logo SBA


Digital archive of theses discussed at the University of Pisa


Thesis etd-08212022-161649

Thesis type
Tesi di dottorato di ricerca
Thesis title
Artificial Intelligence on the edge for satellite on board remote sensing image processing
Academic discipline
Course of study
tutor Prof. Fanucci, Luca
  • ai
  • cnn
  • edge computing
  • embedded
  • low-latency
  • low-power
  • remote sensing
  • satellite
Graduation session start date
Release date
In recent years Artificial Intelligence (AI) has gained a lot of interest thanks to its ability to address a vast variety of problems. In particular, Convolutional Neural Networks (CNNs), a sub-topic of Artificial Intelligence (AI), have obtained astonishing results in the field of computer vision. The design of CNNs for computer vision tasks usually focuses on obtaining the best performance while neglecting the cost in terms of hardware resources. With the increasing proliferation of embedded systems, such as Internet of Things (IoT) devices, the need to move decision-making capability to resource-constrained devices has also increased. The research reported in this thesis focuses on efficient solutions to address the challenges in enabling the use of CNNs on board resource-constrained devices with particular emphasis on remote sensing applications on board satellites. Three research activities are described in this thesis.
The first research activity addresses the issues related to handling onboard generated data in remote sensing applications. Modern satellites for Earth Observation (EO) are remote systems that generate a huge amount of data, e.g., Hyperspectral Images (HSIs). A considerable part of the acquired HSIs may be made useless by clouds covering the observed phenomena. In this activity, a CNN is proposed to identify cloudy HSIs and discard them if needed.
The second research activity focuses on the design and development of a CNN for low-latency oil spill identification directly on board satellite. Oil spills represent one of the main concerns for the marine ecosystem. Early identification of oil spills can speed up the response from authorities and limit damages to the ecosystem. This activity proposes a CNN able to identify oil spills from Synthetic-Aperture Radar (SAR) images leveraging hardware accelerators for embedded applications.
The third research activity is dedicated to the implementation of an innovative solution for accelerating computation demanding algorithms on-board satellites. In recent years, many algorithms have been moved to the edge due to the need for low latency, privacy, etc. This causes an increase in demand for the computation capabilities of remote devices. Nowadays this trend is involving also satellites that can be used to deploy applications for which low latency represents a key factor. In this activity, a scalable and reconfigurable solution is proposed to enable the acceleration of parallelisable algorithms, e.g., CNNs. Moreover, special attention is paid to power consumption and ease of use from the developer’s point of view.