Monitoring Chlorophyll-a concentration in karst plateau lakes using Sentinel 2 imagery from a case study of pingzhai reservoir in Guizhou, China

ABSTRACT Chlorophyll-a concentration (Chla) is an important index for water eutrophication. In this study, retrieval models of Chla were established based on the measured water spectra, spectral response function, measured Chla and the corresponding Sentinel-2 imagery of the Pingzhai Reservoir, the first large-scale trans-regional, trans-basin, and long-distance source reservoir in Guizhou. The retrieved results from 11 Sentinel-2 from 2018 to 2021 were used to analyze the spatiotemporal variations in Chla and the influence of different environmental factors on their spatial differentiation, providing a powerful approach for monitoring Chla in the Pingzhai Reservoir. Our binomial function model based on B8*(B7-B5) of Sentinel-2 yielded acceptable to high fitting accuracies, accounting for 89% of the variation in Chla. Overall, the Chla was relatively low, with a mean value of 10.24 μg/L. Higher Chla were distributed in the catchment area, such as the Nayong River and the dam. Moreover, significant seasonal fluctuations and intra-year changes were observed . Spatio-temporal variations in Chla were influenced by human activities and environmental factors such as Dissolved Oxygen (DO), Total Nitrogen (TN), and Ammoniacal Nitrogen (NH4 +-N). Our work provided compelling evidence that Sentinel-2 could be used for quantitative inversion of Chla in Pingzhai Reservoir.


Introduction
Chlorophyll-a concentration (Chla) is well recognized as an important index for measuring the depth of the euphotic layer, phytoplankton abundance in lakes, water quality, biophysical status and eutrophication (Coz et al., 2017;Mollaee, 2018). Damming can significantly disrupt the continuity of a river by altering natural properties such as water depth, river bed, hydrodynamic force, and river morphology, impacting material circulation, water quality, and aquatic biodiversity. An increasing body of evidence suggests that increasing reservoir age is paralleled by increased eutrophication (H.Q. Yang et al., 2016). In this regard, it has been reported that since the emergence of eutrophication in inland lakes in the 1930s, approximately 40%-50% of the world's lakes and reservoirs have been affected, making it one of the most intractable water environment problems (G.W. Zhu et al., 2018). Given that the rapid growth of algae in the water is a direct manifestation of eutrophication, the concentration of Chla allows an accurate assessment of water eutrophication since it is an important indicator of plankton abundance (Peppa et al., 2020).
The remote lakes on the Guizhou Plateau are surrounded by mountains and influenced by geological tectonic movement, with steep water depth and shore, making it difficult to monitor these lakes regularly. Field sampling and indoor measurement methods of conventional Chla monitoring are time-consuming, laborious, and expensive. Moreover, the findings cannot be generalized to other regions, and spatiotemporal Chla variations in a large area cannot be obtained (S. J. Guo et al., 2021;H. S. Lian et al., 2017). In addition, both natural forces and human activities may affect the distribution pattern of Chla, resulting in significant heterogeneity in the spatiotemporal dynamics and spatial distribution of Chla (Kaire et al., 2016). Ocean Color Remote Sensing has the characteristics of rapid, large range and periodicity, which can better reflect the continuity, spatiality and regularity of water environment changes, and can overcome the lengthy, wide-ranged and expensive elements of conventional monitoring of lake water quality (Arias-Rodriguez et al., 2020;Al-Kharusi et al., 2020;Yuan et al., 2018;. Sentinel-2A/2B is the European space agency multispectral imaging "Copernicus" mission satellite, carrying the payload for 13 spectra of multispectral imaging (MSI), with a temporal resolution of 5-10 days. It is mainly used for high resolution and high revisit cycle monitoring of the world's land coast, biological mapping and hazard assessment, etc. (C. Lanaras et al., 2018;Zbyněk Malenovsky et al., 2012). A study by Al-Kharusi et al. (2020) showed that Sentinel-2 MSI could significantly contribute to color research of lake water in Sydney. In addition, some researchers used MSI data to invert water color parameters such as Chla and suspended matter, substantiating that the MSI sensor has great potential in monitoring the water quality of inland lakes (Cazzaniga et al., 2019;M. Elhag et al., 2019;Gernez et al., 2015;H.Z. Liu et al., 2017;Katja et al., 2016;Kupssinskü et al., 2020;Peppa et al., 2020;S. Poddar et al., 2019;. At present, various visible light band ratio algorithms based on Sentinel-2 MSI images can be used to estimate Chla levels, mostly based on the ratio between the red and green bands or the near-infrared band. For example, Kaire et al. (2016) demonstrated a good correlation between the band ratio algorithm (B5-(B4+ B6)/2) calculated using Sentinel-2 MSI and the estimated Chla concentration in Estonian lake water. Moreover, Thu et al. (2017) found that the ratio of the third and fourth bands of MSI could be used to estimate the concentration of Chla in Lake Ba Be. A study by Grendaitė et al. (2018) demonstrated that the combination of Sentinel 2 MSI band ratio (B5/B4) and three bands (B4, B5 and B8A) yielded the best performance to estimate Chla concentration. However, these algorithms for estimating Chla based on the Sentinel 2 MSI specific band have not been applied; many uncertainties prevail over the successful application for lakes with different geographic environments and Chla concentrations. These findings highlight the need for further studies to explore the best combinations of different bands to estimate Chla concentration.
The Pingzhai Reservoir, located in the karst plateau area, is the core water source project of the Central Guizhou Water Diversion Project in Guizhou. It is a typical Case II Waters, with strong karst development, a complex hydrogeological environment and significant terrain variations. Influenced by factors such as inflow from five Rivers (Nayong River, Shuigong River, Zhangwei River, Baishui River and Hujia River) and human activities, the water color parameters of Pingzhai Reservoir exhibit spatial and temporal heterogeneity, and the optical properties of water bodies are complex and varied. Recently, water eutrophication has occurred in Pingzhai Reservoir.
Chlorophyll-a concentration is an important indicator of plankton abundance and eutrophication. Indeed, obtaining high spatio-temporal resolution information of Chla concentration is essential for understanding water quality dynamics, identifying the driving forces, and protecting aquatic ecosystems. Therefore, it is of great practical significance to use remote sensing technology to monitor the water quality of Pingzhai Reservoir (J. Kong et al., 2021). Y. S. Dan et al. (2020) analyzed the spatial variability of Chla concentration by using Sentinel-2 imagery in Pingzhai Reservoir, but a continuous, large-span and spatiotemporal analysis of Chla was not conducted. In addition, the inversion accuracy and result analysis of Chla were still to be optimized.
Using remote sensing technology and hyperspectral data to break through problems such as difficulty and high cost and poor representativeness in regular monitoring of Chlorophyll-a concentration in karst plateau lakes, which provided a methodological reference and practical application basis for remote sensing inversion of Chlorophyll--a concentration in karst plateau lakes. This study focuses on the key issues such as water quality monitoring, optical characteristics analysis of water and the construction of Chlorophyll a concentration inversion model by remote sensing in karst lakes. Considering Pingzhai Reservoir as the research area, this study aimed to: (i) develop Sentinel-2 MSI-based Chla retrieval models in Karst Plateau, (ii) compare the prediction ability of different Sentinel-2 MSI based Chla retrieval models and validate the model established in this study, and (iii) analyze the correlation between Chla and environmental factors and clarify factors that significantly influence water eutrophication in Pingzhai Reservoir by inversion results of multi-temporal remote sensing images.

Study area
The Pingzhai Reservoir (105°17′3′′E -105°26′44′′E, 26°29′33′′N -26°35′38′′N) is located in the Guizhou Province, at the center of karst area in southwest China, where carbonate rocks have the largest distribution area and karst is well developed. The dam of Pingzhai Reservoir is located in the Sancha River basin of the upper Wujiang River, with tributaries from Hujia, Baishu, Zhangwei, Shuigong and Nayong Rivers. The reservoir area is long and winding (Figure 1), being the source of the Central Guizhou Water Diversion Project and an important water source for the central Guizhou region. The terrain mainly consists of plateau mountains, with an average elevation of 1100 m, dam length and height of 355 m and 1627 m, reservoir area of 14.57 km 2 , lake shoreline of 9.489 × 10 4 m, total storage capacity of 10.89 × 10 8 m 3 , adjusted storage capacity of 4.48 × 10 8 m 3 , normal storage level of 1331 m, and average water depth of 50 m. The water level rises by about 1 m during the wet season and drops by 2-3 m during the dry season, resulting in a discontinuous reservoir water body and shallow shoal. With the aggravation of anthropogenic activities and the massive discharge of domestic sewage, the state of the water environment is relatively serious.

In-situ sampling and laboratory measurement
According to the hydrogeographical characteristics of Pingzhai Reservoir and "Water Quality Technical Regulation on the Design of Sampling Programmes" (HJ 495-2009), eight surface routine monitoring points were set up at five river sections (NY3, SG3, ZW8, BS2, HJ2), river confluence (PZ1, SG1) and reservoir dam (PZ4). Six fieldworks were carried out on 18 November 2018, 12 October 2019, 18 May 2020, 27 July 2020, 14 November 2020, and 28 March 2021, and a total of 48 sites were used to verify the Chla concentration estimated by sentinel imaging.
According to the area of the lake and reservoir, the grid method was adopted to evenly distribute the samples in the Nayong River, Shuigong River, Baishui River, Zhangwei River and Hujia River at intervals of 1 km. The samples sites were added in the Nayong River, dam area and other wide river areas, with a total of 40 sampling points. According to the ratio of 1 to 1 proportion, the sampling points were randomly divided into the training set (20 sampling points) and the validation set (20 sampling points). Among them, the training set contained the extreme value information of Chlorophyll-a. The training set was used to build the inversion model of Chla in Pingzhai Reservoir, and the verification set was used to verify the inversion accuracy of the model and determine the optimal inversion model. Surface spectral measurements and simultaneous water quality sampling were performed between 11:00, 16:00 on 26 August 2020 and 28 March 2021. The sites were geolocalized with a Global Positioning System (GPS) receiver ( Figure 1).
The WTW Multi3430 portable multi-parameter water quality analyzer (Wissenschaftlich-Technische-Werkstaetten of Xylem Inc., German) was used for on-site water quality collection to detect the water temperature (Wt), pH, dissolved oxygen (DO), and electrical conductivity (EC), and simultaneously record the surrounding environmental parameters. The resolution of pH, DO and EC were 0.01, 0.01 mg/L and 1 μs/cm, respectively. A CleverChem380 automatic intermittent chemical analyzer (DeChemTech, Germany) was used to detect ammonia nitrogen (NH 4 + -N) (by salicylic acid spectrophotometry), total nitrogen (TN) (by alkaline potassium persulfate digestion UV spectrophotometry) and total phosphorus (TP) (by ammonium molybdate spectrophotometry). The Secchi depth data (SD) was also analyzed by a standard 20cm plastic Secchi disk (Wildco, USA).
Water samples were collected in a polyethylene plastic bucket at 0.3-0.4 m underwater and then immediately stored in a 500 ml black polyethylene plastic reagent bottle sealed in a dark place. 1% MgCO 3 was added to reduce the decomposition of Chla. The samples were sent to the laboratory for Chla measurement on the same day or the next day. First, the water sample was filtered using a 0.45 μm-filter membrane and freeze-thawed thrice. The extraction solution was then centrifuged, placed in a refrigerator at low temperature (−20°C) for 24 h and filtered with a GF/F glass fiber filter membrane for Chla measurements. 90% acetone (10 ml) was used to remove the filter membrane and extract; then, the supernatant was removed and placed into a cuvette. The absorbance was then measured with an RF-5301 spectrophotometer (Shimadzu, Japan). The concentration of Chla was calculated according to the Department of Environmental Protection (2017). The formula is as follows: where A 630 , A 647 , A 664 and A 750 are the absorbance at the wavelength of 630 nm, 647 nm, 664 nm and 750 nm, respectively. V 1 and V 2 are the volumes of the extract after constant volume and the water sample, and L is the optical path of the colorimeter. Field water spectra were measured using a portable ASD Fieldspec4 (Analytical Spectral Devices Inc., USA) with a spectral range of 350-2500 nm and a spectral sample interval of 1 nm. After calibrating the spectrometer and the observation geometry and the integral measurement time were set, the radiance values of the grey plate, water and sky were measured. The measurement procedure was in accordance with the steps and procedures described by J. W. Tang et al. (2004) in spectrum measurement and analysis of water bodies. Each sampling point was observed 10 times. The ViewSpecPro software (Analytical Spectral Devices Inc., USA) was used to eliminate and filter the spectra. The remote sensing reflectance (R rs ) of each sampling point was calculated with the equation: where L sw ,L sky and L p are the measurement signals from water, sky, and grey plate diffuse reflectance standard, respectively; r is the reflectance of waterair interface; π is the sphericity constant; ρ p is the calibration file of the grey diffuse reflectance standard; E d (0 + ) is the total incident irradiance of water surface (it can also be measured directly using the calibrated ASD cosine receiver RCR).

Satellite images and pre-processing
The remote sensing data used in this study consisted of Sentinel-2 MSI images. For the Pingzhai Reservoir in Guizhou plateau, where the weather is changeable, the high revisit period and spatial resolution of MSI provide more possibilities for water color monitoring. In addition, many cases of inland and evolutionary water quality assessment using Sentinel-2 MSI imagery have been documented domestically and abroad (Casal et al., (Doerffer R et al., 2007;Vitor M et al., 2017).

Sentinel-2 MSI spectral simulation
Based on measured water spectral data and Spectral Response Functions (SRF), the Sentinel-2 MSI remote sensing band reflectance was simulated. The formula is as follows: Where R i is the simulated band reflectance of the band i; R rs is the measured spectral reflectance; SRF (λ) is the spectral response function; λ 1 andλ 2 are the minimum and maximum values of band i.

Normalized difference water index
The Normalized Difference Water Index (NDWI) was proposed by Mcfeeters (1996) based on the strong reflection and absorption characteristics of water body information in green light and near-infrared bands. A data extraction method was established by contrasting these characteristics. According to the wavelength information of band 3 and band 8 of Sentinel-2 image, an appropriate segmentation threshold (0.035) was set for water separation, using the formula: Where B3 is the reflectance of the green light band, and B8 is the reflectance of the near-infrared band.

Accuracy evaluation
The relative error and the absolute error are used to judge the reliability of measurement results, and the average value and standard deviation are used to judge the general level and dispersion degree of data. Formulas are as follows: ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi 1 n ε¼ ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi 1 n where R r is the remote sensing reflectance after atmospheric correction of Sen2cor, R i is the reflectance after spectral response function of band i, R 1 is absolute error, R 2 is a relative error, μ is the average value, σ is standard deviation, ε is the overall relative error, and n is the number of bands.
SPSS was used to carry out a statistical analysis of the data. Two statistical indicators (coefficient of determination, R 2 , and root mean square error (RMSE)) were used to examine whether the estimated and measured values were consistent and to evaluate the accuracy of the inversion model (Equations 11 and 12).
RMSE ¼ ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi P n i¼1 where x i and y i are the measured and the predicted values of sample point i, respectively; n is the sample points of the training and verification sets; and x and y are the average values of the training set and validation set.

Validation of atmospheric correction results
The measured remote sensing reflectance values of each band were calculated based on the SRF. Meanwhile, three synchronous water surface measurement points (PZ4, NY3 and SG1) were selected from the corrected image of 26 August 2020, and calculated using the formulas (6), (7), (8), (9) and (10) to evaluate the correction error of the Sen2Cor algorithm. The atmospheric correction results and error accuracy of synchronous measurement points are provided in Table 1.
The atmospheric correction results of the Sen2Cor model yielded good accuracy, and more than 70% of relative errors were about 20% ( Table 1). The mean relative error (30%) and absolute error (0.3%) of the B5 and B6 bands were relatively high. Comparison between atmospheric correction spectral curves (Figure 2) showed that the atmospheric correction curve and the measured spectral curve exhibited similar variations, and the corrected results were similar to or lower than the measured spectral reflectance, which significantly decreased after B5, in line with the absorption characteristics of the water body.

Validation of water extraction results
During the process of water data extraction, we found that NDWI could extract water data from a large area. However, for small rivers, it was difficult to extract complete water data, resulting in loss of data on small rivers. Shadowing influenced our ability to extract data since it was not easy to distinguish water bodies from shadows. Therefore, morphological expansion and thinning operations (Maragos & Schafer, 1987) were used to connect and fill the disconnected water bodies. Then the water body skeleton line was extracted and processed by an image thinning algorithm.  The extracted water pixel result was compared with a 2 m field of UAV image, and shadow areas were eliminated by visual judgment [ Table 2].

Analysis results and measured spectrum of Chla
After removing the abnormal spectra of 16 samples influenced by strong wind current and unstable light, 64 valid data samples were finally obtained. The statistical analysis (Table 3) showed that the mean value of Chla in all samples was 12.26 μg/L (range 4.12 to 30.79 μg/L), with a standard deviation of 6.01 μg/L, a coefficient of variation of 49%. No significant differences were found between the training (average Chla = 12.86 μg/L, standard deviation = 6.36 μg/L, CV = 50%) and validation (average Chla = 13.77 μg/L, standard deviation = 6.80 μg/L, CV = 49%) datasets.
The measured surface spectrum (350-900 nm) of the Pingzhai Reservoir ( Figure 3a) indicated that aquatic algae were significantly present in the Pingzhai Reservoir, and there was a good correlation between the spectral curve and concentration change.   At 400-500 nm, water reflectance was low due to the strong absorption of Chla, and significant reflection peaks were observed from 550 nm to 700 nm. The weak absorption and cell scattering of algae chlorophyll or carotene accounted for the reflection peak near 560 nm. In comparison, the reflection peak near 700 nm was the most significant feature of watercontaining algae, which is known as one of the basics to determine whether the water contained algae (Gao et al., 2012), while the valley of reflection near 600 nm and 680 nm were caused by the large absorption coefficient of cyanobacteria and the strong absorption of algae Chla, respectively. It has been established that under normal circumstances, a high Chla value tends to produce a high reflectance curve (Coz et al., 2017).
In the present study, the concentration of Chla in August 2020 was significantly higher in March 2021. The correlation between Chla and the Rrs of all the samples was calculated (Figure 3b). We found that Chla positively correlated with Rrs of water in each band, and the sensitive bands were mainly concentrated in the range of 580-750 nm. Reflection peaks were observed at 595 nm, 645 nm and 705 nm; the highest positive correlations were 0.63 at 595 nm on 28 March 2021, and 0.73 at 705 nm on 26 August 2020. According to the central wavelength settings of sentinel-2 MSI sensor bands, bands 4, 5, 7 and 8 were sensitive to Chla levels in the Pingzhai Reservoir.

Construction and validation of Chla inversion model
To identify factors most sensitive to Chla concentration, the four bands of Sentinel-2 MSI sensor were combined through 16 combination forms (Table 4),  and 4, 38, 94 and 33 combination categories were generated respectively. The correlations between them and Chla concentration were then analyzed. The correlation coefficient between Chla concentration and different band combinations ranged from −0.8 to 0.8 (Figure 4). Among these, the correlation coefficients of bands 4, 5 and 8 with Chla were 0.06, 0.27 and 0.65, respectively, which were not suitable for remote sensing estimation of Chla concentration in Pingzhai Reservoir. Among the 169 combinations, there were 3 combinations (B7, B7*B8 and B8*(B7-B5)) with an absolute correlation coefficient greater than 0.7.
B7, B7*B8 and B8*(B7-B5) were selected to transform the 10 windows with different pixel values based on the relationship between chlorophyll concentration and Sentinel 2 MSI band reflectance. Finally, their correlation with Chla was analyzed. The pixel size window with the best correlation was selected for modeling. The combinations with the best correlation with Chla were B8*(B7-B5), followed by B7 and B7*B8, and the average correlation coefficients of 10 windows were 0.72, 0.70 and 0.54, respectively. In general, windows of 1 × 1, 2 × 2, 3 × 3, 4 × 4 . . . 8 × 8, 9 × 9, and 10 × 10 had a good correlation with Chla, and the variation of different windows for the three combinations was consistent. Among them, the 3 × 3 window yielded the best correlation with Chla, especially the 3 × 3 window of the B8* (B7-B5) combination [ Figure 5].
The estimation window and factors with good correlation were selected for regression analysis, and a regression equation was established. As shown in Table 5, the estimated and measured R 2 value in the single-band linear model, two-band (B7*B8) and three-band linear model (B8*(B7-B5)) was 0.512, 0.588, and 0.569, respectively; while the R 2 of the three combined logarithmic estimation models was below 0.5. However, the fitting degree of the above three parameters in the linear and the logarithmic models was not high. The optimal band combination B8*(B7-B5) was selected considering the stability and practicality of the model. The Chla concentration was estimated using the linear, exponential, logarithmic, and binomial models. It was found that the fitting degree of binomial function was the highest and the correlation was best (P < 0.01). Its expression is: where x is the normalized reflectivity value of B8*(B7-B5) and y is the concentration value of Chla. The binomial model (Equation 13) with the highest fitting accuracy B8*(B7-B5) was applied as an independent variable to Sentinel-2 MSI imagery on 26 August 2020, and 19 March 2021, and the predicted values of the model were verified by the test samples ( Figure 6). We found that the measured and predicted values of leaf green concentration in water were distributed around the y = x function. On 26 August 2020, the measured value of Chla concentration in Pingzhai Reservoir was consistent with the predicted value of the model (R 2 = 0.741, RMSE = 6.50), while the

Spatiotemporal distribution of Chla
The final model (Equation 13) was applied to 11 scenes on Sentinel 2 MSI images to obtain the dynamic changes in Chla concentration in the Pingzhai Reservoir over different months (Figure 7). From 2018 to 2021, the distribution of Chla in the study area exhibited a definite spatial distribution pattern, and the average concentration value was relatively stable at about 10.24 μg/L. The Chla values were relatively high in most anthropogenic areas, river junction areas and slow-flow areas in the Pingzhai Reservoir. On 26 August 2020, the concentrations of Chla distributed in 6%, 21%, and 73% of the water body were above 20 μg/L, below 5 μg/L and approximately 10 μg/ L, respectively. The distribution of Chla was significantly different in Pingzhai Reservoir, and the area in front of the dam showed an increasing trend from north to south. As shown in Figure 7, the Chla values in front of the dam and the Nayong River section were significantly higher than in other river sections. In the Nayong River, the Chla ranged from  12.13 μg/L (May) to 20.61 μg/L (November) with large monthly variations, and two significant peaks were observed in July and August 2020 (27.97 μg/L and 29.52 μg/L, respectively), with an annual mean value of 16.05 μg/L. As the largest trunk stream, the upper reaches of the Nayong River pass through Yangchang Town and Nayong County. It is reportedly significantly affected by human activities such as urban sewage and wastewater discharge from key industrial and mining enterprises leading to a high concentration of Chla . It has been established that excessive discharge of nitrogen and phosphorus in sewage leads to phytoplankton proliferation in lakes, resulting in lake eutrophication (J. P. Grover, 2017). However, it should be borne in mind that Chla is an essential component of phytoplankton, and any sewage discharge can significantly increase the Chla concentration. The mean Chla value in front of the dam was 12.57 μg/L, with large monthly fluctuations and several peak values. High Chla values above 16 μg/ L were observed in July, August and November 2020 and December 2018, with a peak value of 19.62 μg/L. This phenomenon may be caused by the increase in Chla in the area in front of the dam due to runoff input from the five rivers and water bodies with different Chla concentrations. In the northeast region of the Pingzhai Reservoir, Zhangjiawan Town of Nayong County and Jichang Town of Zhijin County are close to each other. Zhangwei River is one of the main inflow rivers of the Pingzhai Reservoir, with high Chla values, and has been reported to be significantly affected by point-source pollution (X. M. Liu et al., 2019). The Chla concentrations of Hujia River and Shuigong River were relatively low, which may be attributed to the fact that both rivers are long and narrow, and the water quality in the surroundings is relatively good with little anthropogenic activity. Due to the influence of the topography, water depth and other factors, pollutants mainly came from agricultural non-point sources, and the water quality of most rivers was relatively good The Chla concentration exhibited periodic fluctuations (Figure 7). From March to May 2020, the Chla concentration was relatively stable in the Pingzhai Reservoir, increasing from July 2020 and peaking at 29.52 ug/L in August 2020. It gradually decreased, reaching a nadir of 1.01 μg/L in January 2021, and gradually stabilized until May 2021. In July, August and November 2020, the Chla was widely distributed with high concentration levels, which may be due to fertilization resulting from agricultural activities and the entry of nutrients such as nitrogen and phosphorus into the water that provides a substrate for algae growth.
The Chla inversion in Pingzhai Reservoir was obtained for different seasons (Figure 8). The Chla concentration exhibited seasonal variations, higher in summer and autumn than in spring and winter. The Chla values in the Hujia River were low for all seasons, fluctuating between 1.41 μg/L and 6.96 μg/L. In Baishui River, the Chla concentration was slightly higher during summer and autumn (approximately 15.56 μg/L) and lower in spring (approximately 6.80 μg/L). The Chla values gradually increased from spring to autumn, and a peak value of 13.85 μg/L was observed in autumn in Zhangwei River. In the Shuigong River, the Chla concentration remained at low levels during four seasons, with a peak value of 5.65 μg/L during spring, followed by 5.08 μg/L during autumn, and 2.68 μg/L during summer. Overall, the seasonal variations in Chla concentration in the dam, and Zhangwei, Hujia and Baishui rivers were consistent. Chla concentrations were relatively high during summer and autumn and low during other seasons. However, due to variations in Chla concentrations in the Nayong River during different seasons, it is necessary to identify factors that drive changes in water Chla concentration.

Influencing factors of Chla concentration change
There is ample evidence to suggest that water Chla concentration is closely related to the growth and distribution of phytoplankton and is affected by various environmental factors (Chen et al., 2020). In the present study, Spearman correlation analysis was conducted between the main environmental indicators (TN, TP, NH 4 + -N, Wt, pH, EC, DO, SD) and Chla concentration levels in Pingzhai Reservoir. As shown in Figure 9, a positive correlation was found between the Chla concentration and environmental factors. In spring, Chla concentration was negatively correlated with EC and pH, with correlation coefficients of −0.639 and −0.552, respectively. During summer, Chla concentration positively correlated with EC and DO with correlation coefficients of 0.707 and 0.558, respectively, and negatively correlated with TN and SD with correlation coefficients −0.628 and −0.584, respectively. Moreover, Chla concentrations during autumn positively correlated with all environmental factors and negatively correlated with TN, NH 4 + -N and DO. Finally, a positive correlation was found for TN and SD during winter, with correlation coefficients of 0.699 and 0.669, respectively.
It has been shown that differences in DO, TN and NH 4 + -N lead to variations in phytoplankton growth, yielding differences in Chla concentration (G. Xing et al., 2001). Importantly, Chla positively correlated with DO in all seasons. Generally speaking, higher concentrations of Chla are associated with more phytoplankton growth leading to higher O 2 release via photosynthesis, hence increasing water DO concentration levels. Conversely, phytoplankton respiration requires the participation of DO. When water DO levels are sufficient, it is conducive to phytoplankton growth, such as algae in the water. Nonetheless, the relationship between Chla concentration and nutrients is complex (Wen et al., 2017). In the present study, Chla concentration negatively correlated with TN in spring and summer and positively correlated with TN in autumn and winter. A positive correlation with NH 4 + -N was observed in summer and autumn, but no significant correlation was observed with TN in winter. It is well-recognized that in summer, the high temperature and sufficient light time are beneficial to the microbial activities within the water body, enhancing denitrification with a large amount of inorganic nitrogen absorbed during the rapid growth of algae and other phytoplankton, thus reducing the availability of nutrients in the water body.
To explore the potential impact of different environmental factors during each season on the water body in the study area, TN, TP, NH 4 + -N, Wt, pH, EC, DO and SD were sorted by principal components analysis (PCA) toolbox of SPSS ( Figure 10). Interestingly, the ability of each environmental factor to affect Chla concentration during winter was stronger than during spring, summer and autumn. All environmental factors positively correlated with the Chla concentration in most areas and negatively correlated in the Baishui River during spring. EC exhibited the largest contribution to Chla concentrations, followed by TP, pH, SD, DO, Wt, TN and NH 4 + -N. During summer, the Chla concentration positively correlated with EC, DO, TN and NH 4 + -N, and negatively correlated with pH, SD and Wt. No significant correlation was observed with TP. In the second principal component, the Chla concentration positively correlated with SD, TN, pH and NH 4 + -N, and negatively correlated with Wt, DO, EC and TP. During autumn, the Chla concentration exhibited a strong positive correlation with most environmental factors and accounted for 54.5% of the variance in the first principal component. During winter, the Chla concentration in the first principal component positively correlated with environmental factors, accounting for 45.8% and 30.1% of the variance in PC1 and PC2, respectively, with a cumulative contribution rate of 75.9%. The first principal component included DO, Wt, NH 4 + -N and SD. Algae propagation in water coupled with changes in water temperature are wellrecognized to influence water DO and NH 4 + -N levels. Therefore, the first principal component was called the phytoplankton factor. The second principal component included EC, pH, TN and TP. Indeed, TN and TP are mainly caused by agricultural fertilizers and affect the pH and EC of water by affecting the growth of phytoplankton. Accordingly, the second principal component was called the anthropogenic disturbance factor.
The distribution of each point in the four seasons was scattered and significantly influenced by environmental factors. The first principal component was closely related to BS2, PZ1 and NY3 due to their large vertical values in PC1, indicating that the Chla concentration at PZ1 and NY3 was affected by phytoplankton. In the second principal component, SG3, ZW8 and HJ2 yielded the largest values when projected on PC2. Accordingly, the concentrations of these three on Chla concentrations were affected by human disturbance. NY3, PZ1 and PZ4 were projected on the right side of 0 on PC1 and exhibited a close positive relationship with PC1. Overall, the Chla concentration was mainly affected by TN, NH 4 + -N and DO, and the relationship with environmental factors was most intricate during summer, followed by spring, autumn and winter. Moreover, the Nayong River and the dam (NY3, PZ1, and PZ4) were greatly disturbed by environmental factors during summer, autumn and winter.

Effects of remote sensing model on inversion accuracy of Chla
In this study, the remote sensing inversion model of Chla concentration was successfully applied to Sentinel 2 MSI images of 11 scenes in Pingzhai Reservoir (Figure 7). The Chla values of long-term monitoring points were used to validate the accuracy of the inversion values (Table 6). We found that the model was well-fitted to the data in August 2020, while it yielded significant overestimations for other months, which may be attributed to the fact that the acquisition time of the measured water spectrum and Chla concentrations and Sentinel-2 imaging time occurred on the same day (26 August 2020), which yielded a good inversion effect. Moreover, spatiotemporal scale differences between the models constructed based on field measured data and satellite image data resulted in poor applicability of the model.
The three-band model with [B5-(B4+ B6)/2] as the inversion factor based on Sentinel-2 MSI satellite constructed by Kaire et al. (2016) yielded a model coefficient of determination of 0.83 when applied to the inversion of Chla concentrations in Estonian lakes. Moreover, Dan et al. (2020) established an inversion model of Chla concentrations in the Pingzhai Reservoir based on Sentinel-2 MSI imagery yielding a coefficient of determination of 0.808; however, the RMSE was 10.06%. Most importantly, the accuracy was lower than the binomial function model (RMSE = 9.94%) established in this study using B8* (B7-B5) of Sentinel-2 MSI.  Figure 11. Comparison of inversion accuracy between this model and other three models.
The above-mentioned model was applied to Sentinel-2 MSI images of Pingzhai Reservoir during the same period. The results were compared to determine the reliability and applicability of Sentinel-2 MSI based models developed in this study ( Figure 11). The B8*(B7-B5) binomial model exhibited relatively higher accuracy in remote sensing inversion of the Chla concentration in Pingzhai Reservoir. In contrast, the cubic function model constructed by Y. S. Dan et al. (2020) significantly overestimated the Chla concentration, with an RMSE of 59.67%. Moreover, the Chla concentration used by Y. S. Dan et al. (2020) to establish the model ranged from 0 μg/L to 68.21 μg/L (its inversion time is November 2017, which coincides with the construction period of the phase I project of the central Guizhou Water Diversion Project), which is higher than range (1.01 μg/L-29.52 μg/L) used in this study. Additionally, the inversion effect of the [B5-(B4+ B6)/2] three-band model used to predict Chla concentration in this study was relatively poor, with R 2 and RMSE values of 0.767 and 11.35%, respectively. According to the spectral response function, Kaire et al. (2016) used the fourth and sixth bands of Sentinel-2 MSI to establish a model. As shown in Figures 3(b) and Figure 4, the correlation coefficient between the fourth and sixth bands and Chla was weak, leading to large errors during remote sensing estimation of Chla concentration in the Pingzhai Reservoir.

Effects of environmental factors on Chla
Dissolved oxygen is an important parameter for evaluating water nutrient levels and is also significant condition affecting algae growth (F. Yang et al., 2020). The Chla concentration negatively correlated with DO in most lakes and reservoirs (W. He et al., 2019;A. Ménesguen et al., 2019), while the Chla concentration positively correlated with DO in Pingzhai Reservoir, similar to that in Fuxian (L.J. Wang et al., 2017) and Tianchi Lake (B. Wang et al., 2015). This may be due to the increase in DO content caused by the release of oxygen molecules by algae during photosynthesis, and the lower oxygen consumption during respiration and decomposition of organic matter than dissolved oxygen generated by photosynthesis. Dissolved oxygen significantly affects the Chla concentrations, and some water quality indices such as water temperature (Wt) and pH are important influencing factors. The relationship between Wt and Chla is complex. Generally, Chla positively correlates with Wt (Huang et al., 2021;X. Zhao et al., 2018), and sometimes negatively correlated (Çelik, 2018;Gallina et al., 2013). There was no significant correlation between the concentration of Chla and Wt during all seasons in Pingzhai Reservoir, which may be related to the small seasonal variation of air temperature in Guizhou and the influence of nutrients (X.M. Liu et al., 2019). Wt can affect the enzyme activity in algae cells and the rate of photosynthesis, and alter the Chla concentration. Meanwhile, the change in Wt can also alter other environmental factors in water, such as pH and EC, which in turn can affect the growth and reproduction of algae (Chen et al., 2020). The Chla concentration can be affected by influencing the acid-base reaction of the water body, the absorption and release of nitrogen and phosphorus nutrients, and the rate of algae photosynthesis. Additionally, during photosynthesis, algae can change the pH value by absorbing and releasing CO 2 (Khler et al., 2018). The Chla concentration was positively correlated with pH value, indicating that the increase of Chla concentrations was conducive to the increase of the photosynthetic rate of algae, and the pH increased ( Figure 9).
In most lakes and reservoirs, the growth of algae is mediated by a combination of various factors, such as water power, meteorology, nutrient, trace elements, etc. Studies have demonstrated that nitrogen and phosphorus nutrient input and accumulation are the main causes of eutrophication (Carpenter & Bennett, 2011;Stepanova, 2021;X. Zhang et al., 2021); however, N and P in the water do not always affect the proliferation of algae positively, suggestive of a complicated relationship (Popova et al., 2010). In our study, the water Chla concentration positively correlated with TP, TN, and NH 4 + -N during autumn, negatively correlated with TN and NH 4 + -N in spring, and positively correlated with TP during spring and summer. This is consistent with the research results of Jiang et al. (2019) on Wuliangsu Lake, which showed that Chla is dominated by nitrogen during spring and phosphorus during summer, thereby reflecting the significant influence of seasonal growth of algae on the concentration of TN, TP, and NH 4 + -N in water. In addition, an increasing body of evidence suggests that N and P are significant environmental factors affecting the growth and reproduction of algae and have an important impact on the Chla concentration. For example, Loures et al. (2016) studied the sub-tropical lake group in central Brazil and concluded that TN played a critical role in the distribution of Chla, consistent with our findings that TN was strongly correlated with Chla in this study. Consistently, Smith et al. (2016) advocated that N and P levels accounted for changes in Chla concentration. Moreover, G.W. Zhu et al. (2018) explored the determinants of Chla concentrations in Taihu Lake by laboratory analysis and provided compelling evidence that N and P are crucial to the distribution of Chla concentration.

Effects of the band selection on estimation of Chla
Accurately selecting characteristic bands based on the spectral characteristics of Chla concentration is the key to establishing a remote sensing inversion model of Chla. In 1998, some scholars inverted Chla by establishing a relationship model between Chlorophyll-a concentration and image band data based on water optical theory and achieved ideal research results (Kondratyev et al., 1998). Among them, the reflectance ratio of about 700 nm and 670 nm was the most prominent. Many studies also showed that the determination coefficient of Chla and the ratio of 700 nm to 670 nm in inland water bodies were greater than 0.8, and the inversion range of Chla was 0.1-350 mg/m 3 . Two measurements of the water reflectance spectrum of Pingzhai Reservoir showed obvious reflection peaks at 580 nm and 705 nm, while the correlation between the blue-green band and Chla was weak ( Figure 3). Therefore, the bluegreen band is not suitable for estimating the Chla concentration of Pingzhai Reservoir. In addition, the diversity of water body reflectance measured by different satellite sensors makes it difficult to use the unified estimation algorithm of Chla. However, using the known spectral parameter characteristics and band ratio method to establish the inversion model is ideal because of the interaction between the phytoplankton backscattering and the strong absorption of water. (Carder et al., 2004) Based on the semi-empirical method of band ratio, many studies have used airborne and ground-based hyperspectral data for Chla concentration inversion, with R 2 value of 0.75-0.99 for these models.
Herein, we found that, like many inland lakes, the concentration of Chla in Pingzhai Reservoir varied greatly on a seasonal basis, and the concentration was significantly higher in summer and autumn due to the influence of phytoplankton and human disturbance. Overall, the concentration of Chla is higher than 10 ug/L in Pingzhai Reservoir, and the correlation between water reflectance and Chla concentration is mainly between 580 nm and 750 nm, corresponding to Band4, Band5, Band7, and Band8 of Sentinel-2 MSI. According to the 196 combinations in Table 4, B7, B7*B8, and B8*(b7-b5) are strongly correlated with Chla concentration. Therefore, the above three combinations were used for model calculation when selecting the band ratio to estimate the Chla of Pingzhai Reservoir. Finally, we found that B8*(B7-B5) had the best ability to estimate the concentration of Chla. Among various algorithms for estimating Chla, the relationship between Chla and the "red edge" reflectance of the visible spectrum showed a strong correlation between Chla and the reflectance difference of the near-infrared and red regions (Schiller & Doerffer, 1999) Based on the assumption that backscattering does not vary with the spectrum and can be estimated using near-infrared wavelengths, Gitelson et al. (2003) proposed a three-band model, using the sensitive band (670 nm) and insensitive band (710 nm) of Chla absorption and the 750 nm band less affected by absorption and able to estimate scattering. We corroborated that the three-band model has good inversion accuracy for turbid and high biomass water. In this study, the Chla inversion model constructed by B8, B7, and B5 was successfully applied to Sentinel 2 MSI images of Pingzhai Reservoir in other periods, and the accuracy of the estimated values was verified by using the Chla concentration of long-term monitoring sites as test samples. It was found that good inversion results were achieved.

Conclusions
This study developed Chla concentration retrieval models with simulated Sentinel-2 MSI spectra for insitu Chla concentrations in Pingzhai Reservoir. Our final model was successfully applied to satellite images for Chla retrieval to analyze the spatio-temporal changes of Chla concentrations and influencing factors. Sentinel-2 MSI image reflectance was simulated by measured water spectrum, and the inversion model of Chla concentrations with B8*(B7-B5) as an independent variable was established based on the Bands 5, 7, and 8 of MSI, which had good applicability in the estimation of Chla concentrations in Pingzhai Reservoir. Compared with other Chla models based on Sentinel-2 MSI images, the developed model achieved better results in the estimation of Chla concentration in Pingzhai Reservoir. This model was successfully applied to 11 images of MSI and yielded a clear spatial distribution pattern of Chla concentrations. Moreover, the highest Chla concentrations were observed in the Nayong River and Baishui River, followed by Zhangwei and Baishui rivers. Chla concentrations in Shuigong and Hujia Rivers were relatively low. Interestingly, Chla concentrations were high during autumn and low during winter, and the peak value was observed during August. It has been established that TN, NH 4 + -N, and DO are important environmental factors affecting the spatial variation of Chla concentrations. During summer, autumn, and winter, these factors had the largest contributions in Nayong River, followed by the dam and other regions. Of note, the most significant contribution was observed during autumn. No significant correlation was observed in spring.
This study investigated the remote sensing inversion and spatial-temporal analysis of Chla concentrations in Pingzhai Reservoir. The results showed that Sentinel-2 MSI remote sensing assessment has huge prospects for monitoring Chla concentrations in this area. However, due to the regional complex topography and water quality, some shortcomings were present in our study. Given the lack of accurate data, other lakes on the Guizhou Plateau could not be monitored, emphasizing the need for further studies to increase the robustness of our findings.

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

Funding
The research was supported by the Regional Project of National Natural Science Foundation of China.
[U1612441], Guizhou high-level innovative talent training plan -"best" level talent ([2016] 5674), and study on the scale effect of characteristic crops identification in Karst Mountain based on Multi-source Remote Sensing (QJH YJSKYJJ〔2021〕090), thanks to Guizhou Key Laboratory of Mountain Environment and National Karst Rocky Desertification Control Engineering Technology Research Center, Thanks for their help in spectroscopic measurement and laboratory;Guizhou high-level innovative talent training plan -"best" level talent [[2016] 5674];

Data availability of statement
All data generated or analyzed during this study are included in this published article.