Zhuhai-1 hyperspectral satellite has the characteristics of high spatial, high spectral and high temporal resolution, which effectively promotes the wide application of hyperspectral remote sensing data in the fields of agriculture, forestry and natural resources detection, among which high precision cloud detection is the key step of remote sensing data preprocessing. How to effectively extract features from hyperspectral images and overcome the defects of traditional cloud detection methods, such as complex features, many algorithm parameters, large amount of computation, and poor robustness, is a key issue in the research of hyperspectral cloud detection. In this paper, a U-shaped structure network with multi-scale feature fusion is proposed. The model firstly uses the residual module for feature coding and multi-scale fusion of coding. The coordinate attention mechanism is introduced at the jump junction of the network to extract useful information, and finally the output is obtained by residual decoding. Before the experiment, principal component analysis (PCA) was used to reduce the dimensionality of hyperspectral data to reconstruct the 4D image data. Then, through data annotation and data enhancement, the Zhuhai-1 hyperspectral image cloud detection dataset was established. In this paper, 38-Cloud Cloud data is used to train the initial network parameters, and then the constructed data sets are used for transfer learning. The experimental results show that for the established Zhuhai-1 satellite hyperspectral cloud detection dataset, the pixel accuracy of the proposed method reaches 92.28%, which can realize high precision hyperspectral remote sensing image cloud detection.