Abstract: In optical surveys of space targets and debris, clouds are one of the key factors affecting observation efficiency. Cloud cover can reduce target visibility, complicating target image detection and subsequent accurate position and brightness extraction, thus interfering with the observation of space targets and debris. Efficient and accurate identification of cloud-contaminated images can provide valuable prior information for subsequent data processing, supporting the normal and stable operation of the operational workflow. Currently, mainstream methods based on image segmentation to detect clouds in images have drawbacks such as being time-consuming and vulnerable to noise. This paper combines image feature evaluation and manual screening to establish a cloud-contaminated image dataset based on optical survey data of space debris. We experiment with three machine learning methods: support vector machine, Shufflenet V2, and Resnet 18, to classify cloud-contaminated and normal images. The results show that Shufflenet V2 achieves an overall classification accuracy greater than 97%, while the SVM (Support Vector Machine) model achieves a cloud-contaminated image recognition accuracy of over 98%. Deep learning methods can effectively identify cloud-contaminated images, and the computational speed meets the real-time processing requirements for observational data. In future observations, the proposed method can be integrated with cloud imager and jointly applied to optimize space debris observation plans, reducing the impact of weather on observational equipment performance and promoting more stable operation of observation stations.