Transparency Estimation of Narrow Rivers by UAV-Borne Hyperspectral Remote Sensing Imagery

Urban rivers are often narrow, and general remote sensing data cannot meet the needs of water quality monitoring. In the process of monitoring of river water quality by remote sensing, the spectral and spatial dimension of satellite-borne images cannot be taken into consideration at the same time, making fine pollution monitoring of urban rivers difficult. Transparency is one of the core indicators for evaluating water quality, and hyperspectral remote sensing data are rich in spectral information and can be used for quantitative transparency estimation. The application of unmanned aerial vehicles (UAV)remote sensing effectively makes up for the deficiencies in satellite remote sensing monitoring. Aiming at this problem, this paper proposed the use of the eXtreme Gradient Boosting (XGBoost) regression algorithm for the quantitative inversion of urban river transparency. The spatial resolution of the collected imagery is 18.5 cm, which is suitable for urban rivers that are almost ten meters wide. Compared with five traditional empirical models, integrated algorithms such as gradient regression and random forest get much better results. Moreover, the accuracy of transparency estimation using the XGBoost regression algorithm was significantly improved, and the inversion model R2 in both study areas reached over 0.97. Finally, the established transparency inversion models were used to generate transparency distribution maps of the two study areas. The results showed that the distribution of the water transparency was consistent with the results of the field monitoring, indicating that it is feasible to use the XGBoost algorithm for the inversion of urban river transparency in UAV-borne hyperspectral imagery.

directly affects the depth that light can penetrate [6], [7], and is related to the spatial distribution of benthic plants such as large algae [8]. At certain water depths, the areas with higher transparency tend to have more vegetation. In contrast, in the areas with lower transparency, many aquatic plants cannot photosynthesize, because of the insufficient light. In addition, transparency is also related to planktonic algae, colored dissolved organic matter (CDOM), and suspended particles in the water [9]. Transparency is also a core indicator for evaluating the eutrophication of water bodies [10]. Therefore, it is of great significance to accurately monitor the transparency of water bodies.
Water transparency can be measured by optical sensors (such as transmissometers) and Secchi disks [11]. The Secchi disk, in particular, is widely used to indicate the transparency of water, in part because it is convenient, cheap and easy to obtain. Due to the spatial heterogeneity of water quality distribution, remote sensing technology can accurately, quickly, and effectively cover large areas [12], [13], and can obtain data that are difficult to obtain manually. As a result, remote sensing has become the most popular data source for obtaining regional-scale water body transparency data [14].
With regard to the remote sensing estimation of water body transparency, scholars have done a lot of research, but the research objects are mostly sea areas, rivers, and large lakes, and the data used are mostly space-borne data [4], [7], [15]- [18]. For example, Setiawan et al. [19] used Landsat imagery to estimate long-term lake transparency changes. For airborne hyperspectral data, Thiemann and Kaufmann [20] used airborne hyperspectral data to monitor the transparency of Mecklenburg Lake in Germany. Many scholars use various data to monitor the transparency of water bodies, such as TM [21], MSS [22], IKONOS [23] and even video data [24]. Among them, Landsat data is more widely used, and its single band or band ratio is used to build the transparency measurement algorithm, which has good performance [25]. These researches are aimed at larger scale water body transparency measurements. However, for the transparency monitoring of urban narrow river channels, the spatial and spectral resolutions of the traditional space-borne or airborne remote sensing data cannot meet the accurate monitoring requirements [26].
Unmanned aerial vehicle (UAV)-borne hyperspectral remote sensing images have high spatial and spectral resolutions, and are being used more and more in ecological monitoring [27]. At the same time, UAV-drones have many advantages such as low cost, low loss, reusability and easy operation, which can provide powerful help for environmental monitoring and effectively improve the quality of environmental monitoring. In this study, we applied a UAV remote sensing platform equipped with a miniature hyperspectral sensor to obtain hyperspectral remote sensing image data, which were combined with ground-based measured data, and used eXtreme Gradient Boosting (XGBoost) machine learning algorithm for urban river channel transparency monitoring. The purpose of this study was to combine UAV-borne hyperspectral imagery with a machine learning algorithm, and provide a reference for urban river channel transparency monitoring research.

A. STUDY AREAS
Two inland channels in Wuhan, China, were selected as the experimental areas for this study. The Shahu Port channel (114 • 21' 22.36'' E, 30 • 35' 11.44'' N) is one of Wuhan's many port channels, running from Youyi Avenue in the northeast, to Waisha Lake in the southwest, and Donghu Lake in the southeast. The port channel merges with the Yangtze River and has a total length of about 9.45 km. It is an important channel connecting Donghu Lake and Shahu to the Yangtze River. Due to the poor drainage and environmental conditions, the Shahu Port channel is a typical polluted urban water body. At the time of writing, the black odorous water treatment project at Shahu Port has now been completed, as has a dredging project. However, the water bodies of Shahu Port have been contaminated for a long time, and the problem of low visibility still exists. The selected study area of the port channel runs from Yangyuan South Road to Luojiagang Interchange, in which 40 sampling points are evenly distributed.
The Xunsi River (114 • 18' 0.12'' E, 30 • 29' 58.87'' N) is 16 km long and is located in the southern part of Wuhan City. It flows through Jiangxia District, Hongshan District, and Wuchang District, and finally joins the Yangtze River. With the continuous discharge of untreated domestic sewage into the river, the water body is now seriously polluted. There are many floating objects in the river, and the water quality is very poor. The river is gray-black in color and has an unpleasant odor. The pollution degree of the water body is greater than that of the Shahu Port channel. The research area covers a section of U-shaped river, in which 32 sampling points are evenly arranged in the river channel.

B. DATA COLLECTION
The data collected in this study included the transparency of the river sampling points, latitude and longitude coordinates, and measured ground spectra. A JCT-8 Secchi disk was used for the transparency field measurement. To ensure the accuracy of the measurement results, each measurement was repeated two to three times.
The ground-based measurement of the spectrum of the water was undertaken based on the ''above-water surface method'' [28]. When collecting the ground-based spectrum, calibration was first achieved on a standard board, the water surface spectrum was then recorded, and finally the spectrum of the sky was also collected. In the acquisition process, care was taken to avoid direct sunlight; otherwise, the instrument would be overexposed, resulting in unusable sampling data.

C. UAV-BORNE IMAGE ACQUISITION AND PREPROCESSING
The hyperspectral imagery was obtained with a Nano-HyperSpec mini hyperspectral imager. As shown in Table 1, the wavelength range of the Nano-HyperSpec sensor is 400-1000 nm, the number of spectral channels is 270, and the number of spatial channels is 640. The global positioning system/inertial measurement unit (GPS/IMU) module records the attitude information of the imager (including altitude, longitude, latitude, roll, pitch, and yaw). After loading the attitude information into the post-processing software provided by Headwall, the user can geometrically correct the collected strips. In addition, each Nano-HyperSpec sensor is calibrated with a lens selected by the user before it leaves the factory. The radiation can be directly corrected in the post-processing software provided by Headwall to convert the digital number (DN) data into radiance data. Coaxial reflection spectrum imaging technology is adopted, which has the advantages of a compact size and accurate spectral information. The hyperspectral data collection was performed on May 31 and June 1, 2019, at the Shahu Port channel and the Xunsi River, respectively. The flight platform for the Nano-HyperSpec hyperspectral imaging system was a sixrotor DJI M600 Pro drone. The flight altitude was set to 200 m. The spatial resolution of the collected imagery is 18.5 cm, the wavelength range is 0.4-1um, and the number of spectral bands is 270.
The UAV-borne hyperspectral data processing consisted of sensor radiation calibration, site radiation correction, VOLUME 8, 2020 geometric correction, and water extraction. Sensor radiation calibration involves converting the signal output by the sensor unit into the actual radiation intensity value. In this study, according to the radiation correction model and the conversion parameters provided by the sensor manufacturer, the original image data were converted pixel-by-pixel from the DN value to the radiance value. The remote sensing reflectance of the off-water radiation from the ground sampling point and the geometrically corrected remote sensing image pixel spectrum were then used to construct a linear relationship between the airborne radiation curve and the ground measured spectrum, so as to achieve site radiation correction. Because the drone was flown at a height of only 200 m, the complex atmospheric effects could be ignored. The image information was also geometrically corrected with the GPS/IMU module's attitude information [29].

A. EXTREME GRADIENT BOOSTING
Integrated learning involves building and combining multiple learners to complete the learning task. By combining multiple learners, it is often possible to obtain a significantly better generalization performance than with a single learner, so this approach is widely used in water quality monitoring [30]. The eXtreme Gradient Boosting (XGBoost) algorithm is a typical representative of the integrated algorithms, which was proposed in 2016, based on a regression tree [31]. XGBoost is a massively parallel boosted tree tool. The regression tree and the decision tree are similar in concept, and they are both tree-based algorithms.
The basis of XGBoost is the gradient boosting algorithm. Gradient boosting is a powerful technique for building predictive models. The integrated algorithm builds multiple weak evaluators on the data set and summarizes the modeling results of all the weak estimators to obtain a better regression or classification performance than with a single model. A weak evaluator is defined as a model that performs at least better than random guessing, that is, any model with a prediction accuracy of not less than 50%.
There are many ways to integrate the different weak evaluators. One approach is a bagging method that builds multiple parallel and independent weak evaluators at one time. Its typical representative method is the random forest algorithm [32]. Another approach is the lifting method, whose principle is to build weak evaluators one by one and gradually accumulate the multiple weak estimators after multiple iterations [33].
The most famous algorithms based on the lifting method are AdaBoost [34] and the gradient boosting decision tree (GBDT) [35]. XGBoost was developed from GBDT. There can be regression trees or classification trees in the gradient boosting tree structure, both of which use the classification and regression tree (CART) algorithm as the mainstream tree structure. The gradient lifting regression tree is an integrated model focusing on the regression tree model. Its modeling process is as follows. Firstly, a tree is built and then iterated gradually. A tree is added during each iteration, and a strong evaluator integrated with the many tree models is gradually formed.
For XGBoost, each leaf node will have a prediction score, which is also known as the leaf weight. This leaf weight is the regression value of all the samples on this leaf node on this tree, and is expressed by (x i ) or w, where f k represents the kth decision tree and x i represents the feature vector corresponding to sample i. When there is only one tree, f 1 (x i ) is the result returned by the lifting algorithm, but this result is far from ideal. When there are multiple trees, the regression result of the integrated model is the sum of the predicted scores of all the trees. Assuming that there are k decision trees in the integrated model, then the prediction results given by the entire model on sample i are: i represents the predicted value of the ith sample transparency in this experiment, and x i represents the input sample.
Accurate prediction is a very important factor, and XGBoost implements an algorithm that balances model performance and operation speed. The common loss functions, such as error rate and mean-square error, can only measure the performance of the model and cannot measure the speed of the model. Spatial complexity and temporal complexity are used in many models to measure the computational efficiency of the model. Therefore, XGBoost introduces model complexity to measure the algorithm's operational efficiency, the formula of which is expressed as: where i represents the ith sample in the data set, n represents the total amount of data imported into the kth tree, and K represents all the established trees. The first term represents the traditional loss function, which measures the difference between the true value and the predicted value. The measured difference is usually the adjusted standard error. The second term represents the complexity of the model, which is represented by a transformation of the tree model. This change represents a formula for measuring the complexity of the tree model from the structure of the tree, which is defined as follows: where T is the number of leaf nodes and w is the modulus of the leaf node vector. γ represents the difficulty of the node segmentation and λ represents the L2 regularization coefficient. In the process of iterating each tree, it is minimized in an attempt to obtain the optimal result, so the model error rate and model complexity are minimized at the same time.
According to (2), the objective function of the learning model for a total of t iterations is: According to Taylor's second-order approximate expansion, where: where l(y i ,ŷ (t−1) ) represents the prediction error of the learning model composed of the previous t − 1 trees. g i and h i represent the first and second derivatives of the prediction model for the current model, respectively. The current model iterates in the direction of decreasing prediction error. By combining (1) and (3), ignoring the constant terms, the structure of the tree introduced into the loss function, and the objective function will continue to transform. The purpose of transforming the objective function is to establish a direct relationship between the structure of the tree (the number of leaf nodes) and the size of the objective function, in order to find the optimal tree f t that needs to be solved in the tth iteration, to obtain: where: Find the optimal weight: Bring in the objective function and obtain the optimal solution of the objective function: Equation (10) is a function for scoring a tree structure, where the smaller the score, the better the tree structure. The algorithm searches for the optimal tree structure by constructing different tree structures using (10). All the segmentation points of the sample features are first determined, and each identified segmentation point is segmented. The criteria for the segmentation are as follows: where Gain represents the difference between the single node obj * and the tree obj * of the two nodes after slicing. It traverses the slicing points of all the features and finds the slicing point of the largest Gain, which is the best splitting point. Through this process, the optimal transparency monitoring model can be obtained through the parameter design of the algorithm.

B. PEARSON CORRELATION COEFFICIENTS
Pearson Correlation Coefficients is a way to measure the correlation of variables [36]. Its output range is -1 to +1.0 means no correlation, negative value means negative correlation, and positive value means positive correlation. The larger the absolute value, the stronger the correlation. The calculation formula is: where X represents the spectral characteristics of the sample points. In this experiment, it is the UAV-drone spectral value or the ratio characteristic of the band, Y represents the transparency value of the sample points, and N represents the number of samples.

C. MODEL ACCURACY EVALUATION
In this paper, the following statistical indicators were used to evaluate the transparency inversion models [37], [38].

1) MEAN ABSOLUTE ERROR (MAE)
The mean absolute error (MAE) can measure the stability of the model. The smaller the value, the more stable the model effect. The calculation formula is: where ŷ i − y i represents the absolute error of the second result, and n is the number of samples in the test set. The value range of MAE is [0, +∞). When the predicted value exactly matches the true value, it is equal to 0, which is the perfect model. The larger the error, the larger the MAE value.

2) ROOT-MEAN-SQUARE ERROR (RMSE)
The root-mean-square error (RMSE) is the square root of the ratio of the square of the deviation of the predicted value from the true value to the number of observations n. Its calculation formula is: The RMSE is very sensitive to extraordinarily large or very small errors in a set of data, so that it can reflect the accuracy of the prediction. The value range of RMSE is [0, +∞).

3) R-SQUARED (R 2 )
The R-squared value is used to characterize the quality of a fit through the change of data. Its calculation formula is: whereȳ represents the average of the actual data set. In the formula, the denominator is the degree of dispersion of the original data, and the numerator is the error between the predicted value and the true value. Their ratio can eliminate the effect of the degree of dispersion of the original data. The closer R 2 is to 1, the better the ability of the equation's variables to explain y, and the model also fits the data better.

IV. RESULTS AND DISCUSSION
In total, 40 and 32 spectra were extracted from the UAV-borne hyperspectral images in the first research area of Shahu Port channel and the second research area of Xunsi River, respectively, with a band range of 400-1000 nm. The band range of 400-900 nm (225 bands) was cut to perform the transparency inversion experiments. As mentioned in the overview of the study area in Section II, the river water in the two study areas selected in this experiment is relatively turbid and has low visibility, so this experiment will not be affected by reflections from the bottom of the river. Correlation evaluation is an important part of the quantitative inversion modeling process. In the analysis of the experimental results, Pearson correlation coefficients were used to characterize the correlation between the drone spectrum X_Spectra and the inversion target Y_Transparency [39].   transparency is 71cm. As shown in Table 2, the maximum value of the transparency of all the sampling points is 71cm, the minimum value is 19cm, the average value is 50.85cm, and the standard deviation is 16.34cm. Overall, the transparency of the 40 sample points at Shahu Port channel varies significantly.
The spectral curves extracted from the remote sensing imagery were pre-processed by the band ratio method, and Pearson correlation analysis was performed with the transparency values of the sampling points in the field. The correlation coefficients are shown in descending order in Fig. 4. The maximum correlation coefficient between the original remote sensing reflectance spectrum and the transparency  is 0.72454 (Fig. 4a). Overall, the correlation coefficients of 76 spectra are greater than 0.6, which are mainly concentrated in the ranges of 400-430nm and 680-840nm, indicating that these band ranges are sensitive to changes in transparency.
The band ratio method can eliminate the interference of background noise, such as water surface smoothness and the surrounding environment, which changes temporally and spatially. This is a contrast enhancement operation commonly used in quantitative remote sensing inversion [40], [41]. In this experiment, the exhaustive method was used to   calculate the ratio of the bands. The ratios of the 225 bands of the original spectra were calculated in pairs, and 50,400 characteristic variables were obtained. They are arranged in descending order according to the Pearson correlation coefficient as shown in Fig. 4b. The maximum correlation coefficient is 0.8918, and there are 312 remote sensing reflectance ratio correlation coefficients of greater than 0.8, with the maximum negative correlation being −0.8919. The correlation is significantly improved compared to the original spectra, so these data can be deemed as suitable for the transparency inversion modeling.
In this research, we tried to use the band ratio model to predict the transparency in various common remote sensing inversion models, including exponential function, power function, logarithmic function, linear function, and quadratic polynomial models to explore whether traditional methods are suitable for transparency inversion of urban rivers. In Shahu Port, the maximum correlation coefficient between the band ratio and the transparency ratio is 0.8919. The ratio (B136/B86) is the input variable of five empirical models, and the transparency is the output variable. The results of the inversion accuracy are listed in Table 3, and the fitted results are shown in Fig. 5. These results indicate that the traditional empirical models are not suitable for this experiment in Shahu Port.
Due to the small number of samples in this experiment, it is not necessary to set all the parameters when setting the parameters using the various integrated algorithms for the experiment. The parameters selected by GBDT for adjustment are learning_rate, the number of boosting stages to perform (n_estimators), and the fraction of samples used for fitting the individual base learners (subsample). The parameters that AdaBoost needs to set are n_estimators and learning rate. Random forest only needs to set the parameter n_estimators. The parameters required for XGBoost are n_estimators and maximum depth of the individual regression estimators (max_depth). Through the above three evaluation indexes RMSE, MAE and R 2 , a suitable model is determined for the inversion of transparency.
When the GBDT parameter is set, n_estimators is set in units of 50, and it is finally determined to be 500. To prevent overfitting when setting the other two parameters learning_rate and subsample, the unit is 0.001 and 0.05 respectively, and it is finally determined to be 0.004 and 0.55. AdaBoost set parameters in accordance with GBDT, and finally determined that the parameters n_estimators is 100 and learning_rate is 0.001. When setting parameters in the random forest algorithm, n_estimators is in units of 50, and it is finally determined to be 100. When setting parameters in XGBoost, max_depth is in units of 1 and n_estimators is also in units of 50. Finally, it is determined that max_depth is 4 and n_estimators is 150.
From Table 4, comparing the inversion results of all the models for Shahu Port channel, XGBoost shows the best effect in predicting the transparency. Although the prediction accuracy for the XGBoost validation data set (R 2 = 0.976, RMSE = 2.515cm, MAE = 2.008cm) is lower than the prediction accuracy for the modeled data (R 2 = 0.988, VOLUME 8, 2020 In addition, the accuracy of the transparency inversion prediction of several other integrated methods for Shahu Port channel can also meet the actual needs. The modeling accuracy for the AdaBoost prediction model (R 2 = 0.986, RMSE = 1.908cm, MAE = 1.423cm) and the prediction accuracy for the validation set (R 2 = 0.975, RMSE = 2.545cm, MAE = 2.001cm) are slightly lower than the results of the XGBoost model, as a whole, but this model does obtain relatively good results. Compared with the AdaBoost model, the prediction accuracy of the random forest model is reduced, but it also obtains good prediction results, with the R 2 for the modeling set and the validation set reaching 0.97. Compared with the other methods, the prediction accuracy of the GBDT model in modeling the data (R 2 = 0.973, RMSE = 2.640cm, MAE = 2.285cm) is the same,    set are shown in Fig. 6. The predicted and true values of all samples are evenly distributed on the diagonal, indicating that the inversion results are good, and the model can be used for transparency inversion. Therefore, considering these results, the XGBoost model was selected for the inversion of the hyperspectral imagery of Shahu Port channel.
After saving the determined model and inputting the UAVdrone image, the inversion results of river transparency can be obtained. Fig. 7 shows the results of the inversion of the channel transparency with the hyperspectral imagery using the XGBoost model. According to the inversion results for Shahu Port channel, the maximum value of the inversion result is 70.27cm and the minimum value is 18.92cm. The maximum value of the field measurement of the transparency is 71cm and the minimum value is 19cm. The inversion results are therefore consistent with the actual results. The sampling points in Fig. 7 are marked with the transparency values measured in the field.
Further observation of the map shows that the distribution of the channel transparency is consistent with the results of the field sampling. The transparency distribution shows a trend of low first and then high. Before sampling point No. 12, most of the transparency is below 30cm. During the field sampling, this river section was quite turbid, the water body was black in color, and the pollution was serious. However, starting from the 13th point, the overall transparency increases significantly. Later field investigations revealed that the drainage outlet between points 12 and 13 is the drainage outlet for the nearby waterworks, from which treated purified water is discharged. Therefore, the transparency results for Shahu Port channel appear low at first and then high.

B. SECOND STUDY AREA: XUNSI RIVER
In this experiment, the transparency data for 32 sample points were collected in the field at Xunsi River. The line chart of the transparency distribution is shown in Fig. 8. In contrast with the results for Shahu Port channel, the transparency distribution of the sampling points for the Xunsi River is relatively stable, and the transparency values of each point are all between 35cm and 60cm. Statistics of the transparency information for the 32 sample points in Xunsi River are shown in Table 5. The maximum value of all the sample points is 59cm, the minimum value is 39cm, the average value is 46.67cm, and the standard deviation is 5.28cm. On the whole, the transparency of the sample points at Xunsi River changes only slightly. According to the preliminary judgment, there are no abnormal values in the transparency data from Xunsi River.
The spectral curves extracted from the UAV-borne remote sensing imagery were pre-processed by the band ratio method, and Pearson correlation analysis was performed with the transparency data of the field sampling points from Xunsi River. The correlation coefficients are shown in descending order in Fig. 9. The maximum correlation coefficient between the original remote sensing reflectance spectra and the transparency is 0.6924 (Fig. 9a). Overall, 60 spectra have correlation coefficients of greater than 0.6, which are mainly concentrated in the 400-500nm range, indicating that this band range is sensitive to changes in transparency.
For the Xunsi River experiment, the band ratio enhancement was conducted again, which can eliminate the interference of background noise, such as water surface smoothness and the surrounding environment, which changes temporally and spatially. The exhaustive method was again used to calculate the ratios of the bands. The ratios of the 225 bands of the original spectra were calculated in pairs, and 50,400 characteristic variables were obtained. The Pearson correlation coefficients and the transparency of the sample points are arranged in descending order in Fig. 9b. The maximum correlation coefficient is 0.7887. There are 47 remote sensing reflectance ratio correlation coefficients of greater than 0.75. The correlation is again significantly improved compared to the original spectra. Therefore, these data can be considered as suitable for the transparency inversion modeling.
In this experiment, we also used the band ratio model to predict the transparency in five common remote sensing inversion models. In Xunsi River, the maximum correlation coefficient between the band ratio and the transparency ratio is 0.7887. The ratio (B18/B25) is the input variable. The results of the inversion accuracy are listed in Table 6, and the fitted results are shown in Fig. 10. These results indicate that the traditional empirical models are not suitable for this experiment in Xunsi River.
When the GBDT parameter is set, n_estimators is set in units of, 50, and is finally determined to be 1500. To prevent overfitting when setting the learning_rate and subsample, it is set in units of 0.001 and 0.1 respectively, and it is finally determined to be 0.005 and 0.3. AdaBoost is similar to GBDT when setting parameters, and finally determined that the n_estimators is 400 and learning_rate is 0.001. When setting parameters in the random forest algorithm, n_estimators is also set to 50, which is finally determined to be 600. When setting parameters in XGBoost, max_depth is in units of 1 and n_estimators is also in units of 50. Finally, max_depth is determined to be 4 and n_estimators is 50. From Table 7, comparing the inversion results of all the models for Xunsi River, XGBoost shows the best effect in predicting the transparency. For the Xunsi River research area, the prediction accuracy for the XGBoost validation data set (R 2 = 0.972, RMSE = 0.684cm, MAE = 0.561cm) is comparable to the prediction accuracy for the modeled data (R 2 = 0.971, RMSE = 0.901cm, MAE = 0.594cm). Based on the minimum measured transparency of 39cm and a standard deviation of 5.28cm on the ground for Xunsi River, the inversion results meet the experimental requirements.
In addition, the accuracy of the transparency inversion results of several other integrated methods for Xunsi River shows a significant decrease. The accuracy of the AdaBoost prediction model (R 2 = 0.976, RMSE = 0.799cm, MAE = 0.525cm) is improved, but the prediction accuracy for its validation data set (R 2 = 0.881, RMSE = 1.414cm, MAE = 1.165cm) decreases significantly, so these modeling results are not accurate enough. Similar to the AdaBoost model, the accuracy of the GBDT model (R 2 = 0.974, RMSE = 0.863cm, MAE = 0.774cm) is higher than the test set prediction accuracy (R 2 = 0.871, RMSE = 1.476cm, MAE = 1.152cm). Compared with the other methods, the prediction accuracy of the random forest model for the modeling data (R 2 = 0.932, RMSE = 1.386cm, MAE = 1.233cm) and the prediction accuracy for the verification data (R 2 = 0.858, RMSE = 1.548cm, MAE = 1.274cm) decrease significantly. The inversion results of the training and test data set are shown in Fig. 11. The predicted and true values of all samples are evenly distributed on the diagonal, indicating that the inversion results are good, and the model can be used for transparency inversion. Therefore, the XGBoost model was selected for the inversion of the hyperspectral imagery from Xunsi River.
After the model parameters are determined, the inversion results can be obtained by inputting the UAV remote sensing images from the second study area. Fig. 12 shows the results of the XGBoost model's inversion of the transparency for the Xunsi River hyperspectral imagery. According to the statistics of the inversion results for Xunsi River, the maximum value of the inversion result is 55.75cm and the minimum value is 37.95cm. The statistical results of the field measurement have a maximum value of 59cm and a minimum value of 39cm. Therefore, the inversion result is consistent with the actual result. The sampling points in Fig. 12 are marked with the transparency values measured in the field.
Further observation of the map shows that the distribution of the channel transparency for the Xunsi River is consistent with the results of the field sampling. The transparency distribution shows a trend of low, then high, and then low again. In conjunction with Fig. 2, it can be seen that, at sampling points Nos. 1 to 8, the transparency is mostly above 50cm. Upstream of these sampling points, it was observed during the field sampling that this river section was turbid and severely polluted. However, at the beginning, at sampling point No. 1, the overall transparency increases significantly. Later field investigations revealed that, between the No. 1 and No. 7 sampling points, green floating islands have been set up on both banks of the Xunsi River to purify the water. Green plants, including celery, are planted on the floating islands to purify the water [42]. Therefore, the transparency of Xunsi River appears to be low at first and then high. Because the distribution of the water quality is relatively uniform, this inversion result only reflects the distribution trend of the river transparency, and the prediction results at individual pixels are not considered here. When compared with the actual transparency measurement results, the inversion results can better reflect the transparency distribution trend of the river.

V. CONCLUSION
In this paper, we have described the quantitative inversion of UAV-borne hyperspectral imagery for urban river channel transparency monitoring. In the experiments undertaken in this study, the spectral data in the 400-900nm wavelength range of the hyperspectral imagery were extracted, and the original spectra were used to establish a band ratio model. Compared with five traditional empirical models and other boosting integrated algorithms, the accuracy of the water body transparency estimation using the XGBoost algorithm showed a significant improvement. And in the case of low overall transparency, the XGBoost algorithm works well in both research areas. The training and test set R 2 values in both study areas were higher than 0.97. Finally, the transparency inversion models were used to generate transparency distribution maps of the two study areas. The results confirmed that the distribution of the water transparency was consistent with the results of the field monitoring.
To date, there have been few studies of the application of UAV-borne hyperspectral imagery for water quality monitoring, and the algorithms used have been different. From the above discussion, it has been shown that integrated algorithms such as XGBoost have a high prediction accuracy when applied to the water quality monitoring of UAV-borne hyperspectral images, and the results can meet the actual needs. Therefore, in future research, such algorithms could be combined with UAV-borne hyperspectral imagery to monitor other water quality parameters, such as chlorophyll and suspended solids concentrations. This research involved using the rich features of UAV-borne hyperspectral imagery to build a statistical relationship with water transparency. Moreover, through this study, we confirmed that it is possible to use UAV-borne hyperspectral imagery to achieve the dynamic monitoring of urban river water quality. With the immediacy of UAV-drone data acquisition, this research can be applied in more aspects, such as pollution source monitoring and pollutant transfer. In our future research, we will attempt to make better use of the hyperspectral spectral features.

ACKNOWLEDGMENT
The Intelligent Data Extraction and Remote Sensing Analysis Group of Wuhan University (RSIDEA) provided the datasets. The Remote Sensing Monitoring and Evaluation of Ecological Intelligence Group (RSMEEI) helped to process the datasets.
CAN HUANG received the B.S. degree in computer science and the M.S. degree in cartography and geographic information system from the Faculty of Resources and Environmental Science, Hubei University, Wuhan, China, in 2017 and 2020, respectively. His major research interests include unmanned aerial vehicle-borne hyperspectral remote sensing, and image classification.