Overview: The satellite cloud image can show the characteristics of the cloud system and its evolution process from multiple angles. The research of cloud image retrieval is of great significance for weather monitoring and climate research. In the design and implementation of cloud image retrieval system, feature extraction is the key link. Since different types of cloud image features have their own advantages when portraying cloud images, combining different types of features will help improve the performance of cloud image retrieval system. However, since the combined cloud image features tend to be too high in dimension, the time cost is often high in the similarity measurement phase, and there is redundancy between the feature vectors. In view of the above problems, this paper combines sparse representation and subspace projection technology to propose a dimension reduction method for cloud image combination features. Firstly, the content information of the cloud image is drawn from different angles, that is, the three features of the cloud image color, texture and shape are extracted. After that, the three features are normalized and cascaded into a combined feature vector, and then the combined feature vector is matrixed. The form is arranged, and the matrix is subsequently subjected to block processing. Then, each block is sparsely represented, the variance of each feature block sparse representation coefficient is calculated, and the feature blocks are grouped according to the variance obtained by different blocks and grouped by grouping. It is then possible to separate the initial combined features into salient features and non-significant features. At this point, a subspace will be searched, in which the significant part is preserved, and the non-significant part will be suppressed. A projection matrix can be obtained by learning training to achieve the combined feature dimension reduction. In the retrieval stage, the initial cloud image combination feature vector is projected on the projection matrix, and the dimensionality reduction cloud image feature can be obtained, so that the cloud image retrieval can be realized quickly and accurately. The experimental results show that the cloud image retrieval system achieved by this method is superior to the traditional dimensionality reduction method in the accuracy and recall rate of the cloud image retrieval system, and the time complexity of the retrieval process is low. This indicates that the proposed method has strong dimensionality reduction ability for cloud image combination features, and provides a new idea for cloud image feature dimension reduction and efficient cloud image retrieval system.