Abstract: Due to the inhomogeneous and dynamic characteristics such as flows and internal waves, the sound speed field presents complex spatial-temporal fluctuations, making it difficult to predict the sound speed profile (SSP) time series without sufficient training data. A long short-term memory (LSTM) model based on transfer learning is constructed, and temperature profiles recorded by two thermistor chains in the South China Sea are used to validate the performance of the model. For the two scenarios with and without internal solitary waves, the pre-trained model is finetuned based on transfer learning, and the performance and effectiveness of the deep transfer networks under different amounts of training samples are analyzed. The experimental results show that the deep transfer network can simultaneously meet the requirements of high precision and high efficiency, and the prediction performance is significantly better than that of the non-transfer treatment.