EAPNet: Enhanced Attention-Guided Network for Fine-Grained Semantic Segmentation

Ground-based cloud observation is essential for meteorological research and climate monitoring. Existing methods for cloud detection and classification in whole-sky images (WSIs) often focus solely on binary segmentation (sky and cloud) or image-level classification, which perform poorly in tasks requiring the classification and segmentation of different cloud types. This paper presents an enhanced attention-guided pyramid network (EAPNet) designed to enhance the accuracy and robustness of fine-grained cloud classification in whole-sky images through deep learning techniques. The key innovations include: 1) the utilization of the pyramid spatial pooling (PSP) module to capture multi-scale contextual information; 2) the introduction of the convolutional block attention module (CBAM), which leverages an attention mechanism to enhance feature representation; and 3) an improved decoder module that enhances localization accuracy. Experimental results demonstrate that EAPNet outperforms existing mainstream methods in both cloud classification and segmentation tasks.