A Self-Attention U-Net for Cloud Detection from FY-4B/GIIRS Observations

Accurate cloud detection for geostationary infrared hyperspectral observations is important for the effective use of clear sky radiances in atmospheric retrieval and related applications. In this study, FY-4B/GIIRS observations were used to develop and evaluate three deep-learning cloud detection models, namely a conventional 1D-CNN (GCD-1D), a standard U-Net (GCD-U1), and a self-attention U-Net (GCD-U2). Cloud labels were generated by time-space matching between AGRI Level 2 cloud mask pixels and GIIRS field of views, and model performance was assessed using overall accuracy (OA), probability of detection (POD), and false alarm ratio (FAR) under different seasons, day/night conditions, and surface types. The results show that GCD-U2 achieved the best overall performance, with an OA of 85.88%, a POD of 77.07%, and a FAR of 24.40%, outperforming both GCD-1D and GCD-U1. The learned channel attention pattern was also physically consistent, with high weights assigned to LWIR window channels and selected MWIR bands. In the comparison with the GIIRS L2 operational cloud mask product, GCD-U2 showed higher consistency with the AGRI reference, with an average recognition–performance difference of about 10%. These results demonstrate the potential of attention-enhanced deep learning for operational cloud detection from geostationary infrared hyperspectral sounders.