Image processing machine learning algorithms for interplanetary small-sats images to support Martian rovers navigation

Interplanetary exploration and, in particular, Mars exploration have, nowadays, gained interest in more and more technical and scientific research fields, from aerospace engineering to computer science and data science ones. Indeed, the latter can improve and complement the first, speeding up and making processes more efficient. This is now our science frontier and there is an evident need for this scientific and technical field to penetrate each other and work together. This Master Thesis work fits perfectly into this context, as a matter of fact our aim is to enhance Mars exploration and navigation, processing satellite data and images with Deep Learning algorithms with an ultimate goal of tenfold the Martian rovers traveling speed. This happens in the context of the SINAV project, required by ASI and led by Altec s.p.a., with the participation of Politecnico di Torino and other partners, leaders in the Italian space economy scenario. The processing of satellite images is then approached with semantic segmentation methods. Starting from the supervised approach, going through the challenging labeled dataset issue, we are approaching the unsupervised one. This work presents both the approaches with their strengths and issues, underlining the reasons that led us to the choice of the Unsupervised Semantic Segmentation. Neural Networks are the most suitable tool to approach this kind of processing. Within this work we present the use of CNN (Convolutional Neural Network) firstly, and the use of a Dino-ViT (Visual Transformer) with the aid of the STEGO “head” secondly, to distinguish different types of terrain and help the definition of the path for martian exploration by rover. Both the networks need to be tweaked to adapt to peculiarities of Mars terrain images, rarely approached with these kinds of algorithms. The dataset is provided by HiRise images (High Resolution Imaging Science Experiment) catalog, in particular the images about the Jezero crater, and preprocessed to best fit our neural network and GPUs constraints.