A hybrid CNN–transformer network for cloud detection in multi-source Chinese high-resolution remote sensing imagery

Cloud detection in high-resolution optical remote sensing imagery plays a crucial role in the preprocessing pipeline, directly determining the quality and usability of data for downstream applications. Although various approaches have been developed for cloud detection of Sentinel and Landsat Earth observation imagery, there remains a notable shortage of datasets from multi-source, large-scale, and high-resolution satellites, especially those providing meter-level imagery. To address this limitation, this study develops a Multi-source Chinese High-resolution satellite Cloud Detection dataset (MCHCD). The MCHCD dataset incorporates ten types of high-resolution Chinese satellite images, including Gaofen-1 A-D (GF-1 A-D), Gaofen-2 (GF-2), Gaofen-6 (GF-6), Ziyuan-3 (ZY-3), Ziyuan-302 (ZY-302), and China–Brazil Earth Resources Satellite-04 (CBERS-04), with spatial resolutions ranging from 3 to 10 meters, covering the entire territory of China with a total of 415 scenes. This dataset currently represents the most comprehensive collection of high-resolution Chinese satellite imagery for cloud detection, enabling evaluation of large-scale cloud detection tasks across multiple scenes, scales, and categories. Considering the characteristics of cloud, which form above a variety of underlying surfaces and display features across multiple spatial scales, with diverse shapes and variable thicknesses, a novel algorithm with strong generalization capacity based on a deep learning algorithm called MCHCDNet is proposed. The backbone of MCHCDNet integrates Convolutional Neural Networks (CNNs) and Transformer architectures to effectively capture the local and global features of various cumulus, stratus, and cirriform clouds. Additionally, a feature interaction fusion module (FIFM) is designed to fuse the multi-scale features, enhancing the overall perception of different types of clouds. Moreover, a multi-scale deep supervision mechanism is employed to alleviate spatial detail loss caused by down-sampling during the feature extraction process. Experimental results show that the MCHCDNet achieves improvements in cloud detection IoU across ten satellites by 0.41% and 1.22% compared to state-of-the-art methods HRCloudNet and CDNetv2. It demonstrates superior detection performance for large-scale stratiform and wave clouds as well as small-scale cumulonimbus clouds, while exhibiting strong generalization and robustness across different land-cover scenarios. Therefore, the proposed method holds significant potential for applications in multi-source high-resolution satellite imagery cloud detection.