This study constructs a time-dependent model to predict the nighttime suspended sediment concentration near Wenzhou based on the convolutional neural network U-Net, which integrates the high-resolution Delft3D (version 4.03.01) hydrodynamic model and GOCI satellite observation data. The model’s prediction accuracy is significantly improved by replacing the original tide level with the tide level variation and increasing the temporal resolution of the flow field to 15 min via sensitivity analysis of the model’s input parameters. The validation results show that the model can maintain high consistency with GOCI observations in short-term prediction, with a structural similarity index (SSIM) of 0.82. For multi-hour continuous nighttime predictions, while quantitative uncertainty increases with the forecast horizon, the model successfully captures the spatial evolution patterns and maintains stable structural characteristics. The model effectively provides missing remote sensing nighttime observations as well as a new method for full-cycle prediction of nearshore SSC.
Zhang et al. (Thu,) studied this question.