This article proposes a cloud detection method that considers image radiation and geometric features. By using radiometric correction, sensor correction, sparse sampling, system geometry correction, and histogram stretching, the impact of measurement image quality and format on cloud detection accuracy and efficiency has been reduced. By combining the geometric shapes observed by satellites and the geographic information of cloud areas with U-Net, misjudgments caused by cloud like features were eliminated, and the segmented cloud areas were subjected to morphological processing to optimize the contour of the cloud areas. The results showed that the accuracy of cloud detection increased to 96.63%, and the extreme error rate of cloud detection decreased to 1.19%.