The invention discloses an SAR (Synthetic Aperture Radar) image classification method based on contraction auto-encoders, comprising the steps of: (1) inputting an image; (2) performing stable wavelet decomposition; (3) selecting a training sample; (4) constructing two parallel levels of contraction auto-encoders; (5) training the two parallel levels of contraction auto-encoders; (6) constructing a sample feature set; (7) training a Softmax classifier; and (8) classifying. Compared with present feature extraction methods of multi-level local mode histograms, the SAR image classification method has the advantages of high classification precision, sound area consistency and accurate edge classification, solves the problems of speckle noise influence, area classification confusion and edge untidiness, and can be applied to SAR image object detection and object identification.