Development and Comparative Analysis of Classical and Machine Learning-Based Cloud Detection Algorithms for Real-Time Onboard Processing on CubeSats Using Raspberry Pi Compute Module 4

Abstract

The rapid advancements in space technology necessitate the development of robust and efficient systems for real-time onboard processing on CubeSats, mainly using compact systems like the Raspberry Pi Compute Module 4. This study presents the development and comparative analysis of classical, and modern cloud detection algorithms optimised for these resource-constrained environments. Leveraging large-scale reference datasets with 10–15-meter spatial resolution, the research focuses on enhancing CD capabilities across diverse atmospheric conditions, emphasising system reliability and resource efficiency. Custom tools, including the Cloud Detection Threshold Generator, were developed for in-depth histogram analysis of 4,743 satellite images, establishing critical thresholds. These thresholds were instrumental in generating baseline cloud masks through a purpose-built algorithm, ensuring precise model application and validation. Classical algorithms like SiRTH and CloudTracker.HSV, which employ statistical thresholds and HSV colour space analysis, achieved detection accuracies of 65.36% and 64.61% at 30 FPS and 23 FPS, respectively. In contrast, modern machine learning-based algorithms, such as CloudDet (BBox) and CloudDet (Seg), demonstrated better performance with accuracies of 73.43% and 75.14% at 19 FPS and 13 FPS, respectively, effective in complex environments with variable cloud formations and diverse terrain. While the training of ML models is computationally expensive and demands powerful ground-based resources, their inference on CubeSats is lightweight and efficient, making real-time processing on CubeSats feasible. As remote sensing technology evolves, classical methods may not ensure reliability and operational efficiency on resource-constrained platforms like CubeSats. This study highlights the potential of hybrid approaches, combining classical and modern techniques with modern algorithms, offering greater accuracy and adaptability. Supporting the Clear Shores concept, which aims to enhance water quality monitoring in Aotearoa-New Zealand's dynamic lake and coastal environments, this research advances CD methodologies for onboard processing. By improving CD capabilities, this study contributes to developing reliable, real-time environmental monitoring tools that better inform decision-making processes, particularly in managing New Zealand's vital water resources, while also supporting global sustainability efforts.