Deep Learning for Onboard Real-Time Cloud Segmentation in Satellite Images

The Earth observation satellites have allowed us to monitor the planet at a scale that seemed unattainable up to some years ago. Satellite imagery plays a fundamental role in many applications, like forestry, tracking wildfires, mapping deforestation, railway industry, urban planning or agriculture. Terabytes of image data are produced everyday, making it necessary to have automated methods to recognise anomalies inside the images before using them for further analysis. One of the main obstacles in satellite imagery is represented by the presence of clouds, which regularly cover 66% of the Earth's surface, obscuring the ground objects we are interested in. This thesis studies the use of Deep Neural Networks to design an on-board algorithm able to perform cloud segmentation in real time with the aim of classifying each pixel of the satellite images as cloud or background. This work investigates the design of a deep learning model based on a hybrid recursive attention mechanism that processes images one line at a time. This architecture, through its ability to model dependencies within sequences of complex data while maintaining system simplicity, offers significant advantages in terms of memory efficiency and reduced latency compared to traditional 2D approaches. Specifically, the novel Mamba layer handles the along-track direction of the images, while the 1D convolutional layers of the U-Net process the columns. To assess the prediction capabilities of the developed solution, a comprehensive evaluation of performance metrics, such as accuracy, precision, recall, and computational efficiency is provided. Additionally, a comparative analysis with state-of-the-art approaches is conducted to demonstrate the superiority of the proposed architecture when handling real-time processing with memory constraints.