Machine Learning to Identify Three Types of Oceanic Fronts Associated with the Changjiang Diluted Water in the East China Sea between 1997 and 2021

: Long-term sea surface salinity (SSS) in the East China Sea (ECS) was estimated based on Ocean Color Climate Change Initiative (OC-CCI) data using machine learning during the summer season (June to September) from 1997 to 2021. Changjiang diluted water (CDW) in the ECS propagates northeastward and forms longitudinally-oriented ocean fronts. To determine the CDW’s distribution, three fronts were investigated: (1) a CDW front based on chlorophyll-a concentration (Chl), SSS, and sea surface temperature (SST); (2) a CDW front based on sea surface density (SSD); and (3) a CDW front for nutrient distribution. The Chl fronts matched well with the SSS fronts, suggesting that Chl variation in the ECS is highly correlated with the CDW. Furthermore, the SSD fronts spatially matched well with nitrogen concentration. Sea level anomaly (SLA) variation with SSD was also detected, indicating that CDW had sufﬁciently large effects on SLA so that they may be detectable by altimeter measurements. This result suggests that the inﬂuence of steric height changes and the inﬂow from rivers are signiﬁcant in the ECS. Additionally, the continuous long-term SSD developed in this study enables researchers to detect the CDW front and its inﬂuence on the ECS marine environment.


Introduction
Because Changjiang diluted water (CDW) includes freshwater, nutrients, and sediments, understanding the variations in CDW systems is crucial for analyzing material transport and biogeochemical processes in the East China Sea (ECS) [1][2][3]. CDW detection has been attemptI confirmed using satellite measurements of sea surface salinity (SSS), chlorophyll-a concentration (Chl), and turbidity; however, studies on long-term SSS distribution patterns are scarce, possibly because of limitations in continuous measurements [2,4].
Several recent studies have demonstrated the potential of ocean color satellite remote sensing for estimating the SSS in the ECS [2,5,6]. This is because CDW contains colored dissolved solids or pollutant matter that reflects light at specific wavelengths, such as colored dissolved organic matter (CDOM) and terrestrial particles; the optical feature of CDW is indirectly related to SSS. Based on this principle, Kim et al. [6] estimated the SSS using Geostationary Ocean Color Imager (GOCI, 500 m resolution) data from 2015 to 2020. The application of the same approach also detected SSS variation due to river discharge near the estuaries of the mid-Atlantic [7] and Gulf of Mexico [8], suggesting that SSS variation may be monitored globally using ocean color satellite sensors.
Kim et al. [4] defined CDW based on statistical analysis with the K-means clustering technique using Sea-viewing Wide Field-of-view Sensor (SeaWiFS) Chl in the summer in the ECS from 1998 to 2007. The study results revealed the interannual variation in the CDW indicated by the high satellite Chl and correlated it with the interannual variation in

Satellite Data
To estimate the long-term SSS in the ECS, Rrs measurements derived from the Ocean Color Climate Change Initiative (OC-CCI) version 5.0 were used [14]. The OC-CCI products were developed by merging observations from the SeaWiFS, moderate resolution imaging spectroradiometer onboard the Aqua Earth Observing System (MODIS-Aqua), and Medium Resolution Imaging Spectrometer (MERIS). The products were constructed based on SeaWiFS datasets, and the atmospherically corrected Rrs values of MODIS and MERIS data were shifted to the nearest SeaWiFS bands (412, 443, 490, 510, 560, and 665 nm) based on a bio-optical model [15]. This process generates a set of Rrs values in six bands, which increases the inter-sensor consistency and reduces the bias values. In addition, the OC-CCI Chl dataset was used. OC-CCI uses the OC4V6 algorithm to retrieve Chl. The spatial and temporal resolutions of the OC-CCI used in this study are 4 km and daily, respectively.
MODIS-Aqua global level 3 mapped SST version 2019.0 data from 2002 to 2021 were obtained from the Physical Oceanography Distributed Active Archive Center (PO.DAAC) of National Aeronautics and Space Administration (NASA). Because SST data were available from 2002 to 2021 (20 years), SST front detection and SSD calculations were conducted for 20 years. In addition, altimeter satellite-gridded SLA data from 1997 to 2020 were computed with respect to a twenty-year (1993-2012) mean. These data were estimated by optimal interpolation by merging the measurements from the different available altimeter missions: Jason-3, Sentinel-3A, Haiyang-2A (HY-2A), Saral/AltiKa, Cryosat-2, Jason-2, Jason-1, Topex/Poseidon, Environmental Satellite (ENVISAT), Geostat follow-on radar altimeter (GFO), European Remote Sensing Satellite (ERS) 1 and 2. The spatial and temporal resolutions were 0.25 × 0.25 • and monthly mean, respectively. East/Japan Sea (EJS), South Sea of Korea (SSK), and ECS. This shipboard observation investigates SST, SSS, dissolved oxygen, phosphate, nitrite, nitrate, and silica bi-monthly in the YS, EJS, and SSK, and at three-month intervals in the ECS. Because the model was designed to detect CDW, the southern part of the YS and EJS and the total SSK and ECS station datasets were collected ( Figure 1). The data for these observations in various regions are presented in Table 1. All datasets, including satellite measurements, were obtained during the summer season (June to September) when the CDW is a major factor in SSS variation in the ECS.

Observation Data
Repeated shipboard measurements conducted by the National Institute of Fisher Science (NIFS) from 1997 to 2021 were used to develop and evaluate the SSS model. O servations were conducted near the Korean Peninsula, including the Yellow (YS) a East/Japan Sea (EJS), South Sea of Korea (SSK), and ECS. This shipboard observation vestigates SST, SSS, dissolved oxygen, phosphate, nitrite, nitrate, and silica bi-monthly the YS, EJS, and SSK, and at three-month intervals in the ECS. Because the model w designed to detect CDW, the southern part of the YS and EJS and the total SSK and E station datasets were collected ( Figure 1). The data for these observations in various gions are presented in Table 1. All datasets, including satellite measurements, were tained during the summer season (June to September) when the CDW is a major factor SSS variation in the ECS. The serial shipboard observation stations were marked in green squares. The red, blue, and yell arrows indicate warm, cold, and Changjiang outflow ocean currents, respectively, during summ season from 2002 to 2005 in this region (provided by Korea Hydrographic and Oceanograp Agency). Three bathymetric isobaths (50, 100, and 500 m) are also overlaid.

MPNN Model for SSS Estimation
Because the multi-layer perceptron neural network (MPNN) has been used to estimate SSS based on ocean color measurements in various regions [5][6][7][8], this method was adopted to develop an SSS estimation model for the OC-CCI. Because the shipboard survey stations are far from the Changjiang River mouth, the SSS range of in situ data lacking in the low-salinity conditions (under 25 psu). Thus, the model overestimates the Changjiang River mouth region, where the SSS is extremely low (<20 psu). Figure 2 shows an MPNN schematic diagram including six OC-CCI Rrs bands in the input layer and in situ SSS data in the output layer. To assemble two different datasets, the temporal resolution of in situ measurements was regarded as daily, and OC-CCI data were selected at the nearest location of the in situ survey stations. Thereafter, the datasets were randomly divided into training (80%) and validation (20%) datasets during the training period. The training was repeated 1000 times for each step while changing the number of hidden layers and neurons. Figure 3a shows the performance of the MPNN model based on the validation dataset independent of the training dataset. The scatters were concentrated from 30 to 33 psu and R 2 and root mean square error (RMSE) were 0.56 and 1.06 psu, respectively. Although the RMSE was acceptable, the SSS was corrected by the slope and bias owing to overestimation in the low SSS range (under 30 psu). After linear fitting, the overestimation and underestimation in the low and high SSS ranges were reduced despite the RMSE increasing to 0.34 psu. As the KNIFS serial shipboard survey stations are far from the Changjiang River mouth, an SSS under 25 psu was not obtained. Therefore, the accuracy decreased for low SSS values (under 30 psu). Given that the isohaline of the CDW in previous studies was 30 to 32 over the middle ECS [2,4,16,17], detection of the CDW fronts using our model near Jeju Island and the Korean Peninsula was possible. Moreover, the OC-CCI SSD was validated with KNIFS in situ measurements from 1997 to 2021 (Figure 3b). The R 2 and RMSE between in situ and OC-CCI-based SSD were 0.67 and 1.27 kg m −3 , respectively.  Figure 3a shows the performance of the MPNN model based on the validation dataset independent of the training dataset. The scatters were concentrated from 30 to 33 psu and R 2 and root mean square error (RMSE) were 0.56 and 1.06 psu, respectively. Although the RMSE was acceptable, the SSS was corrected by the slope and bias owing to overestimation in the low SSS range (under 30 psu). After linear fitting, the overestimation and underestimation in the low and high SSS ranges were reduced despite the RMSE increasing to 0.34 psu. As the KNIFS serial shipboard survey stations are far from the Changjiang River mouth, an SSS under 25 psu was not obtained. Therefore, the accuracy decreased for low SSS values (under 30 psu). Given that the isohaline of the CDW in previous studies was 30 to 32 over the middle ECS [2,4,16,17], detection of the CDW fronts using our model near Jeju Island and the Korean Peninsula was possible. Moreover, the OC-CCI SSD was validated with KNIFS in situ measurements from 1997 to 2021 ( Figure 3b). The R 2 and RMSE between in situ and OC-CCI-based SSD were 0.67 and 1.27 kg m −3 , respectively.   The MPNN model architecture based on the OC-CCI Rrs six bands data and KNIFS serial shipboard measurements. Among the 2421 datasets, 80% were used to train, and 20% were used to validate model. The input datasets were fixed, and the number of hidden layers and neurons in hidden layers were varied to derive the best performance. Figure 3a shows the performance of the MPNN model based on the validation dataset independent of the training dataset. The scatters were concentrated from 30 to 33 psu and R 2 and root mean square error (RMSE) were 0.56 and 1.06 psu, respectively. Although the RMSE was acceptable, the SSS was corrected by the slope and bias owing to overestimation in the low SSS range (under 30 psu). After linear fitting, the overestimation and underestimation in the low and high SSS ranges were reduced despite the RMSE increasing to 0.34 psu. As the KNIFS serial shipboard survey stations are far from the Changjiang River mouth, an SSS under 25 psu was not obtained. Therefore, the accuracy decreased for low SSS values (under 30 psu). Given that the isohaline of the CDW in previous studies was 30 to 32 over the middle ECS [2,4,16,17], detection of the CDW fronts using our model near Jeju Island and the Korean Peninsula was possible. Moreover, the OC-CCI SSD was validated with KNIFS in situ measurements from 1997 to 2021 (Figure 3b). The R 2 and RMSE between in situ and OC-CCI-based SSD were 0.67 and 1.27 kg m −3 , respectively.

CDW Front Based on SSS, Chl, and SST
Because CDW rapidly evolves, factors such as gaps from frequent cloud cover, sun glint, and thick aerosols make it difficult to produce a monthly SSS distribution. However, it is possible to combine daily isohaline locations to produce cumulative isohaline footprints. Figure 4 shows the monthly cumulative isohaline footprints and monthly mean Chl The isohalines are nearly parallel to the isobaths, especially the 50 m isobaths in June and the 100 m isobaths in the other three months (Figure 1). The plume shapes were in good agreement with the results of Bai et al. [2]. They presented the SSS distribution in July and August from 1998 to 2010 and identified three major types of extending shapes of the CDW based on the 31 isohalines. In August (Figure 4b and Supplementary Figures), the most apparent shape appeared as a type 2 plume in Bai et al. [2], which extends entirely northeastward, with the majority of the plume water traveling northward to the southern YS and then advected eastward through the Jeju Strait, with less low-salinity water remaining on the middle shelf of the southern ECS. However, the classification of the three types was not conducted owing to the ambiguous criteria. For example, mixed features appeared in August 2016 and the plume extended northeastward and southward. glint, and thick aerosols make it difficult to produce a monthly SSS distribution. However, it is possible to combine daily isohaline locations to produce cumulative isohaline footprints. Figure 4 shows the monthly cumulative isohaline footprints and monthly mean Chl concentrations from the June to September 2016 case as an example. The same figures from 1997 to 2021 are provided in the Supplementary Figures. The background color displays Chl; black dotted and solid lines and gray solid lines indicate 30, 31, and 32 isohaline footprints, respectively. The plume was generally directed northeastward, reaching Jeju Island, the Korean Peninsula, and sometimes the EJS through the Tsushima-Korea Strait. The isohalines are nearly parallel to the isobaths, especially the 50 m isobaths in June and the 100 m isobaths in the other three months (Figure 1). The plume shapes were in good agreement with the results of Bai et al. [2]. They presented the SSS distribution in July and August from 1998 to 2010 and identified three major types of extending shapes of the CDW based on the 31 isohalines. In August (Figure 4b and Supplemental Figures), the most apparent shape appeared as a type 2 plume in Bai et al. [2], which extends entirely northeastward, with the majority of the plume water traveling northward to the southern YS and then advected eastward through the Jeju Strait, with less low-salinity water remaining on the middle shelf of the southern ECS. However, the classification of the three types was not conducted owing to the ambiguous criteria. For example, mixed features appeared in August 2016 and the plume extended northeastward and southward.  The Chl distribution corresponded well with the isohaline footprints, despite not accumulating Chl. A probable reason for this is that the interaction between Chl and CDW had a time lag. Chl directly changes near the river mouth from the river origin Chl, but the rich nutrients in the CDW influence Chl within offshore regions [1,4,18]. This indicates that CDW could affect Chl for approximately one month. In addition, Chl rapidly decreased with increasing distance from the river mouth, unlike the isohalines pattern. This is because of complex situations, such as nutrient consumption, subsurface mixing, and the phytoplankton life cycle, which are difficult to demonstrate quantitatively in this study. Figure 5 shows the climatology of SSS, Chl, and SST from June to September. The SSS and Chl were averaged from 1997 to 2021 (25 years), and SST was averaged from 2002 to 2021 (20 years) because the available data periods were different. Similar to CDW, Remote Sens. 2022, 14, 3574 7 of 15 the SSS and Chl fronts were distributed longitudinally. Chl and SSS showed river plume fronts, whereas SST did not. Near the Changjiang River mouth, SST formed ocean fronts longitudinally but was maintained up to 123 • E. Except for this area, SST generated fronts latitudinally because of an insolation gradient. In addition, the CDW water mass, distinguished by 30 isohalines, was detached from the river mouth and transported to the Jeju Island. This detachment can be generated by the interaction between tide-induced vertical mixing and horizontal wind-driven movement of the CDW [19]. They determined the mechanism for the offshore detachment of the CDW using a three-dimensional numerical model. The detached patches in this study showed a pattern similar to the results of Moon et al. [19].

CDW Front Based on Surface Density
The different patterns of the various fronts suggest a complicated environment in the study area. Therefore, SSD was calculated using SST and SSS to distinguish CDW. Figure  6 shows the climatology of SSD from 2002 to 2021. The overall pattern seems similar to SSS (Figure 5a-d), but over the southeast of the CDW, it is different. This region is known as the high SST and SSS region because the Kuroshio Current is strong in summer. Because SSS is almost sustained in this area, SST increases from June to September, decreasing SSD by approximately 2 kg m −3 . However, in the northern area, isopycnic lines follow isohalines, although the isothermal lines are obviously different. In September, the area of the 1019 isopycnic line (Figure 6a) was smaller than the area of 30 isohalines (Figure 5d) west

CDW Front Based on Surface Density
The different patterns of the various fronts suggest a complicated environment in the study area. Therefore, SSD was calculated using SST and SSS to distinguish CDW. Figure 6 shows the climatology of SSD from 2002 to 2021. The overall pattern seems similar to SSS (Figure 5a-d), but over the southeast of the CDW, it is different. This region is known as the high SST and SSS region because the Kuroshio Current is strong in summer. Because SSS is almost sustained in this area, SST increases from June to September, decreasing SSD by approximately 2 kg m −3 . However, in the northern area, isopycnic lines follow isohalines, although the isothermal lines are obviously different. In September, the area of the 1019 isopycnic line (Figure 6a) was smaller than the area of 30 isohalines (Figure 5d) west of Jeju Island. However, detached patches were still clearly identified in the SSD distribution. During the summer, SSD decreased with increasing SST in the entire study region, but a noticeable reduction was observed in the CDW-dominated area.   Figure 7a). The SSD shows a similar pattern to the SSS rather than the SST, indicating th SSS is a major factor driving SSD during September at this region. The detached SSD pa is also observed in Figure 7. The detached patch (high SSD and high SSS) was formed tidal mixing with the subsurface, and ambient seawater covering due to the wind [1 The Chl is at maximum at the river mouth and decreases toward B. However, the seco peak of Chl appears around 124.3°E, where the SSD phase changes from high to low. T signifies that the Chl increased near the CDW front. Moreover, the SLA showed patter similar to SSD. From the steric height variation perspective, the SLA evolved not only water was added but also as SST or SSS evolved. In this case, the SSS variation cou change the SLA since the SST is almost constant at 26 °C. When the SSD is high (low), S is low (high) because the water volume has a negative relationship with water densi However, the exact locations of the SSD and SLA peaks were slightly different. There three possible causes: (1) the observed values were diminished by the 20-year averagi (2) the spatial resolution of the SSD and SLA observations varied, and (3) the time between the SSD evolutions and the resulting SLA reaction. However, these assumptio are not clear. Consequently, additional studies are required to understand the relationsh between SSD and SLA.   Figure 7a). The SSD shows a similar pattern to the SSS rather than the SST, indicating that SSS is a major factor driving SSD during September at this region. The detached SSD patch is also observed in Figure 7. The detached patch (high SSD and high SSS) was formed by tidal mixing with the subsurface, and ambient seawater covering due to the wind [19]. The Chl is at maximum at the river mouth and decreases toward B. However, the second peak of Chl appears around 124.3 • E, where the SSD phase changes from high to low. This signifies that the Chl increased near the CDW front. Moreover, the SLA showed patterns similar to SSD. From the steric height variation perspective, the SLA evolved not only as water was added but also as SST or SSS evolved. In this case, the SSS variation could change the SLA since the SST is almost constant at 26 • C. When the SSD is high (low), SLA is low (high) because the water volume has a negative relationship with water density. However, the exact locations of the SSD and SLA peaks were slightly different. There are  In addition, the correlation coefficient map and R 2 map showed the regions where Chl was affected by SSD (Figure 8). A relatively high R 2 (negative relationship) between Chl and SSD was observed near Jeju Island, the Yellow Sea, and the Tsushima-Korea Strait (p < 0.05). However, the other regions (particularly further south than 30 • N) showed a low R 2 at insignificant levels. Therefore, the SSD variation caused by SST changes did not influence Chl; however, the SSS changes contributed to regional Chl variation.

CDW Front for Nutrients' Distribution
Because the CDW is rich in nutrients, the CDW extension transports nutrients over the ECS. Therefore, the nutrient conditions differed across the CDW front. In this study, the shipboard observed nitrogen (N) and phosphorus (P) concentrations, and the SSD fronts from satellite data were compared. Figures 9 and 10 show the N and P levels with isopycnic lines in August for nine years (2002,2005,2006,2007,2009,2011,2014, 2019, and 2020) because the shipboard observations were conducted only in August at the ECS stations. In general, N was relatively high (4 to >10 M L −1 ) at a low SSD, close to the Changjiang River mouth. N showed a different pattern across the 1019 isopycnic lines, indicating that the CDW had more N than ambient seawater. Because N was obtained only once at each station, in contrast with SSD, which was a monthly average, the comparison of the spatial distribution had some limitations. Nevertheless, clear differences were revealed in August 2005August , 2007August , 2011August , 2014, and 2020, except for a few stations that showed relatively high (low) N inside (outside) of each 1019 isopycnic line. This indicates that the CDW is relatively rich in N from the river plume, and that the satellite-driven SSD can distinguish the front of N.
However, P did not follow the N and SSD spatial distributions over the nine year period. A high P inside the CDW cases was observed in 2002, 2006, and 2007 ( Figure 10). In particular, P was relatively lower in the CDW interior than in the exterior during 2020, which was the opposite phenomenon in the case of N. Li et al. [20] determined the increasing N trend during the past 50 years in the ECS due to the Changjiang River plume; however, P has been nearly stable since 2000. Therefore, the N/P ratio increased, which contributed to changes in the phytoplankton community's composition. Jiang et al. [21] obtained the same results in August 2009. They determined the factors that control the summer phytoplankton community in the ECS. This study examined phytoplankton blooms that caused enhanced eastward and southward CDW extension in August 2009. The diatoms and dinoflagellates bloomed under high N and low P conditions off the Changjiang Estuary. In addition, Zhou et al. [22] suggested that CDW plays an important role in diatom blooms by enhancing the pycnocline and P stress. Moreover, the coastal area adjacent to the Changjiang River Estuary is a high-risk region for harmful algal blooms [20,[23][24]. Although the analysis of the phytoplankton community's composition depending on the CDW front was not conducted in this study, satellite-driven SSD can be useful for the prediction of N and P distributions or major phytoplankton species before shipboard sur-

CDW Front for Nutrients' Distribution
Because the CDW is rich in nutrients, the CDW extension transports nutrients over the ECS. Therefore, the nutrient conditions differed across the CDW front. In this study, the shipboard observed nitrogen (N) and phosphorus (P) concentrations, and the SSD fronts from satellite data were compared. Figures 9 and 10 show the N and P levels with isopycnic lines in August for nine years (2002,2005,2006,2007,2009,2011,2014, 2019, and 2020) because the shipboard observations were conducted only in August at the ECS stations. In general, N was relatively high (4 to >10 µM L −1 ) at a low SSD, close to the Changjiang River mouth. N showed a different pattern across the 1019 isopycnic lines, indicating that the CDW had more N than ambient seawater. Because N was obtained only once at each station, in contrast with SSD, which was a monthly average, the comparison of the spatial distribution had some limitations. Nevertheless, clear differences were revealed in August 2005, 2007, 2011, 2014, and 2020, except for a few stations that showed relatively high (low) N inside (outside) of each 1019 isopycnic line. This indicates that the CDW is relatively rich in N from the river plume, and that the satellite-driven SSD can distinguish the front of N. Remote Sens. 2022, 14, x FOR PEER REVIEW 13 of 17  However, P did not follow the N and SSD spatial distributions over the nine year period. A high P inside the CDW cases was observed in 2002, 2006, and 2007 ( Figure 10). In particular, P was relatively lower in the CDW interior than in the exterior during 2020, which was the opposite phenomenon in the case of N. Li et al. [20] determined the increasing N trend during the past 50 years in the ECS due to the Changjiang River plume; however, P has been nearly stable since 2000. Therefore, the N/P ratio increased, which contributed to changes in the phytoplankton community's composition. Jiang et al. [21] obtained the same results in August 2009. They determined the factors that control the summer phytoplankton community in the ECS. This study examined phytoplankton blooms that caused enhanced eastward and southward CDW extension in August 2009. The diatoms and dinoflagellates bloomed under high N and low P conditions off the Changjiang Estuary. In addition, Zhou et al. [22] suggested that CDW plays an important role in diatom blooms by enhancing the pycnocline and P stress. Moreover, the coastal area adjacent to the Changjiang River Estuary is a high-risk region for harmful algal blooms [20,23,24]. Although the analysis of the phytoplankton community's composition depending on the CDW front was not conducted in this study, satellite-driven SSD can be useful for the prediction of N and P distributions or major phytoplankton species before shipboard surveys.

Discussion
The SSS estimation model based on the OC-CCI ocean color measurements was developed using NIFS serial shipboard observation data. Although the in situ observation stations cover large regions, including the SSK and central ECS, the lowest SSS in these data was higher than 25 psu (see Figure 3a) because the stations were far from the Changjiang River mouth. Therefore, the model seldom estimated low SSS conditions (<20 psu) ( Figure 5) despite being revealed near the Changjiang River mouth. Hence, the SSD in this area is underestimated. Bai et al. [2] also revealed this limitation. They used a simple analytical algorithm to estimate the SSS using the absorption of CDOM. The range of the in situ SSS was 26-36, and a low SSS (<28 psu) was not observed. However, such handling was not excessive owing to their interest in the variation in the CDW extension between 28, 29, 30, and 31 isohalines. Similarly, in this study, the middle range of SSS (29-32 psu) was used to analyze CDW identification and its impact; thus, the overestimation of low SSS was not significant. The training dataset contained abundant data in the middle SSS range, for which the model performance was relatively accurate ( Figure 5). As shown in Figure 4, the monthly cumulative 31 isohalines were in good agreement with Chl fronts. As the cumulative SSS footprint was not influenced by low SSS (<20 psu), the results were reliable. However, the detached patches in July and August may have impacted the SSS

Discussion
The SSS estimation model based on the OC-CCI ocean color measurements was developed using NIFS serial shipboard observation data. Although the in situ observation stations cover large regions, including the SSK and central ECS, the lowest SSS in these data was higher than 25 psu (see Figure 3a) because the stations were far from the Changjiang River mouth. Therefore, the model seldom estimated low SSS conditions (<20 psu) ( Figure 5) despite being revealed near the Changjiang River mouth. Hence, the SSD in this area is underestimated. Bai et al. [2] also revealed this limitation. They used a simple analytical algorithm to estimate the SSS using the absorption of CDOM. The range of the in situ SSS was 26-36, and a low SSS (<28 psu) was not observed. However, such handling was not excessive owing to their interest in the variation in the CDW extension between 28, 29, 30, and 31 isohalines. Similarly, in this study, the middle range of SSS (29-32 psu) was used to analyze CDW identification and its impact; thus, the overestimation of low SSS was not significant. The training dataset contained abundant data in the middle SSS range, for which the model performance was relatively accurate ( Figure 5). As shown in Figure 4, the monthly cumulative 31 isohalines were in good agreement with Chl fronts. As the cumulative SSS footprint was not influenced by low SSS (<20 psu), the results were reliable. However, the detached patches in July and August may have impacted the SSS (SSD) overestimation (underestimation) (Figures 5 and 6). Thirty isohalines and 1019 isopycnic lines were separated from the Changjiang River mouth at 122.5 • E, where the uncertainty was high. The overestimation (underestimation) of the SSS (SSD) produced a westward extension of the detached patches. Therefore, offshore CDW detachment occurred on the east side, and the area of the patch might have been much smaller than indicated.
To overcome these model limitations, more training datasets near the Changjiang River mouth region are required. Rarely obtained shipboard measurements are valuable for validating satellite-driven SSS products; however, they are insufficient for developing an SSS estimation model. Microwave satellite sensor-based SSS data can solve this problem. Kim et al. [5] and Kim et al. [6] avoided such limitations by using soil moisture active passive (SMAP) SSS data. Their model could estimate an extremely low SSS (<20 psu) near the Changjiang River mouth. However, it was difficult to develop the SSS estimation model for 25 years because SMAP data are available from 2015. Although the Soil Moisture and Ocean Salinity (SMOS) satellite has observed SSS since 2009, the low spatiotemporal resolutions and imprecision in coastal regions (a result of land interference) result in difficulties for its utilization in the ECS. Improving microwave SSS data quality in the ECS and assimilation with in situ measurements are recommended for further research.
In general, oceanic fronts are detected using specific algorithms, such as the histogrambased Cayula-Cornillon algorithm (CCA) and gradient-based Belkin-O'Reilly algorithm (BOA) [25]. However, these algorithms are unsuitable for investigating continuous front detection because they are highly sensitive to different parameter ranges. Therefore, the gradient-based fronts were inconsistent at different times. Therefore, the CDW fronts based on SSS and SSD were identified using thresholds (31 psu for SSS and 1019 kg m −1 for SSD). Although the thresholds were simply determined, previous research [2] supports the value of the threshold (31 psu), and effectively produced CDW fronts for 25 years.
This study was mainly conducted based on qualitative analysis because we focused on the spatial variation in marine environments caused by CDW. The spatial distribution of the CDW fronts explains the influence of CDW on Chl and DIN. However, to understand the mechanism and major factors of CDW behavior, more quantitative analyses are required in further studies. To conduct a quantitative analysis, more accurate SSS estimation and smallscale analyses are required. GOCI and GOCI-II measurements are valuable for small-scale analysis. Using high spatial (<500 m) and temporal (hourly) resolutions, more detailed analyses are possible, such as tides, small-scale eddies, wind, and typhoons.

Conclusions
In this study, an MPNN model for summertime SSS estimation in the ECS was developed based on satellite and in situ data from 1997 to 2021. Using time-series satellite-derived data, the monthly SSS, Chl, SST, and SSD fronts were presented and examined. The concurrence of SSS and Chl fronts indicated that CDW contributed significantly to Chl distribution in the ECS. In contrast, the SST fronts formed different patterns with Chl, indicating that the nutrients in the CDW were a major factor for Chl compared to the SST conditions during the summer months. CDW was identified in 31 isohaline and 1019 isopycnic lines. Detached patches were clearly revealed in the SSS and SSD distributions. In addition, N differences across the SSD fronts were observed. Although such an N distribution is possibly a consequence of CDW extension, more studies are required to explain the interaction between CDW extension and N.