An Interpretable Transformer-Based Framework for Monitoring Dissolved Inorganic Nitrogen and Phosphorus in Jiangsu–Zhejiang–Shanghai Offshore

Highlights

What are the main findings?

• Developed an interpretable transformer-based framework for retrieving DIN and DIP in Jiangsu–Zhejiang–Shanghai Offshore.
• Area of medium-to-high eutrophic waters rose 3.94 × 10² km²/yr (2005–2016) and fell −4.45 × 10² km²/yr (2016–2024).

What are the implication of the main findings?

• Key DIN and DIP drivers are water stratification (MLD), water turbidity (Rrs(667)), and temperature gradients.
• DIN and DIP model achieved high accuracy with MAPE below 33.69% in Jiangsu–Zhejiang–Shanghai Offshore.

Abstract

Anthropogenic increases in nitrogen and phosphorus inputs have intensified coastal water pollution, leading to economic losses and even threats to human health. Dissolved Inorganic Nitrogen (DIN) and Dissolved Inorganic Phosphorus (DIP), as key indicators of water quality, are essential for formulating environmental protection strategies. While deep learning has advanced the retrieval of these nutrients in coastal waters, existing models remain constrained by limited accuracy, insufficient interpretability, and poor regional transferability. To address these issues, we developed a Transformer-based model for retrieving DIN and DIP in the Jiangsu-Zhejiang-Shanghai (JZS) Offshore, integrating satellite observations with reanalysis data. Our model outperformed previous studies in this region, achieving high retrieval accuracy for DIN (R² = 0.88, RMSE = 0.16 mg/L, and MAPE = 33.69%) and DIP (R² = 0.85, RMSE = 0.007 mg/L, and MAPE = 31.59%) with strong interpretability. Based on this model, we generated a long-term (2005–2024) dataset, revealing clear seasonality and spatial patterns of DIN and DIP. Specifically, the concentrations have a distinct seasonal cycle with winter minima and autumn maxima, as well as estuarine-to-offshore decreasing gradient. Water quality assessment further showed that the areal extent of medium-to-high eutrophic waters increased by 3.94 × 10² km²/yr (2005–2016) but decreased by 4.45 × 10² km²/yr (2016–2024). Overall, the proposed Transformer-based framework provided a robust, accurate, and interpretable tool for nitrogen and phosphorus nutrient retrieval, supporting sustainable management of marine water quality in the JZS coastal ecosystems.

1. Introduction

With the rapid development of industry and the continuous expansion of urban areas, anthropogenic disturbances have significantly increased the fluxes of nutrients such as nitrogen and phosphorus entering the ocean, thereby intensifying eutrophication in coastal waters [1,2,3]. Such nutrient enrichment frequently triggers environmental hazards, including algal blooms and red tides, resulting in reduced fishery yields and damaged tourism industries and causing economic losses on a global scale [4]. More seriously, deteriorated water quality may pose direct risks to human health through the bioaccumulation of harmful substances in the food chain [5]. Dissolved Inorganic Nitrogen (DIN) and Dissolved Inorganic Phosphorus (DIP) are widely recognized as key indicators for evaluating water pollution levels and diagnosing the state of marine ecosystems [6,7]. Therefore, accurate monitoring of DIN and DIP and a comprehensive understanding of their spatiotemporal dynamics are of great significance for water quality assessment and the development of environmental management strategies [8,9].

Traditional in situ monitoring of marine nutrients, while providing high accuracy, is constrained by substantial costs and time-consuming operations, especially for large-scale and long-term surveys [10]. In contrast, satellite remote sensing provides extensive spatial coverage and continuous observations, making it a valuable tool for acquiring the dynamic spatiotemporal information on marine nutrients [11]. Taking advantage of satellite observations, early studies employed multivariate linear regression to retrieve nutrient concentrations from optical remote sensing data. For example, Isenstein and Park [12] constructed regression models to estimate total phosphorus (TP) and total nitrogen (TN) in the Lake Champlain and Missisquoi Bay, while Yu et al. [9] applied multiple stepwise regression (MLSR) to estimate the DIN over the Bohai Sea based on the Moderate-resolution Imaging Spectroradiometer (MODIS) observations. Similarly, Mathew et al. [13] employed the MODIS data to assess TP and TN in the northern Arabian Sea. Conventional multivariate regression–based empirical models have ability to estimate the nutrient concentrations, but their retrieval accuracies are inherently limited, as nitrogen and phosphorus nutrients are non-optically active water quality parameters (WQPs) that affect water optical properties indirectly [14].

Recent developments in machine learning (ML), particularly deep learning (DL), have considerably enhanced the capability to retrieve nutrient concentrations from the remote sensing data. DL excels at capturing complex nonlinear relationships and has demonstrated strong performance in estimating WQPs [15,16,17], outperforming traditional empirical approaches. Currently, most ML-based nutrient retrieval studies have focused on inland waters [17,18,19], while several attempts have been made in marine waters. For example, Zhu et al. [6] developed classification and regression trees (CART) ML models to retrieve DIN, soluble reactive phosphate (SRP), and chemical oxygen demand (COD) in the Northern South China Sea (NSCS). However, due to the pronounced spatial heterogeneity of marine systems, regression relationships between nutrients and remote sensing satellite data generally vary by region, necessitating the development of dedicated models for specific marine areas [20]. Wu et al. [7] proposed an optimized deep belief network (DBN) incorporating spatiotemporal information to estimate DIN and DIP concentrations in the Zhejiang Coastal Sea. However, this model suffered from poor interpretability, functioning largely as a “black box”. In addition, despite progress in the Zhejiang coastal region, challenges of model accuracy, interpretability, and generalizability still remain in estimating nitrogen and phosphorus nutrients. Given these limitations, Transformer-based models and their variants have emerged as a promising alternative. Recent studies have increasingly explored their applications in ocean color remote sensing, demonstrating favorable accuracy and strong potential for retrieving water quality parameters [21,22].

Given the above considerations, this study develops an interpretable Transformer-based DL model to retrieve DIN and DIP concentrations in the Jiangsu-Zhejiang-Shanghai (JZS) offshore region, using satellite remote sensing and reanalysis datasets. The proposed framework will retrieve accurate and reliable DIN and DIP concentrations, reveal the spatiotemporal distribution characteristics of DIN and DIP in the JZS Offshore, and provide model interpretability by quantifying the contributions of input features to model outputs. This paper is structured as follows. Section 2 introduces the study area, data sources, processing and match-up methods, and model construction. Section 3 provides the results, including the validations of DIN and DIP, their seasonal characteristics and long-term changes, and water quality classification assessment. Finally, model interpretability, comparisons, and limitations are discussed in Section 4.

2. Data and Methods

2.1. Study Area

JZS offshore is located in the northern part of the East China Sea (ECS), spanning from 120°E to 125°E in longitude and 27.5°N to 32.5°N in latitude (Figure 1c). This area encompasses representative coastal and estuarine systems, including the Yangtze River Estuary and Hangzhou Bay, which have drawn considerable attention due to recurring eutrophication issues [23,24,25]. As the world’s third-largest estuary, the Yangtze River Estuary exhibits distinct hydrographic features such as the maximum turbidity zone and estuarine front [26]. Hangzhou Bay is one of the largest tidal bays globally, characterized by strong tidal dynamics that shape its unique hydrodynamic environment [27,28,29]. With the rapid population growth and industrial expansion in the Yangtze River Basin, nutrient enrichment and seawater eutrophication in the JZS offshore have been increasingly severe. Reports such as the Bulletin on the Ecological and Environmental Status of Zhejiang Province (https://sthjt.zj.gov.cn/art/2025/6/12/art_1229263297_5528655.html, accessed on 15 December 2025), together with relevant studies, consistently identify DIN and DIP as the primary water pollutants exceeding standard thresholds in this region [30,31].

2.2. Data Description

2.2.1. In Situ Measured Data

The in situ water quality observations employed in this study were derived from a comprehensive spatiotemporal dataset covering four decades (1980–2022) across China [32]. This dataset comprises more than 330,000 records of 18 water quality indicators collected from 2384 water quality monitoring sites. These sites are classified by monitoring frequency into daily (244 sites), weekly (149 sites), and monthly (1991 sites), with coverage extending across inland, coastal, and offshore waters.

For this study, the primary dataset used was the national monthly water quality records provided by NMEMC in the period of 2017–2022. The locations of monitoring sites and the mean values of the in situ measured DIN and DIP are shown as Figure 1b,d. Within the study region, DIN concentrations predominantly range between 0 and 2 mg/L (Figure 1a), while DIP concentrations fall mostly within 0–0.08 mg/L (Figure 1e). All data has undergone data cleaning and technical validation to ensure the consistency and scientific reliability. For data cleaning, geographical coordinate inconsistencies of the same monitoring site across different years were resolved by averaging longitude and latitude values, while duplicate records were removed to avoid redundancy. Missing data and empty entries were consistently coded as NA and excluded from the dataset, and values below the detection limits (0.001 mg/L for both DIN and DIP) were standardized and flagged as “<DL”. For technical validation, outliers were identified using the interquartile range (IQR) method where data points falling outside the range of Q1 − 1.5 × IQR to Q3 + 1.5 × IQR (Q1 denotes the first quartile and Q3 denotes the third quartile) were classified as outliers. Given that the number of available samples within the JZS Offshore was relatively limited, additional NMEMC records from 2023 and 2024 were incorporated. These supplemental datasets were exclusively used as training inputs for the retrieval model to alleviate underfitting and enhance model robustness.

2.2.2. Remote Sensing and Reanalysis Data

This study employed Level 3 satellite remote sensing products from the MODIS-Aqua sensor, obtained from the NASA Ocean Color Web (https://oceancolor.gsfc.nasa.gov/, accessed on 15 December 2025). Specifically, the ocean color remote sensing products utilized included CHL, SST, Rrs(412), Rrs(443), Rrs(488), Rrs(555), and Rrs(667), POC, and Kd(490). All datasets were processed under Version 2022.0, with a spatial resolution of 4 km. In addition, ocean reanalysis data were obtained from the Copernicus Marine Service (CMEMS) Data Store (https://data.marine.copernicus.eu/products, accessed on 15 December 2025). The selected variables included SSS, SSH, SSC, and MLD. The CMEMS products were provided at a spatial resolution of 1/12°, ensuring adequate representation of mesoscale oceanographic features relevant to nutrient dynamics. This study integrated in situ measured data, remote sensing data, and reanalysis products to predict DIN and DIP. The data sources and specific variables employed in this study were summarized in

Table 1. Data Sources and Feature Descriptions.

	Data Source	Variables Abbreviation	Variables Definition	Physical Indication	Resolution
Remote sensing data	MODIS-Aqua	Rrs(412) (sr−1)	Remote sensing reflectance at 412 nm	Dissolved organic matter	Monthly 4 km
		Rrs(443) (sr−1)	Remote sensing reflectance at 443 nm	Chlorophyll-a/Phytoplankton biomass
		Rrs(488) (sr−1)	Remote sensing reflectance at 488 nm	Particulate matter scattering
		Rrs(555) (sr−1)	Remote sensing reflectance at 555 nm	Total suspended matter
		Rrs(667) (sr−1)	Remote sensing reflectance at 667 nm	High turbidity, algae bloom
		CHL (mg/m3)	Chlorophyll-a concentration	Phytoplankton biomass
		SST (°C)	Sea Surface Temperature	Thermodynamic effect
		POC (mg/m3)	Particulate Organic Carbon	Biocarbon process
		Kd(490) (m−1)	Diffuse attenuation coefficient at 490 nm	Water transparency
Reanalysis data	CMEMS	SSS (psu)	Sea Surface Salinity	Salinity, freshwater input	Monthly 1/12°
		SSH (m)	Sea Surface Height	Dynamic process (e.g., vortices, fronts, and circulation)
		SSC (m/s)	Sea Surface Current	Control material transport
		MLD (m)	Mixed Layer Depth	Vertical mixing, stratification

.

2.2.3. Data Processing and Match-Up

To address data gaps caused by cloud contamination and atmospheric correction errors in satellite observations, the Data Interpolating Empirical Orthogonal Function (DINEOF) method was applied to reconstruct missing values and ensure spatiotemporal continuity of the dataset [33]. For consistency, MODIS-Aqua remote sensing products (Rrs(λ), Kd(490), CHL, POC, and SST) and CMEMS reanalysis data (MLD, SSS, SSH, and SSC) were resampled to a spatial resolution of 1 km. Since NMEMC water quality measurements, remote sensing products, and reanalysis data are all monthly averages, the in situ DIN and DIP concentrations were matched to the nearest central pixel within a 1 km radius of each sampling point for the corresponding month. To minimize the influence of outliers in satellite-derived variables, an in situ measurement was deemed valid only when (i) more than ten valid pixels were available within a 5 × 5 pixel window, and (ii) the coefficient of variation across the window was less than 0.15. The 5 × 5 window provides a balanced spatial extent that effectively reduces sensor noise and small-scale anomalies while preserving physically meaningful environmental gradients in dynamic coastal waters like the Zhejiang Coastal Sea. The CV threshold of 0.15 filters out highly heterogeneous pixels that may result from sub-pixel contamination or unresolved fine-scale physical processes, thereby enhancing the reliability of the match-up dataset. These criteria are empirically supported and widely adopted in satellite ocean color studies over comparable marginal seas to ensure data quality for model development and validation [34,35]. Consequently, valid central grid cells and corresponding in situ WQPs concentrations were matched to construct the modeling dataset, resulting in a total of 5167 paired samples for DIN retrieval and 5058 paired samples for DIP retrieval, respectively.

2.3. Model Construction

The overall technical workflow of this study (Figure 2) comprises three stages: (i) data preparation, (ii) model construction, and (iii) validation and application. Specifically, data preparation includes two key steps: (1) acquisition of in situ measured data, remote sensing, and reanalysis data, and (2) data processing and matchup procedures. Model construction involves model training, feature engineering, and feature selection. The validation and application stage consists of three components: (1) model evaluation using statistical assessment metrics and interpretability analysis with SHapley Additive exPlanations (SHAP), (2) large-scale data production for the study area, and (3) water quality classification based on the generated outputs and established classification criteria.

2.3.1. Principle and Core Model

As typical non-optically active WQPs, DIN and DIP fail to directly absorb or scatter solar radiation in the visible spectrum (400–700 nm). Therefore, it proves incapable of accurately capturing the distribution patterns of DIN and DIP solely relying on optical characteristics. Instead, their spatiotemporal distributions are synergistically regulated by marine physical dynamic processes, biogeochemical cycles, and spatial geographical locations. To capture these indirect influences, two categories of input features were incorporated: ocean color properties (OCP, including Rrs(λ), Kd(490), CHL, and POC) and physio-chemical properties (PCP, including MLD, SSH, SSS, SST, and SSC). These predictors characterized the optical and physical conditions of the marine environment and were employed to model the nonlinear relationships with the target variables (DIN and DIP).

The core retrieval model is based on the Tabular Prior-data Fitted Network (TabPFN), a Transformer-based foundation model designed for small- to medium- scale tabular datasets [36]. TabPFN has been proven to have superior performance compared to the gradient-boosted decision tree (GBDT) like Categorical Boosting (CatBoost) and eXtreme Gradient Boosting (XGBoost), and the Automated Machine Learning systems (AutoML), achieving competitive regression accuracy while reducing training time by several orders of magnitude [36]. Generally, TabPFN delivers a speedup of up to a 3000× on datasets containing fewer than 10,000 samples and 500 features, alongside robust performance against outliers, uninformative features and missing values that are frequently encountered in real-world tabular regression scenarios. Feature normalization and missing-value handling are internally implemented in TabPFN through z-normalization and indicator-based imputation, respectively. When evaluated on the regression benchmarks, TabPFN improved normalized RMSE by 0.051 under default settings and by 0.093 after tuning compared with the CatBoost [36], underscoring its effectiveness beyond computational efficiency.

Unlike conventional Transformers that primarily process sequential data, TabPFN (architecture is illustrated in Figure 3) leverages a dual-attention mechanism optimized for tabular structures. At the intra-sample level, each feature attends to other variables within the same record to capture contextual relationships across predictors, while at the inter-sample level, attention is applied across rows for each feature, allowing the model to learn population-level patterns [36]. Furthermore, TabPFN was pretrained on millions of synthetic datasets generated through structural causal models, which not only enhanced generalization but also mitigated the risks of privacy leakage and data contamination associated with conventional foundational models [36]. During application, predictions are completed in a single forward pass by conditioning on the available feature set and unmasked samples. TabPFN is highly suitable for this study because the features (OCP and PCP) and the target variable (DIN and DIP) naturally form tabular structures, and the dataset size of the JZS Offshore region falls within the range where TabPFN exhibits its strongest advantages. Hence, TabPFN provides both computational efficiency and predictive robustness, making it an ideal choice for DIN and DIP retrieval. To further validate the advantages of TabPFN over mainstream ML models, this study constructed DIN and DIP retrieval models based on Random Forest (RF), XGBoost and Light Gradient Boosting Machine (LightGBM) using the same datasets.

2.3.2. Feature Engineering and Selection

Feature engineering in this study included classical mathematical transformations and spectrum-based transformations of predictor variables [37]. The former involved reciprocal, logarithmic, exponential, square root, and square operations, while the latter involved band normalization and two-band or three-band combinations. These mathematical operations were applied to both OCP and PCP in this study, thereby enriching the feature space. The algorithms used for the transformation are summarized in

Table 2. Feature mathematical transformation methods, where R denotes OCP and PCP.

Variable	Algorithm
$F_1\left(i\right)$	$\mathit{ln}(R_i)$
$F_2\left(i\right)$	$-R_i^2$
$F_3\left(i\right)$	$R_{i }\times \mathit{cos}\left(\mathrm{Lon}\right)$
$F_4\left(i\right)$	$R_i \times \mathit{cos}\left(\mathrm{Lat}\right)$

.

All parameters processed through feature engineering, along with the raw parameters (i.e., those not subjected to feature engineering), were initially included as candidate inputs for the model construction. To select the most effective features, the mean absolute percentage error (MAPE) is employed to rank and filter the features. The calculation formula of MAPE is as follows:

$$\mathrm{MAPE}(\%)={\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}\left|{\displaystyle \frac{{\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}}}{{\mathrm{y}}_{\mathrm{i}}}}\right|\times 100\%$$

(1)

where ${\mathrm{y}}_{\mathrm{i}}$ and ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ represent in situ measured and estimated values, respectively, and n represents the number of samples. In practice, random perturbations were introduced to a specified feature variable to update the input features, while keeping other features constant. A new model was constructed using these updated features, and its MAPE was calculated. If the resulting MAPE was greater than that of the baseline model, the feature was deemed effective.

The importance scores of all variables were calculated and ranked in descending order. Features were sequentially selected based on the list, ultimately yielding an appropriate set of input features to construct the optimal model. Detailed descriptions of the feature engineering and selection procedures are shown in Text S1 of Supplementary Materials. In this study, input features were randomly partitioned into training and validation subsets using stations as the grouping unit, with a baseline split ratio of 8:2. When a station contained a limited number of samples, the split ratio was reduced to ensure that the observations from each station were represented in both the training and validation datasets. As monthly average CMEMS and MODIS-Aqua products were used and the TabPFN model was trained to learn nonlinear input–output relationships, no additional region-specific outlier filtering was applied in this study.

2.3.3. SHAP Interpretability

Model interpretability was achieved using the SHAP tool, a post hoc explanatory method that quantified the marginal contribution of each feature to the model prediction [38,39]. As one of the most effective approaches for model interpretation [40], SHAP was employed to investigate the explainable mechanisms of TabPFN in estimating DIN and DIP concentrations. The model produced a predictive value for each sample, and the SHAP value was a quantity assigned to each feature within the sample, accounting for specific contribution of the feature toward predictive output of the model.

In addition to quantifying the magnitude of feature contributions, SHAP further revealed the direction (positive or negative) of the relationships between predictors and target variables [41,42]. A positive SHAP value (>0) indicates that the feature increases the predicted value of the target parameter, whereas a negative value (<0) denotes a suppressing effect. Therefore, SHAP enables the disentanglement of both direction and strength of feature contributions [43,44].

2.3.4. Accuracy Assessment

To assess the accuracy of the retrieval model, three widely used statistical metrics indices were adopted: MAPE, root mean square error (RMSE), and coefficient of determination (R²). Their specific mathematical expressions are given in Equations (1)–(3).

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}}{)}^2}$$

(2)

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}}{)}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}{)}^2}}$$

(3)

where ${\mathrm{y}}_{\mathrm{i}}$ and ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ represent in situ measured and estimated values, respectively, $\overline{\mathrm{y}}$ denotes the mean of observed values, and n is the number of samples.

To further assess predictive stability, residuals between in situ observations and model estimates were calculated, and their dispersion was quantified using the standard deviation of residuals (Std Residual) was used to quantify the stability of the model:

$$\mathrm{Std} \mathrm{Residual}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}-1}}\sum _{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{r}}_{\mathrm{i}}-\overline{\mathrm{r}}{)}^2}$$

(4)

where ${\mathrm{r}}_{\mathrm{i}}$ is the residual for the i-th sample, and $\overline{\mathrm{r}}$ is the mean residuals. A smaller Std Residual indicates a narrower residual distribution and, consequently, greater stability of model predictions.

2.4. Water Quality Classification Criterion

This study adopted the single-factor assessment method (https://mee.gov.cn/, accessed on 15 December 2025) to evaluate seawater quality with reference to the National Seawater Quality Standard of the People’s Republic of China (GB 3097-1997). In this method, the classification thresholds were defined separately for DIN and DIP. For DIN: concentrations ≤ 0.2 mg/L correspond to Case I, 0.2 < DIN ≤ 0.3 mg/L to Case II, 0.3 < DIN ≤ 0.4 mg/L to Case III, 0.4 < DIN ≤ 0.5 mg/L to Case IV, and DIN > 0.5 mg/L to Case V. For DIP: concentrations ≤ 0.015 mg/L correspond to Case I, 0.015 < DIP ≤ 0.03 mg/L to Case II or III, 0.03 < DIP ≤ 0.045 mg/L to Case IV, and DIP > 0.045 mg/L to Case V. The overall water quality classification of a given region was determined by the worst-performing indicator among all individual WQPs.

3. Results

3.1. Model Validation

Using remote sensing and reanalysis data as inputs, TabPFN produced DIN and DIP predictions through a single forward propagation. Compared with RF, XGBoost, and LightGBM, TabPFN has been demonstrated to outperform these models on small-to-medium-sized datasets. Specifically, it achieves the highest R² values and lowest MAPE and RMSE values in both DIN and DIP predictions (

Table 3. Accuracy assessment of DIN and DIP retrieval models.

Model	DIN			DIP
	R2	MAPE	RMSE (mg/L)	R2	MAPE	RMSE (mg/L)
RF	0.82	44.06%	0.21	0.76	37.93%	0.009
XGBoost	0.80	46.36%	0.21	0.77	39.45%	0.009
LightGBM	0.84	41.96%	0.19	0.76	37.01%	0.009
TabPFN	0.88	33.69%	0.16	0.85	31.59%	0.007

). Furthermore, the predicted values of TabPFN are uniformly distributed around the 1:1 line in Figure 4, indicating a high level of predictive accuracy. On the training dataset, R² values for DIN and DIP reached 0.95 and 0.91, with RMSE values of 0.11 mg/L and 0.005 mg/L, and MAPE values of 18.79% and 23.71%, respectively. As for the validation dataset, R² values remained high at 0.88 for DIN and 0.85 for DIP, accompanied by RMSE values of 0.16 mg/L and 0.007 mg/L, and MAPE values of 33.69% and 31.59% (Figure 4b,d). Notably, the distribution of scatter points is concentrated in the low-value regions (below 0.02 mg/L) of DIP, whereas higher DIP concentrations (over 0.08 mg/L) tend to be underestimated, with points scattered below the 1:1 line (Figure 4d). Nonetheless, the clustering of high-density scatter points is all located near the 1:1 line, indicating that the TabPFN model effectively captured the nonlinear relationships between WQPs with remote sensing inputs. The slight differences in RMSE and MAPE between the training and independent validation datasets further confirm the model robustness and generalizability.

To assess station-level accuracy, four in situ monitoring sites (Stations A–D) with multi-year observations (2017–2022) were selected for time-series validation (Figure 5). The time-series comparison of the four stations revealed strong agreement between TabPFN retrievals and the in situ measurement, with relatively high R² ≥ 0.79 and RMSE ≤ 0.13 mg/L for DIN and RMSE ≤ 0.007 mg/L for DIP across all stations. Notably, the R² values of DIN and DIP at Station A reached as high as 0.96 and 0.91, respectively, with the corresponding RMSE values of 0.02 mg/L and 0.003 mg/L (Figure 5a). These results underscore the excellent accuracy and reliability of the TabPFN in station-level applications.

To further evaluate the spatial consistency of the remote sensing products for DIN and DIP, we selected the retrieval results for three months to compare with in situ station observations, as shown in Figure 6 and Figure 7. In detail, the months selected for DIN are May 2018, April 2022, and August 2023, and those for DIP are September 2018, April 2020, and July 2022. The selected months ensured adequate station coverage across estuarine, nearshore, and offshore environments. The predicted spatial distributions (Figure 6b,e,h and Figure 7b,e,h) closely matched the in situ observations (Figure 6a,d,g and Figure 7a,d,g), with residuals largely approaching zero. Specifically, the residuals of DIN were concentrated within −0.1 to 0.1 mg/L, and those of DIP within −0.005 to 0.005 mg/L, indicating high retrieval accuracy.

The high-value (>0.2 mg/L) and low-value (<0.2 mg/L) regions in Figure 6c,f,i are mainly localized in the Yangtze River Estuary, southern Hangzhou Bay, and the Zhejiang coast, while offshore stations exhibited minimal residuals, reflecting stronger model performance in open waters. Furthermore, the Std Residual values of DIN at the in situ measurement stations for the three corresponding time points were all less than 0.11, while those of DIP at the in situ stations for its corresponding time points were all less than 0.005, confirming the high stability of TabPFN across different time periods and marine environments.

3.2. Seasonal Variation Characteristics

To investigate the seasonal dynamics and long-term changes in DIN and DIP in the JZS Offshore, TabPFN was applied to reconstruct nutrient concentration values from 2005 to 2024. The long-term time series (Figure 8a,b) reveal distinct seasonal cycles for DIN and DIP, characterized by regular intra-annual fluctuations and interannual variability. DIN concentrations (Figure 8a) typically ranged between 0.08 and 0.30 mg/L, while DIP concentrations (Figure 8b) varied from approximately 0.005 to 0.014 mg/L. The monthly climatologies (Figure 8c,d) show that DIN concentrations were lowest in March (0.08 mg/L) and peaked in May (0.25 mg/L). Seasonally, DIN concentration is the lowest in winter (December to February), while it reaches the highest in autumn (September to November). DIP exhibited a similar but less pronounced seasonal pattern, with minimum values in winter (0.006 mg/L) and maximum values in autumn (0.013 mg/L).

Figure 9 and Figure 10 illustrate the spatial distributions of monthly DIN and DIP retrievals in the JZS Offshore over the past 20 years. Overall, DIN and DIP both exhibited a characteristic estuarine–offshore gradient, with relatively high concentrations (more than 1.5 and 0.06 mg/L for DIN and DIP, respectively) in the estuarine and nearshore waters (highlighted by black ellipses), followed by a gradual decrease toward the open ocean. This spatial pattern reflects the strong influence of terrestrial inputs and estuarine hydrodynamics on nutrient distributions.

As shown in Figure 9, the spatial pattern of monthly DIN shows that in months with high concentrations (mean value over 0.2 mg/L), estuary and nearshore waters exhibit relatively homogeneous distributions, with nutrient levels broadly elevated across the region. In contrast, offshore waters exhibit markedly lower concentrations, with weaker temporal variability. The monthly changing rates of DIN in the estuarine and nearshore areas are considerably higher than that observed in offshore waters, highlighting the dominant role of riverine discharge and anthropogenic inputs in driving temporal fluctuations. Seasonal differences are also evident, with estuarine and nearshore DIN concentration falling below approximately 0.1 mg/L from December to March, whereas offshore low concentrations (less than 0.04 mg/L) are more pronounced during summer months, particularly in June and July, likely due to the intensified stratification and dilution effects.

The spatial distribution of DIP (Figure 10) exhibits pronounced seasonal variability, with higher concentrations (over 0.03 mg/L) mainly confined to estuarine and nearshore waters. Elevated DIP levels (more than 0.04 mg/L) are evident from May to October, particularly around the Yangtze River Estuary and Hangzhou Bay (highlighted by red ellipses), indicating strong terrestrial and anthropogenic influences. In summer (July–September), high DIP patches expand southward along the Zhejiang coast, while in winter (December–February), DIP concentration declines substantially across most coastal regions, reflecting reduced riverine inputs and enhanced vertical mixing. Offshore waters generally maintain low DIP concentration (less than 0.006 mg/L) throughout the year, suggesting limited nutrient transport from nearshore areas. These seasonal patterns collectively suggest that the DIP is predominantly governed by river discharge, coastal circulation, and biological uptake processes.

3.3. Long-Term Trend

Figure 11 presents the spatial and temporal distributions of the long-term trends in model-derived DIN and DIP over the JZS Offshore. From 2005 to 2024 (Figure 11a,d), DIN and DIP both exhibited increasing trends across most coastal waters, particularly along the northern and central coastal waters (regions A and B), indicating sustained nutrient enrichment over the past two decades. The trend patterns in the period from 2005 to 2016 (Figure 11b,e) reveal widespread positive rates of changes in DIN and DIP, with evident increases occurring near the river estuaries and coastal waters, suggesting strong terrestrial influence during 2005–2024. In contrast, the spatial distributions of the variation trends of DIN and DIP during 2016–2024 (Figure 11c,f) present distinct transition, with decreasing trends dominating the offshore and southern waters, reflecting the effectiveness of recent pollution mitigation efforts and the resulting improvement in water quality.

Figure 11g–l illustrate the interannual changes in DIN and DIP concentrations in regions A–C. As for the region A (Figure 11g,h), DIN and DIP both exhibited significant upward trends during 2005–2016, with annual increases of 1.04 × 10⁻² mg/L/yr (p < 0.01) and 3.18 × 10⁻⁴ mg/L/yr (p = 0.01), respectively. In region B (Figure 11i,j), nutrient variations displayed turning pattern, with DIN and DIP increased significantly before 2016 (4.75 × 10⁻³ mg/L/yr and 2.12 × 10⁻⁴ mg/L/yr, respectively) but showed decreasing trends afterward (–7.44 × 10⁻³ mg/L/yr for DIN and –1.42 × 10⁻⁴ mg/L/yr for DIP). For the region C (Figure 11k,l), DIN and DIP exhibited similar temporal trends, with significant increases during 2005–2016 (2.00 × 10⁻³ mg/L/yr for DIN and 8.70 × 10⁻⁵ mg/L/yr for DIP, p < 0.01), followed by declines from 2016 to 2024 (–2.79 × 10⁻³ mg/L/yr and −8.10 × 10⁻⁵ mg/L/yr, respectively). The trend slops and significance levels across the aforementioned regions are detailed in

Table 4. Summary of DIN and DIP trend slopes as well as significance levels in the three Regions.

Region	Time Range	Variable	Trend Slops (mg/L/yr)	Significance Levels
A	2005–2016	DIN	1.04 × 10−2	p < 0.01
		DIP	3.18 × 10−4	p = 0.01
	2016–2024	DIN	1.36 × 10−3	p = 0.53
		DIP	8.30 × 10−5	p = 0.05
B	2005–2016	DIN	4.75 × 10−3	p = 0.02
		DIP	2.12 × 10−4	p < 0.01
	2016–2024	DIN	−7.44 × 10−3	p < 0.01
		DIP	−1.42 × 10−4	p = 0.05
C	2005–2016	DIN	2.00 × 10−3	p < 0.01
		DIP	8.70 × 10−5	p < 0.01
	2016–2024	DIN	−2.79 × 10−3	p = 0.01
		DIP	−8.10 × 10−5	p = 0.06

. The spatiotemporal changes in DIN and DIP indicate that the nutrient enrichment strengthened in the previous twelve years (2005–2016) but has declined in the recent eight years due to the improved pollution control and management.

3.4. Water Quality Classification Assessment

Based on the DIN and DIP derived from remote sensing retrieval, the water quality in the JZS region was classified in accordance with the National Seawater Quality Standard of the People’s Republic of China (GB 3097-1997), and the classification results are presented in Figure 12. Spatially, the water quality levels in the JZS offshore waters (Figure 12a) exhibit a distinct gradient, decreasing progressively from the estuarine and nearshore areas toward the open ocean. Case I waters occupy the largest proportion of the study area, followed by Case V, whereas Cases III and IV, which represent the transitional conditions between polluted and clean waters, cover the smallest areas. In particular, the estuary and nearshore regions were predominantly categorized as Cases III, IV, and V, indicating eutrophication and nutrient enrichment, while the offshore waters were dominated by Cases I and II. These spatial patterns basically align with the annual Bulletin on the Status of Marine Ecological Environment published by NMEMC (https://www.nmemc.org.cn/hjzl/sthjgb/202509/P020250910537280063392.pdf, accessed on 15 December 2025).

For analytical clarity, the water quality classifications were further consolidated into two categories: high-quality waters (Cases I and II) and medium-to-high eutrophic waters (Cases III, IV, and V). Specifically, waters were classified as medium-to-high eutrophic waters when DIN concentration exceeded 0.3 mg/L or DIP concentration exceeded 0.03 mg/L, while the remaining waters were defined as high-quality waters. A long-term temporal assessment (Figure 12b) indicates that the spatial extent of medium-to-high eutrophic waters showed an increasing trend from 2005 to 2016, with a rate of 3.94 × 10² km²/yr (p = 0.02). After 2016, the increasing trend transitioned into a significant decline (−4.45 × 10² km²/yr during 2016–2024, p = 0.02), suggesting notable improvement in coastal water quality during this period. Conversely, the areal extent of high-quality waters (Figure 12c) exhibited an inverse trend, remaining dominant throughout the study period and exhibiting slight expansion in recent years. The observed changes suggest a gradual transition toward better water conditions in the JZS Offshore in the past decade, consistent with recent improvements in coastal environmental management.

4. Discussion

4.1. Model Interpretability

Figure 13 and Figure 14 illustrate the contribution of the first ten most influential input features to the model output, together with the corresponding SHAP value distributions for each feature (Figure 13a–j and Figure 14a–j), respectively. Among these variables, the constructed indices SST × cos(Lat and SST × cos(Lon) effectively capture the role of geographical location in modulating nutrient dynamics. Specifically, SST × cos(Lon) reflects the cross-shore gradient associated with coastal currents, while SST × cos(Lat) characterizes the latitudinal modulation of temperature effects.

For the TabPFN-DIN model (Figure 13), the variations in DIN in JZS Offshore are mainly controlled by MLD, Rrs(667), Rrs(412), SST × cos(Lon), and SST × cos(Lat). MLD exhibits predominantly negative SHAP values (below −0.3), reflecting the effect of water stratification, as deeper mixed layers promote vertical mixing that dilutes surface nutrients and lowers DIN, consistent with the in situ observations [45]. To further quantify the spatial heterogeneity of feature contributions, the study area was divided into 1° × 1° latitude-longitude grid cells, and the dominant positive and negative features influencing DIN were analyzed for each cell (Figure S1a,b in Supplementary Materials). This grid-based quantitative analysis confirms MLD as the consistent negative dominant feature for DIN across the entire study area, with the average negative SHAP value of MLD in each grid cell being lower than −0.23 (Figure S1a). These results quantitatively confirm the robust inhibitory effect of MLD on DIN concentrations, likely driven by enhanced vertical dilution processes. SST × cos(Lon) shows relatively large positive SHAP value (over 0.2), highlighting the impact of temperature gradients induced by ocean currents and at the land–sea interface on the DIN [46,47,48]. The grid-based analysis further reveals distinct spatial patterns in the positive dominant features for DIN (Figure S1b). Specifically, SST × cos(Lon) emerges as the dominant positive feature in nearshore regions, whereas SST × cos(Lat) predominates in the open ocean, with the average positive SHAP values of both features exceeding 0.13. This spatial differentiation suggests that nearshore DIN enrichment is more strongly affected by zonal temperature gradients associated with land–sea interface interactions, while DIN variability in the open ocean is primarily regulated by meridional temperature gradients linked to large-scale ocean circulation. By contrast, Rrs(667) and SST × cos(Lat) exhibit complex relationships with DIN, where lower values negatively affect concentrations, while higher values have mixed positive and negative effects depending on the spatial context, reflecting the regulatory role of water turbidity, biological consumption, and zonal temperature gradients [49,50,51]. Furthermore, in some individual inshore areas, −SSS² and V exert notable influences on DIN (Figure 13f,h), with SHAP values reaching −0.3 and 0.3. By comparison, SST, SSH, and ln(CHL) make relatively weak contributions to DIN concentration, with absolute SHAP values consistently below 0.3. The strong explanatory abilities of the SST × cos(Lat) and SST × cos(Lon) suggest that spatially modulated SST indices can effectively substitute for the direct SST signal, thereby diminishing the standalone contribution of SST in the model.

As shown in Figure 14, the TabPFN-DIP model exhibits influencing factors and effect directions that are largely consistent with those identified for DIN, indicating that the biogeochemical dynamics of these two nutrients are regulated by similar physical and environmental drivers. Consistent with the DIN results, MLD shows predominantly negative SHAP values in the DIP model, reflecting its role in regulating water column mixing and stratification, which controls the vertical dilution of surface phosphate and represents a key mechanism underlying DIP variations [45]. Grid-based analysis (Figure S1c,d) further confirms MLD as a universal dominant negative factor for DIP across the entire study area, with the average negative SHAP value in each grid cell being lower than −0.11 (Figure S1c). Moreover, Rrs(667) and SST × cos(Lat) show bidirectional effects with spatially varying magnitudes, indicating that DIP dynamics are modulated by water turbidity and regional hydrodynamics processes to a certain degree [49,50]. Further grid-based analyses of positive dominant features reveal that Rrs(667) serves as the primary positive contributor in most regions, whereas SST × cos(Lon) and MLD emerge as positive dominant factors in parts of the open ocean, with average positive SHAP values exceeding 0.08 (Figure S1d). This spatial pattern underscores the prominent role of Rrs(667)-related processes, such as variations in water turbidity, in promoting DIP enrichment across most of the study area. Unlike the DIN model, Rrs(412) exerts a stronger influence on DIP than SST × cos(Lat), implying that CDOM-related processes, such as photochemical degradation or interactions with organic phosphorus cycling, contribute more substantially to shaping DIP distributions than thermally driven gradients. Other variables, including V, −SSS², SSH, SST, and U, contribute relatively weakly to DIP, with the absolute SHAP values of most features remaining below 0.3.

4.2. Comparisons and Limitations

In previous studies on the retrieval of WQPs from satellite observations [9,12], traditional multiple regression models have been the dominant approach. However, due to the indirect and highly non-linear relationship between the WQPs and marine optical properties, these models generally suffer from limited predictive accuracy. In addition, ML algorithms such as feedforward neural networks, support vector machines, and random forests still had deficiencies in terms of training accuracy and robustness [52,53,54]. Recently, DL-based regression models have been increasingly applied in the retrieval of WQPs due to their superior capability to capture the non-linear relationships. Despite these advances, the application of DL to the retrieval of DIN and DIP has been primarily restricted to inland waters or estuaries [52,54], where optical characteristics demonstrate a relatively steady state [55]. By contrast, large-scale coastal waters are characterized by obvious seasonality, hydrodynamic variability, and biogeochemical processes [6]. These features reduce the transferability of inland-trained models to coastal environments, underscoring the importance for region-specific model development.

As for the retrieval of WQPs in coastal waters, Zhu et al. [6] constructed retrieval models for DIN (R² = 0.88), SRP (R² = 0.92), and COD (R² = 0.75) in the NSCS using the XGBoost algorithm, which were comparable to those obtained in our study. Unlike the model inputs of Zhu et al. [6], our study reduced the use of spatiotemporal information such as depth, longitude, latitude, and date information to avoid spatial discontinuity. We performed latitude and longitude transformation on SST, which enabled a more stable simulation of the nonlinear relationship between water environmental properties (OCP and PCP) and WQPs, and thus can better reflect the spatiotemporal characteristics of DIN and DIP. Furthermore, as coastal hydrographic and biogeochemical conditions vary significantly among regions, the relationships between WQPs and environmental factors also differ, limiting the direct applicability of models trained in the NSCS for predicting nutrient dynamics in the JZS offshore region.

Wang et al. [55] advanced DIN and DIP retrieval in Zhejiang coastal waters using a DL framework, yet their model accuracy remained limited. In this study, DIN retrieval achieved an R² of 0.88 (0.05 higher) with a lower RMSE (0.16 vs. 0.21 mg/L), while DIP showed a greater improvement with an R² of 0.85 (0.20 higher). Moreover, unlike the black-box model of Wang et al. [55], our model offered higher interpretability, enabling the identification of the key environmental drivers of DIN and DIP. Spatially, our work extended to include the coastal waters of Shanghai, the Yangtze River Estuary, and adjacent East China Sea regions, encompassing diverse ecosystems and a complete nitrogen–phosphorus gradient from estuarine to offshore waters. This broader coverage supports a comprehensive understanding of nutrient dynamics across multiple ecological regions and enhances insights into cross-scale linkages between anthropogenic influences and marine processes.

Nonetheless, several limitations remain in this study. First, the model performance is still constrained by the availability and quality of in situ measured samples [32]. The NMEMC in situ dataset has limited spatial representativeness due to the uneven distribution of sampling stations. Specifically, NMEMC stations are denser in estuarine and coastal regions, whereas offshore coverage is relatively sparse, which may introduce some degree of sampling bias. Second, retrieval accuracy may be affected by uncertainties in satellite-derived inputs arising from high turbidity, imperfect atmospheric correction, or cloud contamination [33,56]. High turbidity can significantly modify water-leaving radiance spectra, while residual atmospheric correction errors and cloud interference may introduce noise or data gaps that propagate into nutrient estimates. These sensor and environmental uncertainties are amplified in optically complex coastal waters, thereby limiting the direct transferability of the current model. Moreover, although the proposed model improves interpretability, it remains region-specific and may not be directly applied precisely in other coastal systems with distinct optical and hydrodynamic characteristics. Future studies will incorporate broader datasets and multi-source observations to enhance model generalization and robustness.

5. Conclusions

In this study, an interpretable DL framework based on the TabPFN model was developed to retrieve high-accuracy DIN and DIP datasets for the JZS Offshore from 2005 to 2024. The proposed framework aimed to improve the retrieval accuracy and model transparency, while facilitating a comprehensive understanding of the spatial and temporal dynamics of nitrogen and phosphorus nutrients in the JZS Offshore. The model demonstrated excellent retrieval performance, achieving R² = 0.88 and RMSE = 0.16 mg/L for DIN, and R² = 0.85 and RMSE = 0.007 mg/L for DIP when validated against independent in situ measurements, outperforming results reported in previous studies. Based on the reconstructed dataset, the spatiotemporal variations in DIN and DIP concentrations were systematically analyzed. Temporally, DIN and DIP exhibited a consistent regularity, reaching minimum in winter and peaking in autumn. Spatially, DIN and DIP exhibited a pronounced decreasing gradient from nearshore regions toward the open ocean, reflecting the influence of terrestrial inputs and offshore dilution. Interpretability analysis revealed that the variability of DIN and DIP was primarily governed by the water stratification (MLD), temperature gradients, and water turbidity (Rrs(667)), with secondary contributions from coastal circulation and freshwater input, reflecting the combined influence of vertical mixing, hydrodynamics, and nutrient gradients. The long-term changes in water quality classification further indicated an overall expansion of medium-to-high eutrophic waters during 2005–2016 (3.94 × 10² km²/yr), decreased at a rate of −4.45 × 10² km²/yr from 2016 to 2024. Our findings highlighted the strong temporal variability of nutrient-driven water quality and its linkage to natural and anthropogenic processes. Overall, this study demonstrated the potential of interpretable Transformer-based DL model in accurately reconstructing large-scale coastal nutrient datasets. This dataset provided scientific evidence for long-term water quality assessment and supported the formulation of coastal environmental protection and management strategies under intensified anthropogenic pressures and climate change.