A nre architecture for onboard change detection based on deep learning

Change detection (CD) from satellite imagery is a task of paramount importance for monitoring Earth, land usage and disaster management. Traditionally, CD is performed at the ground segment where the long product formation pipeline adds significant delays between the time of acquisition and the availability to the end-user. In case of time-sensitive conditions, such as natural disasters, it would be desirable to perform CD directly onboard of the spacecraft, so that alerts could be prioritized for low-latency transmission. However, onboard CD is far from trivial and requires to address a number of issues, including efficient onboard storage, image registration, etc. within the constraints of onboard resources. In this paper, we present a framework towards building an onboard CD pipeline. In particular, the essential operations required by an onboard CD pipeline are efficient storage of the images acquired for a given location during multiple revisits, their geometric registration and, only then, the actual change detection algorithm. We seek to develop a deep-learning (DL) approach that addresses all these issues with a single neural network model that is end-to-end optimized for a desired tradeoff between computational complexity, storage requirements and accuracy or resolution of change detection. This neural network has a modular architecture, conceptually consisting of an image encoder into a compact feature space for storage, an image decoder, a registration module and a change detection module. In particular, one wonders what is the optimal way of storing the image information for the purpose of best detecting change. We argue that instead of saving the images, the optimal approach is to save a compact representation generated by the neural network encoder with the objective of maximizing the downstream CD performance for a given bitrate constraint. In this coding-for-machines philosophy, compression is optimized not for the visual quality of a human observer, but rather for the algorithm performing inference tasks, such as CD, on the compressed data.