Sea Surface Temperature Prediction Method Based on Deep Generative Adversarial Network

Sea surface temperature (SST) prediction plays an important role in ocean-related fields. Therefore, it is increasingly important to be able to make more accurate prediction of SST. In this article, we develop a deep generative adversarial network (DGAN) for generating future maps of SSTs, providing a visual method of predicting SSTs. Our DGAN model consists of a generator and a discriminator. The generator is designed to produce more realistic maps of future SSTs, which uses multiple composite layers to capture the changes of SSTs and generates clear maps of future SSTs. The discriminator uses the structure of patchGAN to obtain more SST features, and distinguishes between real and generated SST maps. In addition, we improve the loss function and perform convergence analysis, and then, obtain that minimizing the loss function is equivalent to minimizing Pearson $\chi 2$ divergence, and the relevant explanations are carried out through experiments. The generator and discriminator are training adversarially during the training stage, eventually reaching a relatively balanced state, and the DGAN is able to produce more reliable visual predictions. Finally, the effectiveness of the DGAN in the prediction of SST is verified experimentally, and it is compared with the generative model-DL model and the long short-term memory-GAN model.