ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

Tesi etd-01172021-203315


Tipo di tesi
Tesi di laurea magistrale
Autore
ZLATKOVA, NATALIJA
URN
etd-01172021-203315
Titolo
Synthesizing a Variable-Resolution Image Dataset for Earth Observation Satellites exploiting Generative Adversarial Networks
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
EMBEDDED COMPUTING SYSTEMS
Relatori
relatore Fanucci, Luca
relatore Bechini, Alessio
supervisore Giuffrida, Gianluca
Parole chiave
  • machine learning
  • synthetic dataset
  • artificial intelligence
  • mantis
  • sentinel-2
  • fake satellite images
  • generation
  • gan
  • satellite images
  • generative adversarial network
Data inizio appello
19/02/2021
Consultabilità
Non consultabile
Data di rilascio
19/02/2091
Riassunto
Nowadays, the number of micro and nano satellites launched in the orbit for earth observation has increased rapidly. Thanks to the recent technological advancements, it is now possible to introduce artificial intelligence directly on board, while maintaining a very low power-consumption. In fact, the recent mission PhiSat-1 demonstrates for the first time the use of embedded low power Commercial Off-The-Shelf Myriad 2 Visual Processing Unit on-board of a satellite. It was successfully tested at CERN, where the Total Ionizing Dose was successfully evaluated, allowing to bring Deep Neural Networks directly on board.
This advancement opens new perspectives and possibilities, but it also introduces new challenges to overcome. An example of the latter is the MANTIS mission. It will use the same technology of the PhiSat-1 to reduce the bandwidth consumption featuring a Deep Neural Network. Unfortunately, the camera on board of the MANTIS satellite has no flight heritage i.e., it has never been used in an official mission. Thus, there is not a sufficient amount of data for training a dedicated Deep Neural Network.
This thesis focuses on the generation of synthetic satellite images for the MANTIS camera manufactured by Satlantis. For this purpose, the Generative Adversarial Networks (GANs) offer a promising solution to synthetic image generation. Recent advancements and studies conducted on novel GAN architectures show their capabilities in producing high-quality, high-resolution and high-diversity images. These capabilities are emphasized when working with relatively small and unbalanced datasets. Hence, the state-of-art GAN for image synthetization called StyleGAN2, was used to generate the synthetic dataset of MANTIS, starting from images taken by a similar satellite, called Sentinel-2. In particular, the best StyleGAN2 model was trained on 512x512 resolution, generating a dataset of about 10k images with the same properties of the MANTIS camera e.g., spatial resolution, colors, light, etc. The result of this synthesized dataset is very promising. In fact, it reached a Frechet Inception Distance of 20.732 and an Inception Score mean of 3.73 with a standard deviation of 0.07.
File