GEMD-UNet: Graph Structure Enhanced Multi-dimensional Learning Unet for Cloud Detection

Abstract

Cloud detection (CD) in remote sensing images is commonly used in satellite imaging and laser communication. UNet-based methods with multi-level feature caching and interaction learning, are popular for superior CD performance. However, most current CD methods focus on spatial feature enhancement through global feature learning, which often overlooks the detailed perception of color, shape, and boundaries due to insufficient channel interactions. Additionally, these methods tend to increase computational requirements. To ensure flexible feature attention, enhancing both structural and detailed understanding, while maintaining low computational complexity. In this paper, we propose Graph Structure Knowledge Enhanced Multi-Dimensional Learning U-shaped Network (GEMD-UNet) for CD. Specifically, we utilize tensor dimension transformations to cleverly employ two fundamental learnable parameters alternately for spatial/channel relationship learning. Furthermore, to enhance feature knowledge capacity and focus more on local detail, we employ local window partitioning to divide hidden features into multiple independent learning groups and inject two different physically meaningful graph structures knowledge in a non-Euclidean space into features. Experimental results demonstrate that GEMD-UNet outperforms previous SOTA methods on both HRC-WHU and L8SPARCS datasets. Code can be available at: https://github.com/mxx123321/cloud_model.