Currently, the primary means of automatic and continuous marine radioactivity monitoring is through sensors based on NaI(Tl) scintillation crystal. The identification of radioactive nuclides in seawater can be accomplished through analysis of the gamma-ray spectrum obtained by the sensor. However, conventional peak-seeking-based nuclide identification methods have problems, such as complex data preprocessing operations and low identification efficiency for weak and overlapping peaks. In light of these challenges, this study proposed a new method based on a convolutional neural network(CNN) to analyze the gamma-ray spectrum of seawater. The technique utilized in this study involves converting one-dimensional spectrum data into grayscale images using the grayscale image conversion method. This method is conducive to convolutional neural networks' powerful and flexible feature extraction and classification learning capabilities. By developing a multi-layer CNN model, we can achieve the goal of accurately identifying radioactive nuclides in the ocean. Simulation experiments have confirmed the effectiveness of this approach in automatically identifying radioactive nuclides in seawater, with satisfactory identification performance for weak or overlapping peaks that meet practical application requirements.