Water Quality Retrieval from ZY1-02D Hyperspectral Imagery in Urban Water Bodies and Comparison with Sentinel-2

: Non-optically active water quality parameters in water bodies are important evaluation indicators in monitoring urban water quality. Over the past years, satellite remote sensing techniques have increasingly been used to assess different types of substances in urban water bodies. However, it is challenging to retrieve accurate data for some of the non-optically active water quality parameters from satellite images due to weak spectral characteristics. This study aims to examine the potential of ZY1-02D hyperspectral images in retrieving non-optical active water quality parameters, including dissolved oxygen (DO), permanganate index (COD Mn ), and total phosphorus (TP) in urban rivers and lakes. We ﬁrst simulated the in situ measured reﬂectance to the satellite equivalent reﬂectance using the ZY1-02D and Sentinel-2 spectral response function. Further, we used four machine learning models to compare the retrieval performance of these two sensors with different bandwidths. The mean absolute percentage errors (MAPE) are 24.28%, 18.44%, and 37.04% for DO, COD Mn , and TP, respectively, and the root mean square errors (RMSE) are 1.67, 0.96, and 0.07 mg/L, respectively. Finally, we validated the accuracy and consistency of aquatic products retrieved from ZY1-02D and Sentinel-2 images. The remote sensing reﬂectance ( R rs ) products of ZY1-02D are slightly overestimated compared to Sentinel-2 R rs . ZY1-02D has high accuracy and consistency in mapping COD Mn products in urban water. The results show the potential of ZY1-02D hyperspectral images in mapping non-optically active water quality parameters.


Introduction
Water resources have a significant role in different functions of cities, such as drinking water, industrial production, and landscape [1]. Rivers and lakes are an important part of urban water bodies. However, with the rapid development of urban society and economy, living needs grow rapidly, and the discharge of domestic water, agricultural water, and industrial water exceeded the self-cleaning capacity of water bodies, causing serious pollution to urban rivers and lakes [2]. Deterioration of urban water quality brings the safety issue of drinking water and destruction of the ecological environment, which in turn affects human health and biodiversity [3]. The city water department needs to regularly formulate policies based on water quality assessment data for the further development of the city.
Water quality monitoring is an important part of water quality evaluation. It aims to understand the water quality of urban water bodies, especially rivers and lakes. Although traditional water quality monitoring methods including manual field sampling and laboratory measurements or automatic in situ measurements have high accuracy, the manual method is labor-intensive, and the construction and maintenance of the automatic station requires expensive costs. Furthermore, both methods can only reflect the water quality In this study, two typical and important rivers and lakes were selected for detailed 104 evaluation and analysis. They are the Huangpu River and its upper tributaries and Di-105 anshan Lake. Huangpu River is the largest river in Shanghai and flows through most of 106 the districts. Dianshan Lake is the largest lake in Shanghai and is located in Qingpu Dis-107 trict. From September 2018 to November 2021, 12 remote sensing reflectance ( ) field 114 measurements were carried out in 90 rivers of Shanghai. The average transparency of 115 these rivers is lower than 0.6m, so the reflectance from the bottom of the rivers cannot 116 affect the measurements of surface water [29]. The main goal of all measurements 117 was to characterize the relationship between non-optically active parameters and optically 118 active parameters by measuring as many rivers as possible at different dates and locations. 119 The average width of these rivers ranges from 10 to 150m. We tried to balance the number 120 of rivers with good and bad water quality, and some rivers measured in 2018 were defined 121 as black and smelly water bodies. Finally, a total of 183 sets of data were obtained. The 122 measurements covered all seasons and nine districts in Shanghai (Table 1)

In Situ Data
From September 2018 to November 2021, 12 remote sensing reflectance (R rs ) field measurements were carried out in 90 rivers of Shanghai. The average transparency of these rivers is lower than 0.6 m, so the reflectance from the bottom of the rivers cannot affect the R rs measurements of surface water [29]. The main goal of all measurements was to characterize the relationship between non-optically active parameters and optically active parameters by measuring as many rivers as possible at different dates and locations. The average width of these rivers ranges from 10 to 150 m. We tried to balance the number of rivers with good and bad water quality, and some rivers measured in 2018 were defined as black and smelly water bodies. Finally, a total of 183 sets of data were obtained. The measurements covered all seasons and nine districts in Shanghai (Table 1).
For each measurement, R rs was measured on the bank or bridge of rivers using a Fieldspec 4 spectroradiometer ranging from 350 to 2500 nm (1 nm interval). Referring to the above-water method and NASA-recommended measurement standards [30,31], specifically, measurements were performed between 9:00 and 15:00 on sunny windless days, and the zenith angle and the azimuth angle were 45 • and 135 • , respectively. Total water surface radiation (L sw (λ)), skylight radiance (L sky (λ)), and the reference plate radiance (L p (λ)) were measured at each site. R rs was calculated by the following equation: R rs (λ) = L sw (λ) − ρ sky (λ)L sky (λ) where the ρ sky is the air-water interface skylight reflectance and related to wind speed and solar altitude. According to field measurement conditions, we used 0.028 [31]. ρ p is the irradiance reflectance of the gray plate (30%). The in situ measured R rs of these sampling sites are shown in Figure 2. For each measurement, was measured on the bank or bridge of rivers using a 125 Fieldspec 4 spectroradiometer ranging from 350 to 2500 nm (1 nm interval). Referring to 126 the above-water method and NASA-recommended measurement standards [30,31], spe-127 cifically, measurements were performed between 9:00 and 15:00 on sunny windless days, 128 and the zenith angle and the azimuth angle were 45° and 135°, respectively. Total water 129 surface radiation ( ( )), skylight radiance ( ( )), and the reference plate radiance 130 ( ( )) were measured at each site. was calculated by the following equation: where the is the air-water interface skylight reflectance and related to wind speed 132 and solar altitude. According to field measurement conditions, we used 0.028 [31].
is 133 the irradiance reflectance of the gray plate (30%). The in situ measured of these sam-134 pling sites are shown in Figure 2. The field-measured DO, CODMn, and TP concentration data were provided by Shang-138 hai Hydrological Station. Table 2 shows the laboratory analysis methods and descriptive 139 statistics of parameters. The dataset shows a large water quality parameter range. The DO, 140 CODMn, and TP concentration measurements were in accordance with the GB 7489-87, GB 141 11892-89, and GB 11893-89, respectively. 142 All data were divided into two parts (Figure 1b). Dataset 1 (183 samples) was con-144 current with radiometric measurements used to develop the algorithm for water quality 145 parameter estimation. Figure 3 shows the distributions of each parameter. All parameters 146 were missing the high concentration part, and CODMn and TP were more concentrated in 147 low concentration areas. Dataset 2 (30 samples) was the near-coincident data obtained 148 The field-measured DO, COD Mn , and TP concentration data were provided by Shanghai Hydrological Station. Table 2 shows the laboratory analysis methods and descriptive statistics of parameters. The dataset shows a large water quality parameter range. The DO, COD Mn , and TP concentration measurements were in accordance with the GB 7489-87, GB 11892-89, and GB 11893-89, respectively. All data were divided into two parts (Figure 1b). Dataset 1 (183 samples) was concurrent with radiometric measurements used to develop the algorithm for water quality parameter estimation. Figure 3 shows the distributions of each parameter. All parameters were missing the high concentration part, and COD Mn and TP were more concentrated in low concentration areas. Dataset 2 (30 samples) was the near-coincident data obtained under Sentinel-2 and ZY1-02D overpasses used to validate the image-retrieved results.
Dataset 2 was mainly distributed in Dianshan Lake and rivers with a width greater than 100 m. According to our field investigation, the water change cycle of Dianshan Lake is 7 days and the flow rate of these rivers is slow. Therefore, in the absence of precipitation and sudden pollution, the water quality of these sites does not change much in a short period of time. under Sentinel-2 and ZY1-02D overpasses used to validate the image-retrieved results. 149 Dataset 2 was mainly distributed in Dianshan Lake and rivers with a width greater than 150 100m. According to our field investigation, the water change cycle of Dianshan Lake is 7 151 days and the flow rate of these rivers is slow. Therefore, in the absence of precipitation 152 and sudden pollution, the water quality of these sites does not change much in a short 153 period of time. We also used in situ and water quality concentrations (Dataset 3) from Taihu 158 Lake in 2009 to validate the suitability of Dataset 1's models across locations and times. 159 Samples were collected monthly at 32 stations in Taihu Lake ( Figure 4).
, DO, CODMn, 160 and TP were determined using the methods described in Section 2.2.1. Overall, we se-161 lected 91 samples distributed across Taihu Lake between February and May.  The multispectral data used in this study was a Sentinel-2 image with a spatial reso-166 lution of 10-60 m. The Multispectral Instrument (MSI) onboard Sentinel-2(A/B) has both 167 13 bands and short temporal resolution. The S2A-MSI Level 1C image of Shanghai ac-168 quired on 13 May 2020 was downloaded from the Copernicus Open Access Hub 169 (https://scihub.copernicus.eu/). 170 Hyperspectral data with coincident Sentinel-2 images were acquired by the Ad-171 vanced HyperSpectral Imager (AHSI) onboard the ZY1-02D satellite of China. The AHSI 172 has 150 bands in the visible and near-infrared (VNIR) and 180 bands in the short-wave 173 infrared (SWIR) with 10 and 20 nm spectral resolution, respectively [32]. Meanwhile, it 174 has a spatial resolution of 30m, which can meet the needs of urban water monitoring. In 175 addition to Band 5 of MSI, AHSI can have at least 2 bands in each of the MSI band's con-176 figurations ( Table 3).

Independent Dataset in Taihu Lake
We also used in situ R rs and water quality concentrations (Dataset 3) from Taihu Lake in 2009 to validate the suitability of Dataset 1's models across locations and times. Samples were collected monthly at 32 stations in Taihu Lake ( Figure 4). R rs , DO, COD Mn , and TP were determined using the methods described in Section 2.2.1. Overall, we selected 91 samples distributed across Taihu Lake between February and May. under Sentinel-2 and ZY1-02D overpasses used to validate the image-retrieved results. 149 Dataset 2 was mainly distributed in Dianshan Lake and rivers with a width greater than 150 100m. According to our field investigation, the water change cycle of Dianshan Lake is 7 151 days and the flow rate of these rivers is slow. Therefore, in the absence of precipitation 152 and sudden pollution, the water quality of these sites does not change much in a short 153 period of time. We also used in situ and water quality concentrations (Dataset 3) from Taihu 158 Lake in 2009 to validate the suitability of Dataset 1's models across locations and times. 159 Samples were collected monthly at 32 stations in Taihu Lake ( Figure 4).
, DO, CODMn, 160 and TP were determined using the methods described in Section 2.2.1. Overall, we se-161 lected 91 samples distributed across Taihu Lake between February and May.  The multispectral data used in this study was a Sentinel-2 image with a spatial reso-166 lution of 10-60 m. The Multispectral Instrument (MSI) onboard Sentinel-2(A/B) has both 167 13 bands and short temporal resolution. The S2A-MSI Level 1C image of Shanghai ac-168 quired on 13 May 2020 was downloaded from the Copernicus Open Access Hub 169 (https://scihub.copernicus.eu/). 170 Hyperspectral data with coincident Sentinel-2 images were acquired by the Ad-171 vanced HyperSpectral Imager (AHSI) onboard the ZY1-02D satellite of China. The AHSI 172 has 150 bands in the visible and near-infrared (VNIR) and 180 bands in the short-wave 173 infrared (SWIR) with 10 and 20 nm spectral resolution, respectively [32]. Meanwhile, it 174 has a spatial resolution of 30m, which can meet the needs of urban water monitoring. In 175 addition to Band 5 of MSI, AHSI can have at least 2 bands in each of the MSI band's con-176 figurations ( Table 3).

Satellite Data
The multispectral data used in this study was a Sentinel-2 image with a spatial resolution of 10-60 m. The Multispectral Instrument (MSI) onboard Sentinel-2(A/B) has both 13 bands and short temporal resolution. The S2A-MSI Level 1C image of Shanghai acquired on 13 May 2020 was downloaded from the Copernicus Open Access Hub (https://scihub.copernicus.eu/ accessed on 5 December 2020).
Hyperspectral data with coincident Sentinel-2 images were acquired by the Advanced HyperSpectral Imager (AHSI) onboard the ZY1-02D satellite of China. The AHSI has 150 bands in the visible and near-infrared (VNIR) and 180 bands in the short-wave infrared (SWIR) with 10 and 20 nm spectral resolution, respectively [32]. Meanwhile, it has a spatial resolution of 30 m, which can meet the needs of urban water monitoring. In addition to Band 5 of MSI, AHSI can have at least 2 bands in each of the MSI band's configurations (Table 3).
Both the Sentinel-2 and ZY1-02D images have less than 10% cloud coverage and a time difference of ±3 days with Dataset 2. To develop retrieval models that can be used on MSI and AHSI images, the in situ R rs should be simulated to satellite band equivalent reflectance. To achieve this, convolution via Equation (2) based on the spectral response function (SRF) of MSI and AHSI sensors was calculated as follows: where SRF i (λ) is the relative SRF of the MSI and AHSI ith band, and the Gaussian function was used to describe SRF of the AHSI sensor [33]. λ max and λ min are the wavelength ranges in this band. The MSI and AHSI equivalent reflectance spectra are displayed in Figure 5. Figure 5c shows the average of in situ R rs and the equivalent reflectance simulated by the two sensors. For the hyperspectral sensor AHSI, the average equivalent reflectance of each band and in situ R rs basically coincide. However, some bands' average equivalent reflectance of the multispectral sensor MSI deviates significantly from the in situ R rs . To develop retrieval models that can be used on MSI and AHSI images, the in situ Rrs 183 should be simulated to satellite band equivalent reflectance. To achieve this, convolution 184 via equation (2) based on the spectral response function (SRF) of MSI and AHSI sensors 185 was calculated as follows: where ( ) is the relative SRF of the MSI and AHSI ith band, and the Gaussian func-187 tion was used to describe SRF of the AHSI sensor [33]. and are the wavelength 188 ranges in this band. 189 The MSI and AHSI equivalent reflectance spectra are displayed in Figure 5. The Pearson-based correlation analysis was used to describe the correlation between 199 the water quality parameters and the satellite equivalent reflectances. The Pearson corre-200 lation coefficient (r) ranges from −1 to +1. When the r is close to -1 or +1, this indicates a 201

Model Development
The Pearson-based correlation analysis was used to describe the correlation between the water quality parameters and the satellite equivalent reflectances. The Pearson correlation coefficient (r) ranges from −1 to +1. When the r is close to −1 or +1, this indicates a strong inverse or positive correlation between the variables, respectively. However, the r close to zero indicates no correlation between the variables [34]. As shown in Figure 6, the correlations varied significantly for different water quality parameters, but they are both at a low level (highest |r| = 0.41).
Remote Sens. 2022, 14, x FOR PEER REVIEW 7 of 18 strong inverse or positive correlation between the variables, respectively. However, the r close to zero indicates no correlation between the variables [34]. As shown in Figure 6, the correlations varied significantly for different water quality parameters, but they are both at a low level (highest | | = 0.41). Figure 6. The absolute value of r between water quality parameters and MSI (a) and AHSI (b) equivalent reflectance.
In this study, there were four machine learning methods that were considered to develop retrieval models: support vector regression (SVR), partial least squares regression (PLSR), K-nearest neighbor (KNN), and XGBoost. SVR can solve the non-linear problems in low dimensional feature space by transforming the input data to a high dimensional space with a non-linear function, then seeking a linear regression hyperplane in high dimensional feature space [35]. PLSR model the covariance relations between features and targets by finding the latent variables, so it can reduce the multicollinearity among input values [36]. KNN predicts the target by local interpolation of the targets associated with the nearest neighbors in the training set [37], and XGBoost is a gradient boosting tree model, which predicts the sum of scores in multiple regression trees [38]. All methods were implemented by scikit-learn of Python.

Satellite Data Preprocessing
The main preprocessing of the satellite images included atmospheric correction (AC) water body extent extraction, and remote sensing reflectance calculation [39]. First, the radiometric calibration coefficients of AHSI were used to convert the digital number values to top-of-atmosphere radiances. Second, the Sen2Cor processor was used to obtain the Sentinel-2 Level 2A Bottom of Atmosphere reflectance product [40]. The FLAASH atmospheric correction module in ENVI was used to retrieve surface reflectance for AHSI images [41]. In this study, the mid-latitude summer atmosphere and rural aerosol were selected in FLAASH.
The modified normalized difference water index (MNDWI) and the OTSU [42] method were used to segment the water bodies from images. Based on the assumption that the minimum surface reflectance in the SWIR band of turbid water is composed only of residual aerosol scattering, skylight, and sun glint, a remote sensing reflectance estimation method [43] was used for MSI and AHSI surface reflectance images to correct the skylight effect and retrieve remote sensing reflectance as follows: where ( ) represents the remote sensing reflectance, ( ) represents the surface reflectance, and min( ) indicates the minimum surface reflectance of the SWIR band in MSI and AHSI, where of AHSI use the average of R in the 1530-1630 nm bands. In this study, there were four machine learning methods that were considered to develop retrieval models: support vector regression (SVR), partial least squares regression (PLSR), K-nearest neighbor (KNN), and XGBoost. SVR can solve the non-linear problems in low dimensional feature space by transforming the input data to a high dimensional space with a non-linear function, then seeking a linear regression hyperplane in high dimensional feature space [35]. PLSR model the covariance relations between features and targets by finding the latent variables, so it can reduce the multicollinearity among input values [36]. KNN predicts the target by local interpolation of the targets associated with the nearest neighbors in the training set [37], and XGBoost is a gradient boosting tree model, which predicts the sum of scores in multiple regression trees [38]. All methods were implemented by scikit-learn of Python.

Satellite Data Preprocessing
The main preprocessing of the satellite images included atmospheric correction (AC), water body extent extraction, and remote sensing reflectance calculation [39]. First, the radiometric calibration coefficients of AHSI were used to convert the digital number values to top-of-atmosphere radiances. Second, the Sen2Cor processor was used to obtain the Sentinel-2 Level 2A Bottom of Atmosphere reflectance product [40]. The FLAASH atmospheric correction module in ENVI was used to retrieve surface reflectance for AHSI images [41]. In this study, the mid-latitude summer atmosphere and rural aerosol were selected in FLAASH.
The modified normalized difference water index (MNDWI) and the OTSU [42] method were used to segment the water bodies from images. Based on the assumption that the minimum surface reflectance in the SWIR band of turbid water is composed only of residual aerosol scattering, skylight, and sun glint, a remote sensing reflectance estimation method [43] was used for MSI and AHSI surface reflectance images to correct the skylight effect and retrieve remote sensing reflectance as follows: where R c rs (λ) represents the remote sensing reflectance, R(λ) represents the surface reflectance, and min(R SW IR ) indicates the minimum surface reflectance of the SWIR band in MSI and AHSI, where R SW IR of AHSI use the average of R in the 1530-1630 nm bands.

Accuracy Assessment
The coefficient of determination (R 2 ), mean absolute percentage error (MAPE), and root mean square error (RMSE) were used to assess the performance of water quality retrieval models and the agreement between in situ data and image retrievals [44].
where n is the number of samples, M i and E i represent the measured values and estimated values, respectively.

Spectral Response to Non-Optically Water Quality Parameter Variation
The mean values of in situ spectra for DO, COD Mn , and TP concentrations in different value ranges are shown in Figure 7. Overall, the spectral reflectance between 400 and 700 nm is inversely proportional to the DO concentration. In 2-8 mg/L and 8-12 mg/L regions, the spectral reflectance is similar from 700 to 900 nm. When the DO concentration is higher, the spectral reflectance trough at 675 nm and peak at 705 nm are more obvious. These indicate that for water with lower DO, suspended sediment accounts for the largest proportion, and for water with higher DO, the spectral variability as a response to Chl-a and algae is more obvious.
Similarly, the reflectance between 400 and 700 nm is inversely proportional to the COD Mn concentration. When the COD Mn concentration is greater than 3 mg/L, the spectral reflectance is similar from 700 to 900 nm. The water with COD Mn concentration over 6 mg/L is usually considered to be polluted. As shown in Figure 7b, the spectral reflectance peak at~700 nm of high COD Mn concentration moves toward the longer wave.
In 0.1-0.4 mg/L regions of TP concentration, the spectral reflectance between 400 and 700 nm is inversely proportional to concentration, and the spectral reflectance between 700 and 900 nm is similar. When the TP concentration is lower than 0.1 mg/L, the spectral reflectance also shows the characteristics of a high Chl-a concentration, which indicates that in one water body the correlation with Chl-a might be with P or N. The spectral reflectance peak of high TP concentration is also obvious at 705 nm. The coefficient of determination (R 2 ), mean absolute percentage error (MAPE), and 239 root mean square error (RMSE) were used to assess the performance of water quality re-240 trieval models and the agreement between in situ data and image retrievals [44].
where n is the number of samples, and represent the measured values and esti-242 mated values, respectively.

245
The mean values of in situ spectra for DO, CODMn, and TP concentrations in different 246 value ranges are shown in Figure 7. Overall, the spectral reflectance between 400 and 700 247 nm is inversely proportional to the DO concentration. In 2-8 mg/L and 8-12 mg/L regions, 248 the spectral reflectance is similar from 700 to 900 nm. When the DO concentration is 249 higher, the spectral reflectance trough at 675 nm and peak at 705 nm are more obvious. 250 These indicate that for water with lower DO, suspended sediment accounts for the largest 251 proportion, and for water with higher DO, the spectral variability as a response to Chl-a 252 and algae is more obvious. 253 Similarly, the reflectance between 400 and 700 nm is inversely proportional to the 254 CODMn concentration. When the CODMn concentration is greater than 3 mg/L, the spectral 255 reflectance is similar from 700 to 900 nm. The water with CODMn concentration over 6 256 mg/L is usually considered to be polluted. As shown in Figure 7b, the spectral reflectance 257 peak at ~ 700nm of high CODMn concentration moves toward the longer wave. 258 In 0.1-0.4 mg/L regions of TP concentration, the spectral reflectance between 400 and 259 700 nm is inversely proportional to concentration, and the spectral reflectance between 260 700 and 900 nm is similar. When the TP concentration is lower than 0.1 mg/L, the spectral 261 reflectance also shows the characteristics of a high Chl-a concentration, which indicates 262 that in one water body the correlation with Chl-a might be with P or N. The spectral re-263 flectance peak of high TP concentration is also obvious at 705nm.  To compare the aquatic products retrieved from AHSI and MSI, the band selection range in this study was the nine visible and near-infrared bands of MSI and the AHSI bands of the corresponding wavelength range (Table 3). Although other bands of AHSI could be more suitable, we did not consider the bands out of range for the MSI sensor's spectral configuration. The input of each model included nine bands and three band ratios. Previous studies have confirmed that adding band ratios can improve the performance of water quality retrieval models [45][46][47]. The second correlation analysis between water quality parameters and equivalent reflectance ratios was carried out to find better model inputs. According to Figure 6 and Table 4, the most appropriate band composition of each water quality parameter was used to develop SVR, PLSR, KNN, and XGBoost models. For band ratios, all of the variables were significantly correlated at the 95% confidence level, with higher correlation coefficients compared to the single-band variables. To examine that adding variables to inputs is not just fitting noise, we tested the performance of XGBoost on the validation dataset using three combinations: (1) five bands without red-edge bands; (2) total nine bands; (3) nine bands and three band ratios. Table 5 shows that the input variables with 12 variables produced the best performance.

Performances of Machine Learning Models
Dataset 1 was randomly divided into a training dataset (N = 128) and a validation dataset (N = 55). For the development of the model, all machine learning models used the same training dataset, and the hyperparameters were determined by the strategy of grid search. Comparing the water quality parameters estimated by the machine learning models with the validation dataset, the optimal model of each water quality parameter for two satellites was selected as follows: For DO retrieval models, XGBoost had the best performance for Sentinel-2 (R 2 = 0.53, MAPE = 22.66%, RMSE = 1.69 mg/L) and ZY1-02D (R 2 = 0.53, MAPE = 24.28%, RMSE = 1.67 mg/L) (Figure 8).
Remote Sens. 2022, 14, x FOR PEER REVIEW 10 of 18 same training dataset, and the hyperparameters were determined by the strategy of grid search. Comparing the water quality parameters estimated by the machine learning models with the validation dataset, the optimal model of each water quality parameter for two satellites was selected as follows: For DO retrieval models, XGBoost had the best performance for Sentinel-2 (R 2 = 0.53, MAPE = 22.66%, RMSE = 1.69mg/L) and ZY1-02D (R 2 = 0.53, MAPE = 24.28%, RMSE = 1.67mg/L) (Figure 8).  For TP retrieval models, SVR had the best performance for Sentinel-2 (R 2 = 0.46, MAPE = 37.81%, RMSE = 0.08 mg/L) and XGBoost had the best performance for ZY1-02D (R 2 = 0.47, MAPE = 37.04%, RMSE = 0.07mg/L) ( Figure 10). same training dataset, and the hyperparameters were determined by the strategy of grid search. Comparing the water quality parameters estimated by the machine learning models with the validation dataset, the optimal model of each water quality parameter for two satellites was selected as follows: For DO retrieval models, XGBoost had the best performance for Sentinel-2 (R 2 = 0.53 MAPE = 22.66%, RMSE = 1.69mg/L) and ZY1-02D (R 2 = 0.53, MAPE = 24.28%, RMSE = 1.67mg/L) (Figure 8).  For TP retrieval models, SVR had the best performance for Sentinel-2 (R 2 = 0.46 MAPE = 37.81%, RMSE = 0.08 mg/L) and XGBoost had the best performance for ZY1-02D (R 2 = 0.47, MAPE = 37.04%, RMSE = 0.07mg/L) ( Figure 10). For TP retrieval models, SVR had the best performance for Sentinel-2 (R 2 = 0.46, MAPE = 37.81%, RMSE = 0.08 mg/L) and XGBoost had the best performance for ZY1-02D (R 2 = 0.47, MAPE = 37.04%, RMSE = 0.07 mg/L) ( Figure 10). Table 6 shows the performance comparison of all machine learning models. All models did slightly underestimate high concentrations because the training dataset lacks high concentration data. Remote Sens. 2022, 14, x FOR PEER REVIEW 11 of 18 Figure 10. Performance evaluation of TP retrievals using the SVR for Sentinel-2 (a) and the XGBoost for ZY1-02D (b). Table 5 shows the performance comparison of all machine learning models. All models did slightly underestimate high concentrations because the training dataset lacks high concentration data.

Further Validation on Taihu Lake
In situ data from turbid and eutrophic shallow Taihu Lake were used to estimate DO, CODMn, and TP, and were compared with measured data to determine the potential of the machine learning models for producing spatial and temporal products. The derived CODMn values were similar to the in situ data. However, the -derived DO values were less than the in situ data and the -derived TP values were more than the in situ data (Figure 11).

Further Validation on Taihu Lake
In situ R rs data from turbid and eutrophic shallow Taihu Lake were used to estimate DO, COD Mn , and TP, and were compared with measured data to determine the potential of the machine learning models for producing spatial and temporal products. The R rs -derived COD Mn values were similar to the in situ data. However, the R rs -derived DO values were less than the in situ data and the R rs -derived TP values were more than the in situ data ( Figure 11).  Figure 10. Performance evaluation of TP retrievals using the SVR for Sentinel-2 (a) and the XGBoost 312 for ZY1-02D (b). 313 Table 5 shows the performance comparison of all machine learning models. All mod-314 els did slightly underestimate high concentrations because the training dataset lacks high 315 concentration data. 316 In situ data from turbid and eutrophic shallow Taihu Lake were used to estimate 319 DO, CODMn, and TP, and were compared with measured data to determine the potential 320 of the machine learning models for producing spatial and temporal products. The -321 derived CODMn values were similar to the in situ data. However, the -derived DO val-322 ues were less than the in situ data and the -derived TP values were more than the in 323 situ data (Figure 11). 324 325 Figure 11. Comparison between AHSI R rs -derived and in situ DO (a), COD Mn (b), and TP (c).

Water Quality Mapping
The water quality maps derived from ZY1-02D AHSI and Sentinel-2 MSI are shown in Figures 12-14. The waters in the images are Dianshan Lake and the upper reaches of the Huangpu River which provide more than 60% of the domestic water for residents [48]. Huangpu River is densely populated with boats, making it difficult to retrieve water quality.
ity. 332 The range of the color bar was retained consistently for a better comparison. DO de-333 rived from Sentinel-2 and ZY1-02D images had similar spatial distribution characteristics; 334 the DO concentration in the eastern part of Dianshan Lake and the outflow rivers was low 335 (Figure 12). CODMn derived from Sentinel-2 and ZY1-02D images are relatively average in 336 the study area, which is consistent with the distribution of the in situ measured value. 337 ZY1-02D overestimated the CODMn concentration in the aquaculture area in the northeast-338 ern part of Dianshan Lake ( Figure 13). The agreement is relatively strong for TP retrievals 339 and DO retrievals of Sentinel-2, while the TP derived from the ZY1-02D image showed 340 some limitations and higher errors associated ( Figure 14). ity. 332 The range of the color bar was retained consistently for a better comparison. DO de-333 rived from Sentinel-2 and ZY1-02D images had similar spatial distribution characteristics; 334 the DO concentration in the eastern part of Dianshan Lake and the outflow rivers was low 335 (Figure 12). CODMn derived from Sentinel-2 and ZY1-02D images are relatively average in 336 the study area, which is consistent with the distribution of the in situ measured value. 337 ZY1-02D overestimated the CODMn concentration in the aquaculture area in the northeast-338 ern part of Dianshan Lake ( Figure 13). The agreement is relatively strong for TP retrievals 339 and DO retrievals of Sentinel-2, while the TP derived from the ZY1-02D image showed 340 some limitations and higher errors associated ( Figure 14).  We compared the average of Sentinel-2 and ZY1-02D images at Dataset 2 sites. 350 As shown in Figure 15, the ZY1-02D spectra are in good agreement with Sentinel-2 spectra 351 in terms of both shape and magnitude. However, ZY1-02D are brighter than those of 352 Sentinel-2 in these sites. 353 The accuracy of the image-derived affected the accuracy of the estimated water 354 The range of the color bar was retained consistently for a better comparison. DO derived from Sentinel-2 and ZY1-02D images had similar spatial distribution characteristics; the DO concentration in the eastern part of Dianshan Lake and the outflow rivers was low Remote Sens. 2022, 14, 5029 13 of 18 ( Figure 12). COD Mn derived from Sentinel-2 and ZY1-02D images are relatively average in the study area, which is consistent with the distribution of the in situ measured value. ZY1-02D overestimated the COD Mn concentration in the aquaculture area in the northeastern part of Dianshan Lake ( Figure 13). The agreement is relatively strong for TP retrievals and DO retrievals of Sentinel-2, while the TP derived from the ZY1-02D image showed some limitations and higher errors associated (Figure 14).

Comparison of R rs Products between ZY1-02D and Sentinel-2
We compared the average R rs of Sentinel-2 and ZY1-02D images at Dataset 2 sites. As shown in Figure 15, the ZY1-02D spectra are in good agreement with Sentinel-2 spectra in terms of both shape and magnitude. However, ZY1-02D R rs are brighter than those of Sentinel-2 in these sites.
The accuracy of the image-derived R rs affected the accuracy of the estimated water quality parameter concentration. Therefore, the agreement evaluation of ZY1-02D and Sentinel-2 image-derived R rs was conducted using r and RMSE ( Figure 16). Here, 704 nm had the highest r and lowest RMSE, which was the band with the highest agreement between ZY1-02D R rs and Sentinel-2 R rs . Similarly, 560 and 665 nm also had a high agreement. The energies of these bands were higher than that of other wavelengths, and they were relatively less affected by noise. Additionally, 443 and 492 nm were affected by atmosphere scattering. The band after 740 nm was lower in energy; therefore, their agreement was relatively low.

Comparison of Products Between ZY1-02D and Sentinel-2
We compared the average of Sentinel-2 and ZY1-02D im As shown in Figure 15, the ZY1-02D spectra are in good agreement in terms of both shape and magnitude. However, ZY1-02D ar Sentinel-2 in these sites.
The accuracy of the image-derived affected the accuracy quality parameter concentration. Therefore, the agreement evalu Sentinel-2 image-derived was conducted using r and RMSE nm had the highest r and lowest RMSE, which was the band with between ZY1-02D and Sentinel-2 . Similarly, 560 and 66 agreement. The energies of these bands were higher than that of o they were relatively less affected by noise. Additionally, 443 and 4 atmosphere scattering. The band after 740 nm was lower in energy ment was relatively low.  The biggest challenge of applying models to satellite images was the performance of 372 AC methods [49][50][51]. The optimal model of each water quality parameter was used to 373 retrieve water quality parameter concentration for Sentinel-2 and ZY1-02D images of the 374 same day. Given the lack of in situ for the direct test of AC, Dataset 2 was used to 375 further validate the accuracy and stability of the models on the satellite images. Spatial 376 windows (3 × 3 pixels) were applied to extract the average concentration at the location of 377 Dataset 2 stations. Figure 17 shows the comparison of concentration between Dataset 2 378 and image retrievals. The satellite retrievals and the in situ CODMn had the highest agree-379 ment, but they still had significant deviations at high and low concentrations. The DO and 380 TP retrievals of ZY1-02D had large differences with in situ measurements. This can be 381 attributed to the three-day difference between satellite overpasses and in situ measure-382 ment as the distribution of the parameters may have some variations. 383 The comparison of water quality products between ZY1-02D and Sentinel-2 is shown 384 in Figure 18. The agreement of the retrieved concentration is examined by comparing the 385 values extracted from the water quality products. There is a high agreement between the 386 CODMn retrievals from ZY1-02D and Sentinel-2. Nevertheless, the DO retrievals from ZY1-387 02D are underestimated compared with those of Sentinel-2. The TP retrievals from ZY1-388 02D are not successful in this image. TP is the parameter most affected by the atmosphere. 389 Figure 16. Comparison of R rs data between ZY1-02D and Sentinel-2 (a-i) represent the band-center wavelength of Sentinel-2, respectively.

Comparison of Water Quality Products between ZY1-02D and Sentinel-2
The biggest challenge of applying models to satellite images was the performance of AC methods [49][50][51]. The optimal model of each water quality parameter was used to retrieve water quality parameter concentration for Sentinel-2 and ZY1-02D images of the same day. Given the lack of in situ R rs for the direct test of AC, Dataset 2 was used to further validate the accuracy and stability of the models on the satellite images. Spatial windows (3 × 3 pixels) were applied to extract the average concentration at the location of Dataset 2 stations. Figure 17 shows the comparison of concentration between Dataset 2 and image retrievals. The satellite retrievals and the in situ COD Mn had the highest agreement, but they still had significant deviations at high and low concentrations. The DO and TP retrievals of ZY1-02D had large differences with in situ measurements. This can be attributed to the three-day difference between satellite overpasses and in situ measurement as the distribution of the parameters may have some variations.
The comparison of water quality products between ZY1-02D and Sentinel-2 is shown in Figure 18. The agreement of the retrieved concentration is examined by comparing the values extracted from the water quality products. There is a high agreement between the COD Mn retrievals from ZY1-02D and Sentinel-2. Nevertheless, the DO retrievals from ZY1-02D are underestimated compared with those of Sentinel-2. The TP retrievals from ZY1-02D are not successful in this image. TP is the parameter most affected by the atmosphere. According to the statistical results of accuracy and agreement, even though the mod-395 els of ZY1-02D and Sentinel-2 trained by in situ data have the same accuracy, the final 396 water quality products are still affected by atmospheric correction, radiometric sensitivity, 397 and signal-to-noise ratio of ZY1-02D and Sentinel-2 sensors. 398

399
Whether the image-derived modeling methods or in situ observation modeling meth-400 ods, the applicability of water quality parameter inversion models are limited by the rep-401 resentativeness of the measured data. The revisit time of ZY1-02D is 55 days and the cloud 402 cover is common in Shanghai, resulting in few scenes per year. Our twelve field radio-403 metric measurements surveys included 80 rivers in Shanghai and covered all seasons. For 404 measured rivers, (550) ranged from 0.0103 to 0.079 sr−1, which is a relatively broad 405 spectral range. Dataset 1 fully reflects the principal relation between water quality param-406 eters and remote sensing reflectance of the study rivers in Shanghai. Therefore, the models 407 developed by Dataset 1 would be applicable to the water quality parameter retrievals of 408 rivers in the study area though some rivers have not been measured. Compared with the 409 non-optically active water quality parameter (such as TP or TN) retrieval models estab-410 lished by Qiao et al. [10], Gao et al. [18], and Lu et al. [22], their application is limited to 411 specific lakes or rivers and our models have wider applicability. Compared with the DO 412 retrieval model established by Al-Shaibah et al. [16] using a linear model, the DO retrieval 413 model in this study based on XGBoost has higher accuracy. 414 However, our data did not include water quality in large inland lakes. Here, we used 415 data and water quality data from Taihu Lake in 2009 to examine the further univer-416 sality of models. The models developed by Dataset 1 had slightly lower performance on 417 data of Taihu Lake than that of the validated dataset described in Section 3.2.2 ( Figure 11). 418 This can be attributed to the ten-year difference between the data in Shanghai and Taihu 419 Lake. Moreover, precipitation and human factors are important factors affecting river wa-420 ter quality. Precipitation increases river flow and makes it easier for pollutants to transfer 421 According to the statistical results of accuracy and agreement, even though the mod-395 els of ZY1-02D and Sentinel-2 trained by in situ data have the same accuracy, the final 396 water quality products are still affected by atmospheric correction, radiometric sensitivity, 397 and signal-to-noise ratio of ZY1-02D and Sentinel-2 sensors. 398

399
Whether the image-derived modeling methods or in situ observation modeling meth-400 ods, the applicability of water quality parameter inversion models are limited by the rep-401 resentativeness of the measured data. The revisit time of ZY1-02D is 55 days and the cloud 402 cover is common in Shanghai, resulting in few scenes per year. Our twelve field radio-403 metric measurements surveys included 80 rivers in Shanghai and covered all seasons. For 404 measured rivers, (550) ranged from 0.0103 to 0.079 sr−1, which is a relatively broad 405 spectral range. Dataset 1 fully reflects the principal relation between water quality param-406 eters and remote sensing reflectance of the study rivers in Shanghai. Therefore, the models 407 developed by Dataset 1 would be applicable to the water quality parameter retrievals of 408 rivers in the study area though some rivers have not been measured. Compared with the 409 non-optically active water quality parameter (such as TP or TN) retrieval models estab-410 lished by Qiao et al. [10], Gao et al. [18], and Lu et al. [22], their application is limited to 411 specific lakes or rivers and our models have wider applicability. Compared with the DO 412 retrieval model established by Al-Shaibah et al. [16] using a linear model, the DO retrieval 413 model in this study based on XGBoost has higher accuracy. 414 However, our data did not include water quality in large inland lakes. Here, we used 415 data and water quality data from Taihu Lake in 2009 to examine the further univer-416 sality of models. The models developed by Dataset 1 had slightly lower performance on 417 data of Taihu Lake than that of the validated dataset described in Section 3.2.2 ( Figure 11). 418 This can be attributed to the ten-year difference between the data in Shanghai and Taihu 419 Lake. Moreover, precipitation and human factors are important factors affecting river wa-420 ter quality. Precipitation increases river flow and makes it easier for pollutants to transfer 421 According to the statistical results of accuracy and agreement, even though the models of ZY1-02D and Sentinel-2 trained by in situ data have the same accuracy, the final water quality products are still affected by atmospheric correction, radiometric sensitivity, and signal-to-noise ratio of ZY1-02D and Sentinel-2 sensors.

Strengths and Limitations of the Models
Whether the image-derived modeling methods or in situ observation modeling methods, the applicability of water quality parameter inversion models are limited by the representativeness of the measured data. The revisit time of ZY1-02D is 55 days and the cloud cover is common in Shanghai, resulting in few scenes per year. Our twelve field radiometric measurements surveys included 80 rivers in Shanghai and covered all seasons. For measured rivers, R rs (550) ranged from 0.0103 to 0.079 sr −1 , which is a relatively broad spectral range. Dataset 1 fully reflects the principal relation between water quality parameters and remote sensing reflectance of the study rivers in Shanghai. Therefore, the models developed by Dataset 1 would be applicable to the water quality parameter retrievals of rivers in the study area though some rivers have not been measured. Compared with the non-optically active water quality parameter (such as TP or TN) retrieval models established by Qiao et al. [10], Gao et al. [18], and Lu et al. [22], their application is limited to specific lakes or rivers and our models have wider applicability. Compared with the DO retrieval model established by Al-Shaibah et al. [16] using a linear model, the DO retrieval model in this study based on XGBoost has higher accuracy.
However, our data did not include water quality in large inland lakes. Here, we used R rs data and water quality data from Taihu Lake in 2009 to examine the further universality of models. The models developed by Dataset 1 had slightly lower performance on data of Taihu Lake than that of the validated dataset described in Section 3.2.2 ( Figure 11). This can be attributed to the ten-year difference between the data in Shanghai and Taihu Lake. Moreover, precipitation and human factors are important factors affecting river water quality. Precipitation increases river flow and makes it easier for pollutants to transfer into rivers [52]. The agricultural activities directly transport nutrients into the rivers. These factors mean the actual retrievals of non-optically active water quality parameters in small rivers may have potential uncertainties.

Conclusions
In this study, we examined an in situ observation modeling approach based on machine learning for inversion of non-optically active water quality parameters from the hyperspectral ZY1-02D imagery at rivers and lakes in Shanghai. The machine learning models of ZY1-02D have better performance than those of Sentinel-2 because of finer spectral resolution. We conducted analyses on the applicability in different times and spaces of models, and the results showed that the models based on the in situ data in Shanghai can be applied to Taihu Lake. Finally, we validated the accuracy and consistency of R rs and water quality products derived from ZY1-02D compared to those derived from Sentinel-2 images. The comparison of R rs data showed strong agreement at bands of high energies. The COD Mn products showed stronger agreement than DO and TP.
We performed the field-based models' application to a ZY1-02D image which had been atmospherically corrected. However, the FLAASH model used in this study was not for water bodies. Future studies will be dedicated to exploring accurate AC methods for preprocessing ZY1-02D images in the context of quantitative water quality. Overall, the results show the high potential of ZY1-02D hyperspectral imagery in aquatic-oriented applications, though retrieving reliable non-optical water quality parameters is still challenging and further developments are needed.