• CNN
• Dario Amorosi
• earth observation
• ESA
• GAN
• Machine Learning
• New Space Economy
• PhiSat
• Python
• Space
• Tensorflow

The dramatic increase in the number of satellites in orbit in recent years has brought progressive interest even from companies that did not operate in sectors directly related to space. To fully exploit this area, however, years of flight heritage are required, which not all companies can afford, because of that many rely on services offered by others. Deploying a new remote sensing service, however, is not simple, especially considering that the advancing interest does not keep pace with the advancement of artificial intelligence for space applications, which although on earth have had incredible improvements in recent times, on satellites had never been used by european missions. For this reason, several companies in collaboration with the European Space Agency and the University of Pisa have created the PhiSat-1 mission, demonstrating that artificial intelligence could be used on-board even for sensors never adopted in orbit, using a training dataset composed of images from the Sentinel-2 mission. Thanks to systems like the Intel Myriad Movidius Stick 2 it has been possible to reduce the time to market of neural networks for embedded satellite systems, allowing to meet the strict requirements of size and power consumption, still obtaining good performances. The use case examined in this thesis is the application of Artificial Intelligence to filter out images that are too cloudy to be usable, in order to send on-ground only those with a cloud presence considered acceptable. Given that the neural network deployed on PhiSat-1 had great limitations on its training set, as it was not very large in size and with a very unbalanced distribution in the cloudiness of the images, it has been used an architecture of Generative Adversarial Network partly customized to produce artificial images to join the real ones during a new training of the classifier network employed on PhiSat-1, both to increase the size of the dataset and to rebalance it. A custom Frechet Inception Distance has been defined to assess the quality of the produced samples. Several scenarios were defined to test in which situation the artificial images were useful. In particular, there was a statistical increase in the ability of the network to detect cloudy images when artificial samples were used to rebalance the cloudiness of the training set.