HyTC-Net: A Multi-Scale, Multi-Branch Hybrid Attention Cloud Detection Network for High-Resolution Image Cloud Texture Recognition

In response to issues with high-resolution imagery from the ZY-3 satellite—such as blurred cloud boundaries and incomplete segmentation of thin clouds following training on the OmniCloudMask (OCM) baseline model, as well as the tendency for clouds detected against high-brightness backgrounds such as snow-covered terrain to blend with the background—this paper proposes a hybrid CNN–Transformer network, HyTC-Net, designed to enhance cloud texture recognition capabilities. Based on the ConvNext-Tiny backbone, this model innovatively incorporates the CloudTextureAttention (CTA) module and a lightweight Transformer module, LightweightMobileViT (LM), while combining ASPP with a deep supervision (DS) mechanism to enhance the perception of cloud textures and boundaries, thereby improving segmentation capabilities for complex scenarios such as thick clouds, thin clouds, clouds in high-brightness scenes, and patchy clouds. On the ZY-3 dataset, HyTC-Net achieved accuracy of 94.73 ± 0.18% and a precision of 92.28 ± 0.21%, significantly outperforming the baseline model OCM by 1.12 percentage points in accuracy and 2.4 percentage points in precision. Compared with mainstream models, such as UNet and DeepLabV3+, the accuracy improved by up to 6.6 percentage points and the precision by up to 17.9 percentage points. The results demonstrate that HyTC-Net can effectively address challenges arising from uneven textures in high-resolution imagery and limited spectral information, such as unclear boundaries, misclassification of highly reflective features, and failure to detect thin clouds.