On-board cloud screening algorithms for satellite imaging

Remote sensing data acquired through multi-spectral satellite sensors provide an opportunity to monitor and understand the Earth's physical, chemical, and biological systems. However, approximately 70% of the sky is often covered by clouds, which can lead to the processing and storage of images that are not useful for Earth observation tasks, such as monitoring vegetation, ice, or water bodies. To better understand which images should be processed and how, cloud detection algorithms have been designed to identify cloudy pixels, generating cloud segmentation masks that can be used to compute cloud coverage. Accurate detection is essential to filter out highly cloud-covered images for reliable downstream processing. The variability in cloud characteristics makes this identification process complex, as clouds can vary in shape, density, and other spectral properties, typically resembling bright surfaces like snow, ice, deserts, or rocks in certain spectral bands. Moreover, on-board cloud detection is a real-time application that requires algorithms to identify cloud-covered data, either to discard or to exploit cloudy pixels to efficiently compress data for downstream tasks. However, the primary challenges in on-board systems stem from the limited computational resources available on satellites, including power, memory, and processing capabilities. Balancing the trade-offs between accuracy, false positive rates, and efficiency is the main challenge in this context. This thesis explores multiple approaches to cloud detection, including a threshold-based cloud detection algorithm, machine learning algorithms, and a deep learning model. The first approach is a physical cloud detection algorithm used in Sen2Cor, which involves a threshold-based test on image bands that, in this thesis, was optimized to minimize false positive rates (FPR). Support vector machines and random forests were also trained for cloud segmentation, despite being computationally heavy, to provide a better benchmark for comparison. Finally, a lightweight model for cloud segmentation is proposed, designed to be competitive with other deep learning models in terms of accuracy, while remaining efficient and comparable to threshold-based algorithms. The results obtained using these methods are presented and compared at the end of the thesis.