A deep neural network for CrIS cloud detection and transfer to HIRAS via implicit contrastive learning

Accurately identifying whether a field of view of satellite hyperspectral infrared sounders is clear or cloudy is critical for geophysical parameter retrieval and meteorological applications. In this study, a deep neural network is developed for cloud detection in Cross-track Infrared Sounder (CrIS) data and subsequently transferred to the High-spectral Resolution Infrared Atmospheric Sounder (HIRAS) onboard the Fengyun (FY)-3D satellite. Specifically, a deep convolutional neural network is designed for CrIS cloud detection, leveraging the cloud mask products from the collocated Visible Infrared Imaging Radiometer Suite (VIIRS) onboard the same platform as labels for supervised learning. To mitigate the inherent discrepancies between CrIS and HIRAS observations, an implicit contrastive learning strategy is introduced to facilitate model transfer to HIRAS. The proposed model achieves an overall accuracy of 93.32% on the global CrIS data test set, demonstrating highly consistent performance across daytime (94.30%)/night-time (92.34%) and land (93.24%)/ocean (93.36%) samples. Evaluations using HIRAS data indicate robust transferability, despite underlying differences in instrument calibration and noise characteristics. Case studies further illustrate the model’s capacity to reliably capture cloud features across diverse atmospheric conditions. Furthermore, the implicit contrastive learning significantly improves detection performance for challenging samples in complex environments, such as regions affected by haze or sun glint.