Abstract:Sea surface temperature (SST) is an important parameter for ocean and atmospheric forecasting systems and climate change research. The National Satellite Meteorological Centre (NSMC) develops the Fengyun-4A (FY-4A)/AGRI (advanced geostationary radiation imager) SST products using the split-window nonlinear SST (NLSST) algorithm. However, the traditional regression algorithm is difficult to meet the needs of higher accuracy SST retrieval. To solve this problem, this paper proposes a FY-4A/AGRI sea surface temperature retrieval method based on deep learning, aiming to improve the retrieval accuracy of SST and provide more accurate data support for meteorological research. FY-4A/AGRI satellite data, SST climatology data, and in situ SST observations are used to construct the retrieval dataset according to quality control standards and spatio-temporal matching rules. The NLSST algorithm is used to select features, including 10.7 μm band brightness temperature, 12 μm band brightness temperature, satellite zenith angle, and SST climatology data. According to the ratio of 8∶2, the feature data are divided into a training dataset and a validation dataset, which are used for training and validation respectively. A SST retrieval model based on a deep neural network is obtained through experiments. Finally, the FY-4A/AGRI satellite data are retrieved by the DNN model to generate SST products. The model-retrieved SST products are evaluated from two dimensions of accuracy and long-term series stability based on in situ SST, and also compared with the FY-4A/AGRI official SST products. By applying the quality levels of FY-4A/AGRI official SST products to the model-retrieved SST products, the performance of model-retrieved SST products under different quality levels (excellent, good, and bad) in three periods of day, night, and dawn is analysed. The statistical results show that when the quality level is excellent, the mean bias of the model-retrieved SST products is -0.19 ℃, the root mean square error (RMSE) is 0.67 ℃, and the correlation coefficient reaches 0.992. However, the mean bias of FY-4A/AGRI official SST products is -0.49 ℃, the RMSE is 0.99 ℃, and the correlation coefficient is 0.985. The mean bias and RMSE of the model-retrieved SST products are 0.3 ℃ smaller than those of the FY-4A/AGRI official SST products, and the correlation coefficient also indicates a good correlation between the model-retrieved SST products and in situ SST. In addition, the temporal stability of the model-retrieved SST products over an extended period outperforms that of the FY-4A/AGRI official SST products. This study provides a new approach for the SST retrieval from the next-generation geostationary satellite.
