HYPERBOLA: HYPerspEctRal onBOard cLoud AI: preliminary results

Abstract

Cloud and cloud-shadow masking in hyperspectral (HS) remote sensing is a critical step for reliable Earth Observation data. However, performing this task on-board a satellite remains challenging. Traditional approaches, such as thresholdbased tests or CNN classifiers, often fail to capture complex spectral–spatial dependencies and typically treat cloud detection as a binary problem, overlooking cloud-shadow and multi-class distinctions. We propose HYPERBOLA (HYPerspEctRal onBOard cLoud AI), an AI framework that integrates Vision Transformers (ViTs), spectral–spatial Graph Neural Networks (GNNs), and hybrid CNN-Transformer architectures to improve cloud and cloud-shadow segmentation. In this work, we present preliminary results focusing on a single hybrid CNN-Transformer model (Hybrid Model 2). This architecture shows promising performance in segmenting Sea and Land classes, while highlighting current challenges in detecting thin or sparse clouds. These findings establish a baseline for onboard hyperspectral cloud masking and demonstrate the feasibility of transformer-based methods in constrained environments. Future work will extend the evaluation to the full set of proposed models (ViTs, GNNs, and hybrid variants), incorporate advanced optimization techniques such as structured pruning and 8-bit quantization, and benchmark performance across a broader set of test scenes. The ultimate goal of HYPERBOLA is to deliver an efficient and accurate onboard AI framework for real-time hyperspectral cloud and cloud-shadow segmentation.