Mask-Net: A U-Net architecture with spatial and channel attention mechanisms for improved cloud detection in sea surface temperature imagery

Sea surface temperature (SST) is an essential climate variable, yet its estimation by infrared remote sensing remains fundamentally limited by cloud cover. Existing cloud masking approaches rely heavily on non-standardized thresholding algorithms subject to significant bias, particularly in dynamic coastal regions such as upwelling zones, where cold ocean structures can easily be confused with clouds. To address this challenge, we propose Mask-Net, a novel cloud segmentation architecture based on U-Net, integrating spatial and channel attention mechanisms through an Inter-Level Fusion Module (ILFM). This design enables fine-grained discrimination between clouds and cold oceanic structures, where conventional methods show more limitations. Evaluated on one year of daily AVHRR sensor data over the Tropical Atlantic at resolutions of 18 km and 9 km, Mask-Net achieves an overall accuracy (OA) of 89%, an mIoU of 83%, and a Dice score of 87%, significantly outperforming existing models. The results demonstrate better spatial continuity of thermal fields and a 29% increase in usable SST observations during the cold season. An ablation study confirms that the synergy between spatial and channel attention is a key factor in this robustness, enabling Mask-Net to correct systematic over-masking biases inherent to expert annotations. The generic nature of the U-Net framework and the universality of infrared cloud-detection challenges suggest potential applicability beyond the Tropical Atlantic, while acknowledging that further validation on independent datasets would be required.