Compression and cloud screening for satellite images

The increasing volume of Earth Observation (EO) data generated by the high-resolution satellites' sensors poses critical challenges to onboard storage, transmission, and real-time processing. The images collected during the space missions need to clearly show the surface of the Earth in order to be used in multiple contexts and fields: land cover classification, vegetation, ice and water analysis, atmospheric correction and mineral mapping. Unfortunately, the sky is not always clear of clouds, which degrade the information quality when appearing in the satellite images, often making major portions of the data unfit for further analysis and losing the very purpose the images were meant to serve. Accurate cloud detection algorithms are required to discriminate cloudy pixels from cloud-free ones directly onboard the satellite to generate binary segmentation masks, exploited during the onboard compression stage. According to the CCSDS 123.0-B-2 standard for multispectral and hyperspectral image compression, it is possible to assign different compression parameters depending on pixel classes. Specifically, cloudy pixels, being less informative, can be more compressed than cloud-free ones. This particular compression approach reduces the volume of data to be transmitted, saving on onboard resources. The thesis addresses the challenge of detecting clouds directly onboard through a lightweight U-Net, and create the segmentation masks used in the compression stage. Satellites are equipped with MultiSpectral Imagers (MSI) Instruments which are provided with scanners acquiring data line by line to create an image. For this reason, standard image-level processing is impractical in onboard scenarios, as it would require loading entire high-resolution images into memory. To address this constraint, the proposed lightweight U-Net architecture is trained using two memory-efficient strategies: chunk-based and sliding-window-based learning. These approaches allow the network to operate on smaller portions of the image, significantly reducing the memory footprint during inference. This design choice ensures that segmentation masks can be generated progressively during onboard execution, enabling real-time cloud screening. Particular emphasis is placed on the trade-off between segmentation accuracy and computational efficiency. To further reduce computational complexity and adapt the network for efficient deployment on hardware platforms, the lightweight U-Net used to create binary cloud masks is optimized through quantization, specifically weight ternalization. To this end, several quantization strategies have been investigated and empirically compared using a consistent set of performance metrics. The presence of ternary weights simplifies hardware implementation by lowering complexity in arithmetic operations, resulting in a reduction in both memory usage and inference latency while, for the majority of methods, maintaining high segmentation accuracy. Experimental results confirm the effectiveness of the proposed methods in both compression efficiency and cloud screening accuracy.