Abstract: Sea surface temperature (SST) is a crucial factor in the interaction between the ocean and the atmosphere, controlling the changes of global atmospheric and oceanic ecosystems. Accurately predicting the evolution of SST is of great significance in managing global atmospheric and oceanic systems. To accurately model the spatial autocorrelation of SST data, this paper proposes a deep neural network based on global cross-scale spatiotemporal attention (GCSA-DNN) for predicting SST. The model consists of three parts: a temporal modeling module that extracts temporal dependent features from long time series data, a multi-scale local spatial modeling module that extracts spatial distribution pattern features from SST series mean values, and a global cross-scale spatiotemporal attention fusion module that models the global autocorrelation of each grid point. In this study, the data of the East China Sea and South China Sea with different spatial distribution patterns are selected, and the national oceanic and atmospheric administration (NOAA) data from September 1, 1981 to April 7, 2022 are predicted and analyzed with a total of 14829 data, of which 70% of the data from September 1, 1981 to August 31, 2021 were used for training, 30% for validation, and the data from September 1, 2021 to April 7, 2022 were used for testing. The results show that the model can accurately capture the evolution pattern of SST data in spatial-temporal processes in different experimental conditions, with an accuracy improvement of 14.07% and 14.18% on the East China Sea and South China Sea SST datasets, respectively, compared with the convolutional long short-term memory (ConvLSTM) model, achieving an improvement in SST prediction accuracy.