<p>Sea surface temperature (SST) is an essential physical parameter of the ocean and serves as a critical indicator of energy exchange between the sea and the atmosphere. Accurate prediction of SST is vital for protecting the marine ecosystem and providing effective guidance to several industries. Traditional numerical methods are limited in application due to the complexity of the physical models, and general deep learning models struggle to fully utilize the spatial-temporal dynamics inherent in SST data. Recently, diffusion models have been introduced into the field of SST prediction due to their ability to capture remote features naturally. However, they still suffer from numerous parametric quantities in the model and insufficient extraction of spatial features. This paper proposes a lightweight diffusion model (LWDF) for SST prediction. It compresses the number of model parameters while enhancing the capture of spatial correlation to improve the SST prediction performance. Specifically, the LWDF first utilizes the designed Depthwise and Weight-Standardized Conv2d (DWSC) module to reduce the number of parameters. The DWSC performs convolutional operations for each input channel independently and combines weight standardization to smooth the loss for improved computational efficiency. Second, the LWDF performs channel and spatial attention operations sequentially on the dual-path feature maps through the Enhanced Channel Spatial Shuffle (ECSS) residual attention module and introduces channel shuffle at the end of ECSS to enhance the feature interaction. The module further reduces the number of model parameters and improves the ability to represent spatial features of SST data. Finally, we validate the performance of the model through a 7-day SST prediction experiment. The experimental results show that LWDF has the best overall performance compared with other baseline models. It can reduce the number of parameters and size of the diffusion model by 71% and further improve the prediction accuracy.</p>