Abstract: Firstly, we selected the world ocean atlas 2013 (WOA13) ocean data  as the experimental data, the unequal distance differential method and vertical gradient method were applied to the preprocessing of ocean data, the division of ocean area and analysis of thermocline. Through the performance analysis of various neural networks based on the threedimensional WOA13 ocean data in the binary classification experiment, we chose the residual network as the network model of the binary classification experiment, and added the Dropout retention layer on the basis of the threelayer residual network model to prevent over-fitting. Secondly, the residual network model was used for thermocline analysis and determination, and the comparative tests such as the super parameters optimization, the residual unit improvement and  the retention rate adjustment were carried out for the improved model. The experimental results show that the improved ResNet26 network is effective for the thermocline data classifica
tion of WOA13 ocean area data, and the classification accuracy is more than 94%.

Key words: convolutional neural networks (CNN）, residual network, thermocline, WOA13 data