AI-Based On-Board Image Compression for Earth Observation Satellites

The growing demand for high-resolution imagery from Earth Observation (EO) satellites, driven by the need for better environmental monitoring, resource management, and disaster response, necessitates innovative approaches to on-board data compression to optimize downlink transmissions to ground stations. This thesis investigates the application of Artificial Intelligence (AI) techniques for on-board image compression. Unlike traditional algorithms, such as JPEG 2000, which have been the standard for many missions, AI-based models offer the potential for both greater compression efficiency and higher reconstruction quality. This research focuses on developing a lightweight Convolutional Autoencoder (CAE), a type of Convolutional Neural Network (CNN), for single-band image compression capable of operating within the limited resources available to satellite systems. A baseline CAE is established and further optimized through various modifications to enhance its performance. A large and well-curated Sentinel-2 (S2) Level-1C (L1C) dataset is used to train and evaluate the models, which are then subjected to rigorous and extensive comparisons with the JPEG 2000 standard, employing both quantitative and qualitative metrics. The results reveal that while the proposed CNN-based compression methods generally provide good image reconstruction quality, their performance is not yet on par with the well-established JPEG 2000 standard for images with intricate patterns or high-contrast edges. However, for simpler images, the CAE achieves better results than JPEG 2000, underscoring its advantages in specific scenarios. Moreover, the recent trend of integrating high-performance, space-ready hardware accelerators on-board EO satellites, coupled with the capability of AI models to leverage such increased computational power, makes these models appealing when reduced encoding time and low power consumption are prioritized over reconstruction quality. These findings highlight both the current benefits and the need for further advancements in AI-driven solutions to close the gap in reconstruction performance compared to traditional methods. Future work should focus on optimizing the model architecture and refining the training procedure to ensure faithful reconstructions for complex scenes.