Combined U-Net and LSTM approach to detect safeguards-relevant changes in Sentinel-2 images

Satellite imagery provide a valuable source of information for monitoring and verifying nuclear activities. By comparing images acquired at different time intervals, changes can be identified, which may indicate potential non-compliance with safeguards obligations. The use of deep learning algorithms enhances the accuracy and efficiency of change detection by automating the process and reducing the reliance on manual assessment. In this work a combined U-Net and Long Short-Term Memory (LSTM) network able to detect safeguards-relevant changes in high-resolution satellite imagery is presented. The U-Net architecture, a deep learning model commonly used for image segmentation, is well-suited for change detection in satellite images [1]. It leverages both contracting and expanding paths to capture fine-grained details and contextual information. Through its encoder-decoder structure, the U-Net model can effectively segment and identify changes, such as new constructions, modifications to buildings, or alterations in land cover within nuclear sites. LSTM, a type of recurrent neural network, is particularly useful for analyzing sequential data, making it applicable to temporal analysis in satellite imagery [2]. By incorporating LSTM into the change detection process, it becomes possible to model the temporal relationships and dependencies present in a sequence of satellite images. LSTM can effectively capture the dynamics and patterns of changes occurring over time, thus providing valuable insights into long-term trends and behavior within nuclear facilities. The U-Net model is trained using the INRIA Aerial Image Labeling dataset which has emerged as a fundamental resource in the field of remote sensing for studying land cover classification, object detection, semantic segmentation, and change detection [3]. The trained model is then used for change detection to identify new structures, modifications, or any other relevant changes within the nuclear site or its surrounding areas. LSTM complements the change detection process by considering the temporal aspect of the data.  By combining the capabilities of U-Net and LSTM, the change detection process in satellite images for nuclear safeguards can be significantly improved. The integration of these deep learning techniques enhances the accuracy, efficiency, and automation of the process, aiding in the timely detection of relevant changes within nuclear facilities.