Boundary position attention network for enhanced cloud detection in mixed cloud and snow scenes

Cloud detection is a fundamental first step in the preprocessing of optical remote sensing images and a pivotal element for subsequent analytical tasks. High-resolution remote sensing images often require cropping into smaller patches to facilitate effective cloud detection. However, these segmented patches often lack rich features and exhibit high similarity between classes, which complicates accurate classification, particularly in intricate scenarios such as those involving both clouds and snow. To tackle this challenge, we introduce a novel cloud detection approach that employs boundary position attention, based on the CloudS26 cloud-snow dataset. This technique utilizes deformable convolution to generate boundary-focused attention, proficiently delineating the interfaces between clouds and snow in coexistent environments. Our approach has demonstrated robust detection capabilities in both the CloudS26 and CSWV datasets, showcasing its effectiveness in handling complex meteorological phenomena.