In this study, two sets of SMAP satellite data were utilized, namely daily L3 Tb data and daily SSS product. The L3 Tb data include microwave brightness temperature at both H-pol and V-pol, with a spatial resolution of 36 km. The daily SSS product has a spatial resolution of 40 km, represented as an 8-day running average, which helps to smooth out short-term variability and offers a more consistent dataset for analysis (Meissner et al., 2018). Specific variables of the SMAP data used in this study are H-pol Tb (Tbh), V-pol Tb (Tbv) and SSS in the Yellow and East China Seas (25°N to 37°N and 119°E to 130°E) from January 1, 2016 to December 31, 2020. The Tb data can be publicly accessed from the National Snow and Ice Data Center (NSIDC) at https://nsidc.org/data/spl3smp/versions/8, and the SSS product can be downloaded from the website: https://www.remss.com/missions/smap/salinity/.

The HYCOM reanalysis dataset serves as a reliable reference for SSS retrieval in this study. HYCOM is a data-assimilated hybrid isobaric-sigma-pressure (generalized) coordinate ocean model (Cummings and Smedstad, 2014). It integrates observational data with numerical models to generate high-resolution (1/12°×1/12°) simulations of ocean dynamic processes. The reanalysis dataset has been proven to capture fine-scale features and temporal variations of salinity fields in different ocean regions, making it widely applied in oceanographic research (Ning et al., 2019; Song et al., 2024). Recently, Sam-Khaniani (2022) compared HYCOM SSS with buoy measurements in coastal regions and obtained an average root mean square error (RMSE) of 0.57. This high level of accuracy further confirms that HYCOM reanalysis data is a reliable reference for training and validating the salinity retrieval models. The dataset is available through the HYCOM data server at ftp://ftp.hycom.org/.

The environmental data utilized in this study include Cross-Calibrated Multi-Platform (CCMP) Level 4 (L4) Version 3.0 (V3.0) wind speed data, the Group for High-Resolution Sea Surface Temperature (GHRSST) L4 SST data, and the NOAA Climate Data Record (CDR) of the CPC Morphing Technique (CMORPH) high-resolution global precipitation data.

The CCMP L4 wind dataset provides information on global sea surface wind speed and direction at a 10-meter altitude above sea surface, generated by fusing data from multiple satellite platforms (e.g., QuikSCAT, MetOp) with long-term time series (Mears et al., 2022). The dataset has a spatial resolution of 0.25°×0.25°, and is available at: https://www.remss.com/measurements/ccmp/.

The GHRSST Level 4 product is a SST analysis produced by NASA’s Physical Oceanography Distributed Active Archive Center (PO.DAAC) using wavelets as a basis function and optimal interpolation on a global 0.01°×0.01° grid (Reynolds et al., 2007). The product uses SST data from various instruments including Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E), Advanced Microwave Scanning Radiometer 2 (AMSR-2), Moderate Resolution Imaging Spectroradiometer (MODIS), WindSat, Advanced Very High Resolution Radiometer (AVHRR), and in-situ SST data (Canadian Meteorological Centre, 2017). This dataset can be downloaded from: https://www.ncei.noaa.gov/data/oceans/ghrsst/L4/.

The CMORPH CDR consists of satellite precipitation estimates that are bias-corrected and reprocessed using the Center for Climate Prediction’s Morphing Technology to form a global precipitation analysis. In this paper, we use the daily CMORPH data with a spatial resolution of 0.25°×0.25° (Xie et al., 2019). The dataset is available at: https://www.ncei.noaa.gov/data/cmorph-high-resolution-global-precipitation-estimates/access/daily/0.25deg/.

The in-situ data includes salinity measurements from NSO and Argo float observations. The NSO data, collected in the waters around Korea, features a comprehensive measurement framework with 207 stations, 14 standardized vertical layers, and 25 lines. Measurements of 17 parameters, including water temperature, salinity, meteorological factors, etc., are conducted six times annually (February, April, June, August, October, and December). These data are accessible through the National Ocean Data Center of Korea (https://www.nifs.go.kr/kodc/).

Since its inception in 2000, the Argo program has deployed over 15,000 automated profiling floats globally, accumulating more than 2.15 million temperature and salinity profiles, with some profiles also including biogeochemical parameters. In this study, we use Global Ocean Argo Dataset (V3.0) (Li et al., 2019). The Argo data can be obtained from ftp://ftp.argo.org.cn/pub/ARGO/global/.

As shown in Figure 1 , the integration of NSO and Argo observations helps to capture salinity variations in both coastal and open oceans. In total, approximately 6,000 in-situ data points were acquired. These in-situ datasets, spanning from January 1, 2016, to December 31, 2020, enables an independent evaluation of the performance of the deep learning model, particularly in addressing the challenges of coastal salinity retrieval.

To develop the DNN SSS retrieval model, the following preprocessing was performed on all types of gridded data within the study area from 2016 to 2020. (1) Eliminate outliers in the SMAP Tb dataset affected by sea ice and sea spray based on the quality flags of L3 data. (2) Considering the different spatial resolutions of datasets from multiple sources, all datasets were spatiotemporally matched and then resampled onto a unified 0.25° grid using bilinear interpolation, enabling direct comparisons and analysis. (3) Grid points with an instantaneous rainfall rate exceeding 0.15 mm/h were excluded, since heavy rainfall significantly influences the brightness temperature observed by microwave radiometers. (4) Calculate the shortest offshore distance of grid points based on latitude and longitude. (5) Remove all NaN values. (6) Normalize the input values including Tb and environmental data to standardize the feature range by transforming the data to have a mean of 0 and a standard deviation of 1, which is a critical step in ensuring consistency across different datasets and preventing performance biases caused by varying input feature scales.

Finally, we obtained a total of 996,063 data samples. These data were divided into two parts: the data from 2016 to 2019 were utilized for training and validation, and that in 2020 were used as independent test set. To further evaluate the performance of the DNN framework, in-situ measurements were collocated with the model retrieval results and SMAP SSS products using strict temporal and spatial criteria. Specifically, we used a 12 h time window and a 25 km spatial search radius. This matching process results in a total of 1,540 collocated samples.

The architecture of the DNN model (Yuan et al., 2020) for SSS retrieval is depicted in Figure 2 , which consists of an input layer, three hidden layers, and an output layer. The performance of a DNN is significantly influenced by several factors, including the number of hidden layers, the number of neurons in each hidden layer, the learning rate, the choice of optimizer, and the activation function (Liu et al., 2017; Zhang et al., 2023). To identify the optimal set of network parameters, a series of sensitivity experiments were conducted by evaluating the model performance under various configurations: 1) The number of neurons in the three hidden layers was tested within the ranges of 64 to 1024, 32 to 256, and 32 to 128, for the first, second and third layers, respectively. 2) The batch size varied between 256 and 2048. 3) The learning rate was randomly sampled between 0.0001 and 0.01. 4) The dropout rate was set between 0 and 0.3. 5) The optimizers tested include SGD, RMSprop, and Adam, both with and without batch normalization.

The network architecture uses the ReLU activation function and mean squared error (MSE) loss function. The remaining hyperparameters are optimized using a combination of grid search and random search. A total of 100 grid searches were performed, with an epoch setting of 500. The use of random search is specifically designed to efficiently explore the continuous hyperparameter space, particularly for parameters such as the learning rate by sampling values within a predefined interval rather than testing every discrete option in great detail through grid search. This approach, in combination with grid search, allows us to reduce computational costs while ensuring a comprehensive search for high-performance configurations. The results reveal that the optimal network parameters consist of 1024, 32, and 32 neurons, and dropout rates of 0.3, 0, and 0 for the three hidden layers, respectively, a batch size of 512, a learning rate of 0.008, and the Adam optimizer without batch normalization. Meanwhile, the DNN model was trained using K-Fold Cross Validation (Nti et al., 2021) to enhance its generalization ability.

During the development of the DNN model, the SHAP method (Lundberg and Lee, 2017) was employed to evaluate the relative importance of input parameters in SSS estimation. By calculating the Shapley value for each parameter, this approach effectively addresses the limitations of traditional feature importance methods, which often rely on linear assumptions. Specifically, the SHAP method is capable of accurately capturing the nonlinear interactions and coupling effects among multidimensional parameters.

We first conducted SHAP analysis for the entire study area and time period. As illustrated in Figure 3 , the importance scores provide a clear representation of the contribution of each input parameter to the model performance. The significance of Tbv is evident, as it plays a pivotal role in revealing sea surface emissivity changes driven by salinity (Entekhabi et al., 2010). While the overall importance of Tbh is relatively low, possibly because it is generally less sensitive to SSS-induced emissivity changes than Tbv (Song and Wang, 2017). This may also explain why v-h ranks between Tbv and Tbh. But further analysis in different regions and seasons (Section 5) shows that the contribution of Tbh is more significant in offshore areas (50 km<dis ≤ 200 km) or in summer and autumn. Lat and lon rank among the top four, reflecting the spatial pattern in SSS, especially its more pronounced meridional gradient (Reul et al., 2014a). Theta also contributes a lot, which reflects the temporal variations of SSS to some extent. The overall importance of distance (dis) is relatively low, possibly because lat/lon features already implicitly contains distance information, thereby reducing the model’s need for explicit distance-based corrections. But it still shows greater impact in nearshore areas (dis<50 km) ( Figure 4a ), where freshwater input and land-sea interactions strongly affect salinity patterns. As for the environmental factors, the SHAP analysis results reveals the substantial contribution of SST, reflecting the thermodynamic interactions between temperature and salinity (Alory et al., 2012; Hernandez et al., 2014; Reul et al., 2014b; Umbert et al., 2015). However, the effects of wind speed and direction are much weaker, possibly because the linear combination of vertical and horizontal polarization brightness temperatures (v-h) introduced in this study already implies a relatively strong wind direction signal, and thus reduces the importance of wind-related information (Han and Hong, 2023; Soisuvarn et al., 2007). Rainfall exhibits a negligible effect, which should be due to the data preprocessing step excluding rain-contaminated data to mitigate the influence of RFI (Ouyang et al., 2023).

Incorporating all parameters into DNN may lead to increased training time. Moreover, the inclusion of redundant information may potentially reduce the model performance. Based on the results of the SHAP analysis, this study designed a series of feature combination experiments to evaluate the cumulative effects of parameter contributions and identify the optimal set of input parameters. All sensitivity experiments were conducted throughout the entire study area and time period, using consistent three-layer DNN architecture based on the previously optimized network configuration, which ensures that any variations in performance could be attributed to differences in input parameters.

As presented in Table 1 , the first stage involves the construction of four DNN models: 1) Model 1 (one parameter) uses only the primary feature Tbv as the control group. 2) Model 2 (two parameters) examines the impact of incorporating the marine environment factor, SST. 3) Model 3 (three parameters) assesses the spatial representation capability of geographic coordinates (lat and lon). 4) Model 4A or 4B (four parameters) further quantifies the combined effects of geographic coordinates and SST or dis, respectively. In the subsequent stages, from Model 5 to Model 9, the next most important parameter identified by the SHAP analysis was sequentially incorporated. This approach allows for a detailed evaluation of the incremental contributions of each parameter to model performance.

The performance of the DNN model was evaluated using three key metrics: the RMSE, bias, and Pearson’s correlation coefficient (CORR). Additionally, to highlight the advantages of the DNN model, we further conducted comparative analysis with three other machine learning models, namely K-Nearest Neighbors (KNN) (Taunk et al., 2019), Random Forest (RF) (Breiman, 2001), and Extreme Gradient Boosting (XGBoost) (Chen and Guestrin, 2016).

Figure 5 illustrates the performance of each baseline DNN model listed in Table 1 . The RMSE between model estimations and HYCOM values is 1.27 when using only Tbv as input in Model 1, indicating that the radiation signal from SMAP contains important information on sea surface salinity. Additional inclusion of SST in Model 2 improves model accuracy, albeit to a limited extent. However, the incorporation of latitude and longitude in Model 3 significantly reduces the RMSE by 33% and increases the CORR to 0.87. Comparing Models 4A and 4B with Model 3, further adding SST or dis exhibits a small contribution to model performance. This suggests that geographic information effectively reflects the inherent regional characteristic of SSS.

As the number of input parameters continues to increase, the accuracy of salinity estimation gradually improves. Model 7, which combines the first seven most important input parameters-Tbv, lat, SST, lon, theta, v-h, and dis, achieves the highest accuracy, with RMSE of 0.69, bias of -0.11 and CORR of 0.91. Further inclusion of wind speed (ws) information or Tbh in Models 8 and 9 produces salinity estimation results comparable to Model 7, indicating that high-accuracy SSS estimation can be achieved without inputting these parameters. This minimizes data redundancy while retaining the essential physical factors. The results also confirm the SHAP analysis findings that the contribution of wind or Tbh to the baseline DNN model is not significant. Consequently, the seven parameters in Model 7 were selected as the optimal input set. When comparing DNN with other machine learning models in the next section, all models use the same set of input parameters.

Table 2 shows the RMSE, bias and CORR values between HYCOM reanalysis and sea surface salinity derived from the baseline DNN, KNN, RF, and XGBoost models. To ensure optimal performance of each model, all models underwent hyperparameter optimization using grid search, with parameter spaces customized for each algorithm. Specifically, for RF and XGBoost, key parameters such as tree depths, feature selection methods, and sample splitting criteria were adjusted to enhance accuracy. For KNN, the number of neighbors, search algorithms, distance weighting method, and leaf size were optimized to ensure robust retrieval while balancing model complexity. In general, the independent validation using HYCOM data from 2021 demonstrates that RF, XGBoost, and DNN achieve comparable levels of high accuracy. These models significantly outperform KNN with RMSE approximately 17% lower.

To further investigate the robustness of the models, we compared model estimations against in-situ measurements. The accuracy of SMAP salinity product was also evaluated based on the same dataset. As shown in Table 3 and Figure 6 , when compared with in-situ observations, the tree-based models (RF and XGBoost) exhibit degraded performance and is inferior to that of the SMAP products, with the maximum RMSE value exceeding 1.05 and minimum CORR value dropping to 0.55. In contrast, the DNN model remains high performance. The RMSE is 0.78, which is 38.6% lower than that of XGBoost and also the lowest among all evaluated models. Compared to the SMAP products, the RMSE has decreased by 25.7%, further highlighting its superior performance.

The DNN model’s enhanced performance can be attributed to its ability to learn hierarchical representations from SMAP brightness temperature and ancillary parameters, effectively capturing the non-linear interactions between oceanographic or geographic factors and salinity changes. Additionally, the DNN model achieves a CORR value of 0.78, slightly higher than that of the SMAP products, indicating a strong linear relationship with in-situ measurements. This reveals the high ability and robustness of the DNN model to learn complex feature interactions.

Table 4 compares DNN model estimation results and SMAP salinity products with in-situ measurements in the nearshore region less than 50 km away from land and offshore region 50 km to 200 km away from land, respectively. Obviously, the baseline DNN exhibits varying performance in different regions. In the nearshore region, compared to the SMAP products, the RMSE (0.88) of DNN derived SSS has significantly decreased by 33.8%. This indicates that the DNN model is more effective in capturing coastal salinity changes, where factors such as coastal processes and river runoff significantly affect the magnitude and spatiotemporal distribution of salinity. In this region, the relatively low accuracy of SMAP products is likely due to noise and interference from land-based sources.

In offshore areas, the accuracy of both DNN estimation and SMAP products has been greatly improved. Especially for the SMAP products, the RMSE has decreased from 1.33 in coastal areas to 0.73. On the one hand, the impact of RFI is much smaller here. On the other hand, the salinity patterns of offshore waters become more stable and consistent. Compared with the SMAP products, the DNN estimation still maintains slightly higher accuracy, which is likely due to its superior ability to learn and generalize complex patterns and relationships in data more effectively than traditional methods.

Due to the different characteristics of SSS in nearshore and nearshore regions, can we further improve the accuracy of salinity retrieval by establishing separate regional models? To address this issue, input parameter importance analysis and sensitivity experiments described in Section 3.2 were conducted in each region. As shown in the SHAP analysis results in Figure 4 , key factors such as Tbv and geographic coordinates still maintain high importance scores in both regions, although their rankings are slightly different. Compared with Figure 3 , the role of the distance is more prominent in nearshore areas, which may reflect the more significant impact of coastal processes on salinity and the complex interactions between land and sea. In offshore regions, its impact is greatly reduced, but subsequent sensitivity experiments show that incorporating it into the model input can substantially reduce salinity retrieval errors. In addition, Tbh emerges as an important factor in offshore areas. Based on these findings, optimal input parameters were determined through sensitivity assessment of nearshore and offshore DNN models. For DNN-nearshore, the optimal input parameters are the same as the baseline DNN model, i.e., Lon, Lat, Tbv, dis, v-h, SST and theta. When considering Tbh additionally, DNN-offshore demonstrates the best performance.

Figures 7 , 8 compare the performance of the baseline DNN, DNN-nearshore and DNN-offshore models, as well as the quality of SMAP products in different regions. Overall, both two regional models outperform the baseline DNN in terms of RMSE and correlation coefficient. In the nearshore region, DNN-nearshore derived RMSE is reduced to 0.81, about 8.0% lower than that of baseline DNN, and the correlation coefficient increases by approximately 12.5% from 0.72 to 0.81. In offshore areas, DNN-nearshore achieves a RMSE of 0.64, 5.9% smaller than that of baseline DNN, and the correlation coefficient is as high as 0.90. On the contrary, the bias of the two regional models has increased, but still less than 0.3. These results demonstrate that regional modeling helps to further improve the accuracy of salinity retrieval. This may be attributed to the fact that the regional sample data more accurately reflect the spatiotemporal distribution characteristics of salinity in each area, allowing the model to focus more specifically on the unique data features.

This section further evaluates the robustness of two regional models by analyzing their performances in different seasons. As shown in Figure 9 , the RMSE of both models and SMAP products show some seasonal variation, but the magnitude of change is much smaller in the offshore region, which may be related to the relatively more stable distribution of salinity in this area. In nearshore areas, compared to SMAP salinity products, DNN-nearshore significantly reduces salinity estimation error in each season, especially in spring and winter, with the RMSE 60.3% and 53.6% lower than that of SMAP products, respectively. In offshore, the model error shows a more stable seasonal variation trend, with values ranging from 0.5 and 0.7, also much smaller than that of SMAP products in spring and winter. This indicates that the regional salinity retrieval models can accurately capture changes in coastal salinity over time.

Considering that the performance of DNN-nearshore is slightly lower than that of DNN-offshore, and nearshore salinity often exhibits more pronounced seasonal variations due to the influence of dynamic processes or estuarine runoff. In nearshore areas, we attempt to investigate whether seasonal adaptive models can further improve salinity retrieval accuracy. SHAP analysis reveals that the top ten input parameters remain consistent across all four seasons, although their rankings vary slightly. Among them, Tbv and geographic information (lat and lon) consistently rank in the top four, highlighting their fundamental role in salinity retrieval, and dis remains a key parameter, reflecting its crucial impact on the spatial variation of nearshore salinity caused by land-sea interactions. In summer and autumn, the importance of Tbh rises, indicating an increased correlation with salinity during these periods, which may be due to seasonal fluctuations in ocean thermal characteristics and radiation signals. Sensitivity experiments further determine the optimal input parameters for each seasonal model (DNN-nearshore-season). During spring and winter, the selected input parameters align with those of DNN-nearshore or baseline DNN. In summer and autumn, the seasonal models incorporate one additional parameter Tbh.

Figure 10 compares the performance between four seasonal DNN models (DNN-nearshore-season) and the non-seasonal nearshore model (DNN-nearshore). The results reveal that accounting for seasonal variations significantly enhances salinity retrieval accuracy across all seasons. Compared to the non-seasonal DNN-nearshore model, the average RMSE is reduced by 12.3% from 0.81 to 0.71, and the most substantial improvements in accuracy appear in summer and autumn, with RMSE decreasing by 13% and 18%, respectively. This suggests that the seasonal modeling approach effectively captures temporal variations in the relationship between input parameters and nearshore salinity. By dynamically adjusting the weights of input features across different seasons, the DNN-nearshore-season models achieve superior performance, highlighting the importance of incorporating seasonal dynamics in salinity estimation.