Xenia Ivashkovych, Vito Remote Sensing
Bart Beusen, Vito Remote Sensing
Nick Witvrouwen, Vito Remote Sensing
Tanja Van Achteren, VITO Remote Sensing
Xenia Ivashkovych, Ms, Vito Remote Sensing
The disparity between data capture in space and Earth’s downlink capabilities is well-documented and poses a significant challenge for space industry stakeholders, academics, and public institutions alike. The consequences of this disparity are manifold and far-reaching. Chiefly, it limits the development of applications reliant on data-intensive sensors and missions, such as hyperspectral missions, where the imbalance between the maximum data capture rate and downlink capabilities exceed a factor of three hundred.
The CORSA technology addresses these challenges on multiple fronts. Firstly, its achieve up to three-hundredfold data compression for hyperspectral instruments. Secondly, as foundation models, they enable the training of AI-based applications with minimal labelled data. With its compact design and compression capabilities, CORSA functions as an edge foundation model, tailored for multi-purpose use in contexts where weight, power, and computation capabilities are constrained, including but not limited to space applications.
For background, CORSA is a type of AI model called a Hierarchical Vector Quantized Variational Auto-Encoder (VQVAE). This deep neural network architecture combines both powerful expressivity with high compression efficiency. While typically optimised for reconstructing and interpreting a specific sensor, the model can be readily adapted to others. In fact, the closer the sensors, the less data the CORSA model—already pre-trained on the first sensor—requires to be fine-tuned for the second sensor. Its compression performance has been validated for instruments such as Sentinel-2 RGB-NIR, on ENMAP hyperspectral [400-900 nm], on PRISMA hyperspectral [400-900 nm], and on simulated APEX hyperspectral data [400-900 nm]. Downstream applications already developed include classification, super-resolved parcel delineation, land-cover mapping, change detection, and flood monitoring.
At the SmallSat Europe Conference, we will focus on the implementation of CORSA-based object detection on the Nvidia Jetson platform. This requires the deployment of the edge foundation model itself alongside an application-specific head. Object detection will serve as a case study, demonstrating how CORSA-based applications can be optimised for on-board operations. Performance evaluation will consider classical AI metrics such as accuracy, as well as efficiency-related measures like throughput and resource utilisation. We will investigate the limits of the application in constrained environments, assessing real-time viability and, where not feasible, the trade-offs between performance, power consumption, and resource usage.
We extend our thanks to ESA for sponsoring the Smart-Connect project for advanced space technology for disaster response, as part of their Civil Security from Space programme, without whom the research necessary to refine CORSA would not have been possible.