JCSNN: Joint multi-channel and multi-scale neural networks for sea surface current prediction

Abstract The sea current has significant impact on human production activities and resource development. But most of studies employ traditional spatiotemporal data mining model like LSTM, which requires large amounts of data and is poor at transferability. Therefore, based on the form of two-dimensional seawater motion equation, this paper proposes a neural network model (JCSNN) that combine CNN and self-attention mechanism. The model extracts multi-channel mixed features through three-dimensional convolution and self-attention mechanisms, and calculates the intrinsic associations at different spatial positions. To improve the transferability of JCSNN, we apply improved mean square error as the training loss function, which increases the model’s attention to the complex area. The MSE of the predicted results with the actual data is 0.0114. Compared with the prediction results of VIT, MLP, and LSTM, the MSE is improved by 5.3%, 62.3%, and 6.1%, respectively. The experimental results indicate that the prediction model proposed in this paper has high accuracy and strong practicality, providing a new method for predicting sea surface current. The code for our model and experiments is made publicly available on github: https://github.com/qingxufish/JCSNN.