and offer a transferable solution for global nuclear monitoring applications.