Design and Evaluation of the Reliability of Convolutional Neural Networks for Earth Observation Applications

Convolutional Neural Networks for object detection are increasingly executed onboard satellites to enable real-time, at-the-edge inference for Earth observation applications. In this context, the memory devices used to store these models are exposed to space radiation, which can induce bit flips in the stored weights, degrading system dependability and, in particular, the reliability of the neural networks. A promising approach to mitigate such effects is to perform a reliability assessment of neural networks prior to deployment, allowing for more informed decisions regarding their use in space missions. This thesis presents RADRELAX, a tool designed to assess the reliability of Convolutional Neural Networks performing object detection tasks. The tool simulates radiation-induced faults by injecting multiple bit flips into the weights of convolutional and linear layers, following a configurable fault model. By evaluating how networks respond to these faults across different severity levels, RADRELAX enables quantitative comparison of model architectures based on their intrinsic resilience to radiation effects. The tool was used in a real use case: the development of an airplane detection model destined for onboard deployment. Experimental results demonstrate that the tool measures and characterizes expected fault propagation behaviour: as fault severity increases, performance degrades accordingly. The experiments reveal different fault manifestation modes, ranging from catastrophic failures to silent degradations. These findings contribute to a deeper understanding of model robustness and can guide the design of more resilient architectures for such spaceborne applications.