Sea Surface Temperature Prediction Based on Frequency Domain Amplitude Awareness and Temporal Semantic Representation

Abstract

Sea surface temperature (SST) is one of the important parameters in marine environments. Accurate prediction of SST plays an important role in marine ecological environment protection and marine disaster prevention and mitigation. The nonstationarity of the SST series makes accurate prediction challenging. To address this problem, we propose an SST prediction model based on frequency domain amplitude awareness (FDAA) and temporal semantic representation (PST). First, the proposed FDAA module transforms the SST series into the frequency domain and adaptively enhances the periodic components that are strongly correlated with SST prediction while suppressing nonstationary interference via a learnable gating mechanism, thereby effectively mitigating the interference of nonstationarity. Then, the prediction model is constructed by combining the bidirectional long short-term memory (LSTM) with strong temporal feature extraction capability and the multihead attention with strong temporal semantic PST capability. Finally, the 15-day SST sequence is fed into the prediction model, which outputs the SST for the next five days. Extensive experiments are conducted to demonstrate the effectiveness of our proposed method. Experimental results demonstrate that our proposed method achieves competitive prediction accuracy, effectively mitigates the effects of nonstationarity in SST, and yields accurate predictions of future SST.