OTCFM: A Sea Surface Temperature Prediction Method Integrating Multi-Scale Periodic Features

Sea surface temperature (SST) is a critical factor in the interaction between the ocean and the atmosphere, directly influencing global climate patterns and the dynamic changes in marine ecosystems. Accurate prediction of SST is of great significance for assessing and managing global climate change and maintaining marine ecological balance. However, existing SST prediction methods face challenges such as low accuracy, short prediction periods, and significant errors. This paper proposes an innovative deep learning prediction method, Ocean Temperature Cycle Fusion and Analysis Model (OTCFM), constructed based on datasets from the South China Sea and the East China Sea. This approach aims to accurately capture and predict the cyclical variations and variability in ocean temperature data to provide more precise forecasts of ocean temperatures. Firstly, based on observations of SST’s seasonal and periodic variations, we present a periodic partitioning strategy to decompose complex temperature changes into intra-period and inter-period variations. Secondly, we propose the Ocean Unit to capture both long-term and short-term small-scale changes, moving beyond the inherent attributes of the dataset’s frequency and time domain characteristics to extract intra-period and inter-period feature changes simultaneously. Finally, by stacking the Ocean Units using residual connections, we alleviate the gradient vanishing problem and achieve more accurate long-term and short-term SST predictions. In this study, data from the East China Sea and the South China Sea with different spatial distribution patterns are selected for predictive analysis of the National Oceanic and Atmospheric Administration (NOAA) data from September 1, 1981, to June 7, 2023, with a total of 15,408 data. The experimental results show that OTCFM can accurately capture the evolution patterns of SST data in the spatial and temporal processes under different experimental conditions. The MAE values on the East China Sea and South China Sea SST datasets are improved by 19.08% and 19.52%, respectively, compared with the convolutional long short-term memory neural network (ConvLSTM), which improves the accuracy of the long and short-term prediction of SST time series and has a far-reaching impact on the subsequent promotion of sustainable marine resource management and environmental protection.