Spatiotemporal Dynamic Analysis of Eutrophication Status Based on Machine Learning-Based Retrieval Algorithm: Case Study in Liangzi Lake, Hubei, China

Abstract

The routine monitoring of eutrophication is an important measure for observing the variation in water quality and protecting the ecological health of lakes. However, in situ information reflects eutrophication levels within a limited distance and period. In this study, we retrieved the trophic level index (TLI) based on Landsat 8 remote sensing images and using a machine learning (ML) method in Liangzi Lake in Hubei Province, China. The results showed that random forest (RF) outperformed other ML algorithms in estimating the TLI, evaluated by its higher fitness through the Monte Carlo method (median values of R², RMSE, and MAE are 0.54, 0.047, and 0.037, respectively). In general, 8% of the areas of Liangzi Lake presented an increasing eutrophication level from 2014 to 2022, and 20.1% of the areas reached a mild eutrophication level in 2022. In addition, we found that temperature and anthropogenic activities may impact the eutrophication conditions of the lake. This work uses remote sensing imagery and a ML method to monitor the dynamics of the lake’s eutrophication status, thereby providing a valuable reference for pollution control measures and enhancing the efficiency of water resource management.

1. Introduction

The deterioration of water quality and eutrophication of lakes, driven by the influx of sewage and wastewater from extensive human industrial and agricultural activities, significantly contribute to the proliferation of plankton and algae, which intensely strain the aquatic ecosystem [1,2]. Consequently, the routine monitoring of lake eutrophication is crucial for ensuring the safety of aquatic ecosystems and preventing adverse impacts on human health [3]. In eutrophic water bodies, chlorophyll-a (Chl-a), nitrogen (N), and phosphorus (P) are essential elements for the growth of algae and plankton and serve as key indicators of eutrophication [4,5]. In addition, Chl-a directly affects the photosynthetic potential of algae and the spectral characteristics of lakes [6]. An elevated value of Chemical Oxygen Demand (COD) usually indicates increased organic pollution, possibly due to the accumulation of algae and other biomass [7]. Water clarity tends to decrease due to heightened concentrations of algae and suspended solids [8]. In order to enhance the assessment of lakes’ eutrophic status, the trophic level index (TLI) derived from the water quality index has been proposed to assess the eutrophic status of lakes and is widely applied in China [9].

Traditionally, the TLI is deduced based on water quality measurements, including sampling, in situ measurements, and laboratory analysis. However, the traditional method consumes a considerable amount of labor power, material resources, and time costs, which only reflect the water quality of the sampling points within a limited distance and period [10]. Additionally, uncertainties in the monitoring process, such as sampling methods, locations, and laboratory measurements, could lead to biases in the results [11]. Despite extensive enhancements in the routine monitoring of lake eutrophication, most lakes around the world still lack comprehensive daily monitoring systems [12]. Therefore, monitoring measurements with convenient data acquisition and short-term and geographically dispersed feedback are required. Remote sensing technology has been developed and widely applied to monitor the spatial distribution and dynamic changes in eutrophication in lakes because of its cost-effective measurements at extensive spatiotemporal scales [13].

At present, satellite-based water quality retrieval algorithms are classified into four broad categories: analytical, empirical, semi-empirical, and machine learning (ML) methods [14]. As an optical active component, the relationship between Chl-a and the radiance or reflection spectrum of off-water radiation was derived using analytical methods such as bio-optical models and radiation transmission models [15]. However, the complex relationship between water composition and radiation transmission limits the application of the analytical method, resulting in the need for more parameters to be measured, such as the inherent optical characteristics and surface tourism characteristics [16]. Empirical methods estimate water quality parameters directly using correlation statistical regression models based on measured values and the reflectance of remote sensing bands or combinations of bands [17]. In practice, empirical models present uncertainty in water quality inversion due to the complex composition of water quality variables affecting the related spectral characteristics [18]. Semi-empirical methods combine paradigms derived from rigorous analytical models with a small number of in situ data and radiance values to calculate water quality parameters, which are widely used to retrieve water quality components [19]. However, semi-empirical methods still struggle with capturing nonlinear relationships and face challenges in accuracy and complexity [20].

In recent years, machine learning (ML) methods have been increasingly used to retrieve water quality parameters based on satellite-measured reflectance due to their heightened ability to capture nonlinear relationships compared to other methods [21]. ML improves the retrieval accuracy of water quality and reduces the uncertainty caused by complex physical processes [22,23]. A considerable number of researchers have employed ML models, including neural networks (NNs), random forest (RF), and support vector machines (SVMs), to achieve relatively satisfactory outcomes in the field of water quality retrieval [24]. Comparing the effectiveness of traditional regression models with ML models for water quality inversion, ML models were shown to effectively improve accuracy and reduce bias, as well as providing an effective means of monitoring water quality on a large scale in resource-constrained areas [25]. Additionally, ML models have been applied to assess lake eutrophication by the TLI directly [26,27]. However, relatively less research has focused on the selection of optimal ML models based on remote sensing and the performance of direct TLI retrieval.

In this study, different ML algorithms were compared for direct TLI retrieval based on in situ monitoring records and Landsat 8 remote sensing images in Liangzi Lake, Hubei, China. The principal aims were to (i) establish an optimal ML model for eutrophication assessment, (ii) retrieve reliable images with a high spatial and temporal resolution of the whole lake, and (iii) evaluate and analyze the spatiotemporal patterns and evolution characteristics of the eutrophication status from 2014 to 2022 in Liangzi Lake. In summary, we applied the ML model to water quality retrieval based on the satellite–ground synchronization observations and used this method for the spatiotemporal analysis of the lake.

2. Data and Methods

2.1. Study Area

Liangzi Lake (30°03′–30°19′N, 114°26′–114°38′E) is located in the middle reaches of the Yangtze River, southeast of Hubei Province (Figure 1). It is the second largest shallow freshwater lake in Hubei Province, with a mean depth of approximately 4 m [28]. In 2000, the water area of Liangzi Lake was recorded at 302 km², with the watershed covering about 2082 km² and sloping from higher elevations in the south to lower ones in the north [29]. The freshwater supply of Liangzi Lake primarily comes from regional surface runoff and precipitation. Over 30 rivers contribute to the lake’s water volume, with the largest being the Gaoqiao River in the south. Notably, Liangzi Lake has only one outlet located in the east, known as Modaoji, which channels water into the Yangtze River via the Changgang River. Liangzi Lake is generally recognized as three lake areas: West Liangzi Lake, Niushan Lake, and East Liangzi Lake. Moreover, Liangzi Lake is situated in the northern subtropical monsoon region, with the annual temperature here being about 17 °C and the average annual rainfall being about 1282.8 mm [30].

2.2. Sample Preparation and Data Collection

Water quality data from four lake monitoring stations in Liangzi Lake (located at 114.5382°E, 30.2623°N; 114.4466°E, 30.1954°N; 114.4886°E, 30.2368°N; and 114.5396°E, 30.3221°N) were selected as the computing sets for constructing TLI models, as shown in Figure 1. These sections are parts of the water quality monitoring network established, operated, and managed by the China National Environmental Monitoring Centre, who have been operating since 30 April 2020 and upload water quality measurements every 4 h. After data screening and proofreading, inaccurate and abnormal values were removed by statistical outlier detection, and the remaining data were applied to calculate the TLI.

Furthermore, multispectral data from Landsat 8 served as the basic inputs for the TLI retrieval model. Surface reflectance data from Collection 2 and Tier 1 levels of Landsat 8 from 2014 to 2022 were procured via the Google Earth Engine platform [31]. The surface reflectance data, which effectively mitigated atmospheric influence on the remotely sensed imagery using the LaSRC algorithm, were directly employed for the retrieval of the eutrophication status of Liangzi Lake.

2.3. Trophic Level Index

The trophic level index (TLI) is a modified version of the Trophic State Index (TSI), which was developed based on Secchi disk transparency (SD), chlorophyll-a (Chl-a), and total phosphorus (TP) [32]. In addition, as a widely utilized indicator of eutrophication status by Chinese monitoring agencies, the TLI needs to take into account the total nitrogen (TN) and potassium permanganate (COD_Mn) indices, shown as follows:

$$\begin{gathered} \mathrm{TLI}=0.266\ast \mathrm{TLI}\left(\mathrm{Chl}-\mathrm{a}\right)+0.1879\ast \mathrm{TLI}\left(\mathrm{TP}\right)+0.1790\ast \mathrm{TLI}\left(\mathrm{TN}\right)+0.1834 \\ \ast \mathrm{TLI}\left(\mathrm{SD}\right)+0.1834\ast \mathrm{TLI}\left(\mathrm{COD}\right) \end{gathered}$$

(1)

$$\mathrm{TLI}\left(\mathrm{Chl}-\mathrm{a}\right)=10\left(2.5+1.086 \mathrm{lnChl}-\mathrm{a}\right)$$

(2)

$$\mathrm{TLI}\left(\mathrm{TP}\right)=10\left(9.436+1.624 \mathrm{lnTP}\right)$$

(3)

$$\mathrm{TLI}\left(\mathrm{TN}\right)=10\left(5.453+1.694 \mathrm{lnTN}\right)$$

(4)

$$\mathrm{TLI}\left(\mathrm{SD}\right)=10\left(5.118-1.94 \mathrm{lnSD}\right)$$

(5)

$$\mathrm{TLI}\left(\mathrm{COD}\right)=10\left(0.109+2.661 \mathrm{lnCOD}\right)$$

(6)

where SD measurements at the lake monitoring stations were not available, so we derived them by extrapolation from turbidity values with the equation developed by USGS [33]:

$$\mathrm{SD}=3.39029{ \ast \mathrm{Turbidity}}^{-0.637}$$

(7)

The TLI method classifies the water trophic state into five levels [12]: Oligotrophic (TLI < 30), Mesotrophic (30 ≤ TLI ≤ 50), Light eutrophic (50 < TLI ≤ 60), Medium eutrophic (60 < TLI ≤ 70), and Heavy eutrophic (TLI > 70).

2.4. Retrieval Approach of TLI

2.4.1. Selection of Input Features

We constructed the ML models based on a variety of spectral features, as shown in

Table 1. A list of the equations and references of the input features.

Features	Equation	Reference
Individual spectral bands	Blue (B), Green (G), Red (R), NIR, SWIR 1, SWIR 2	-
Blue–green ratio	B/G	[34]
Red–green ratio	R/G	[35]
Red–blue ratio	R/B	[36]
Ratio of NIR to red	NIR/R	[37]
Ratio of NIR to green	NIR/G	[38]
Difference between NIR and green	NIR-G	[39]
NDVI	(NIR − R)/(NIR + R)	[40]
NGRDI	(G − R)/(G + R)	[41]
Plant pigment ratio	(G − B)/(G + B)	[42]
Blue NDVI	(NIR − B)/(NIR + B)	[43]
GNDVI	(NIR − G)/(NIR + G)	[44]
Green–blue NDVI	(NIR − G + B)/(NIR + G + B)	[45]
Green–red NDVI	(NIR − G + R)/(NIR + G + R)	[46]
Red–blue NDVI	(NIR − B + R)/(NIR + B + R)	[45]
Pan NDVI	(NIR − B + R + G)/(NIR + B + R + G)	[45]

. The features included the original spectral information, difference, and normalized difference in the spectral values. These features were treated as inputs for the training and testing of ML models.

2.4.2. Machine Learning Algorithms

This study employed ML algorithms including NNs, SVMs, and RF and compared their performance with traditional multiple linear regression (MLR) (more details are given in the Supplementary Materials). The NN was designed to model complex patterns in data through learning from layers of interconnected nodes, making them suitable for handling nonlinear relationships [47]. The SVM is another robust method that focuses on finding the optimal hyperplane to separate different classes with the maximum margin, ensuring good generalization properties [48]. RF is a type of ensemble learning technique based on the bagging method, which creates multiple decision trees during training and outputs the regression result as the mode of the individual trees, offering advantages in terms of accuracy and controlling overfitting [49]. The models were trained and tested on Python 3.9.

2.4.3. Evaluation of Model Performance

All the algorithms were evaluated according to the fitness coefficient (R²), mean absolute error (MAE), and root mean square error (RMSE). Specifically, the models were randomly run 500 times according to the Monte Carlo algorithm, and then the optimum model was chosen by comparing all the evaluation coefficients. The hyperparameters of the optimum model were tuned based on the grid search method. R², MAE, and RMSE are the most widely used accuracy evaluation indices of the retrieval model and are defined as follows:

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

(8)

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

(9)

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

(10)

where $y_i$ is the observed TLI value, ${\widehat{y}}_i$ is the model-inverted prediction value, and $\overline{y}$ is the average of the observed TLI values.

3. Results

3.1. Descriptive Statistical Analysis Based on In Situ Measurements

The Mann–Kendall (MK) test and descriptive statistics for the TLI at four national stations during the observed period (December 2020 to December 2022) are detailed in Figure 2. The results highlighted that TLI values exceeding the critical threshold of 30 were the most frequent during the summer months at these monitoring stations. This pattern underscored that Liangzi Lake was particularly vulnerable to eutrophication, with a significantly increased risk during warmer periods. Further analysis using the MK test revealed distinct trends in TLI dynamics across the four monitored locations. Positions 1 and 3 exhibited a noticeable increasing trend in TLI values over the two-year observation period, suggesting a worsening eutrophication issue at these sites. In contrast, positions 2 and 4 showed no significant change in TLI levels, indicating a relatively stable eutrophication status at these locations.

The box plot further elucidated these findings by illustrating the distribution of TLI values across the monitoring sites. It distinctly shows that the median TLI values at positions 1, 2, and 3 were elevated compared to the median TLI at position 4. This disparity in median TLI values indicated that eutrophication was more pronounced at positions 1, 2, and 3. The severity at these locations could be attributed to localized factors such as nutrient runoff or insufficient water circulation, which foster conditions conducive to algal blooms and nutrient accumulation. Notably, although position 2 did not exhibit an increasing trend in TLI values, it displayed a relatively higher median and narrower fluctuation range of TLI values in the box plot compared to the other locations. This suggests that position 2 is consistently at a relatively higher risk of eutrophication. These observations underscored the complex spatial and temporal dynamics of eutrophication within Liangzi Lake.

3.2. Model Development and Performance Evaluation

This study utilized widely recognized machine learning regression models including RF, SVM, MLP, and MLR to develop a robust inversion model for assessing the eutrophication status of lakes. The models were each implemented and run 500 times, employing a Monte Carlo algorithm for the random allocation of train datasets, test datasets, and systematic hyperparameters. This methodology ensured a comprehensive evaluation of each model’s performance under varied conditions. Among the tested models, the RF model emerged as the most effective algorithm for accurately predicting the TLI. This superiority was determined through a rigorous comparison of the distribution of evaluation metrics across the different models (as shown in Figure 3). Specifically, the median values of the R², MAE, and RMSE of the RF model are 0.543,0.374, and 0.047, respectively. Except for the median value of the MAE being slightly higher than that of the MLR model, the RF model showed the best performance in terms of the other evaluation indices. Meanwhile, the RF model presented more concentrated interval distributions of R², MAE, and RMSE than the MLR model, which also had better robustness.

A grid search (GS)-optimized RF model was proposed to search for the optimum hyperparameters of the model to rapidly evaluate the eutrophication status in Liangzi Lake. The RF model was particularly suited for this task as it can avoid overfitting, thereby ensuring better generalization performance and accurately calculating the TLI. This model leverages 21 band features derived from Landsat 8 data, which are crucial for long-term monitoring and the comprehensive assessment of eutrophication across the entire lake surface. In order to tune the optimal hyperparameters of the RF model, the dataset was employed in a GS-optimized RF framework through a 5-fold cross-validation approach. The dataset was randomly divided into five equal subsets, with each subset serving once as a validation dataset, while the remaining four subsets were combined into a new training dataset. By applying the GS technique and conducting five iterations of calculations, the optimal hyperparameters for the RF model were successfully identified and are shown in Figure 4. The results indicated that the R² of the train set is higher than 0.89, and the MAE and RMSE are around 0.02; the best R², MAE, and RMSE of the test sets are 0.71, 0.03, and 0.04, and the worst are −0.06, 0.04, and 0.06. This phenomenon is caused by the uneven distribution of the test set. In summary, the optimal trained RF model was selected to analyze the spatiotemporal change in the eutrophication status in Liangzi Lake.

3.3. Spatiotemporal Changes in TLI in Liangzi Lake

The eutrophication level of the entire Liangzi Lake region was retrieved and analyzed based on the trained RF model and the advantages of the extensive monitoring capabilities of satellite images. It expanded the monitoring of the lake’s TLI from discrete points to a comprehensive spatial distribution via the ensemble of the pixel. Figure 5 illustrates the spatiotemporal variations in the eutrophication status of Liangzi Lake over the period from 2014 to 2022 based on the annual average values of the TLI data derived from Landsat 8 images. The establishment of the retrieval model compensated for the lack of a lake monitoring section in the East Liangzi Lake area. Additionally, this advancement significantly enhanced the ability to assess eutrophication across the entire lake surface, providing a more detailed and holistic view of the lake’s eutrophication status.

Notably, as shown in Figure 5, the eutrophication levels were effectively controlled in the years 2016, 2017, and 2020. However, the lake exhibited a mild state of eutrophication in 2022. Examining the proportions of the mild eutrophication area further, the area of the lake showing mild eutrophication remained below 8.5% from 2014 to 2021, with the lowest incidence occurring in 2018 at just 1.6%. Nevertheless, this ratio escalated dramatically to 20.1% in 2022, indicating a significant increase in eutrophication levels. Spatially, the regions closer to the shore of Liangzi Lake consistently exhibited higher levels of eutrophication compared to the more central and offshore areas. This pattern was particularly pronounced in the lake’s complex branching regions, where the annual average TLI consistently ranged between 25 and 35. Moreover, the southeastern part of the lake faced more severe eutrophication threats compared to other areas. This region displayed a state of mild eutrophication even during years like 2020 and 2021, which otherwise demonstrated better overall water quality in other parts of the lake.

In addition, we performed the MK test on the ensemble prediction results of pixel points over the whole lake. Figure 6a suggests that the TLI in most areas of Liangzi Lake displayed an overall fluctuating trend from 2014 to 2022. Among them, about 8% of the lake area showed an increasing trend of eutrophication which is mostly concentrated in East Liangzi Lake and part of the lakeside areas. In addition, the Theil–Sen slope analysis (Figure 6b) demonstrated that about 38% of the area in Liangzi Lake showed a decreasing trend, mainly in the central part of Liangzi Lake and Niushan Lake. On the contrary, around 62% of the lake area showed positive slopes, mainly in East and West Liangzi Lake, and about 5% of the area presented slope values greater than 0.5, which showed a faster growth rate of eutrophication level in part of the lakeside areas.

4. Discussion

4.1. The Advantages and Limitations of the Model

The random forest model proposed in this study demonstrates good performance in the inversion of the eutrophication index in the Liangzi Lake area, outperforming other representative algorithms. The MLP exhibits the poorest performance in the retrieval of eutrophication status, likely due to its high data requirements [50]. In this study, the limited number of data may have resulted in lower inversion accuracy for the MLP. SVMs are particularly more well suited for classification tasks compared to regression methods due to their design for maximizing margin, which offers a more defined separation between classes; therefore, the method has mostly been used for classification tasks [51]. The regression model for TLI detection may limit the effectiveness of the SVM algorithm. The robustness of MLR’s model predictions is not well behaved, and the distribution of the various prediction indices is relatively wide, indicating that the subparameters are dependent on the data selection and confirming that the dataset in this study does not apply to the MLR algorithm.

The performance of most traditional ML models is contingent upon the accuracy of feature extraction in feature engineering, such as the widely used Pearson correlation coefficient. However, Pearson’s method assesses linear relationships between data, not nonlinear ones, thus imposing certain limitations and impacting the accuracy of the final model. The random forest algorithm circumvents the need for the prior selection of TLI features and provides feature importance metrics to characterize the statistical relationships between inputs and outputs. According to Figure 7, the top four rankings of correlation coefficients and feature importance are the same, yet the rankings of other correlation coefficients and feature importance do not completely coincide. This indicates that features with linear correlations are also prioritized by the random forest algorithm. However, correlation coefficients alone should not be solely relied upon for feature selection in constructing a random forest model for the eutrophication index.

Previous studies have shown that time windows within several days are oftentimes used for the matching-up of images and water quality measurements, and they are generally considered to produce reliable results [52]. According to the analysis of the monitoring data, the TLI can exhibit different variations within a few days. As shown in Figure 8, within intervals of one to four days, approximately 80% of the TLI values change within a range of 20%. There is a notable increase in the frequency of changes within the range of 20% to 40%, whereas changes greater than 40% are relatively infrequent. This suggests that while the TLI can fluctuate substantially over short periods, the majority of these changes remain within a moderate range. The observed pattern indicates that rapid and large-scale shifts in the trophic conditions of Liangzi Lake are uncommon, implying a degree of reasonableness for time windows within several days between the satellite transit time and actual measurement time. However, these findings also underscore the fact that star–earth synchronous observation can enhance the compatibility and accuracy of the input data for the construction of TLI inversion models.

4.2. Impacts of Meteorological Conditions on Lake Eutrophication

According to the analysis of two remote sensing images, which capture over 90% of the lake area in a single revisit cycle of Landsat 8, this study utilized remote sensing retrieval data to assess the effect of natural conditions on the lake’s eutrophication change. Concurrently, hourly meteorological data including the U-component of wind (U), the V-component of wind (V), 2 m temperature (T), surface solar radiation (SSR), surface pressure (SP), and total precipitation (TP) were sourced from the ERA5 database [53]. We screened out eight events that could display two satellite images with an interval period of 16 days (the revisiting period of Landsat 8) and calculated the correlation coefficients and p-values of the TLI and the mean value for each index of the meteorological data (Figure 9). The correlation analysis revealed a strong positive relationship between air temperature and eutrophication and a negative correlation with surface atmospheric pressure. This finding suggested that higher temperatures may worsen eutrophication by accelerating the growth rates of algae and phytoplankton, as well as reducing the dissolved oxygen content in the water [54]. Moreover, the present study found minimal evidence to support a direct correlation between atmospheric pressure and eutrophication. However, lower atmospheric pressures generally accompany higher temperatures, which could explain the observed higher correlation coefficients in the statistical analysis.

Additionally, this study indicates that variations in eutrophication status are not strongly linked to precipitation patterns. While rainfall can transport nutrients such as nitrogen and phosphorus from the watershed into the lake, potentially exacerbating eutrophication [55], significant rainfall can also dilute nutrient concentrations in water. Thus, changes in rainfall patterns may not consistently contribute to reductions in eutrophication in aquatic environments. This complexity underscores the need for further research to disentangle the impacts of hydrological dynamics on lake eutrophication, highlighting the complex interplay of environmental factors affecting aquatic ecosystem health.

4.3. The Impact of Human Activities on Lake Eutrophication

Based on the multi-year satellite retrieval results, East Liangzi Lake consistently exhibited a relatively high eutrophication status compared to other sub-lakes, particularly in the southeastern region, where the annual average value of the TLI reached the level of mild eutrophication. As the largest watershed area of sub-lakes, the major tributary Gaoqiao River and other tributaries flow into the East Liangzi Lake area, with an upstream watershed area of approximately 1271.3 km² (derived from HydroBASINS, https://www.hydrosheds.org/products/hydrobasins, accessed on 15 December 2023). The area of tributary watersheds has been proven to significantly influence the degree of lake eutrophication because larger watershed areas of tributaries tend to introduce more nutrients into lakes, such as nitrogen and phosphorus [56]. These nutrients are the primary drivers of algal blooms, which further exacerbate the eutrophication process [55]. This phenomenon is also corroborated by the satellite retrieval results of Liangzi Lake, where the influx of nutrients into the lake through the Gaoqiao River led to mild eutrophication in the southeastern part of East Liangzi Lake.

Moreover, variations in watershed management and land use types may also impact the eutrophic status of lakes. Generally, agricultural fields can generate excessive nutrients due to the application of chemical fertilizers by humans, which subsequently enter lakes through hydrological processes or drainage systems [57]. Additionally, the discharge of municipal sewage contributes to the accumulation of nutrients in rivers, ultimately reaching and promoting eutrophication in recipient lakes [58]. We extracted the land use data from the China Land Cover Dataset (CLCD) to characterize human activities on lake eutrophication [59]. As shown in Figure 10, the areas of agricultural land and urban areas (impervious surfaces) of the East Liangzi Lake basin are the highest among the sub-lakes, amounting to 689.94 km² and 18.21 km², respectively. This could be one of the reasons for the mild eutrophication observed in the southeastern region of the lake.

In summary, the area of tributary watersheds and land use have a significant impact on the long-term eutrophication status of Liangzi Lake, but it is also emphasized that this impact is determined by a combination of various environmental and anthropogenic factors. Considering the detrimental effects of eutrophication on lake ecosystems, various mitigation strategies have been proposed. These include the implementation of nutrient management plans, the restoration of wetlands adjacent to lakes to serve as natural biofilters, and the enforcement of stricter regulations on industrial discharges and agricultural runoff. Additionally, ongoing monitoring and adaptive management strategies are essential to effectively mitigate eutrophication and preserve aquatic biodiversity in East Liangzi Lake. Future research needs to explore these interactions more deeply to more accurately predict and manage lake eutrophication issues.

4.4. The Eutrophication Level of Lakes in Wuhan

We attempted to use the trained model for promoted applications in areas of the lake without monitoring sites or stations. Therefore, the trained model was utilized to evaluate the eutrophication status of typical lakes in Wuhan in 2022. As shown in Figure 11, the results indicate that the water quality of Xiajiaci Reservoir is significantly better in terms of eutrophication compared to other lakes, with the mean TLI values being 23 and there being a maximum value of only 31 in 2022. The area of the eutrophication in this lake is predominantly near the shoreline. This phenomenon can be attributed to the reservoir’s location within a tourist area, far from densely populated regions, thus less affected by human activities. In contrast, Dong Lake is also situated within a tourist area that has consistently been under government attention and management and exhibits a less favorable eutrophication status than Xiajiaci Reservoir. This may be because East Lake is located in the center of the city and collects some of the domestic sewage and pollutants from the urban subsurface with the drainage of precipitation. However, its water quality has been steadily improving due to the sustained efforts of water management [60]. Furthermore, Futou Lake has been extensively used for fish farming and demonstrates a distinct boundary by the TLI. It indicated that aquaculture activities have a direct and significant impact on lake eutrophication status. In conclusion, the application of this model for water quality inversion not only demonstrates its effectiveness but also highlights its potential for providing significant guidance and broader applicability in the management and conservation of lake water quality.

5. Conclusions

This study compared different ML methods of the remote sensing retrieval algorithm to extract the distribution of the TLI in Liangzi Lake from 2014 to 2022 based on Landsat 8 satellite images. The results showed that RF is the most appropriate machine learning model to be used directly to reflect the TLI and indicated a simulation accuracy that far exceeds the level of other models. According to the results of the quantitative assessment of the eutrophication level by the TLI, the overall eutrophication level of the lake was controlled below the mild eutrophication level but showed a slow increasing trend year by year. Spatially, East Liangzi Lake exhibits a perennial mild eutrophication level and shows a higher risk of eutrophication than the other two sub-lakes. In addition, the combination of historical meteorological information and land use data suggests that lake eutrophication is likely to be influenced by temperature and human activities, especially temperature, which shows a high correlation with lake eutrophication levels. In the future, better high-resolution satellite imagery and multispectral data will drive further improvement in retrieval methods and extensive spatiotemporal monitoring measurements on eutrophication status.