FA-CloudSeg:Feature-Aligned Architecture for Ground-Based Cloud Image Segmentation

As a vital component of the Earth’s system, clouds play a crucial role in climate change research and meteorological forecasting. Due to the blurred edges and highly variable shapes of cloud layers, existing deep learning methods for cloud image segmentation often suffer from boundary misjudgments. To address these challenges, this paper proposes a convolutional neural network (CNN) model named FA-CloudSeg, which effectively optimizes segmentation performance for complex cloud boundaries through the integration of an Image Embedding Layer, Self-Attention Module, and Feature-Aligned Module. The proposed network exhibits the following key characteristics: First, Image Embedding Layer: Extracts cloud-related features from images, enhancing the network’s ability to distinguish cloud regions. Second, Self-Attention Module: Captures global dependencies to improve feature representation under varying lighting and morphological conditions. Finally, Feature-Aligned Module: Integrates shallow and deep-level information while maintaining cross-scale consistency, refining segmentation accuracy for intricate cloud boundaries. Experimental results on public datasets demonstrate that the model outperforms existing methods in core metrics such as recall rate, offering a novel solution for ground-based cloud image segmentation.