Abstract: Sea surface temperature (SST) is not only a key indicator of oceanic state but also directly affects atmospheric dynamics, thereby impacting global and regional weather systems and climate change. Therefore, accurate SST predictions of inestimable value for maintaining ecological balance, promoting economic development, and addressing climate challenges. In order to achieve precise SST predictions, this study proposes a deep learning model based on multi-scale channels and spatio-temporal fusion ( MCSF) on the basis of ConvLSTM. The MCSF model first employs a multi-scale feature extraction module to extract capture regional and edge features from SST data across multiple scales. Then, it utilizes a convolutional block residual attention module to effectively filter the information, followed by a convolutional long-short memory network to extract longtime-dependent features. Finally, a convolutional 2D layer outputs a one-day prediction, with multi-day prediction generated through rolling prediction. The MCSF model was applied to the South China Sea for short-term (5 d) and medium-term (20 d) SST forecasting. Experiments with varying time windows determined optimal input periods of 20 d and 40 d for short- and medium-term forecasts, respectively. Results demonstrate that compared to the ConvLSTM model, the MCSF model reduces the RMSE by 11.9% and 7.6%, and the MAE by 16.7% and 16.9% for the 5^th- and 20^th-day forecasts, respectively. These findings confirm that the proposed MCSF model significantly improves SST forecasting accuracy.