Aquarius sea surface salinity retrieval in coastal regions based on deep neural networks

A sea surface salinity (SSS) retrieval model for coastal regions (0–200 km) is here proposed using a deep neural network (DNN) which demonstrates the capability of the DNN model to retrieve SSS. The rain filters are applied to the Aquarius observation data and SSS data used in this paper when there is heavy rainfall. Then, Aquarius V5 L2 observation data and hybrid coordinate ocean model SSS (HYCOM SSS) data are used to train the model, and the grid search and 10-fold cross-validation method is used to obtain optimal model parameters. After screening, a DNN model with three hidden layers and 30, 40, and 50 neurons in each layer is constructed. The performance of the proposed DNN model is then estimated and analyzed using independent testing data. We compared the SSS retrieval results with different types of salinity data sources. Compared to in-situ SSS, the retrieved SSS root mean squares error (RMSE) is 0.35 and the mean bias (MB) is −0.01. The RMSE and MB of retrieved SSS are reduced by 0.14 and 0.03 than those of the Aquarius Remote Sensing Systems (Aquarius RSS) SSS, and 0.31 and 0.03 than those of the Aquarius Combined Active-Passive (Aquarius CAP) SSS. Results show that our DNN model had better performance than Aquarius official in most coastal regions. Our results provide a dependable reference for SSS retrievals in global coastal regions. •Coastal SSS of Aquarius is explored using a deep neural network for the first time.•Optimize DNN model parameters to reduce information redundancy and training time.•Grid search and 10-CV are used to determine the network structure of the DNN model.•DNN model improves Aquarius coastal SSS retrievals a lot.•DNN model still has good performance in the area within 100 km of the coastline.