High precision lightweight prediction of short term sea surface temperature in the East China Sea: TFLinear model

Sea surface temperature (SST) serves as a critical indicator for assessing marine ecosystem health. Given the increasing human exploitation of marine resources, accurate SST prediction has garnered significant attention. While existing neural network based approaches effectively capture spatio temporal dependencies within SST data, they often suffer from high computational complexity. To address this, we propose TFLinear, a lightweight SST prediction model that incorporates a novel Residual Temporal Frequency (RTF) module which combines residual linking, depthwise separable convolution, and fast Fourier transform (FFT) into the DLinear framework. The method operates in three key stages: spatial feature extraction via depthwise separable convolution, time frequency decoupling of SST sequences using FFT to isolate trend, seasonal, and transient components, and multi step prediction through dedicated linear channels followed by component wise fusion. We evaluated TFLinear using OSTIA SST data from the East China Sea for 1 to 10 days forecasts, comparing it against state of the art benchmarks. Results show that TFLinear achieves superior performance in MAE, RMSE, and R², with improvements of 7.9% to 23%, while maintaining significantly lower computational cost — demonstrating strong potential for efficient and accurate SST forecasting in practical scenarios.