Currently, synthetic aperture radar (SAR) images are generally used for ship identification. However, the SAR can only be used to classify surface ships without any underwater target. In contrast, since sonar can receive radiated noises from both vessels and underwater targets, its images can be used to effectively identify targets at different depths. However, due to the shortage of underwater target data and the difficulty in modeling, sonar images are hardly used for deep learning (DL). To solve such problems, this letter proposes a compound convolutional neural network (CSDN) based on a shared latent sparse feature (SLS) and a deep belief network (DBN) to learn striation-based sonar images. This letter has three contributions. First, this letter uses striation images to overcome the lack of training data for CNNs. Second, the DBN is applied to optimize fuzzy or discontinuous fringes, and an SLS feature is proposed to represent the interference fringe. Finally, the two features are exploited to separately train CNNs, combined with weight, to enhance the accuracy of classifying water targets. The experimental results indicate that, compared with the other DL-based models-VGG, SSD, RFCN, and SCDAE-a CSDN is more stable for different datasets and has the highest accuracy of up to 93.34%.