A novel water optical types framework for Chinese inland waters with the application of multitype satellite sensor

ABSTRACT

Optical Water Type (OWT) analysis is crucial for comprehending water composition and quality, key factors in assessing water quality over extensive areas. However, China’s inland waters lack a standardized system for such analysis. To quantitatively analyze the classification results, our study compared three K-means clustering methods, for analyzing 1310 spectral data from various Chinese lakes and reservoirs, thereby addressing this gap. The innovative split-merge K-means method identified 13 distinct OWTs that more closely adhere to the principles of minimizing intra-class distance and maximizing inter-class distance. These were categorized into four groups: clear water, turbid water, eutrophic water, and special type water. Additionally, we developed a method based on Spectral Angle Distance (SAD) to evaluate the classification capabilities of 12 satellite sensors. The results show that Sentinel-3 OLCI (Ocean and Land Color Instrument), MERIS (Medium Resolution Imaging Spectrometer), and Sentinel-2 MSI (Multispectral Instrument) have the best water classification capabilities, making them well-suited for large-scale monitoring of OWT changes. Conversely, other sensors, such as the Sustainable Development Scientific Satellite-1 (SDGSAT-1), Landsat-8, GaoFen-6, GaoFen-1, GaoFen-2, Landsat-5, Landsat-7, Moderate Resolution Imaging Spectroradiometer (MODIS), and HuanJing-1, necessitate the consolidation of water types for effective categorization, indicative of their more limited classification capabilities.

KEYWORDS:

• Optical water type (OWT)
• inland water
• water quality
• remote sensing reflectance

1. Introduction

Water is classified into two types for traditional watercolor remote sensing: Case I and Case II water. Case I water mainly consists of oceanic water, while Case II water includes coastal water as well as inland water such as lakes, rivers, and reservoirs (Morel and Prieur 1977). Inland water is a crucial component of the natural ecosystem that humanity depends on for survival and development (Feng et al. 2016). Remote sensing parameter inversion of watercolor represents a pivotal technological tool for monitoring, assessing, and predicting changes in aquatic ecosystems and the water environment (Cao et al. 2017; Cui et al. 2020; Hu, Lee, and Franz 2012; Wang and Yang 2019; Yin et al. 2023). While optical water type (OWT) provides a macroscopic reflection of water constituents and comprehensive water quality status, offering an effective indicator for understanding the current state and changing trends of water quality across extensive areas (Du, Song, and Liu 2022). OWT also serves as a foundation for selecting classification-based models for fine water quality parameter inversion and optimizing those parameters (Eleveld et al. 2017; Fangfang et al. 2023; Nazeer and Nichol 2016; Shen et al. 2015; Sun et al. 2011; Xiong et al. 2020).

Water classification relies on the optical properties of water. These properties reveal the water’s composition, reflecting the concentrations of optically active constituents including suspended solids, chlorophyll, and colored dissolved organic matter. The earliest study by Jedov and Koczv in 1951 classified global water into eight groups based on optical indicators (Jerlov and Koczy 1951). Later, Morel’s 1977 classification divided water into Case I oceanic and Case II nearshore and inland waters, which is widely accepted (Morel and Prieur 1977). In recent decades, scholars have further explored optical water classification, studied different waters, and contributed to water classification. For example, Mélin classified the Yellow Sea (Mélin et al. 2011), Spyrakos’ studied the Adriatic (Spyrakos et al. 2011), Vantrepotte worked on the Iberian Coastal Classification (Vantrepotte et al. 2012), Moore (Moore et al. 2014), Mélin and Vantrepotte (Mélin and Vantrepotte 2015), Jackson (Jackson, Sathyendranath, and Mélin 2017), Spyrakos (Spyrakos et al. 2018), Wei (Wei et al. 2022), and Men (Men et al. 2023) conducted global classification of ocean, coastal, and inland water. These studies provided insights into the optical properties and water quality conditions of various regions, supporting water resources management and environmental protection. However, previous studies have primarily focused on Case I ocean waters or global waters and have not addressed the classification of black and odorous waters. Furthermore, these studies have relatively limited actual measurement data in China, highlighting the necessity of assessing the suitability of these classification systems within the Chinese context. In recent years, there has been an increase in studies focusing on the classification of inland water. Le studied certain lakes and reservoirs in China (Le et al. 2011), Shen classified several optically complex waters in China (Shen et al. 2015), Botha analyzed inland waters in Australia (Botha et al. 2020), and Bishun explored the classification of inland and coastal waters in China (Bi et al. 2022). Du Yunxia and colleagues conducted a study on the classification of lake reservoirs in northeastern China and the eastern part of the Mengxin Lake District (Du, Song, and Liu 2022). However, most of these studies are limited to specific lakes or regions and lack a comprehensive optical classification system for a wide range of inland waters in China.

There are three primary bases for water classification: intrinsic optical quantities, water quality parameters, and remote sensing reflectance. Classification based on intrinsic optical quantities and water quality parameters provides intuitive characterizations of the optical properties and water quality conditions of different waters. However, the inversion of intrinsic optical quantities and water quality parameters using remote sensing remains challenging and currently feasible only with in-situ data, limiting its application in remote sensing images (Binding, Bowers, and Mitchelson-Jacob 2005; Sampsa Koponen, Kallio, and Hallikainen 2002). Water classification based on remote sensing reflectance relies on differences in remote sensing reflectance ${\mathrm{R}}_{\mathrm{rs}}\left(\lambda \right)$ and its spectral shape across various water. This method is commonly referred to as water optical classification, reflecting both apparent and intrinsic optical properties. It has the advantage of direct classification based on images, which facilitates large-scale application and comprehensive reflection of different water quality parameter components and concentrations. Methods for optical water classification based on ${\mathrm{R}}_{\mathrm{rs}}\left(\lambda \right)$ include supervised classification (Ye et al. 2016), unsupervised classification (Ben Mustapha et al. 2014; Feng et al. 2005; Zhang et al. 2015), spectral shape classification (Le et al. 2011), fuzzy clustering (Bi et al. 2019; Moore, Campbell, and Hui 2001; Xue et al. 2019), and neural network modeling (Kajiyama, D'Alimonte, and Zibordi 2019). Within the category of unsupervised classification methods, there are several subdivisions, including hierarchical clustering (Shi et al. 2014), K-means clustering (Palacios, Peterson, and Kudela 2012), ISODATA clustering (Mélin and Vantrepotte 2015), and self-organizing mapping (Ben Mustapha et al. 2014). The K-means clustering algorithm not only classifies water into different types but also calculates the distances between these types, enabling quantitative analysis and enhancing classification precision. Therefore, this study will use the K-means clustering algorithm for water classification.

The main goal of this study is to create a comprehensive water classification system for inland waters in China. We compared three different K-means clustering implementation strategies using a dataset of 1310 in-situ hyperspectral measurements. We conducted a comparative analysis to identify the optimal K-means optimization strategy for classifying inland waters in China. This resulted in the creation of a comprehensive classification system applicable to large-scale lakes and reservoirs nationwide. Additionally, we evaluated the classification capabilities of commonly used sensors in inland waters, providing essential technical support for the development of large-scale, long-term water type products.

2. Study area and data acquisition

2.1. Study area

In our research, we conducted 57 experiments across 30 typical lakes and reservoirs in China, collecting a total of 1310 in-situ data points. The data spans from 2006 to 2022, offering a comprehensive perspective over time. These 30 waterbodies encompass a diverse range of environments, including clean, turbid, eutrophic, and black-odor, representing various water quality types. Our study area is evenly distributed in North China, Northeast China, the middle and lower reaches of the Yangtze River, Southwest China, and the Tibetan Plateau, with broad regional representation, and the distribution of the study area is shown in Figure 1.

Figure 1. Geographic distribution of study areas.

2.2. Water surface spectral acquisition

In this research, we employ the ‘above-water surface method’ to measure the surface spectra of the water (Tang 2004). We use the ASD FieldSpec® ProFR portable field spectrometer, which covers a spectral wavelength range of 350–2500 nm and offers a spectral resolution of 3 nm. A gray plate with a reflectivity ranging from 20% to 50% serves as our reference plate. For optimal measurements, the best observation azimuth angle was set at 135° away from the direction of the sun, and the optimal zenith angle for measuring radiance on the water surface was 40°. Our measurement process involves capturing spectra at each site in the following sequence: first the reference plate, then the skylight, and again the reference plate. Finally, we calculate ${\mathrm{R}}_{\mathrm{rs}}\left(\lambda \right)$ spectra using Equation (1) (Lee et al. 2016). (1) ${\mathrm{R}}_{\mathrm{rs}}={\displaystyle \frac{{\mathrm{L}}_{\mathrm{sw}}-\mathrm{r}\ast {\mathrm{L}}_{\mathrm{sky}}}{{\mathrm{L}}_{\mathrm{p}}\ast {\displaystyle \frac{\pi }{{\rho }_{\mathrm{p}}}}}-\mathrm{\Delta }}$(1) where ${\mathrm{L}}_{\mathrm{sw}}$ is the total radiance of the water; $\mathrm{r}$ is the reflectance of the air–water interface; ${\mathrm{L}}_{\mathrm{sky}}$ is for the radiance of the sky’s diffuse scattering; ${\mathrm{L}}_{\mathrm{p}}$ is the radiance of the standard gray plate; ${\rho }_{\mathrm{p}}$ is the reflectance of the standard gray plate.

2.3. Water quality parameter acquisition

In this research, we considered key water quality parameters, including Chlorophyll-a concentration (Chla), measured in milligrams per cubic meter (mg/m³); Total Suspended Matter (TSM), measured in milligrams per liter (mg/L); Colored Dissolved Organic Matter (CDOM) absorption coefficient ($a_{\mathrm{CDOM}}\left(440\right)$). Chla concentration was determined through a spectrophotometric method. We filtered water samples through Whatman GF/F glass fiber filter membranes with a pore size of 0.7 µm and a diameter of 25 mm. Then, we extracted chlorophyll using hot ethanol and calculated Chla concentration based on absorbance measurements at two wavelengths: 665 and 760 nm. To determine TSM concentration, we first weighed a dried Whatman GF/F glass fiber filter membrane with a diameter of 47 mm and pore size of 0.7 µm after drying at 105°C for 6 h. After filtering the water sample, we re-weighed the filter membrane after drying at 105°C for 4 h to calculate the TSM concentration. For CDOM, we created a mixed solution by filtering the water sample through a filter membrane with a pore size of 0.2 µm. The absorbance of the resulting solution ($\mathrm{O}{\mathrm{A}}_{\mathrm{CDOM}}\left(\lambda \right)$) was determined using a spectrophotometer, enabling the calculation of the CDOM absorption coefficient, denoted as ${\mathrm{a}}_{\mathrm{CDOM}}\left(\lambda \right)$.

3. Methods

This section offers a systematic overview of the research methodology, addressing three key aspects: spectral smoothing and normalization, methods for optical classification, and the assessment process for the classification capabilities of commonly used sensors. Firstly, spectral smoothing employs the Saviztky-Golay method, with a specific focus on normalizing spectra within the 400–900 nm range. Subsequently, we introduce three optical classification methods compared in this study: the traditional K-means clustering method, the stepwise iterative K-means clustering method, and the split-merge K-means clustering method. Finally, the section outlines the steps involved in evaluating the classification capabilities of commonly used sensors, including the calculation of equivalent observed spectra to assess the satellite sensor’s classification performance Figure 2.

Figure 2. The technical roadmap of this study.

3.1. Spectral smoothing and normalization processing

Before classifying the data, it’s important to preprocess the remote sensing reflectance curves. In practical measurements, spectral data often contains noise, leading to irregularities or spikes in the reflectance spectra. Additionally, inherent measurement errors can affect subsequent analysis. To improve the quality and reliability of the spectra, we use smoothing on the in-situ ${\mathrm{R}}_{\mathrm{rs}}\left(\lambda \right)$ data. We achieve smoothing by applying the Saviztky-Golay method with a window size of 15 and a polynomial degree of 2 (Ruffin, King, and Younan 2008).

To eliminate the influence of environmental lighting and ensure that there is no scale difference between the spectral curves, focus solely on their shape characteristics. We selected ${\mathrm{R}}_{\mathrm{rs}}\left(\lambda \right)$ spectra in the 400–900 nm wavelength range and normalized them using the following equation: (2) ${\mathrm{R}}_{\mathrm{rs}}\left({\lambda }^{\mathrm{\prime }}\right)={\displaystyle \frac{{\mathrm{R}}_{\mathrm{rs}}\left(\lambda \right)}{{\displaystyle \frac{1}{n}{\sum }_{400}^{900}{\mathrm{R}}_{\mathrm{rs}}\left(\lambda \right)}}}$(2) where ${\mathrm{R}}_{\mathrm{rs}}\left({\lambda }^{\mathrm{\prime }}\right)$(nm⁻¹) is the normalized spectrum derived by dividing each wavelength position by the average of the spectra ranging from 400 nm to 900 nm.

3.2. Methods for optical classification

In this study, we utilized the K-means clustering algorithm due to its dual capabilities. It effectively categorizes water into distinct types and quantitatively measures the distances between these types, significantly enhancing classification precision. This adaptability makes K-means particularly apt for the detailed task of classifying water types. The decision to use K-means was also driven by its proven effectiveness in managing large datasets and its simplicity and efficiency, which are essential in handling the extensive data typical in water-type studies (Spyrakos et al. 2018). The algorithm’s track record of success in similar fields further supports its suitability for our research. To bolster the scientific rigor and precision of our clustering results, we utilized spectral angle distance (SAD) as the similarity metric. We compared three clustering methods – traditional K-means clustering, stepwise iterative K-means clustering, and split-merge – to assess their effectiveness in this context. This methodical approach, integrating K-means’ robustness with the detailed analysis provided by SAD, forms the basis of our accurate and comprehensive water classification system. (3) $\mathrm{SAD}=1-{\displaystyle \frac{{\mathrm{x}}_{\mathrm{s}}{\mathrm{x}}_{\mathrm{t}}^{\mathrm{T}}}{\sqrt{\left({\mathrm{x}}_{\mathrm{s}}{\mathrm{x}}_{\mathrm{s}}^{\mathrm{T}}\right)\left({\mathrm{x}}_{\mathrm{t}}{\mathrm{x}}_{\mathrm{t}}^{\mathrm{T}}\right)}}}$(3) where ${\mathrm{x}}_{\mathrm{s}},{\mathrm{x}}_{\mathrm{t}}$represent two spectral reflectance vectors, and ${\mathrm{x}}_{\mathrm{s}}^{\mathrm{T}},{\mathrm{x}}_{\mathrm{t}}^{\mathrm{T}}$ denote the transpose of these spectral reflectance vectors. Numerically, a smaller SAD indicates a higher similarity between the two spectral profiles.

3.2.1. Traditional K-means clustering methodology

MacQueen introduced the K-means clustering method in 1967. This unsupervised learning algorithm remains one of the most extensively utilized techniques for cluster analysis (MacQueen 1967). The process of the K-means clustering method is as follows:

1. For a dataset X containing n data objects (n > K), randomly select K spectra as initial cluster centers.

2. Calculate the spectral angular distance between each sample in the dataset and these K initial centers, assigning each sample to the class whose initial center is closest.

3. Calculate the mean of all data objects in that class to obtain new cluster centers.

4. Continuously iterate through steps 2 and 3 until the cluster centers no longer change or until a convergence condition is met, and then output the clustering results.

3.2.2. Stepwise iterative K-means clustering methodology

The stepwise iterative K-means algorithm is an enhancement of the original K-means method, incorporating a progressive iterative approach. This methodology introduces an element of gradual refinement, making it a straightforward and practical solution. As the number of iterations increases, the final results exhibit a notable trend towards stability and consistency. The workflow of the stepwise iterative K-means clustering method is as follows:

1. Treat each of the n samples as an individual cluster, resulting in n clusters.

2. Calculate the spectral angular distances between each pair of samples $x_i$ in the dataset, and merge the two clusters with the smallest distance, one pair at a time.

3. Repeat step 2 until no further merging of clusters is possible or until the maximum iteration limit is reached, and then output the clustering results.

3.2.3. Split-Merge K-means clustering methodology

In this study, building upon the traditional K-means clustering method, we introduced a split-and-merge strategy to better capture the differences in spectral information. In comparison to the traditional K-means approach, this method divides the dataset into multiple categories and carries out both intra-class splitting and inter-class merging during the clustering process. Additionally, the clustering performance can be further optimized by iteratively adjusting parameters such as thresholds and maximum iteration counts. The workflow of the split-merge K-means clustering method is as follows:

1. Initially, use the traditional K-means clustering method to partition the dataset into K clusters. Calculate the intra-class distance for each category using the formula $\mathrm{D}=\sqrt{{\displaystyle \frac{1}{N_i}{\sum }_{k=1}^{N_i}\mathrm{SA}{D}^2\left({\mathrm{X}}_k^i,{\mathrm{X}}^i\right)}}$, where $N_i$ represents the number of samples in the category, ${\mathrm{X}}_k^i$ denotes the spectral reflectance vector for the k-th sample in category $i$, and ${\mathrm{X}}^i$ is the mean spectrum for category $i$. Examine the spectra of K categories and determine the splitting threshold.

2. Based on the threshold set in the first step, split the categories with intra-class distances exceeding the threshold, resulting in a greater number of categories.

3. For each category, compute the mean value of all samples within that category, substitute the category with the mean value, and then recalculate the inter-class spectral angular distances.

4. Merge the two categories with the smallest distance in each iteration, repeating this process until no further merging is possible or until the maximum iteration count is reached, and then output the clustering results.

3.3. Methods for evaluating classification performance of commonly used sensors

In general, hyperspectral sensors like ASD provide reflectance values for every 1 nm, whereas satellites typically have only a few spectral bands. Due to the unique spectral response characteristics, the trends in the spectral reflectance curves from ground-based hyperspectral sensors like ASD may not align perfectly with those from satellite sensors. Therefore, in this study, spectral response functions will be used to perform spectral equivalence calculations on the reflectance data acquired by the ground-based ASD sensor. This will enable the derivation of remote sensing reflectance values that are representative of satellite measurements. (4) ${\mathrm{R}}_{\mathrm{rs}}\left({\mathrm{B}}_i\right)={\displaystyle \frac{{\int }_{{\lambda }_{\min }}^{{\lambda }_{\max }}{\mathrm{F}}_0\left(\lambda \right)\ast \mathrm{SRF}\left(\lambda \right)\ast \mathrm{Rrs}\left(\lambda \right)\mathrm{d\lambda }}{{\int }_{{\lambda }_{\min }}^{{\lambda }_{\max }}{\mathrm{F}}_0\left(\lambda \right)\ast \mathrm{SRF}\left(\lambda \right)\mathrm{d\lambda }}}$(4) where ${\mathrm{R}}_{\mathrm{rs}}\left({\mathrm{B}}_i\right)$ is the equivalent reflectance in the i-th satellite band, ${\lambda }_{\min }$ and ${\lambda }_{\max }$ are the upper and lower bounds of the spectral response function, $\mathrm{Rrs}\left(\lambda \right)$ is the in-situ remote sensing reflectance, F₀ is the extraterrestrial solar irradiance, and $\mathrm{SRF}\left(\lambda \right)$ is the spectral response function at wavelength λ.

To quantitatively evaluate the discriminative capability of common satellite sensors for classifying average spectra of water types, we propose an approach for the evaluation of the classification capabilities of commonly used sensors:

1. Compute Initial SAD: Calculate the SAD between the in-situ average spectra of different water types.

2. Establish Threshold: Determine the minimum SAD value among these water types and use it as a threshold for merging water spectra by satellite sensors.

3. SAD Calculation for Satellite Observations: Calculate the SAD values between water types as observed by satellite sensors.

4. lassification and Merging: If the SAD value between two types is smaller than this threshold, consider them ineligible for separate classification and merge them. Conversely, if the SAD value exceeds the threshold, treat them as distinct categories.

5. Evaluate Merged Types: If sensors result in an equal number of merged types, calculate J_SADs, the sum of the minimum SAD between the average spectra of these final water types. A higher J_SADs value indicates superior classification capability of the sensor, following the ‘maximum inter-class distance’ principle. (5) $\mathrm{J}\mathrm{\_}\mathrm{SADs}={\sum }_{I=1}^K\mathrm{SA}{D}_{\min }^i$(5) where $\mathrm{SA}{D}_{\min }^i$ denotes the minimum SAD between the average spectrum of the i-th water type and the average spectrum of other water types, and k is the total number of water types.

4. Results and analysis

4.1. Comparative analysis of three classification methods

By comparing the final clustering results obtained through three methods, it is evident that both traditional K-means and stepwise iterative K-means algorithms are relatively sensitive to the random selection of initial cluster centroids, leading to instances of misclassification in the final results. In contrast, the split-merge clustering method optimizes clustering performance by iteratively adjusting parameters such as thresholds and maximum iteration counts. This approach successfully separates spectral types that may have been misclassified, reducing spectral misclassification errors and yielding relatively improved classification results. Even when one cluster has a significantly larger number of samples compared to others with fewer samples, the split-merge method excels in obtaining high-quality classification results. We use J_SADs and N_SADs to evaluate the effectiveness of the three clustering methods. (6) $\mathrm{N}\mathrm{\_}\mathrm{SADs}={\sum }_{I=1}^K{D}_i$(6) where N_SADs represents the sum of intra-class distance for all types in the final water type set, $D_i$ is the intra-class distance of the i-th water type, and k is the total number of water body types. The results, as shown in Table 1, the split-merge K-means clustering method achieves smaller N_SADs and simultaneously larger J_SADs compared to the other two methods. This implies that the Split-Merge clustering method effectively adheres to the principles of minimizing intra-class distance and maximizing inter-class distance, making it particularly advantageous in optical classification studies of water.

Table 1. Comparison of SAD Values Across Different K-means Clustering Methods. J_SADs represents the sum of the minimum SAD between the average spectra of final water types; N_SADs represents the sum of intra-class distances for all types in the final water types.

	J_SADs	N_SADs
Traditional K-means	0.37	4.24
Stepwise Iterative K-means	0.34	4.26
Split-Merge K-means	0.4	4.22

4.2. Optical water types

In this study, 13 optical water types (OWTs), identified using the split-merge k-means clustering method, make up the final classification system for inland water. Figure 3 visually represents the spectral profiles of these 13 different water types, both before and after the normalization process. Figure 4, on the other hand, presents the definitive spectral characteristics that define these 13 distinct OWTs. Notably, within this array of water types, OWT3, OWT5, OWT7, OWT8, OWT9, and OWT10 emerge with a notable abundance of samples, each surpassing the 100-spectral records. OWT10, in particular, distinguishes itself with the largest sample set, encompassing an impressive 325 individual spectral records. Conversely, seven water types, namely OWT1, OWT2, OWT4, OWT6, OWT11, OWT12, and OWT13, exhibit a comparatively limited number of samples.

Figure 3. Average spectra of 13 OWTs. (a) In-situ spectra; (b) Standardized spectra.

Figure 4. Average spectra of 13 OWTs and included sample Data. Each type’s ${\mathrm{R}}_{\mathrm{rs}}$ is represented by a light gray curve, and the 12th type is represented by a dark gray curve. The average spectra of each water type are shown in black curves.

4.3. Analysis of spectral characterization in various water types

The total absorption coefficient of water can be expressed as the linear summation of absorption coefficients from various components such as pure water, non-pigmented particulate matter, phytoplankton, and yellow substance. These components exhibit distinct absorption characteristics, leading to variations in the spectral shapes of different water. In this study, 13 OWTs were broadly grouped into four major categories based on their optical properties: Clear Water, Turbid Water, Eutrophic Water, and Special Type Water. From the perspective of these four categories, the characteristics of each type of water are as follows:

4.3.1. Group I: clear water

Group Ⅰ represents clear water, characterized by a low total absorption coefficient, mainly influenced by pure water. These waters typically show a peak in reflectance between 400-600 nm, a rapid decrease between 550-600 nm, and a gradual decline in the 600-700 nm range, with reflectance values approaching zero after 700 nm. These spectral features are due to the dominance of pure water’s absorption characteristics. Pure water exhibits a relatively weaker absorption in the blue–green spectral range, consequently resulting in higher reflectance values. In contrast, in the red spectral range, pure water exhibits more pronounced absorption, leading to a consequent decrease in reflectance values. However, different water compositions can lead to variations in reflectance curves among different water. For example, OWT1 water has a continuous decline in reflectance from around 400 nm. In contrast, OWT2 water exhibits an initial increase in the $\mathrm{Rrs}\left(\lambda \right)$ spectral curve from 400 nm, followed by a swift decrease after reaching a peak near 470 nm. OWT3 water displays a spectral curve shape akin to OWT2, commencing at 400 nm, showcasing an initial ascent followed by a decline. However, the upward trend in OWT3 continues for a longer range, with a maximum around 540 nm, followed by a rapid decrease.

4.3.2. Group II: turbid water

Group II represents turbid water, where the total absorption coefficient is primarily influenced by non-pigmented particulate matter. A common feature observed in the spectral curves within this group is the transition of non-pigmented particulate matter absorption from strong to weak within the 400-560 nm range, resulting in an increase in water reflectance. Simultaneously, within the 500-700 nm range, where non-pigmented particulate matter absorption is relatively subdued, water reflectance significantly rises. However, beyond 720 nm, reflectance gradually declines due to the increased absorption of pure water in the near-infrared spectrum and the diminishing influence of non-pigmented particulate matter diminishes. Differently, when examining the spectral profiles of OWT4, OWT5, and OWT6, it becomes evident that the reduction in reflectance between 560-720 nm for these three water types gradually becomes less pronounced. The nearly straight-line shape of OWT6’s spectral curve between 560-720 nm is a result of the increasing influence of pure water absorption.

4.3.3. Group III: eutrophic water

Group III represents eutrophic water, where phytoplankton primarily dominates the total absorption coefficient. The decreased absorption coefficient of phytoplankton leads to an increase in reflectance between 400-550 nm, with a peak typically occurring around 550 nm. Within the 550-670 nm range, the heightened absorption by phytoplankton causes a reduction in water reflectance. Beyond 750 nm, there are minimal spectral changes due to weak absorption and scattering by water constituents, resulting in a relatively flat spectral curve. All three water types show reflectance peaks around 550 and 700 nm, mainly due to the high algal content, leading to a noticeable chlorophyll reflectance peak. Simultaneously, phycocyanin absorption near 620 nm and chlorophyll absorption near 675 nm lead to a reflectance peak near 650 nm. The 700 nm peak correlates with Chla concentration, with OWT9 having the highest concentration, resulting in the most significant differences in peak values around 700 nm and valley values near 675 nm compared to OWT7/OWT8.

4.3.4. Group IV: special type water

Group IV encompasses four special water types. OWT10’s spectral curve resembles that of OWT5 but exhibits more pronounced peaks and valleys due to the co-dominance of non-pigmented particulate matter and phytoplankton in OWT10’s water. OWT11’s reflectance values are slightly lower than typical water due to the influence of organic matter and microorganisms, resulting in altered light transmission and scattering, and its spectral reflectance curve is flatter with no distinct characteristic peaks and valleys. OWT12 and OWT13 exhibit spectral shapes akin to OWT9, marked by a characteristic peak between 500-600 nm. However, OWT12 and OWT13 exhibit increased reflectance in the infrared range, primarily due to the substantial presence of cyanobacteria and suspended particles. Notably, when compared to OWT13, OWT12 exhibits a marginally milder increase.

4.4. Analysis of water quality parameters in various water types

Water quality parameter analysis involves the quantitative measurement and statistical analysis of a range of chemical substances and biological indicators in water, serving to evaluate its environmental quality. For this study, our primary focus was on the collection and analysis of three key water color parameters: Chla, TSM, and CDOM absorption coefficients, across the 13 OWTs. Figure 5 illustrates the distribution of water quality parameters among these 13 OWTs, highlighting notable variations in water color parameters across distinct water types. Now, we’ll analyze the characteristics of water quality parameters within the 13 OWTs, categorized into four main groups.

Figure 5. Distribution of water quality parameters for 13 OWTs. Boxplots with probability density of TSM (black), Chla (red), $a_{\mathrm{CDOM}}\left(440\right)$ (blue) for 13 OWTs. The sample median is indicated by a horizontal line within the box while squares indicate mean values.

4.4.1. Group I: clear water

OWT1 primarily includes 19 spectra from Nam Co, and OWT2 consists of 32 spectra mainly from lakes such as Bangda Co, Longmu Co, Guozha Co, Selin Co, Daze Co, and others. Although these two types have limited data, it’s evident that both exhibit low Chla concentrations, resembling clear seawater. OWT3 comprises 120 spectra from various lakes and rivers like Danjiangkou Reservoir, Xiaolangdi Reservoir, Qinghai Lake, Bamu Co, and Zigetang Co. These waters are characterized by high transparency and relatively low concentrations of water constituents. They are mainly lakes and river waters with good ecological conditions, maintaining high water clarity.

4.4.2. Group II: turbid water

OWT4 primarily consists of 16 spectra from Taihu Lake and Penglai Area. These waters exhibit relatively low transparency and not particularly high concentrations of water constituents. OWT5 mainly includes 178 spectra from Taihu Lake. In comparison to OWT4, the waters in OWT5 have lower transparency, higher TSM concentrations, and lower Chla concentrations and CDOM absorption coefficients. OWT6 comprises 13 spectra primarily from Taihu Lake, Three Gorges Reservoir, and Luhun Reservoir. These waters have the lowest transparency among the 13 OWTs, the highest TSM concentrations, and relatively low concentrations of other water constituents.

4.4.3. Group III: eutrophic water

OWT7 primarily includes 230 spectra from Xiaolangdi Reservoir, Guanting Reservoir, Luhun Reservoir, Qiandao Lake, and Taihu Lake. Among these three water types, OWT7 has the lowest Chla concentration. OWT8 comprises 206 spectra from Taihu, Guanting Reservoir, Yuqiao Reservoir, and Baiyang Lake. OWT8 exhibits higher Chla concentrations compared to OWT7 but lower than OWT9. Suspended sediment concentrations are relatively low, but CDOM absorption coefficients are relatively high in this type. OWT9 consists of 113 spectra from Guanting Reservoir, Taihu Lake, and Baiyang Lake. The defining characteristic of this type is the high concentration of water constituents, particularly Chla.

4.4.4. Group IV: special type water

OWT10 includes 325 spectra primarily from Taihu Lake, Baiyang Lake, Luhun Reservoir, and Guanting Reservoir, representing turbid waters with high Chla and CDOM. OWT11 consists of 17 spectra, primarily from the Xiaotaihou River, which predominantly falls under the type of black and odorous waters. The Chla concentration in this type does not exhibit significant differences compared to other water types. However, waters in this type display relatively higher concentrations of TSM and CDOM absorption coefficients due to organic pollutants. OWT12 mainly includes 29 spectra from Lake Taihu, featuring high Chla, TSM, and CDOM levels (though lower than OWT13). OWT13 primarily consists of 11 spectra from Taihu Lake and is characterized by extremely high Chla concentrations, indicating the presence of very high cyanobacteria content near or on the water surface. Furthermore, this type’s waters exhibit high TSM and CDOM.

4.5. Assessment of classification capabilities of common sensors

The in-situ spectral reflectance varies continuously across the entire spectral range. However, multispectral sensors on satellites typically cover only a few discrete bands, and their spectral measurement intervals are often higher than the satellite’s resolution. This implies that some spectral features might be missed by lower-resolution sensors. Therefore, we selected twelve commonly used satellite sensors to investigate their discriminative capabilities for water types. Figure 6 displays the central wavelengths and bandwidths of the twelve satellite sensors, while Table 2 presents a comprehensive overview, including the year of launch, spatial resolution, temporal resolution, number of visible and near-infrared (VNIR) bands, and their wavelength range.

Figure 6. Bandwidth and center wavelengths of 12 commonly used satellite sensors.

Table 2. General information of 12 commonly used satellite sensors (Sources: (Adjovu et al. 2023; Dörnhöfer and Oppelt 2016; Wang and Yang 2019; Yang et al. 2022)).

Satellite Sensor	Year Launched	Spatial Resolution (m)	Temporal Resolution (Day)	Spectral Resolution Band
				Number of VNIR Bands	Wavelength (µm)
GF1 PMS	2013	8	4	4	0.45-0.89
GF2 PMS	2014	4	5	4	0.45-0.89
GF6 WFV	2018	16	2	8	0.45-0.89
HJ1 CCD	2008	30	4	4	0.43-0.90
Landsat-5 TM	1984	30–120	16	4	0.45–2.35
Landsat-7 ETM+	1999	15–60	16	4	0.45–12.5
Landsat-8 OLI	2013	30	16	5	0.43–12.51
Terra/Aqua MODIS	1999	250–1000	1–2	4	0.41–14.5
SDGSAT-1 MII	2021	10–40	11	7	0.38–0.90
ENVISAT MERIS	2002	300–1200	1	15	0.39–1.04
Sentinel-2 MSI	2015	10–60	5	12	0.44–2.19
Sentinel-3 OLCI	2016	300	1.1	21	0.40-1.02

To quantitatively assess the discriminative capability of the twelve commonly used satellite sensors, our initial step involved identifying the minimum SAD value within the in-situ spectral averages of the 13 water types, denoted as the ‘y’ value in Figure 7. This ‘y’ value was then designated as the threshold for amalgamating water types by commonly used satellite sensors. Subsequently, we calculated the minimum SAD values among the 13 water types for each satellite sensor, as illustrated in Figure 7. Overall, these twelve sensors can effectively distinguish the four water types: OWT3, OWT11, OWT12, and OWT13, as the SAD values for each sensor in these four types surpass the y-values. However, when differentiating nine other water types (OWT1, OWT2, OWT4, OWT5, OWT6, OWT7, OWT8, OWT9, and OWT10), aside from Medium Resolution Imaging Spectrometer (MERIS), Sentinel-2 Multispectral Instrument (MSI), and Sentinel-3 Ocean and Land Color Instrument (OLCI), several other sensors exhibit minimum SAD values below the y-values.

Figure 7. Inter-class minimum SAD for 12 commonly used satellite sensors, with the amalgamation threshold for water types. The horizontal red line represents the threshold for amalgamating water types, which is determined by the minimum SAD derived from the in-situ spectral averages across 13 water types.

If the SAD value between two water types is smaller than this threshold, they are considered ineligible for separate classification and should be merged. Conversely, if the SAD value is greater than the threshold, they should be classified separately. Employing this method, we obtained the final water classification results for each satellite sensor. Based on the number of merged water body types in the end, we identified four scenarios, as depicted in Figure 8.

Figure 8. Average spectrum of 12 commonly used satellite sensors after consolidation.

4.5.1. I: 8 categories: GF-1 PMS, GF-2 PMS, HJ-1 CCD, Landsat-5 TM, Landsat-7 ETM+, and MODIS

The first scenario involves a final water classification into 8 categories, including six satellite sensors: GaoFen-1 (GF-1) Panchromatic and Multispectral Sensor (PMS), GF-2 PMS, HuanJing-1 (HJ-1) Charge-Coupled Device (CCD), Landsat-5 Thematic Mapper (TM), Landsat-7 Enhanced Thematic Mapper Plus (ETM+), and Terra/Aqua Moderate-resolution Imaging Spectroradiometer (MODIS). Firstly, the clear water types OWT1 and OWT2 were combined into one category. Figure 4 demonstrates that the primary distinction between these two water types lies in their reflectance values within the 400-500 nm spectral range. However, these six satellite sensors have only one band within this range, resulting in an equivalent spectrum that cannot accurately differentiate the spectral reflectance features of the two water types in the 400-500 nm range. Secondly, OWT4, OWT7, and OWT8 water types were merged into a single category. Similarly, as observed from Figure 4, the distinguishing feature among these three water types lies in the spectral shape characteristics between 400-550 nm and the chlorophyll peak around 700 nm. However, these six sensors lack bands near 700 nm, leading to the loss of chlorophyll peak features within this range, rendering it impossible to differentiate the chlorophyll peak features among these three water types. Finally, OWT5, OWT9, and OWT10 water types are merged into one category. A closer examination of the spectra of these three water types reveals that their distinguishing feature primarily revolves around the chlorophyll peak near 700 nm. However, these six sensors lack bands near 700 nm, making them unable to differentiate among these three water types within this specific spectral range.

4.5.2. II: 10 categories: GF-6 WFV and Landsat-8 OLI

The second scenario involves a final water classification into 10 categories, mainly associated with the GF-6 Wide Field View Multispectral Camera (WFV) and Landsat-8 Operational Land Imager (OLI) satellite sensors. With regard to GF-6, OWT4 and OWT7 are initially merged. The differences between these two water types mainly manifest in their spectral growth trends between 400-560 nm and the decline trends between 600-700 nm. Unfortunately, GF-6’s band settings reduce these differences, making it less effective in distinguishing between OWT4 and OWT7. Meanwhile, OWT5, OWT9, and OWT10 are combined into a single category. The primary distinctions among these three water types are in the declining trends between 570-700 nm and the chlorophyll reflection peak near 700 nm. Regrettably, GF-6’s band settings also weaken these distinctions. As for the Landsat-8 satellite sensor, it initially combines OWT5, OWT9, and OWT10. Analyzing Landsat-8’s band settings and the distinctions among these water categories, similar to GF-6, Landsat-8’s band settings diminish the spectral differences among OWT5, OWT9, and OWT10. In the case of OWT7 and OWT8, the primary difference lies in the chlorophyll peak near 700 nm. However, Landsat-8 lacks bands near 700 nm, resulting in the loss of the chlorophyll feature peak near 700 nm.

4.5.3. III: 11 categories: SDGSAT-1 MII

The third scenario involves a final water classification into 11 categories, primarily Sustainable Development Goals Science Satellite 1 (SDGSAT-1) Multispectral Image for Inshore (MII). Similar to the GF-6 satellite sensor, SDGSAT-1 MII combines OWT4 and OWT7 due to band configuration diminishing their spectral differences. However, unlike GF-6, SDGSAT-1 combines OWT5 and OWT10 as one category while keeping OWT9 separate. A comparison of the band settings between these two satellite sensors reveals that the SDGSAT-1’s inclusion of two bands within the 570-700 nm range better highlights the differences in the declining trends among OWT5, OWT10, and OWT9 within that range. In contrast, the band configuration of the GF-6 makes these differences less noticeable.

4.5.4. IV: 13 categories: Sentinel-2 MSI, Envisat MERIS and Sentinel-3 OLCI

In the fourth scenario, the final water classification consists of 13 categories, involving three satellite sensors: Sentinel-2 MSI, Envisat MERIS and Sentinel-3 OLCI. These three sensors effectively maintain the original water types. To assess their classification capabilities, we calculate the J_SADs for the final categories. Sentinel-3 achieves the highest J_SADs value at 0.3764, signifying that its water classification results better align with the principle of maximizing inter-class distances. Conversely, Sentinel-2 displays comparatively weaker classification capabilities, with its lowest recorded J_SADs value at 0.3284. Thus, Sentinel-3 demonstrates the strongest classification capability, followed by MERIS, while Sentinel-2 has the weakest classification capability. However, it’s worth noting that Sentinel-2 offers advantages in spatial resolution, particularly for smaller inland waters.

From the above research, it’s evident that different sensors, owing to their distinct spectral response functions and resolutions, exhibit variations in their ability to identify and differentiate water types. Based on the final consolidation of results, we find that Sentinel-3, MERIS, and Sentinel-2 are the most capable sensors in distinguishing among the 13 OWTs. Following these, SDGSAT-1, Landsat-8, and GF-6 exhibit decreasing levels of discrimination ability. GF-1, GF-2, Landsat-5, Landsat-7, MODIS, and HJ-1 demonstrate relatively weaker discrimination capabilities.

5. Discussion

5.1. Comparative analysis of our water classification framework with established methods

Our study presents a sophisticated classification system that excels in scalability and sensor applicability, allowing it to be effectively adapted to various remote sensing platforms. Unlike the study by Bi (Bi et al. 2022) (Bi2022), our research is based on actual in-situ data, making our classification method not only more precise but also highly scalable and applicable to various sensors. While Bi2022’s dataset encompasses both inland and coastal waters, its methodology of equating in-situ data to OLCI wavebands for classification may have led to the loss of certain spectral characteristics. This approach could result in less distinct differentiation between certain water types, exemplified by the high similarity observed between Cluster 1 and Cluster 2, and a notable absence of clean water samples. This restricts the scalability of their method. Furthermore, compared to Spyrakos’s system (Spyrakos et al. 2018) (SR2018), our system provides comprehensive coverage of diverse water types, including challenging conditions like turbidity and algal blooms. Additionally, our approach effectively addresses the challenge of high similarity between water types, a common issue in broader classification systems. This enhanced precision and adaptability make our classification system especially valuable for detailed and precise monitoring of water Figure 9.

Figure 9. Comparative analysis of spectral reflectance clustering for inland waters. (a)Bi2022; (b) SR2018.

5.2. Analysis of split-merge K-means classification strategy

The traditional K-means algorithm typically converges to a locally optimal solution and is sensitive to the initial clustering center. In contrast, the stepwise iterative K-means algorithm improves clustering results by iteratively adjusting cluster centers and reallocating samples, but it may still get stuck in local optima. In contrast to the two aforementioned approaches, the split-merge K-means algorithm refines clustering results by splitting and merging clusters, allowing it to explore areas near local optima. This significantly increases the chances of discovering global optima, leading to higher-quality and more robust clustering results. Additionally, traditional K-means requires the upfront specification of the number of clusters, which can be a challenging task. In contrast, the split-merge K-means algorithm adaptively adjusts the cluster count by splitting dissimilar clusters and merging similar ones, making it more flexible and suitable for different dataset characteristics. Furthermore, traditional K-means struggles with imbalanced clusters, while the split-merge K-means can divide larger clusters into smaller subclusters. This capability effectively captures subgroups within the dataset and improves clustering performance, especially on imbalanced data distributions. In summary, the split-merge K-means algorithm outperforms traditional K-means and stepwise iterative K-means by reducing local optima, adapting cluster numbers, and handling imbalanced cluster sizes, among other advantages.

5.3. Factors affecting sensor classification capabilities

The classification capability of sensors is influenced by several factors. One crucial factor is the number of spectral bands, which determines the richness of spectral information that sensors can capture. A greater number of spectral bands can improve classification accuracy by providing additional spectral details. However, a sensor’s classification capability is not solely determined by the number of bands; it’s also influenced by its ability to capture the essential optical properties of water. In practice, as long as a sensor can capture these key optical features of water, it can achieve accurate water type classification. For example, in this study, Sentinel-3, MERIS, and Sentinel-2 satellite sensors possess different numbers of bands within the 400-900 nm range. Despite these differences, they all exhibit high water type classification accuracy because they capture the critical optical information of water. Besides the number of bands, the width and central wavelengths of spectral bands also impact water type classification accuracy. As an example of GF-6 sensor, we observed exhibit significant overlap between OWT9 and OWT5, making it challenging to accurately distinguish between these two water types. In contrast, the band configuration of the SDGSAT-1 increases the spectral differentiation between OWT9 and OWT5, thus improving classification accuracy. Therefore, when evaluating the classification capabilities of sensors, it’s essential to consider various factors, including the number of bands, the sensor’s ability to capture critical optical features of water, central wavelengths, and the degree of spectral overlap, to improve the accuracy of results.

6. Conclusions

We conduct a classification study of inland water using in-situ hyperspectral data and the K-means clustering method. By comparing the effectiveness of different K-means clustering approaches, it identifies the most suitable clustering method for optical classification of inland water. Additionally, we analyze the spectral characteristics and water quality parameter characteristics of various types of water. Finally, this classification system is applied to several commonly used sensors to develop water classification systems suitable for each sensor. The following conclusions are drawn from the research analysis:

1. The split-merge K-means clustering method better adheres to the principles of minimizing intra-class distance and maximizing inter-class distance, making it the most suitable clustering method for optical water classification.

2. Typical inland waters in China encompass 13 OWTs, which can be classified into four major categories: clear, turbid, eutrophic, and special typical water. There are significant differences in the water quality parameters and optical properties of each type of water.

3. We observe substantial differences in optical classification capabilities of 12 common satellite sensors. Among them, Sentinel-3, MERIS, and Sentinel-2 exhibit the highest classification capabilities. Following closely is SDGSAT-1. However, Landsat-8 and GF-6 exhibit declining classification abilities. GF-1, GF-2, Landsat-5, Landsat-7, MODIS, and HJ-1, on the other hand, display relatively weaker classification capabilities.

Building upon the key findings of our study, particularly the strengths and limitations identified in the use of various satellite sensors, we acknowledge the centrality of the OLCI sensor in our current research. Recognizing the importance of a more comprehensive analysis, we aim to expand our research to include a broader array of satellite sensors, thereby extending its applicability to global water bodies. Our future efforts will focus on developing extensive long-term time series products and conducting detailed spatiotemporal variability analyses. This progression is expected to not only deepen our understanding of optical water classification but also to make significant contributions to global environmental monitoring and management practices. The comprehensive classification system we have established provides a robust foundation for future research in this field, with the potential to inform and shape environmental policies and sustainable water management strategies worldwide.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The participants of this study did not give written consent for their data to be shared publicly, so due to the sensitive nature of the research supporting data is not available.

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

This work was supported by the National Key R&D Program of China under Grant number 2022YFC3204101; Henan Polytechnic University Youth Backbone Teacher Support Program (2023XQG–12).