Abstract: To improve the performance and applicability of modulation classification of underwater acoustic communication signal in impulse noise environment of shallow sea, an approach based on denoising automatic-encoder (DAE) and convolutional neural network (CNN) modulation recognition is proposed. First, a DAE-based noise reduction module is built to reduce the effects of the alpha stable distribution noise on modulation characteristics; Second, the power spectrum of the output signal is recognized by CNN, so as to complete the modulation recognition. Meanwhile, the idea of data migration is used to solve the problem that small sample target data cannot support neural network training. Simulation results and practical signal test results demonstrate the effectiveness of the proposed method. Compared with the existing algorithms, the proposed method reduces the requirement of professional knowledge, improves the recognition rate in impulse noise environment, and still has a better recognition rate under the condition of small sample target data.