Studying Long-Term Nutrient Variations in Semi-Enclosed Bays Using Remote Sensing and Machine Learning Methods: A Case Study of Laizhou Bay, China

Abstract

Semi-enclosed bays are greatly affected by human activities and have undergone drastic changes in their ecological environment, which requires our continuous attention. Laizhou Bay (LZB) is a typical semi-closed bay that is greatly influenced by human activities, and it is highly representative on a global scale. Investigating the multi-scale variation in nutrient concentrations in semi-enclosed bays can provide scientific support for environmental management and policy adjustments. To address the limitations of in situ data and the high cost of field surveys, this study utilizes machine learning methods to construct MODIS remote sensing models for quantitatively analyzing the concentrations of dissolved inorganic nitrogen (DIN) and dissolved inorganic phosphorus (DIP) in the surface water of LZB, as well as the spatiotemporal factors influencing them. Among various methods tested, the Support Vector Machine Regression (SVR) algorithm demonstrated the best performance in retrieving nutrient concentrations in LZB. The R² values of the DIN and DIP retrieval results based on the SVR algorithm are 0.91 and 0.92, respectively, while the RMSE values are 5.43 and 0.08 μmol/L, respectively. The retrieval results indicate that nearshore nutrient concentrations are significantly higher than those in offshore areas. Temporally, from 2003 to 2024, the DIN concentration in l has decreased at a rate of 0.4 μmol/L/yr, while the DIP concentration has remained relatively stable. Given sufficient observation data, the proposed machine learning and remote sensing approach can be effectively applied to other bays, offering the advantages of long time series, high spatial resolution, and a low cost.

1. Introduction

Semi-enclosed bays are strategic areas for sustainable human development, providing ongoing support for marine aquaculture, transportation, and economic trade [1]. However, these human activities release large amounts of nutrients and organic matter [2,3], significantly altering the structure and functions of bay ecosystems [4]. Nutrients such as dissolved inorganic nitrogen (DIN) and dissolved inorganic phosphorus (DIP) are essential for the growth and development of marine phytoplankton and microorganisms [5,6]. An appropriate N/P ratio promotes the uptake and utilization of nutrients by phytoplankton, enhancing marine primary productivity and possibly ecosystem health [7,8,9]. However, excessive nutrients can lead to serious ecological problems such as eutrophication, hypoxia, and harmful algal blooms [10,11,12], thereby impacting human livelihoods. In recent years, changes in bay nutrient concentrations have become a global concern [13,14,15], and clarifying the spatiotemporal trends of coastal nutrient concentrations can help inform environmental policy and promote public health.

Currently, studies on the spatiotemporal variation of nutrient concentrations in semi-enclosed bays are still mainly based on field observations and analyses [16,17]. Given the sparse spatial and temporal resolution of traditional monitoring, the ability of conventional methods to accurately capture long-term trends in nutrient concentrations is questionable. Therefore, methods such as remote sensing and machine learning are increasingly being applied to studies related to the spatiotemporal variation of coastal nutrient concentrations [11,15]. Satellite remote sensing is an important tool for monitoring marine environments (such as suspended sediment concentration, sea surface temperature, chlorophyll-a concentration, etc.), characterized by wide spatial coverage and long time series [18,19]. Although there is no direct physical relationship between nutrient concentrations and ocean color remote sensing, and thus nutrient concentrations cannot be directly monitored [9,20], it is possible to indirectly monitor water nutrient concentrations by using machine learning methods to invert marine nutrient data based on the relationship between nutrient concentrations and satellite-derived information [9,11,20]. At present, the use of remote sensing and machine learning to study bay nutrients has been applied in areas such as the Yangtze River Estuary, Shenzhen Bay, and the Yellow Sea [9,20].

Laizhou Bay (LZB), located in the southern part of the Bohai Sea, China (Figure 1e), is a typical semi-enclosed bay surrounded by land on three sides and fed by more than ten rivers, including the Yellow River, which transport large quantities of land-derived substances into the bay [21]. Due to its nutrient-rich river inflows, LZB serves as a key natural spawning and feeding ground in China [22,23] and is an important area for aquaculture. However, rapid industrial and agricultural development and population growth along the LZB coastline have resulted in increased discharge of land-based pollutants and aquaculture feed inputs, significantly altering the concentration and ratios of nutrients in LZB [16]. These changes have caused considerable damage to the ecological environment and biological communities in LZB [24]. Therefore, refined monitoring of nutrient concentrations in LZB is necessary to provide scientific evidence for environmental regulation and to achieve ecological balance between humans and nature.

Previous studies on temporal variations in LZB nutrient concentrations have mainly focused on short-term changes [25] or inferred long-term trends from limited observations [16]. Earlier research indicated that DIN and DIP in LZB initially increased and then declined over time, with the limiting factor shifting from nitrogen to phosphorus, with this turning point occurring around 2010 [16]. However, recent human activities have made DIN the dominant pollutant in LZB, and from 2019 to 2021, the bay reached a mildly eutrophic state, with deteriorating water quality [17]. These studies are all based on the analysis of measured data and still cannot solve the problem of low spatiotemporal resolution. Therefore, this study developed a nutrient retrieval method that is suitable for LZB by combining a large amount of field observation data with MODIS remote sensing data. Additionally, various machine learning algorithms were compared for their applicability to LZB based on measured data. Using this method, we present 22 consecutive years (2003–2024) of DIN and DIP distributions in LZB and analyze their variation patterns.

The retrieval algorithm proposed in this study can provide long-term trends in nutrient evolution in LZB, offering essential data support for water quality management and pollutant control in LZB. Additionally, this study provides valuable references for developing nutrient remote sensing retrieval algorithms and long-term trend analysis for other similar semi-enclosed bays worldwide.

2. Study Area

LZB is located in the southern part of the Bohai Sea (Figure 1e) and is one of the three major bays in the Bohai Sea, with an area of over 6000 square kilometers, making it the largest bay in Shandong [18]. The bay receives several rivers, including the Yellow River and the Xiaoqing River, which bring rich sediments and nutrients into the sea. Furthermore, this sea area serves as a spawning and nursery ground for species such as crabs and clams, with numerous mariculture zones. River inputs and human activities have increased the variation in nutrient concentrations in LZB. Additionally, LZB is a semi-enclosed bay with a relatively weak hydrodynamic environment, slow exchange rates between the inner and outer seas, and long nutrient residence times, which can easily lead to ecological issues like red tides [9].

3. Data and Methods

3.1. In Situ Measurement Data

The measured data used in this study were mainly obtained from two sources: cruise data provided by the National Natural Science Foundation of China (NSFC) and independently collected observational data. The station locations and observation times for the NSFC cruises are shown in Figure 1b–d, totaling 427 data points. Details of the sampling and testing methods can be found in the references [26,27]. The independently collected observational data were acquired from in situ monitoring buoys. The buoy data were collected through high-resolution (4–6 h) in situ observations using multi-parameter water quality instruments. The main monitoring parameters include DIN, DIP, temperature, turbidity, chlorophyll, and others. The specific buoy locations, observation times, and temporal resolutions are shown in Figure 1e and

Table 1. Measurement information for the independent buoys.

Station Name	Time	Time Resolution
BZ1	19 March 2024–19 April 2024	4 h
DY1	26 May 2024–18 November 2024	6 h
DY2	26 May 2024–18 November 2024	6 h
DY3	26 May 2024–21 November 2024	6 h
WF1	1 July 2022–16 September 2022	6 h
WF2	1 July 2022–19 September 2022	6 h

, with a total of 2900 data points. These buoy monitoring data, along with the NSFC’s measured data, will be primarily used to establish the remote sensing retrieval model. In summary, this study collected a total of 3327 measured data points, providing a richer dataset than those in previous studies [9,11].

3.2. Satellite Data

The remote sensing data used in this study were obtained from the Aqua sensor onboard NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS). The data were downloaded from https://ladsweb.modaps.eosdis.nasa.gov/, covering the time period from 2002 to 2024. The raw data were preprocessed using NASA’s SeaWiFS Data Analysis System (SeaDAS 8.1.0) software to perform atmospheric correction and cloud removal, resulting in Level-2 remote sensing reflectance (Rrs) products. When matching remote sensing reflectance to in situ measurements, the time difference between the two observations was less than 3 h, and the spatial distance was within 2 pixels. This matching method has been widely accepted and applied in water quality studies in the Bohai and Yellow Seas [18,19]. It should also be noted that in addition to conventional methods, sea ice can occur in LZB during winter, which significantly affects remote sensing reflectance (Rrs). In this study, the following method was used to remove sea ice, and the de-icing performance was significantly improved [18,28].

ρ_ωN(859) > 0.048

(1)

0.7 < ρ_ωN(555)/ρ_ωN(412) < 3.0

(2)

ρ_ωN(412) > 0.046

(3)

ρ_ωN is the water-leaving reflectance value (ρ_ωN = Rrs × π), ρ_ωN(555) denotes the water-leaving reflectance values of the green band with a wavelength of 555 nm, and Rrs is the band reflectance in the satellite data.

3.3. Machine Learning Method

Three popular machine learning algorithms (Support Vector Regression, Random Forest Regression, and the Back Propagation Neural Network) were used to establish the nutrient (DIN, DIP) machine learning models. These methods have been widely applied in nutrient retrieval studies [9,11].

Support Vector Regression (SVR) is based on the principle of structural risk minimization. The core idea is to use nonlinear functions to map samples from a low-dimensional feature space to a high-dimensional feature space, effectively solving nonlinear problems in the original space. SVR performs well with small sample sizes and is highly effective in tackling nonlinear, high-dimensional pattern recognition problems [29]. The SVR model established in this study was configured with a penalty factor of 4 and a radial basis function parameter of 0.8.

Random Forest Regression (RFR) is an ensemble model composed of multiple decision trees. Rather than constructing a large tree using the entire dataset, RFR builds several smaller trees using different subsets of the data and feature attributes, and then integrates them into a stronger overall model. This method has been successfully applied in hydrological research [30]. The RFR model was set with 100 decision trees and a minimum leaf size of 5.

The Back Propagation Neural Network (BP) is a multi-layer feedforward neural network trained using the error backpropagation algorithm. It is one of the most widely used neural network types. A typical three-layer BP structure includes an input layer, hidden layer, and output layer, making it well-suited for retrieving multiple ocean environmental parameters [11]. The BP model was trained with 1000 iterations, a learning rate of 0.01, and a minimum training error target of 0.000001.

3.4. Design of Machine Learning Models for Nutrient Concentration in the Sea Surface

In constructing the machine learning models, the training dataset included measured data (longitude, latitude, date, DIN, DIP) and remote sensing reflectance (10 individual bands and 10 combinations of bands). The specific parameters and modeling process are shown in Figure 2. All input features were standardized to eliminate the influence of outliers. SVR, RFR, and BP algorithms were used to train the nutrient retrieval models. The prediction accuracy of the three algorithms was compared, and the most suitable model for LZB’s marine environment was selected to derive the spatiotemporal variation in surface DIN and DIP concentrations in LZB.

3.5. Evaluation of Machine Learning Models

To verify the effectiveness of the machine learning models, we evaluated model accuracy using four metrics: the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and normalized root mean square error (NRMSE). The specific calculation methods are as follows:

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

(4)

where ${\widehat{y}}_i \mathrm{represents} \mathrm{the} \mathrm{predicted} \mathrm{value}$; ${\stackrel{\mathrm{-}}{y}}_i$ denotes the average of the measured values; n is the number of samples.

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

(5)

where $y_i\mathrm{represents} \mathrm{the} \mathrm{measured} \mathrm{values}$.

$$\mathrm{MAE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|y_i - {\widehat{y}}_i\right|$$

(6)

$$\mathrm{NRMSE}={\displaystyle \frac{\mathrm{RMSE}}{\max \left(y_i\right) - \min (y_i)}}$$

(7)

4. Results

4.1. Machine Learning Model Evaluation

This study assessed the applicability of three machine learning algorithms (SVR, RFR, BP) in nutrient research in LZB (Figure 3). Overall, the performance of the three models was satisfactory (

Table 2. Verification results.

Parameter	RF		BP		SVR
	DIN	DIP	DIN	DIP	DIN	DIP
R2	0.89	0.7	0.87	0.82	0.91	0.92
RMSE	5.94	0.16	5.13	0.16	5.43	0.08
MAE	4.42	0.13	5.12	0.1	2.53	0.04
NRMSE	0.105	0.145	0.09	0.145	0.096	0.073

). The R² for the DIN model was greater than 0.87 for all models, the RMSE was less than 5.9 μmol/L, the MAE was less than 5.12 μmol/L, and the NRMSE was less than 0.105. For the DIP model, R² values were all greater than 0.7, RMSE values were all less than 0.16 μmol/L, the MAE was less than 0.13 μmol/L, and NRMSE was less than 0.145. Among these three algorithms, SVR achieved the highest accuracy (Figure 3e,f), followed by BP (Figure 3c,d). Specifically, the SVR algorithm produced a DIN model with R², RMSE, MAE, and NRMSE values of 0.91, 5.43 μmol/L, 2.53 μmol/L, and 0.096, respectively, while the DIP model had R², RMSE, MAE, and NRMSE values of 0.92, 0.08 μmol/L, 0.04 μmol/L, and 0.073, respectively. In conclusion, the trained SVR model for the reconstruction of LZB nutrient concentrations is reliable and can be used for subsequent analytical research.

4.2. Regional and Seasonal Variations in Nutrients in LZB

DIN and DIP in LZB exhibit significant seasonal variations (Figure 4 and Figure 5). In general, the spatial distribution of DIN shows a decreasing trend from the coast to the offshore (Figure 4a–i). High DIN concentration areas are mainly concentrated in the western part of LZB near the Yellow River mouth and the southwest part near the Xiaoqinghe River mouth, while the northeast shows lower concentrations. However, there are significant differences in DIN values between seasons. In winter and spring, DIN is higher, with average values of 38.42 μmol/L and 34.55 μmol/L, respectively. Autumn shows intermediate levels, with an average DIN concentration of 30.4 μmol/L, while summer has the lowest values, with a DIN concentration of only 26.6 μmol/L.

DIP in LZB is characterized by higher concentrations in spring and autumn, and lower concentrations in summer and winter (Figure 5). In spring, the average DIP concentration in LZB can reach 0.77 μmol/L, with high values occurring along the eastern coast of LZB, while the area near the Yellow River mouth shows lower DIP concentrations (Figure 5c–e). The average DIP concentration in autumn is also around 0.77 μmol/L, with the high-concentration areas concentrated in the northeastern part of the study area (Figure 5i–k). In winter, the average DIP concentration is 0.68 μmol/L, with high-concentration areas mainly concentrated in the southern part of Laizhou Bay (Figure 5a,b,l). Summer shows the lowest average DIP concentration at 0.62 μmol/L, with high values found near the Yellow River mouth.

4.3. Interannual Variation in Nutrients in LZB

To analyze the interannual variation trends of nutrient concentrations in LZB from 2003 to 2024, this study presents the annual average results over the years and calculates the annual average concentration for LZB and the Yellow river estuary (YRE). As shown in Figure 6 and Figure 7, there are differences in the distribution areas of high-value regions for DIN and DIP in LZB. High DIN values are mainly concentrated on the western side of the bay (Figure 6), showing little interannual variation, while the eastern side exhibits larger year-to-year fluctuations in DIN (Figure 8a). In contrast, DIP concentrations have remained consistently high on the eastern side of the bay (Figure 7), while the Yellow River estuary area has shown increasing variability (Figure 8b). DIN in LZB showed a clear decreasing trend, with a decline of 0.4 μmol/L/yr (Figure 8c); in comparison, the decrease rate at the Yellow River estuary was lower, only 0.22 μmol/L/yr (Figure 8c). In contrast, DIP showed no significant interannual trend, with only a slight decrease of 0.002 μmol/L/yr (Figure 8d).

5. Discussion

5.1. Factors Affecting Seasonal Variations in Nutrient Concentrations in LZB

Previous studies have shown that LZB is a typical phosphorus-limited sea area [31], with an N/P ratio far exceeding the optimal threshold for phytoplankton nutrient uptake and utilization (16:1). Since the 21st century, the DIP concentration in LZB has shown little change (Figure 8d), so this study focuses on the mechanisms affecting DIN variation in Laizhou Bay.

River inputs are considered the primary contributors to marine environmental pollution [32]. This study has compiled total nitrogen data from 11 major rivers (Yellow River, North Jiaolai River, Dehuixin River, Majia River, Mi River, Tuhai River, Wei River, Xiaoqing River, Ze River, Zhangweixin River, and Zhimai River) flowing into LZB from 2018 to 2023 (Figure 9). The results show that, from a seasonal perspective, riverine total nitrogen flux was negatively correlated with DIN in LZB (Figure 9a), indicating that river nutrient input is not the main controlling factor for DIN in LZB. During the summer, the total nitrogen input from rivers to LZB was the highest, yet DIN concentrations in LZB were the lowest (Figure 9a). This was due to rapid phytoplankton growth and reproduction in the summer, where competitive effects among algae result in a significant uptake of nitrogen by plants from seawater, leading to the lowest DIN levels in the summer [33]. During the winter, although riverine total nitrogen input was lower (Figure 9a), the strong winter winds enhance the vertical mixing of water in LZB [18,34]. The disturbance of the sediment releases large amounts of DIN [35], bringing nutrients from the bottom layers to the surface, significantly increasing surface nutrient concentrations. Additionally, in winter, phytoplankton in LZB were light-limited [9,36], leading to a relatively weak absorption of nutrients by plankton, making nutrient concentrations in LZB higher in winter compared to other seasons.

5.2. Factors Affecting Interannual Variation in Nutrient Concentrations in LZB

Previous research on the factors controlling interannual variation in seawater nutrient concentrations has mainly focused on changes in riverine nutrient fluxes [11,17] and government environmental policy regulation [11,16]. However, there has been a lack of discussion regarding the influence of hydrodynamic changes on nutrient concentration variations in LZB, which was an important addition to previous studies. The strong correlation between riverine total nitrogen flux and DIN in LZB (Figure 9b) suggests that river input was indeed a controlling factor for interannual DIN variations in LZB. With the implementation of national marine policies and government control over key polluted areas, river nutrient concentrations are expected to decrease, and the eutrophication index and N/P ratio in LZB will show a downward trend, leading to significant ecological improvements [16,37]. In addition, with the advancement of agricultural non-point source pollution control in Shandong Province, the input of nutrients into LZB via groundwater from agricultural sources has decreased in recent years. This is also one of the reasons for the reduction in DIN concentrations in the bay.

In addition to the impacts of river inputs and environmental policies, changes in the marine dynamic environment are also important factors influencing nutrient concentrations in LZB. Previous studies have shown significant changes in wind conditions in LZB, with a decreasing wind speed trend [38,39], and an increasing proportion of northeast winds [40]. The decrease in wind speed reduces the resuspension of sediment [34], thereby reducing the release of nutrients from sediments [35]. The increasing proportion of winter winds inhibits the upward diffusion of bottom materials [40], thus reducing nutrient concentrations in LZB.

5.3. Future Work

This study assessed the applicability of SVR, RFR, and BP algorithms for LZB water quality retrieval. Previous research on the application of machine learning algorithms to marine nutrient salts has also included ANN algorithms [9]. Therefore, it is necessary to evaluate the applicability of other machine learning algorithms in LZB to improve the retrieval accuracy of DIN and DIP in LZB. Additionally, this study used remote sensing band data as input for machine learning, but future work should include marine environmental parameters (e.g., temperature, salinity, turbidity) as input conditions, which is expected to improve retrieval accuracy.

Due to remote sensing data limitations, this study only reconstructed surface DIN and DIP. With the development of numerical simulation technology in the future, spatially three-dimensional nutrient concentrations can be reconstructed. We have preliminarily developed a three-dimensional hydrodynamic–ecological coupling model that is suitable for LZB and have achieved some promising results [21], which can be further applied to simulate water quality in LZB across different temporal scales. Furthermore, the influence of nutrient concentration variations on the ecological environment and biodiversity at different time scales in LZB can be further analyzed.

6. Conclusions

This study assessed the applicability of different machine learning algorithms for the retrieval of surface nutrients in LZB based on field data and used satellite remote sensing data along with the most suitable SVR machine learning model to reconstruct the surface DIN and DIP in LZB from 2003 to 2024. Based on the reconstruction results using the SVR algorithm, the following conclusions can be drawn:

(1)

Three machine learning algorithms were evaluated for their suitability in the retrieval of surface nutrients in LZB, with SVR showing better performance than BP and RFR. The DIN and DIP retrieval results based on the SVR algorithm achieved R² values of 0.91 and 0.92, with RMSE values of 5.43 and 0.08 μmol/L, respectively.

(2)

Seasonal variations in nutrient concentrations in LZB show higher concentrations in the winter half of the year compared to the summer half, which is hypothesized to be mainly due to the uptake of nutrients by phytoplankton growth and reproduction.

(3)

From 2003 to 2024, DIN concentrations in LZB decreased at a rate of 0.4 μmol/L/yr, mainly due to changes in riverine nutrient flux and the implementation of environmental policies by the government.

(4)

Changes in hydrodynamic conditions also significantly affected nutrient concentrations in LZB.

The combined machine learning and remote sensing approach proposed in this study can be effectively applied to other bays, offering advantages such as long time series, high spatial resolution, and a low cost. This method provides a new approach for assessing the long-term evolution and ecological effects of marine nutrients.