Machine Learning-Based Remote Sensing Inversion and Spatiotemporal Characterization of Chl-a Concentration in the Leizhou Peninsula Coastal Waters

Abstract

The Leizhou Peninsula, located in the northern South China Sea, features coastal waters with dual functions as both marine ranch demonstration zones and ecological protection areas. Remote sensing monitoring of Chlorophyll-a (Chl-a) concentration in this region holds strategic significance for assessing primary productivity, red tide risk, and the sustainability of the blue food economy. This study integrates in situ survey data from four cruises conducted between 2020 and 2024 with Sentinel-3 OLCI remote sensing imagery, constructing and comparing the performance of six machine learning inversion models. The results show that for the inversion scenarios of the Leizhou Peninsula waters, the GBDT model performs best among the evaluated models (R² = 0.79, RMSE = 0.36 mg/m³, MAE = 0.30 mg/m³). Based on the GBDT model, pixel-by-pixel inversion maps with 300 m spatial resolution were generated for four seasons in 2024, revealing a spatial gradient of “high nearshore–low offshore, high in the east–low in the west” and a seasonal pattern of “low in spring–rise in summer–stable in autumn–high in winter.” In addition, the study verified the operational potential of machine learning in complex type-II waters, analyzed the distribution characteristics and influencing factors of Chl-a concentration in this region, and provided scientific and technical support for marine ranch carrying capacity assessment, eutrophication early warning, and carbon sink accounting in the Leizhou Peninsula.

1. Introduction

The Leizhou Peninsula, one of the three major peninsulas in China, is located at the southernmost point of the Chinese mainland, bordering the northern South China Sea. Its coastline is highly indented, with numerous bays and islands. To the east lie natural harbors such as Liusha Bay, Zhanjiang Bay, and Leizhou Bay; to the south is the Qiongzhou Strait and to the west is the Beibu Gulf [1]. The region hosts diverse coastal ecosystems and rich fishery resources. In recent years, the combined impacts of riverine input, industrial pollution, and agricultural and aquacultural wastewater discharge have led to increasingly complex water quality conditions, with elevated concentrations of nitrogen, phosphorus, and other nutrients that support phytoplankton proliferation [2,3,4]. As the marine economy and marine ranching continue to expand, the impact of human activities has intensified, and the assessment of nearshore water quality has become an issue of growing concern.

Marine ecological monitoring has long been an important research focus in hydrological remote sensing and ocean science. Among its indicators, Chlorophyll-a (Chl-a), as the core driving factor of marine primary productivity, is commonly used to track changes in phytoplankton [5,6]. Phytoplankton provide energy for marine ecosystems through photosynthesis and form the base of the marine food web. Their role in fixing atmospheric or dissolved CO₂ into organic carbon constitutes a critical process in the global carbon cycle [7]. Therefore, real-time and accurate monitoring of Chl-a concentrations is essential for the early identification of eutrophication or algal bloom risks. It also serves as a key basis for decision-making in feed management, capacity control, and ecological restoration in marine ranches, and is fundamental to ensuring the sustainable use of the “blue food economy” and maintaining nearshore ecological security.

The remote sensing inversion of Chl-a concentration has evolved through three major stages: empirical models, semi-analytical models, and machine learning approaches [8]. Early OCx-series algorithms performed well in open ocean waters [9] but showed significant errors in turbid nearshore waters dominated by CDOM and suspended matter [10]. Later, semi-analytical models such as QAA and GIOP incorporated the decomposition of inherent optical properties [11], improving their applicability to optically complex (Type II) waters. In recent years, machine learning algorithms such as Random Forest, XGBoost, and GBDT have demonstrated strong capabilities in modeling high-dimensional nonlinear relationships [12,13]. These approaches have been shown to improve Chl-a inversion accuracy by more than 15–30 percent in complex estuarine and coastal waters such as the Pearl River Estuary and Bohai–Yellow Sea [4]. In the field of ocean color satellite remote sensing inversion, multi-source sensors such as Sentinel-3 OLCI, Sentinel-2 MSI, GF-6 WFV, and HY-1C/D COCTS have been widely applied to the retrieval of surface Chl-a concentration, CDOM, suspended matter, and nutrients [14,15,16].

For the Leizhou Peninsula, existing studies have mainly focused on individual bays such as Zhanjiang Bay and Liusha Bay, relying on short-term and scattered cruise surveys. There is a lack of systematic and continuous spatiotemporal Chl-a concentration products covering the broader Leizhou Peninsula–Qiongzhou Strait region in recent years. Moreover, the advantages of machine learning approaches under optically complex environments have not been fully explored [17,18,19]. Additionally, the synergistic variations of Chl-a concentration with CDOM, suspended matter, temperature, salinity, and turbidity—as well as the quantitative influence of anthropogenic activities—remain unclear. The overall technical framework of this study is illustrated in Figure 1. The main objectives are as follows: (1) to conduct seasonal offshore surveys of Chl-a concentrations, water quality parameters, and spectral data in the Leizhou Peninsula coastal waters; (2) to explore various machine learning methods and develop seasonal remote sensing inversion models of surface Chl-a concentration and (3) to investigate the seasonal patterns and influencing factors of Chl-a concentration in relation to sea surface temperature, turbidity, and salinity. Through this study, we aim to obtain a comprehensive understanding of the dynamic variations in Chl-a concentration across the Leizhou Peninsula region and provide scientific support for its ecosystem management and protection.

2. Materials and Methods

The overall technical framework of this study is shown in Figure 1. Taking Chl-a in the coastal waters of the Leizhou Peninsula as the research object, Sentinel-3 OLCI remote sensing imagery was used and subjected to preprocessing operations such as atmospheric correction and resampling. At the same time, in situ measurements obtained from water quality instruments and spectrometers were combined to construct six machine learning inversion models. Subsequently, the best-performing inversion model among the evaluated ones in this study was applied to conduct seasonal remote sensing inversion, thereby obtaining the Chl-a distribution patterns across different seasons. Finally, analysis and discussion were carried out with the aim of achieving machine learning-based remote sensing inversion and spatiotemporal feature analysis of Chl-a concentration in the nearshore waters of the Leizhou Peninsula.

2.1. Study Area

The coastal waters of the Leizhou Peninsula (19.8° N–21.8° N, 109° E–111° E) are situated in the northern tropical marine monsoon climate zone (Figure 2a) and exhibit transitional ecological characteristics between tropical and subtropical marine systems. The region experiences a warm and humid climate, with an annual average temperature ranging from 22 to 26 °C and abundant precipitation. Combined with substantial freshwater inflow from nearby rivers, these conditions contribute to a eutrophic aquatic environment that provides ideal conditions for phytoplankton growth [20]. As an important land–sea interaction zone in the northern South China Sea, this area is influenced by a complex hydrodynamic regime involving small river systems, coastal currents, and the South China Sea circulation. These multiple dynamic drivers result in a coupled hydrodynamic structure that regulates the physicochemical environment of the euphotic zone through vertical mixing processes, while also affecting the spatial redistribution of nutrient fluxes [21,22]. According to the water body classification criteria of the International Ocean Colour Coordinating Group (IOCCG), the coastal waters of the Leizhou Peninsula are predominantly classified as optically complex (Type II) waters. Other studies, based on the National Seawater Quality Standards of China, have provided more detailed classifications: Zhanjiang Bay and the Jianjiang River estuary are designated as Class IV waters; southeastern Donghai Island, northeastern Xuwen, and Liusha Bay are classified as Class III; while the remaining areas are designated as Class I or II [23]. Therefore, this region is characterized by both nearshore turbid waters dominated by suspended particles and offshore clearer waters with open-ocean spectral features, making it an ideal case study area for ocean color remote sensing inversion and dynamic tracking of coastal Chl-a concentrations.

2.2. In Situ Sampling and Data Processing Procedures

This study utilized field sampling data collected during four oceanographic cruises in the coastal waters of the Leizhou Peninsula between September 2020 and July 2024. A total of 135 matched in situ measurements of remote sensing reflectance (Rrs) and Chl-a concentrations were obtained. Sampling stations were distributed along the eastern, southern, and western coasts of the Leizhou Peninsula, as shown in Figure 2b. Remote sensing reflectance spectra were acquired using a TriOS RAMSES-SW surface hyperspectral radiometer (TriOS GmbH, Oldenburg, Germany), with a spectral range of 320–950 nm and a spectral resolution of 0.3 nm. Measurements were conducted above the water surface following the observation geometry recommended by NASA: the azimuth angle between the instrument’s viewing plane and the solar plane was set to 90–135°, and the sensor zenith angle was maintained at 30–40° relative to the nadir direction of the water surface [24]. The above-water remote sensing reflectance data were ultimately obtained, with the specific calculation formula as follows:

$$R_{\mathrm{rs}}\left(\lambda \right)={\displaystyle \frac{L_{\mathrm{u}}\left(\lambda \right)-L_{\mathrm{sky}}\left(\lambda \right)\ast r_{\mathrm{sky}}}{E_{\mathrm{d}}\left(0^{+}\right)}}$$

(1)

In the above equation, $L_{\mathrm{u}}\left(\lambda \right)$ represents the water–leaving radiance, $L_{\mathrm{sky}}\left(\lambda \right)$ denotes the sky radiance, and $r_{\mathrm{sky}}$ is the surface reflectance factor for sky light at the air–water interface, typically ranging from 0.022 to 0.028; a value of 0.025 was adopted in this study. $E_{\mathrm{d}}\left(0^{+}\right)$ is the total downwelling irradiance incident at the water surface.

Water quality parameters were measured in situ using the RBRmaestro multiparameter water quality instrument (RBR Ltd., Ottawa, ON, Canada). This study mainly recorded four water quality data points: Chl-a concentration, salinity, temperature, and turbidity. The Chl-a sensor module operates on a fluorescence excitation principle, with an excitation wavelength of 470 nm, a detection wavelength of 685 nm, and a resolution of 0.02 mg/m³. Surface Chl-a concentrations were calculated as the average of multiple measurements within the upper 1 m water column. The RBRmaestro system employs a dynamic baseline calibration technique for its sensors, which effectively minimizes interferences caused by turbidity and colored dissolved organic matter (CDOM) in the water. This enhances the measurement accuracy under optically complex coastal water conditions [25].

2.3. Remote Sensing Data Acquisition and Preprocessing

Sentinel-3 Ocean and Land Colour Instrument (OLCI) remote sensing data are freely accessible through the European Space Agency (ESA) Copernicus Open Data Platform (https://dataspace.copernicus.eu). Launched in February 2016, the Sentinel-3 satellite provides a spatial resolution of 300 m and a swath width of 1270 km, enabling daily global coverage. The satellite is equipped with seven sensors, among which the OLCI, an advanced successor to the Medium Resolution Imaging Spectrometer (MERIS) onboard the ENVISAT satellite, is optimized for ocean color observations in coastal regions. The OLCI provides 21 spectral bands ranging from visible to near-infrared wavelengths (400–1020 nm) (

Table 1. Sentinel-3 OLCI spectral band parameters and primary applications of each band.

Band	Centre Wavelength (nm)	Primary Application
Oa1	400	Aerosol correction
Oa2	412.5	Retrieval of CDOM and terrestrial substances
Oa3	442.5	Chl-a absorption retrieval
Oa4	490	Retrieval of high Chl-a in inland waters
Oa5	510	Monitoring of HABs in marine waters
Oa6	560	Benchmarking of Chl-a retrieval in aquatic systems
Oa7	620	Retrieval of SPM concentration
Oa8	665	Inversion of water-quality parameters and phytoplankton
Oa9	673.75	Utilizing the 665 and 681nm spectral bands
Oa10	681.25	Provision of Chl-a fluorescence peak information
Oa11	708.25	Provide fluorescence peak detection baseline values
Oa12	753.75	Vegetation monitoring
Oa13	761.25	Aerosol retrieval
Oa14	764.375	Atmospheric correction
Oa15	767.5	Terrestrial fluorescence intensity information
Oa16	778.75	Participation in aerosol correction
Oa17	865	Cloud detection
Oa18	885	Water vapor absorption reference and vegetation monitoring
Oa19	900	Vegetation monitoring
Oa20	940	Water vapor absorption reference
Oa21	1020	Participation in aerosol correction

), supporting applications in numerical ocean forecasting, maritime safety, atmospheric monitoring, and coastal environmental management [20,26].

Additionally, specific bands, such as B4 (490 nm) for the absorption valley, B10 (681.25 nm) for the fluorescence peak, and B11 (708.25 nm) for compensation in turbid waters, are optimized for the remote sensing characteristics of Chlorophyll-a (Chl-a). These bands provide a reliable data source for the quantitative retrieval of Chl-a in nearshore waters, enhancing the operational accuracy of remote sensing applications, including red tide monitoring and eutrophication assessment.

To investigate the spatial dynamics of Chl-a concentrations in the coastal waters of Leizhou Peninsula, this study utilized 20 high-quality Level-2 Full Resolution Water and Atmospheric Geophysical Products derived from the Sentinel-3 Ocean and Land Colour Instrument (OLCI) sensor, covering the entire year of 2024. The dataset includes five images per season, representing spring, summer, autumn, and winter, to capture seasonal variations in remote sensing data. Since the spectral resolution of Sentinel-3 satellite bands differs from that of the measured spectral data, the experiment resampled the above-water remote sensing reflectance from 135 stations to the central wavelengths of the satellite channels by applying bandwidth-weighted averaging based on the spectral response function of each satellite band. This approach preserves the spectral characteristics while significantly improving data consistency and usability.

The cloud masking of Sentinel-3 Level-2 (OL_2_WFR) data mainly relies on the flags CLOUD, CLOUD_AMBIGUOUS, and CLOUD_MARGIN provided in the WQSF (Water Quality and Science Flags) file. By identifying these flags, valid pixels without cloud contamination can be selected. The core of atmospheric correction is based on the main neural network algorithm, whose logic involves multilayer iterative computations that combine multi-band raw data and atmospheric parameters to separate and remove atmospheric contributions, ultimately outputting above-water remote sensing reflectance, which is particularly well-suited for marine and coastal water scenarios [27,28].

In addition, the OLCI sensor includes a special band at 1020 nm designed to improve atmospheric correction. By exploiting the weak water reflectance and strong aerosol backscattering at this wavelength, aerosol backscattering information can be captured to help eliminate its interference with water reflectance spectra and restore more realistic optical characteristics of the water body. To further suppress noise, the algorithm removes bands B13, B14, B15, B19, and B20 affected by atmospheric gas absorption after correction, significantly improving the accuracy and reliability of above-water reflectance data in the Level-2 products [29].

During the atmospheric correction validation process, the time interval between satellite image acquisition and in situ measurements often exceeded 24 h. Since water bodies are subject to hydrodynamic and environmental influences, they exhibit dynamic variability, which may cause discrepancies between the two observations [30]. To mitigate the impact of such discrepancies on validation results, this study followed existing research approaches [31,32] and prioritized star–ground match-up points with low cloud coverage and minimal sun-glint interference, thereby reducing the influence of external environmental noise on data reliability [33]. After atmospheric correction, the surrounding areas of the selected match-up points showed good water quality homogeneity, which further provided favorable conditions for accurate validation.

Considering the typical spectral absorption and reflection characteristics of Chl-a in water and the wavelength range of in situ spectral measurements in the Leizhou Peninsula, bands B1 (400 nm) to B11 (708.25 nm) were selected as the study range. Figure 3 compares near-synchronous remote sensing images with field survey data. Although the atmospheric-corrected Sentinel-3A OLCI reflectance shows certain degrees of overestimation or underestimation compared with in situ reflectance, the overall atmospheric correction results are satisfactory, allowing for subsequent inversion model construction.

2.4. Development of Machine-Learning-Based Algorithmic Models

Remote sensing inversion of Chl-a concentration aims to derive the Chl-a content in water bodies from spectral data, revealing the dynamic changes in phytoplankton and water quality characteristics. However, the complex nonlinear relationships between marine environmental parameters and spectral data often limit the performance of traditional inversion methods [34]. In contrast, machine learning, through a data-driven approach, effectively captures multidimensional nonlinear associations between spectral data and Chl-a concentration, significantly enhancing the accuracy of inversion models. Within the machine learning framework, supervised learning focuses on establishing a mapping from input features to output labels by learning from labeled data, thereby generating a predictive model. This core principle aligns closely with the requirements of Chl-a concentration inversion models, enabling improved robustness and accuracy through the construction of nonlinear models [35,36]. This study compares six supervised learning methods, including Backpropagation Neural Network (BP), Multilayer Perceptron (MLP), Support Vector Machine Regression (SVR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Gradient Boosting Decision Tree (GBDT), and evaluates their accuracy and applicability for water color remote sensing inversion in the coastal waters of the Leizhou Peninsula.

To construct robust machine learning inversion models, the dataset is initially split using a standard eighty-twenty ratio, with eighty percent allocated to the training set and twenty percent reserved for the test set, ensuring stable and reliable performance in real-world applications. All feature selection and hyper-parameter tuning were performed exclusively within the training data using inner K-fold cross-validation: in each inner fold, we recomputed the correlation ranking, selected the top-k combinations, tuned model hyper-parameters via GridSearchCV (MAE), and evaluated on the inner validation split. The final model was refit on the full training set with the selected features and parameters, and evaluated once on the held-out test set. Model construction serves as a key step in analyzing the spatiotemporal distribution characteristics of Chl-a concentration in the Leizhou Peninsula waters. Related studies have shown that systematically analyzing the reflectance of all spectral bands using two-band and three-band combinations is an effective strategy to improve the accuracy of inversion models [13,37]. Building on the selection of bands B1–B11 as the spectral data study range, correlation analysis was conducted between the spectral combination data of resampled bands and corresponding in situ Chl-a measurements. The formula for calculating the correlation coefficient is as follows: The Pearson correlation coefficient r is calculated as follows:

$$r={\displaystyle \frac{{\sum }_{i=1}^n(x_i-\overline{x})(y_i-\overline{y})}{\sqrt{{\sum }_{i=1}^n{(x_i-\overline{x})}^2{\sum }_{i=1}^n{(y_i-\overline{y})}^2}}}$$

(2)

where n represents the number of in situ samples, $\overline{x}$ and $\overline{y}$ are the sample means.

The band combinations were derived from arithmetic operations on the resampled reflectance data. Given the large number of bands, the experiment used Python algorithms to traverse all bands and construct combinations of band sums, band differences, band ratios, and multi-band parameters. In total, 19,339 band-combination correlation datasets were generated, with each combination evaluated by its correlation coefficient. The experimental results are shown in Figure 4 (the polar coordinate plot is used only for intuitive visualization of the correlation coefficient magnitude; the radial distance from the origin to the data point represents the absolute value of the correlation coefficient, while the angle has no physical or statistical meaning). Notably, some band combinations exhibited correlation values with Chl-a concentration exceeding 0.55, more than twice the correlation achieved by single bands with measured Chl-a, further confirming that band combinations can effectively enhance model prediction accuracy.

Based on the correlation analysis of band combinations, this study selected 12 highly correlated band combinations as candidate input features for the subsequent experiments (

Table 2. Candidate band combinations and their correlation coefficients for machine learning modeling.

No.	Band Combination	Correlation Coefficient	No.	Band Combination	Correlation Coefficient
1	B6/B5	0.55	7	(B6 + B2)/B3	0.58
2	(B2 + B6)/B4	0.65	8	(B6 − B5)/B3	0.58
3	(B6 + B1)/B3	0.60	9	(B6 − B5)/B4	0.56
4	(B6 + B1)/B4	0.64	10	(B6 − B4)/B3	0.56
5	(B6 + B4)/B5	0.57	11	(B5 − B6)/B2	0.55
6	(B6 + B3)/B4	0.59	12	(B5 − B6)/B5	0.55

). During training, K-fold cross-validation was used for feature selection to avoid the risk of data leakage, and machine learning models for Chl-a concentration inversion in the Leizhou Peninsula waters were constructed to improve inversion accuracy and robustness.

2.5. Model Performance Evaluation Metrics

To better evaluate the predictive performance of the machine learning inversion model, this study employed three commonly used metrics for error analysis: the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). These metrics quantify the model’s goodness of fit and prediction error from different perspectives. R² is typically used to assess the model’s predictive performance, while RMSE and MAE evaluate the consistency between measured and predicted values. Their respective formulas are as follows:

The coefficient of determination (R²) measures the degree of fit between the model’s predicted values and the actual values, with values typically ranging from 0 to 1. A value closer to 1 indicates a superior predictive model. The coefficient of determination $R^2$ is defined as follows:

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

(3)

$\mathrm{Root}\mathrm{Mean}\mathrm{Square}\mathrm{Error}\left(\mathrm{RMSE}\right)$: reflects the standard deviation of the prediction errors between predicted and true values; smaller values indicate higher prediction accuracy.

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

(4)

$\mathrm{Mean}\mathrm{Absolute}\mathrm{Error}\left(\mathrm{MAE}\right)$: represents the average magnitude of the absolute errors between predicted and true values, reflecting the average prediction error.

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

(5)

where n represents the number of samples, $y_i$ is the observed value of the i-th sample, ${\widehat{y}}_i$ is the predicted value by the model, and $\overline{y}$ is the mean of all observed values.

These three metrics provide a comprehensive evaluation of the model from multiple dimensions, including overall fitting capability, error variability, and average error level. They offer a quantitative basis for assessing the performance of different machine learning algorithms in the Chl-a concentration retrieval task, thereby facilitating the selection of the best-performing inversion model among the evaluated ones for subsequent application studies.

3. Results

3.1. Descriptive Statistical Analysis of In-Situ Survey Data

The plotted remote sensing reflectance data (Supplementary Materials) indicate that the nearshore waters of the Leizhou Peninsula exhibit the optical characteristics of mixed turbid water. The in-situ measured reflectance spectra show significant fluctuations within the 400–700 nm wavelength range, reflecting the combined influence of multiple water quality constituents. In the violet-blue band around 400–450 nm, absorption is mainly dominated by colored dissolved organic matter (CDOM) and Chl-a [38], resulting in generally low reflectance and noticeable spectral differences across stations. In the green band from 500 to 580 nm, reflectance increases significantly, with most curves showing distinct reflectance peaks around 490 nm or 570 nm. This green spectral region is influenced by both phytoplankton and CDOM absorption characteristics, as well as backscattering from suspended sediments [39]. However, due to spatial variations in Chl-a concentrations, the peak height and position still exhibit certain fluctuations within this range. In the red band from 600 to 700 nm, reflectance drops significantly due to enhanced Chl-a absorption, forming a typical Chl-a absorption trough around 670 nm, indicating strong phytoplankton biomass absorption. Subsequently, a characteristic fluorescence peak appears near 680 nm due to the phytoplankton fluorescence effect [40]. Beyond 690 nm, reflectance decreases as a result of increased pure water absorption. Near 760 nm, a narrow reflectance trough occurs due to strong water molecule absorption and low or absent phycocyanin concentrations. In the near-infrared spectrum, the reflectance remains non-zero due to the continuous contribution of backscattering by suspended particulates, further confirming the turbid nature of the nearshore waters in the study area [41,42,43].

Overall, the spectral reflectance characteristics of remote sensing are significantly influenced by three main water quality factors: CDOM, suspended matter, and Chl-a, with the green and red wavelength bands being the most sensitive. Specifically, in the green band, the peak height, width, and position of reflectance vary distinctly with differences in Chl-a and CDOM concentrations. In the red band, the strong absorption valley and fluorescence peak of Chl-a further reveal the spatial distribution patterns of phytoplankton biomass. Based on this analysis and related studies, the spectra of the study area can be classified into four types (Figure 5): the first type typically exhibits a high, narrow reflectance peak in the 500–580 nm range; the second type, with a broad reflectance peak, indicates increased suspended matter content and enhanced optical complexity; the third type shows a flat reflectance curve with no distinct peak or valley and the fourth type displays a monotonous decrease with increasing wavelength. These classifications effectively reflect the optical properties and spatial variations of water quality elements in the Leizhou Peninsula coastal waters [44,45].

As shown in

Table 3. Statistical ranges and averages of in situ water quality parameters during different sampling periods.

Sampling Period	Number	Chl-a (mg/m3)		Temp (°C)		Salinity (PSU)		Turbidity (NTU)
		Range	Mean	Range	Mean	Range	Mean	Range	Mean
Spring	33	0.24–1.03	0.57	25.68–27.74	26.55	30.16–33.52	32.19	0.48–19.59	3.99
Summer	35	0.06–3.47	0.96	29.23–32.67	31.20	25.85–32.60	30.93	0.24–32.16	5.22
Autumn	37	0.12–2.53	0.81	28.61–31.25	30.17	28.38–33.76	31.89	0.06–29.63	5.86
Winter	30	0.73–2.45	1.42	18.5–21.20	20.33	31.22–33.69	32.74	0.48–10.16	3.99

, a preliminary statistical analysis of in-situ water quality parameters collected during four cruises from September 2020 to July 2024 indicates a clear seasonal variation in Chl-a concentrations in the coastal waters of the Leizhou Peninsula, reflecting the response of phytoplankton primary productivity to varying environmental conditions. In spring (April 2021 Voyage cruise), the average Chl-a concentration across 33 sampling stations was 0.57 mg/m³, with a generally low level and narrow fluctuation range. The in-situ water temperature was 26.55 °C, indicative of a typical early spring stage characterized by low biomass and oligotrophic conditions [46]. During the summer (July 2024 Voyage), the average Chl-a concentration at 35 stations rose to 0.96 mg/m³, with values ranging from 0.06 mg/m³ to 3.47 mg/m³. The significantly widened fluctuation range and an average temperature of 31.20 °C suggest that high temperature, strong irradiance, and stratified waters in summer jointly stimulated rapid phytoplankton growth, resulting in elevated biomass levels. In autumn (September 2020 Voyage), the 37 surveyed stations exhibited an average Chl-a concentration of 0.81 mg/m³, with a maximum of 2.53 mg/m³ and a minimum of only 0.12 mg/m³, indicating a spatially uneven distribution. Some stations may have experienced concentrated phytoplankton blooms, typical of a post-summer biomass release period when phytoplankton flourish again [47]. In winter (January 2022 Voyage), the 30 sampled stations recorded the highest annual average Chl-a concentration of 1.42 mg/m³, with a maximum of 2.45 mg/m³ and a minimum of 0.73 mg/m³, indicating that phytoplankton biomass remained relatively high even under low-temperature marine conditions [48,49].

Along with changes in Chl-a concentrations, environmental factors also exhibited variations. According to the ranges and averages shown in Table 3, the average salinity generally remained high, ranging from 30.93 to 32.74 PSU. Turbidity peaked in summer and autumn, reaching 5.22 NTU and 5.86 NTU, respectively. The highest turbidity occurred in summer, likely influenced by riverine input and coupled with the influx of land-derived suspended matter. These findings indicate that phytoplankton primary productivity in the coastal waters of the Leizhou Peninsula responds sensitively to seasonal environmental changes, exhibiting notable spatiotemporal heterogeneity and ecological indicative value.

3.2. Machine Learning Inversion Model Performance Evaluation

The experimental results in Figure 6 demonstrate that machine learning models significantly outperform traditional inversion models in terms of accuracy on both training and test datasets. Among the evaluated algorithms, the Gradient Boosting Decision Tree (GBDT) was identified as the best-performing model among the evaluated algorithms, achieving a correlation coefficient (R²) of 0.79, RMSE of 0.36, and MAE of 0.30 on the test set. GBDT is an ensemble learning algorithm based on the boosting concept, which constructs a series of weak learners and aggregates the results of multiple decision trees to produce the final predictive output, thereby achieving a high-precision approximation of complex nonlinear relationships.

The advantages of the GBDT model primarily stem from its multi-model ensemble learning strategy and its ability to handle complex data relationships [50]. First, the iterative additive model structure enables GBDT to effectively capture high-order nonlinear interactions between bands when dealing with the complex relationships between Chl-a concentration and reflectance, without requiring additional construction of interaction terms [51]. Second, GBDT exhibits high robustness to noise and outliers, naturally suppressing these interferences. Compared with neural networks and Support Vector Regression (SVR), GBDT is less sensitive to noise, making it particularly suitable for the small dataset used in this study (135 samples). GBDT can progressively improve model predictions through weighted trees without relying on large data samples, as required by deep learning methods. Another advantage of GBDT is its high computational efficiency on small to medium-sized datasets compared with neural networks, support vector machines, and random forests. By optimizing learning rate and iteration counts, GBDT achieves smooth convergence of the error curve [52]. Although optimized gradient boosting algorithms like XGBoost offer improved performance, GBDT provides satisfactory results for small datasets with fewer hyperparameters, making tuning simpler.

Overall, GBDT’s advantages—its ensemble tree structure, nonlinear modeling capability, robustness to noise, and high training efficiency—make it widely applicable in remote sensing inversion. Unlike the parallel ensemble approach of Random Forest (RF), GBDT trains tree models sequentially, optimizing residuals from the previous round in each iteration, which enhances its ability to correct model errors. Consequently, GBDT demonstrates stronger adaptability and explanatory power for handling hierarchical and correlated spectral data structures [53,54].

3.3. Spatiotemporal Distribution of Chl-a Concentration Retrieved in the Study Area

Based on the GBDT algorithm, this study conducted pixel-by-pixel inversion of Chlorophyll-a (Chl-a) concentrations in the Leizhou Peninsula coastal waters for the year 2024 and constructed seasonal spatiotemporal distribution maps. A total of 20 high-quality Sentinel-3 OLCI images were selected for the year, with five images per season. Within each season, the inversion results of the five images were averaged pixel-by-pixel to generate the seasonal spatial distribution map, effectively suppressing transient noise such as clouds and sun glint. As shown in Figure 7, the Chl-a concentration exhibits significant seasonal differences in both temporal and spatial dimensions, reflecting the spatiotemporal dynamic patterns of phytoplankton biomass with seasonal succession.

From the spatial distribution perspective, Chl-a concentrations in the Leizhou Peninsula waters generally exhibit a pronounced “higher nearshore, lower offshore” gradient pattern, with distinct differences between the eastern and western marine areas. Higher Chl-a concentrations are observed in nearshore regions such as the entrance of Zhanjiang Bay, Yingluo Port, and both sides of the Qiongzhou Strait [3]. The east–west disparity is characterized by significantly higher Chl-a concentrations in the eastern waters, where some high-value areas exceed 3 mg/m³, while the western waters mostly show concentrations below 1 mg/m³ with relatively uniform distribution. Terrestrial input and human activities are the key drivers of this spatial heterogeneity between the eastern and western parts of the Leizhou Peninsula waters [55].

Overall, the Chl-a concentration in the Leizhou Peninsula waters follows a temporal pattern of “low in spring, rising in summer, stable in autumn, and high in winter,” reflecting the seasonal variation and cyclical fluctuations in phytoplankton biomass within the water body.

4. Discussion

4.1. Applicability of Machine-Learning-Based Inversion Models

The core applicability of machine learning-based inversion models in complex water bodies lies in their adaptive capability to represent nonlinear optical coupling relationships. The nearshore waters of the Leizhou Peninsula are jointly influenced by terrestrial suspended matter and colored dissolved organic matter (CDOM). Traditional semi-analytical algorithms rely heavily on single bands or fixed assumptions, making it difficult to resolve the spectral overlap between Chl-a concentration and water quality parameters such as suspended sediments and CDOM. In contrast, machine learning models like Random Forest, XGBoost, and neural networks map data into high-dimensional feature spaces, automatically capturing nonlinear interactions among multiple spectral features of Chl-a, including reflectance peaks and troughs, spectral slopes, and near-infrared backscattering. This enables spatially continuous remote sensing inversion in water bodies with complex optical properties and transitional zones [56,57]. In this study, the hyperparameters of the GBDT model were tuned using GridSearchCV, with candidate parameters including n_estimators (number of trees), learning_rate (learning rate), and max_depth (maximum depth). Grid search employed K-fold cross-validation to minimize the Mean Absolute Error (MAE), with a random seed set to 60 to ensure reproducible results. The machine learning model demonstrated a robust response to local optical variations, providing a scalable framework for high-frequency monitoring of phytoplankton dynamics [58].

However, the applicability boundaries are constrained by two types of uncertainty: One of these is conceptual uncertainty caused by the combined influence of various factors on optics. When suspended sediments and CDOM concentrations increase simultaneously, their combined absorption and scattering in the visible to near-infrared bands compress the independent variance of Chl-a sensitive features, causing the model to misclassify high CDOM as high Chl-a. Second, spatiotemporal heterogeneity leads to extrapolation uncertainty. Extreme turbid waters or algal bloom events are underrepresented in training samples, making the model prone to empirical extrapolation beyond the optical-biogeochemical space, resulting in physically meaningless negative or saturated output values [59,60]. In the prediction results shown in Figure 6f, the shaded area represents the confidence interval of the predictions, indicating the uncertainty of the model’s results. It can be observed that the model exhibits greater prediction uncertainty and more pronounced errors on the test set, particularly in high-value regions where the confidence interval widens, reflecting the model’s generalization error on new data. Therefore, in the future, actively incorporating extreme scenario samples through active learning strategies and embedding semi-analytical constraints into the network’s loss function as a form of regularization will be necessary. This approach can enhance the interpretability and long-term stability of machine learning models in complex water bodies while maintaining their flexibility.

4.2. Spatiotemporal Distribution Characteristics of Chl-a Concentration and Their Driving Factors in the Leizhou Peninsula Waters

To investigate the driving mechanisms of water quality parameters on the spatiotemporal patterns of surface Chl-a concentration in the Leizhou Peninsula coastal waters, this study selected quasi-synchronously obtained in situ sea surface temperature, salinity, and turbidity data (Figure 8) to systematically assess their potential influence on Chl-a variability.

From a spatial distribution perspective, the eastern coastal waters of the Leizhou Peninsula are dominated by irregular semi-diurnal tides, which promote vertical transport of bottom materials through mixing. Additionally, the diluted water from the Pearl River forms a low-salinity plume front, combined with inputs from rivers such as the Jianjiang and nitrogen-phosphorus loads from industrial wastewater and high-density aquaculture in Zhanjiang Port. These factors collectively create a eutrophic environment, continuously stimulating phytoplankton proliferation [61]. In contrast, the western coastal waters, located in the semi-enclosed Beibu Gulf, are influenced by regular diurnal tides. The enclosed topography of the Beibu Gulf leads to water retention, resulting in significantly lower nutrient replenishment efficiency compared with the eastern region, which limits biological productivity. Due to minimal river runoff and low human activity interference, nutrients in this area are primarily from natural sources, suppressing biological productivity [20]. The northwest coastal waters of the Leizhou Peninsula maintain high Chl-a concentrations throughout the year, particularly during the rainy summer and autumn seasons, when abundant runoff carries terrestrial nutrients and suspended matter, forming a persistent high-concentration zone [62]. Hydrological dynamic processes in the coastal waters of the Leizhou Peninsula further amplify the east-west differences by regulating nutrient transport [63]. During summer (June to September), when the southwest monsoon prevails, peak coastal currents transport large amounts of nutrients to the eastern nearshore, but strong seawater stratification hinders offshore nutrient diffusion, resulting in high Chl-a concentrations concentrated nearshore. In winter, strong northeast monsoons drive vertical water movement, effectively uplifting nutrient-rich bottom waters, while an enhanced mixed layer depth further strengthens vertical nutrient replenishment, promoting phytoplankton growth and significantly expanding the high Chl-a concentration zones [64]. In spring and autumn, high wind speeds and lower average temperatures compared with summer limit phytoplankton growth, resulting in lower Chl-a concentrations. Notably, a persistent “high-salinity, low-temperature, low-chlorophyll” cold water mass exists at the bottom near the entrance of the Zhanjiang Port channel, primarily due to stable stratification induced by topography. Limited light penetration and low transparency suppress phytoplankton growth, while the absence of tidal mixing and upwelling leads to insufficient vertical nutrient supply [65,66]. This cold water mass forms a sharp vertical contrast with the high Chl-a surface waters, highlighting the complexity of coupled physical and biological processes in this region [67].

From a temporal perspective, climate and atmospheric circulation dominate the physical-biological coupling processes in the Leizhou Peninsula coastal waters [64,68]. In spring, rising sea surface temperatures suppress phytoplankton growth rates, and transitions in wind speed and direction weaken vertical water mixing, resulting in relatively low Chl-a concentrations and reduced phytoplankton biomass. In summer, higher water temperatures and abundant sunlight directly stimulate phytoplankton photosynthetic efficiency, significantly increasing Chl-a concentrations, particularly in nearshore areas where increased river runoff brings more nutrients. Enhanced mixed layer dynamics driven by monsoons further amplify nutrient supply and light utilization, leading to widespread increases in Chl-a concentrations [69]. In autumn, sea surface temperatures decline, with an average Chl-a concentration dropping to 0.81 mg/m³. However, nearshore areas such as estuaries and bays exhibit significant biomass accumulation due to front effects, showing a secondary peak with a pattern of “offshore decrease, nearshore increase,” indicating a phytoplankton growth peak in autumn. In winter, the northeast monsoon enhances mixed layer thickness, and although low temperatures suppress phytoplankton growth rates, sufficient nutrient supply through vertical mixing maintains relatively high Chl-a concentrations, reflecting stable phytoplankton growth or a more balanced distribution [70].

As shown in Figure 9, sea surface temperature, salinity, and turbidity exhibit clear regulatory patterns on surface Chl-a concentrations: temperature shows a significant positive correlation, salinity consistently shows a negative correlation, and turbidity shows a negative correlation only in summer [71]. In spring, sea surface temperatures gradually rise from 25 °C to 30 °C, directly enhancing phytoplankton enzyme activity and photosynthetic efficiency, acting as a positive driver for Chl-a concentration increases. Meanwhile, coastal runoff carries suspended sediments, increasing nearshore turbidity, which provides attachment substrates but weakens the underwater light field. However, light availability is not a limiting factor at this time, and Chl-a concentrations rise synchronously with temperature. In summer, sea surface temperatures reach an annual peak of 28–33 °C, with both salinity and turbidity showing negative correlations with Chl-a concentrations. High-concentration zones begin to increase significantly along the coast, spreading in a plume-like pattern on both sides of the Leizhou Peninsula, with increasing spatial variability leading to front formation [72]. In autumn, sea surface temperatures remain at 28–32 °C, temperature stress weakens, and monsoon transitions drive high-salinity water intrusion, increasing surface salinity and maintaining its negative correlation. Reduced wind speeds and lower turbidity enhance light availability, and Chl-a concentrations form a secondary peak along the coast [55]. In winter, sea surface temperatures drop to 18–20 °C, low temperatures suppress enzyme activity, strong vertical mixing homogenizes salinity to 30–34 PSU, and reduced wind and waves lower turbidity, leading to uniform optical properties and a more homogeneous spatial distribution of Chl-a concentrations.

In summary, the spatiotemporal variations in Chl-a concentrations in the Leizhou Peninsula coastal waters are not a simple response to a single environmental factor but rather the result of multi-scale coupling driven by climate stress, terrestrial runoff, and hydrological dynamic mixing.

4.3. Investigating Chl-a Concentration Frontal Dynamics and Their Driving Factors on Both Sides of the Leizhou Peninsula

The Leizhou Peninsula coastal waters, located in the midsection of the northern South China Sea, separate the Qiongzhou Strait and the Beibu Gulf into two semi-enclosed marine systems. Influenced by monsoon transitions, coastal runoff, and complex topography, the sea surface Chl-a concentrations on both sides of the peninsula exhibit pronounced frontal structures in terms of location, intensity, and gradient, particularly in summer and autumn. Seasonal rainfall, temperature variations, hydrological dynamics, and climate events collectively drive the dynamic changes in these fronts, profoundly impacting the regional ecosystem and fishery resource distribution [73,74].

The Chl-a front on the eastern side originates from the topography and monsoon-driven dynamics of the Qiongzhou Strait. Influenced by irregular semi-diurnal tides, water exchange is limited, facilitating nutrient accumulation and promoting phytoplankton proliferation, resulting in high Chl-a concentration zones [75]. In spring, lower sea temperatures and reduced wind speeds weaken vertical mixing, leading to the lowest Chl-a concentrations of the year and a gentler frontal gradient. In summer, the prevailing southwest monsoon rapidly warms surface waters and establishes a thermocline. The narrow channel of Zhanjiang Bay restricts material exchange with the open sea, further exacerbating nutrient accumulation, forming a plume-like front as the eastern band-shaped front advances eastward [76]. In autumn and winter, nutrient distribution becomes more uniform, reducing the concentration difference between the eastern and western sides. The high Chl-a concentration zones near the coast contract compared with summer, and the frontal intensity weakens. On the western side, the Beibu Gulf and Qiongzhou Strait are influenced by regular diurnal tides and ocean currents. The Beibu Gulf exhibits lower Chl-a concentrations due to strong water mixing, but nearshore areas have relatively abundant nutrients, resulting in distinct fronts [77]. In winter, low-salinity, high-turbidity freshwater accumulates at the bay’s apex, moving southward along the western coast of the Leizhou Peninsula under the Coriolis force, converging with high-salinity, low-turbidity offshore waters, leading to simultaneous changes in salinity and turbidity and the formation of a composite front.

Overall, these fronts not only restrict horizontal diffusion but also uplift nutrient-rich deep waters to the surface, triggering rapid phytoplankton proliferation and significantly higher Chl-a concentrations in narrow band-shaped zones compared with surrounding waters [78]. Climate phenomena such as El Niño or La Niña may alter rainfall and ocean current patterns, further affecting the position and intensity of the fronts, necessitating further research with long-term meteorological data. In summary, the two fronts divide the Leizhou Peninsula coastal waters into a nearshore high-concentration diluted water zone, a frontal high-concentration zone, and an oligotrophic offshore zone, with their seasonal shifts directly influencing fishery distribution and red tide risks.

5. Conclusions

This study conducted research on the spatiotemporal distribution of Chl-a concentration in the Leizhou Peninsula sea area. Based on in situ survey data and machine learning algorithms, an inversion model for the region was established. The model successfully retrieved the spatiotemporal distribution of Chl-a concentration for the four seasons of 2024 in the Leizhou Peninsula sea area. Further analysis was conducted on the seasonal variation characteristics and influencing factors in the region, leading to the following main research conclusions:

• Marine survey and analysis: From 2020 to 2024, four field campaigns were conducted in the coastal waters of the Leizhou Peninsula with simultaneous in-situ observations and remote sensing reflectance measurements. Analysis showed that the optical characteristics of the nearshore waters are jointly influenced by Chl-a, CDOM, and suspended matter. The remote sensing reflectance spectra exhibited typical turbid multi-peak and trough structures. Chl-a concentrations were higher in summer and even higher in winter, with significant spatial heterogeneity. These concentrations were coupled with salinity, turbidity, and terrestrial inputs, revealing the sensitive response and spatiotemporal heterogeneity of phytoplankton primary productivity to environmental changes in this area.
• Regarding the machine learning inversion model and its accuracy: The model fitting and performance evaluation results demonstrate that combining spectral bands with machine learning algorithms can effectively improve prediction accuracy, particularly with Gradient Boosting Decision Tree (GBDT). The model achieved a correlation coefficient (R²) of 0.79, RMSE of 0.36, and MAE of 0.30 on the test set, indicating strong applicability and robustness in the coastal waters of the Leizhou Peninsula.
• Spatiotemporal distribution of Chl-a concentration: Using the GBDT algorithm, seasonal maps of Chl-a concentration in the Leizhou Peninsula waters were generated. The spatial distribution showed a clear gradient of “high nearshore, low offshore” and “higher in the east, lower in the west.” Seasonal variation followed the pattern of “low in spring – increasing in summer – stable in autumn – high in winter,” reflecting that the spatiotemporal pattern of phytoplankton biomass results from the combined effects of monsoon, runoff, and tidal-driven nutrient transport and mixed layer evolution.

Although this study achieved relatively satisfactory results, there remains room for further improvement: First, there is potential for enhancement in the precision of data processing and model simulation accuracy; second, the spatiotemporal scale coverage of the study could be expanded to improve the generalizability of the conclusions; third, the current study only reports model-level confidence intervals and has not yet generated pixel-level uncertainty maps, indicating a deficiency in the detailed characterization of result uncertainty. Future research can focus on targeted optimization and innovation in these areas, which will lay a more solid foundation for more accurately revealing the patterns of water quality changes in the Leizhou Peninsula waters and providing scientifically effective guidance for marine ecological management.