We introduced a new cloud detection method using multi-scale feature extraction (MFE). This new method focused on extracting features across or in different scales and orientations of image for classification rather than designing a sophisticated classifier. In the first step of MFE, the steerable pyramid decomposition was used to decompose a remote sensing image (RSI) into two scales and six orientations in each scale. Then, a 62-dimension-feature vector was computed from the original image and the twelve derived images (two scales, six orientations) to represent the original sample counterpart. At last, the popular classifier, SVM, was used to test the discrimination of the 62-dimention-feature vectors in RSIs. The experimental results showed that the new method has a good performance and robustness.