Semantic Segmentation with Deep Learning for Safeguards Imagery

The analysis of structures in satellite images is a key step in safeguards verification of State Declarations. Over the past few years, the volume and variety of satellite imagery has increased significantly. In order to take advantage of this increasing amount of imagery, a deep-learning-based system is presented for detecting infrastructure changes in satellite imagery for safeguards. Recent applications of deep-learning building-detection models to satellite imagery have focused on urban environments. However, the variety of building sizes and types at nuclear fuel cycle facilities makes the application of these models more challenging. Specifically, we study deep-learning methods for semantic segmentation of buildings, which allows multiple classes (e.g. building, road, water) to be labeled in satellite imagery. Our system is trained using open-source datasets of medium-resolution images with high revisit time. We also study the use of training imagery with additional spectral bands to improve the performance of the segmentation models. By applying these techniques to imagery of facilities and comparing the results to State Declarations, changes in infrastructure can be detected. Using this system for preliminary screening can greatly improve the workflow of image analysts.