Hyperspectral Land-Sea-Cloud Segmentation for Onboard Processing via Spectral Band Selection

Cloud detection and land-sea discrimination are fundamental preprocessing steps for the analysis of satellite imagery, particularly in hyperspectral missions where atmospheric effects and spectral variability pose significant challenges. In this work, we investigate the impact of spectral band selection on land-sea-cloud segmentation for onboard processing, using a lightweight NU-Net architecture applied to hyperspectral data from the HYPSO-1 mission. The original hyperspectral images, composed of 120 bands spanning wavelengths from 388 nm to 810 nm, are subdivided into fixed-size patches to increase data availability and reduce computational complexity. Four configurations are evaluated: a 3-band setup, a 4-band RGB-NIR (near-infrared) configuration, and two 6-band variants incorporating additional near-infrared and red-edge information. The models are trained using a combined loss function based on weighted cross-entropy and Dice loss to address class imbalance. Performance is assessed using Intersection over Union (IoU) and pixel accuracy on independent validation and test sets. Experimental results demonstrate that multi-band configurations, particularly those including multiple near-infrared channels, significantly improve cloud discrimination and land-sea boundary delineation, achieving a mean IoU of up to 0.81 on the test set. These findings highlight the importance of spectral analysis for effective semantic segmentation in resource-constrained remote sensing.