Towards Storage-Aware Onboard Change Detection

Change Detection is typically performed at the ground segment of satellite systems, where images can be carefully post-processed and registered, and complexity is not an issue. We investigate a framework entirely based on neural networks to perform change detection directly onboard of satellites in order to minimize the latency of the detection pipeline. The proposed framework accounts for image compression, registration and detection of change in order to manage the limited storage of the satellite and the lack of image alignment between revisits, while minimizing the computational complexity of the design. Preliminary results on a lightweight design show good change detection performance as a function of compression rate, while managing geometric misalignment.