Change detection is an interesting task in the field of remote sensing, thanks to many useful applications that range from land cover studies to anomalies' observation (landslips, snowslides, wide firewoods, floods, etc.). Satellites like Sentinel-2 provide a full coverage of our planet every few days, but transmitting multispectral images of the same region multiple times within a small time interval is not always an efficient operation. At the same time, the analysis of each image on ground requires a considerable amount of time and efforts that might be reduced if knowing in advance that a portion of the new data do not contain any additional information with respect to data acquired in a previous time. Therefore, the idea of comparing onboard a new image with an older one of the same region represents a powerful tool that can help to both reduce the bottleneck effect occurring during the transmission of data to the ground stations and organize the post-processing analysis in a more efficient way.

In this study, deep learning methods are used to perform the change detection task with Sentinel-2 multispectral images. A pre-existing dataset focused on urban changes is exploited for training and validation purposes, while adopting two different approaches (semantic segmentation and classification). In addition, a benchmark test is conducted on a low-power consumption GPU, the NVIDIA Jetson AGX Xavier, to investigate throughput and speed performance with two different inference frameworks, TensorFlow and NVIDIA TensorRT, as this energy-efficient platform is suitable for the installation onboard the satellites in future missions.