To train the CNN model, we used sea surface data from multi-satellite observations as input data, such as sea surface temperature (SST), sea level anomaly (SLA), sea surface salinity (SSS), and sea surface wind (SSW), combined with geographical information (longitude and latitude). We used SST data from the operational SST and sea ice analysis data (OSTIA L4 SST), obtained from the NASA Physical Oceanography Distributed Active Archive Center (PO.DAAC) (https://podaac.jpl.nasa.gov/dataset/OSTIA-UKMO-L4-G LOB-v2.0). The data are based on a combination of satellite and in-situ measurements. They are currently available from 2007 to the present day, with a daily temporal resolution and a spatial resolution of 0.05° × 0.05°. The altimeter satellite gridded SLA downloaded from the CMEMS was combined with various altimeter missions. The SLA was computed for a twenty-year (1993–2012) mean. The dataset had a spatial resolution of 0.25° × 0.25° and a daily temporal resolution (https://data.marine.copernicus.eu/product/ SEALEVEL_GLO_PH Y_L4_MY_008 _047/services). The SSW data, specifically the eastward (U-wind) and northward (V-wind) components of the 10 m wind datasets were provided from the European Center for Medium-range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5), which have a spatial resolution of 0.25° × 0.25° and an hourly temporal resolution (https://www.ecmwf.int/en/forecasts/dataset/ecmwf-reanalysis-v5). The SSS data were obtained from the Soil Moisture Active Passive (SMAP) satellite. The SMAP level 3 product has a 25 km spatial resolution and a daily temporal resolution (8-day running average) (https://podaac.jpl.nasa.gov/SMAP). The dataset covers the period from April 2015 to present, except from June 19th to July 26th, 2019, during which there were missing values due to a safe mode event that caused all instruments to shut down.

To investigate the influence of precipitation on CDW volume, we used an integrated multi-satellite retrieve for GPM (IMERG) daily final run (GPM_3IMERGDF) product, with a daily resolution of 0.1° × 0.1° (https://gpm.nasa.gov/data/imerg). The IMERG combines information from the global precipitation measurement (GPM) satellite, microwave satellite, and gauge observations to estimate precipitation over most of the Earth’s surface. It has an advantage in oceans without ground level precipitation-measuring instruments. All sea surface products (SST, SLA, SSS, and SSW) were resampled into 25 km spatial and daily temporal resolutions for model training. This study focuses on the summer period (May to September) from 2015 to 2021, specifically targeting the middle ECS area (119–131°E, 29–37°N) to investigate the effect of CDW volume changes.

The National Institute of Fisheries Science (NIFS) has been conducting serial hydrographic cruises around the Korean Peninsula four to six times annually. The datasets are collected using conductivity-temperature-depth (CTD) sensors at each station for specific depths, such as 0, 10, and 20 m, etc. The observation stations consist of 207 points in 25 lines. In this study, we used 138 points in 17 lines, excluding the East Sea (Japan Sea), because we focused on the CDW’s effects on the ECS. The CDW exists at a depth of approximately 0–20 m during summer (Lie et al., 2003; Moon et al., 2019; Hong et al., 2022; Zhu et al., 2022).

Table 1 shows the number of in-situ salinity measurements recorded by the NIFS at three different depths (0, 10, and 20 m) from 2015 to 2021. Of the total 1,892 data points, 284 (15.01%) were identified as CDW (salinity< 31 psu) at a depth of 10 m, while only 62 (3.28%) were identified as CDW at a depth of 20 m. Therefore, the salinity at a depth of 10 m (S_10m) was considered more suitable for the CDW compared to 20 m. Of the 1,892 data points, we used 1,310 in-situ salinity data points at 10 m, matching the input data as ground truth data for model training.

Figure 1 shows the location of each station used for this study. Other validation datasets for S_10m were collected from the Ieodo Ocean Research Station (I-ORS) and the Korea Meteorological Administration (KMA). The geographical location of the I-ORS (32.07° N, 125.10° E) makes it an ideal observation site for monitoring the expansion of low salinity water from the Changjiang River. The I-ORS estimates salinity at various depths, including 10 m. Therefore, we used the I-ORS S_10m from 2020 to validate the model performance. The KMA provided CTD observation data around the Yellow Sea and ECS. Since that area plays an important role in meteorology and climatology, serial observations using shipboard were conducted every two months since January 2016 to observe changes in the marine environment; thus, we used summertime data (May to September) from 2016 to 2021. The observation point was selected to monitor the influence of the Yellow Sea Bottom Cold Water and CDW.

In addition, to investigate the effect of Changjiang River discharge (CRD) on CDW volume, the daily flow rate of the Changjiang River, measured by the Datong Station, was collected from 2015 to 2021 (www.cjh.com.cn). The Datong station is approximately 624 km from the Changjiang River Estuary. It is the first hydrometric station in the mainstream of the Changjiang River to the estuary and the uppermost boundary of the ocean tide. The discharge at the Datong hydrometric station generally represents the CRD streaming down to the ECS (Erfeng et al., 2003). Table 2 summarizes the datasets used in this study.

In this study, we developed a CNN model to estimate the S_10m. The CNN method has been demonstrated to be superior to other deep learning methods in remote sensing applications, owing to its ability to capture spatial feature information and achieve high accuracy in spatial distribution (Yamashita et al., 2018; Meng et al, 2021a; Zuo et al., 2021; Shin et al., 2022). We constructed a CNN structure for feature extraction with one input layer, three convolution layers, three max pooling layers, three dropout layers, and one regression layer as the output layer. Rectified linear unit (ReLU) layers were used as activation functions ( Figure 2 ).

For the training dataset of CNN model, we generated patch pairs between sea surface parameters and geographical factors (latitude and longitude) as input and the corresponding in-situ S_10m as the ground truth data. Each patch was generated with a size of 8 × 8 × 7 pixels utilizing the nearest pixels of the satellite gridded dataset (25 km at daily) extracted from the corresponding in-situ salinity. Patches that did not fill at least half of the 8 × 8 pixels were excluded from the training set. The z-score standardization method was applied to process the sea surface data to weight all variables equally. A total of 1,310 coupled patches of in-situ salinity and multi-satellite data were randomly divided into two groups for training and test. The proportions of patch pairs for the training and the validation were 70 and 30%, respectively. Additionally, in-situ data sources (i.e., I-ORS and KMA) were used to evaluate the effects of different data sources on model performance. We evaluated the performance of the CNN model with in-situ salinity using various statistical values such as the determined coefficient (R²), root mean square error (RMSE), and relative RMSE (RRMSE).

We used the SHAP method to investigate the contribution between sea surface parameters and S_10m. The SHAP is one of the eXplainable Artificial Intelligence (XAI) methods developed to interpret complex black-box, artificial intelligence models. The SHAP values quantify the impact of each input feature on the model output, explaining the individual predictions of how much each input contributes to the prediction. It has the advantage of providing not only the relative importance of each input variable but also the positive or negative relationship on the output (Lundberg and Lee, 2017; Mangalathu et al., 2020; Tian et al., 2022). SHAP has been widely used in machine learning in marine science to interpret nonlinear or indirect relationships between input and output variables. It also helpful for examining complex interactions between multiple variables in such systems (Jang et al., 2021; Jang et al., 2022; Tian et al., 2022). The SHAP value was calculated using the following equation (Eq. 1):

where z indicates the subset of input parameters, N is all input parameters, M is the number of input parameters, f z ∪   { i } ( x z ∪   { i } )   is output with i^th parameter, and f z ( x z ) is output without i^th parameter (Lundberg and Lee, 2017; Mangalathu et al., 2020; Jang et al., 2021). We can better understand the relationships between input and output by comparing the SHAP value for each input.

The CDW volume (V_CDW ) was calculated using Eq. 2:

where d is the depth (here d = 10 m), A(z) is CDW area (defined as< 31 psu). We used SMAP data for the A(0) and the results of the CNN model for the A(10). The density-based spatial clustering of applications with noise (DBSCAN) method was used to isolate the area (A(z)) of CDW in gridded data. DBSCAN is a data clustering algorithm that determines the noise points or outliers and detects dense spatial points (Ester et al., 1996). We applied this method to remove outliers and extract the A(z) at each depth. To calculate the V_CDW , we used the alpha shape method, a computational geometry algorithm that creates a bounding area or volume around a given set of points (Edelsbrunner and Mücke, 1994). Since the Changjiang River plume spreads eastwards over the broad area of the ECS, reaching as far as Jeju Island, we estimated the V_CDW over a limited area affected by CDW (119–128°E, 29–35°N).

We analyzed the relationship between the CDW volume and CRD, which is the main controlling factor for CDW ( Figure 9 ) (Beardsley et al., 1985; Lie et al., 2003; Chen et al., 2008; Siswanto et al., 2008). In spring (May to early June), the CDW volume was close to zero and peaked in July as the CRD increased. It should be noted that there was a time lag of 34 ± 15 days between the CDW volume and CRD. This is consistent with previous studies demonstrating that the CDW extends eastward in June with increasing CRD and reaches Jeju Island after 1–2 months (Chen et al., 2008; Kim et al., 2009; Son and Choi, 2022). Moreover, the years with the maximum volume in July (2016, 2017, 2018, and 2019) had an average time lag of 24.3 days, while the years with the maximum volume in August (2015, 2020, and 2021) had a longer time lag of 46.3 days. It does seem to depend on different mechanisms affecting the movement of CDW, such as wind and ocean currents. Particularly in 2021, CRD appeared to be different from other years, and it seemed to be artificially adjusted. It can be occurred by the impact of the Three Gorge Dam.

We conducted a comparison between the monthly cumulative sum of CRD (Q_CRD) and the monthly average volume to examine the variation in CDW volume with respect to the inflow of CDW ( Table 3 ). Q_CRD revealed an increasing trend and reached its maximum (9.73×10¹⁰ m³) in July and then decreased from August. With a 1,000-times difference in units between CDW volume and Q_CRD, we converted the values into a range between 0 and 1 using minimum-maximum (min-max) scaling and conducted regression analysis. Figure 10A displays the scatter plot for the entire period. It was observed that CDW volume and CRD have a positive correlation (Beardsley et al., 1985; Lie et al., 2003; Chen et al., 2008; Siswanto et al., 2008; Bai et al., 2014). We conducted the monthly regression coefficients to compare the impact of each month ( Figure 10B ). The influence of Q_CRD on volume increased over time, with a maximum impact (1.32) in August. This may be due to the one-month time lag from when CRD reached its maximum in July to the study area. It suggested that the rapid decrease in Q_CRD after July contributed significantly to the decrease in CDW volume after August.

Another factor, precipitation was calculated using Eq. 3:

where A is CDW area, which is the region with SSS< 31 psu, and P represents the precipitation that falls in area A. Precipitation immediately enters the ocean; thus, the inflow of precipitation to the CDW area was calculated as the monthly cumulative sum of precipitation (Q_PRE). As shown in Table 3 , the highest values of Q_PRE were observed in August (5.65×10⁷ m³) which can be attributed to heavy precipitation caused by typhoons. During the entire study period, 12 typhoons passed through the area, and 10 of them occurred in August and September (reported by KMA). To investigate the influence of Q_PRE on volume fluctuations, we applied min-max scaling and conducted a regression analysis ( Figure 10C ). In Figure 10D , the monthly contribution was highest in July, with a coefficient of 1.14, followed by August, with 1.01. From the results, it is clear that the precipitation, particularly during the monsoon season, significantly affects CDW volume fluctuations.

In summary, while the monthly trends showed a one-month time lag between CRD and CDW volume changes, precipitation had a more immediate effect on CDW volume due to its direct entry into the ocean ( Table 3 ). Regression analysis determined the extent to which each factor affected CDW volume. By comparing monthly regression coefficients, the CRD had the most significant influence in August, while precipitation had the greatest impact in July. These results demonstrated that the influence of CRD and precipitation on CDW volume variation differs from month to month. Furthermore, the Q_CRD was 1,000 times larger than the Q_PRE. However, the study only measured precipitation within the CDW region and did not consider horizontal dispersion, which may have led to an underestimation of the influence of freshwater input. To obtain a more comprehensive understanding, further investigation into the impact of freshwater input and additional factors is required.

When the highest CRD was recorded in 2020, we compared the I-ORS S_10m with the CNN S_10m ( Figure 11A ). The variation patterns of S_10m (blue lines) were similar for the entire period, and a rapid decrease in S_10m occurred in August 2020 due to a large amount of freshwater flowing into the ECS. It was confirmed that the CDW volume (orange line) also increased accordingly. To identify the effect of CDW near the I-ORS, the S_10m spatial distributions were generated before the salinity decrease, during the intrusion of CDW, and after salinity recovery following Typhoon Bavi ( Figures 11B–D ). During August 24–26 (shown as gray shading in Figure 11A ), Typhoon Bavi moved northward and passed through near the I-ORS as a super typhoon with a strong wind speed (maximum of 66.1 m/s reported by the KMA). A rapid increase in salinity occurred in SSS (8 psu reported by KMA) and S_10m (2.61 psu) after the typhoon passed on August 26. Regarding CDW volume, an extreme decrease (0.77×10¹² m³) occurred during the same period. This result demonstrated that extreme vertical mixing induced by the typhoon as reported by Hong et al. (2022). In addition, it should be noted that CDW was eliminated and divided into two patches along the typhoon track ( Figure 11D ). Lee et al. (2017) suggested that the impact of typhoons inhibited the spread of CDW and induced vertical mixing, preventing its persistence. Therefore, although the CRD was highest in 2020, the relatively small CDW volume compared to 2016 ( Figure 9 ) could be due to the dissipation of CDW caused by the typhoon.