Abstract
Today there are more satellites in orbit than ever, and their number has been increasing exponentially in the last decade. With more and more data being gathered by these and other remote sensing missions, downlink capacity and data storage on land are an ever more crucial bottleneck. Data compression is, therefore, a key technology for the viability of these missions that enables the transmission and long-term archiving of the valuable information they gather.A revolution has taken over the state of the art of image compression in recent years: \ac{ML}. Advances in architecture design with the introduction of variational autoencoders, as well as the availability of computing power to train these models, have led to a breakthrough in which these novel approaches can today compete with sophisticated algorithms that are the result of decades of fine-tuning and innovation. Impressive as these results are, and despite missions such as $\Phi$-Sat 1 and $\Phi$-Sat 2 pioneering the usage of ML in space, several key hurdles remain in the way to the widespread adoption of these technologies. Two of the most important barriers to adoption are addressed in this thesis: complexity, and rate/quality allocation.
Computational cost is perhaps the biggest barrier to adoption of neural networks in remote sensing, in particular for compression. Indeed, most improvements in state-of-the-art image compression performance have been achieved by introducing ever more computationally costly architectures, which could not be used in the low-power environments that remote sensing missions typically operate with, especially spaceborne applications. While this issue had been addressed for single-band remote sensing data at the beginning of this thesis, no proposals had been made to that effect that were viable at a wide range of compression ratios and scalable on the number of bands, in particular for reduced-complexity compression of hyperspectral data. To compress multispectral and hyperspectral data, it is common to use a transform along the spectral dimension to decorrelate redundancy between components, a spectral transform. In conventional spectral transforms, applying spectral transforms in clusters of bands is a proven approach, which has been further developed into multi-scale spectral transforms such as the \ac{POT}. Following that precedent, the usage of ML compression architectures in clusters of bands to compress hyperspectral data is investigated. First, the usage of a learned clustered linear spectral transform followed by a learned 2D transform is proposed and evaluated. In a second contribution, the usage of a multi-band ML compression architecture in clusters of bands is evaluated, studying the trade-off between performance and complexity resulting from using different numbers of input bands. Both approaches show competitive performance in lossy compression with an also clustered spectral Karhunen-Loève Transform (KLT) followed by JPEG~2000.
These proposals, as most contributions to the broader field of ML image compression, are with fixed-rate models: models that are optimised for a particular rate-distortion trade-off and can only compress a given image at a fixed rate and quality. Besides the obvious limitation of having to train multiple models to compress at different rates, this is a significant barrier for remote sensing missions, where having control over the compression ratio or recovery quality is necessary to ensure the data being captured is adequately reconstructed down on Earth. To that end, a modification of a successful reduced-complexity architecture for panchromatic data compression is made to allow for continuous variable-rate compression. In a first contribution, it is shown that this variant performs on par with its equivalent multi-model baseline, competitively with current standards such as JPEG~2000 and CCSDS 122.0-B-2, and a practical method is proposed to compress at a user-defined rate, a novel feature for ML image compression. A second contribution describes a method with that same architecture for fixed-quality image compression, that is compressing an image at a user-defined reconstruction quality. This method is shown to be accurate in local fixed-quality compression (recovering a given region at a user-defined quality), and it is found that using local fixed quality achieves the same rate-distortion performance than global fixed-quality compression.
Date of Award | 25 Jul 2024 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Joan Serra Sagrista (Director) & Joan Bartrina Rapesta (Director) |