Duration: 36 months
Of late, edge-AI has received tremendous interest; indeed, onboard image generation and processing enables detection of significant events with latency much smaller than in the case that the data are downlinked to the ground segment and processed there. Some onboard applications have been designed, tested and will soon fly in space, notably cloud screening for data reduction in the CHIME mission; on the other hand, bringing other applications to edge-AI is more challenging. This is clearly the case of change detection, which represents a compelling use case because it would enable low-latency identification of changed regions for fast decision-making aimed at disaster prevention and management. Change detection, however, poses significant challenges for onboard implementation because of the need to use registered image products, as well as to keep “memory” of the previously seen scenes; hence, it cannot be accommodated as a short-term project. This activity proposes a co-funded doctoral program with the objective to develop deep learning methods for onboard change detection in order to overcome the challenges mentioned above. The goal is to design a change detection system suitable for edge-AI, centered around a deep learning architecture having complexity suitable for modern onboard data processing systems, and providing the basic functionalities needed to generate suitable geometric information about the acquired images, and compare them with previous scenes in order to calculate a change map. The activity will procure representative datasets, define the onboard change detection architecture, develop the required tools, train the system on the datasets and validate its performance also in comparison to change detection systems operated on the ground. Finally, it will demonstrate an implementation of the deep learning part of the system on a suitable low-power accelerator.
