Journal of Cleaner Production 385 (2023) 135681

Contents lists available at ScienceDirect

Journal of Cleaner Production

journal homepage: www.elsevier.com/locate/jclepro

Inversion prediction of COD in wastewater based on

hyperspectral technology
Danping Huang a, *, Ying Tian a, Shaodong Yu a, Xiaomei Wen a, Siyu Chen a, Xiang Gao a,
Luotong Ren b, Jia Zhen c, Xiaoguang Chen b, **
a
Sichuan Provincial Key Lab of Process Equipment and Control, School of Mechanical Engineering, Sichuan University of Science & Engineering, Zigong, 643000, China
b
College of Environmental Science and Engineering, State Environmental Protection Engineering Center for Pollution Treatment and Control in Textile Industry, DongHua
University, Shanghai, 201620, China
c
Wuliangye Yibin Co., Ltd, Yibin, 644007, China

A R T I C L E I N F O A B S T R A C T

Handling Editor: Zhen Leng COD is an important detection index in the field of water environmental treatment. At present, COD detection
methods and instruments have shortcomings such as long detection time, complicated detection process and high
Keywords: consumption of chemical agent. In this paper, a rapid detection method for COD index of wastewater based on
Wastewater index detection hyperspectral technology and machine learning is developed, and a complete non-contact detection scheme is
Hyperspectral technology
formed. In the method, the characteristic spectral data of standard COD in the sample spectral curve are
Regression analysis
extracted by successive projections algorithm (SPA) and genetic algorithm (GA) to establish the regression model
COD
Machine learning of the index. The model is applied to the rapid detection of COD index in the textile desizing wastewater
treatment process, and the accuracy of the model in real environment is comprehensively evaluated by root mean
square error (RMSE), relative analysis error (RPD) and determination coefficient (R2). Our results indicate that
the model has high prediction stability and strong generalization ability (the RMSE is 40.4489 mg/L, RPD is 9.37,
and R2 is 0.97) and is a green method for COD detection in wastewater.

1. Introduction unstable and the reliability is poor when detecting the wastewater of
refractory organics.
Water resource protection has always been one of the key and hot In recent years, hyperspectral technology has been widely used in the
research topics in the world. Rapid and effective detection of wastewater quantitative detection of water quality parameters. Water quality
indicators is of great significance to the protection of natural waters. The detection based on hyperspectral technology mainly adopts empirical
detection indicators of wastewater mainly include chemical oxygen method, semi-empirical method and machine learning method. The
demand (COD), turbidity, total phosphorus, total nitrogen and pH (Xing semi-empirical method is most commonly used at present. It can quan­
et al., 2019). COD describes the degree of organic pollution in water, and titatively detect indicators by combining the known spectral charac­
is an important detection index in the process of industrial wastewater teristics of water quality indicators with relevant statistical models and
treatment and supervision. Environmental engineering researchers want has been applied in the detection of suspended matter, chlorophyll,
to get a rapid method of obtaining COD indicators of water quality in turbidity and eutrophication index. For example, S.Chande et al.
order to detect and stop water pollution in time. Traditionally, COD (Chander et al., 2019) estimated turbidity, suspended sediment con­
detection mainly adopts dichromate methods, which have defects such centration and chlorophyll in lake water by using Aviris-NG hyper­
as long detection period, high consumption of chemical agent and easy spectral reflectance data. Zhang et al. (Fernandez-Beltran et al., 2018)
to produce secondary pollution. The most critical problem is that this used the combined method of depth factor decomposition machine,
method of detecting COD by oxidation reaction has its own defects. spatial distribution pattern analysis and probability analysis, and used
Since it is hard to control the dosage of oxidant in order to ensure the hyperspectral reflectivity data of water quality to quantitatively esti­
complete reacts of organics. Therefore, the COD measurement is mate water quality parameters. The biggest disadvantage of

* Corresponding author. Lingang Economic and Technological Development Zone, 188 University Town, Yibin City, Sichuan Province, 644000, China.
** Corresponding author. Room 4161, No.4 Academic Building, No. 2999, North Renmin Road, Songjiang District, Shanghai, 201620, China.
E-mail addresses: [email protected] (D. Huang), [email protected] (Y. Tian), [email protected] (X. Chen).

https://doi.org/10.1016/j.jclepro.2022.135681
Received 29 August 2022; Received in revised form 8 November 2022; Accepted 16 December 2022
Available online 20 December 2022
0959-6526/© 2022 Elsevier Ltd. All rights reserved.
D. Huang et al. Journal of Cleaner Production 385 (2023) 135681

Fig. 1. PVA wastewater treatment process using 2SSSAB-AFB.

semi-empirical method is that prior knowledge is required, which limits

Table 1
its application scope. COD has the characteristics of both optically
Standard physical and chemical index value of wastewater.
sensitive water quality indexes and non-optically sensitive water quality
indexes, and its spectral characteristic response is mainly affected by the Sample I inlet I inlet II inlet II inlet Aerobic outlet

spectral characteristics of organic pollutants in wastewater (A thesis COD (mg/L) 3622.31 989.35 614.46 569.71 457.51
submitted for the Degree of Doctorate at the Graduate School of the
Chinese Academy of Sciences n.d., 2007). Therefore, empirical method
The main pollutants are polyvinyl alcohol (PVA), acid and a small
and semi-empirical method cannot be used for COD indexes of
amount of starch (Song et al., 2021a). In the experiment, wastewater
wastewater.
samples were taken from the SSSAB-AFB treatment reactor, and treated
At present, the method based on machine learning is the most
by the two-stage spiral symmetrical flow anaerobic reactor-anaerobic
promising non-contact detection method for COD indicators. Gated
fluidized bed (2SSSAB-AFB). The treatment process is shown in Fig. 1
Recurrent Neural Network (GRU), improved from CNN(Tu et al., 2019),
(Dai et al., 2016; Zhang et al., 2018). Firstly, the wastewater is pumped
can predict COD indicators by using hyperspectral data of water quality.
from the inlet bucket into the SSSAB I and the outlet bucket 1 by the
However, GRU model has a complex structure which requires a large
peristaltic pump at the flow rate of φ and 1/5 φ respectively. Then the
amount of training sample data, and takes a long time to calculate.
wastewater will be treated by SSSAB I and then flows into the bucket 1;
Zhang (Zhang et al., 2021) established an intelligent water quality
Secondly, the sample liquid from outlet bucket 1 is pumped into SSSAB
monitoring system based on fusing Random Vector Functional Link
II and flows into outlet bucket 2 after treatment by SSSAB II. Finally, the
network (RVFL) and Group Method of Data Handling model (GMDH),
sample liquid is pumped into the anaerobic fluidized bed (AFB) and
but this method could only conduct qualitative analysis and could not
discharged from the outlet pipe after AFB treatment (Song et al., 2021b;
achieve quantitative detection. Deng (Deng et al., 2019) proposed
Zhou et al., 2021). In Fig. 1, ①~⑤ marks the sampling points of textile
modified Capsule network, which can quantitatively predict COD index
desizing wastewater, and the water samples are defined as I inlet, I
concentration in Baiyangdian lake of China by using one-dimensional
outlet, II inlet, II outlet, aerobic outlet.
hyperspectral data. However, the detection accuracy of this method is
Forty groups of samples were extracted from each sampling point,
low and cannot meet the detection requirements.
the standard COD index of the samples was determined by instruments
In this study, a fast non-contact COD index online monitoring
for detecting COD produced by HACH. Its principle is dichromate
method is intended to develop for monitoring the purification effect of
method (Huang et al., 2022), and the discrete points in the test results
industrial wastewater treatment reactor. The realization of quantitative
were eliminated by Grubbs test method. The final standard COD test
detection of COD index in wastewater mainly involves extraction of
results of the wastewater samples are shown in Table 1.
characteristics for hyperspectral data and analysis of quantitative
regression model. Extraction of spectrum characteristics is applied by
dimensionality reduction of hyperspectral data using statistic methods. 2.1. Hyperspectral data acquisition and processing of wastewater
Characteristics spectrum of COD index is then extracted from the data.
Firstly, using correlation coefficient, data characteristics of hyper­ The wastewater hyperspectral data were collected by FX10 hyper­
spectral data of wastewater is analyzed. According to the characteristics spectral camera (400–1000 nm) and FX17 hyperspectral camera
of the data and combined with experimental comparison, the best (900–1700 nm) developed by SPECIM In, Finland. The hyperspectral
dimensionality reduction method is chosen. Then, using several data of wastewater in the visible-near infrared (VNIR) and near infrared
regression models, different quantitative regression models are built (NIR) bands were obtained. The light transmittance of wastewater is
based on extracted characteristics spectrum and standard COD value. generally strong, and the material of the container will affect the
After analysis the best regression model is selected and has the stability hyperspectral data of wastewater (Yang et al., 2022). In this study, the
tested in the testing system build in this study. Finally, a stable testing hyperspectral data of wastewater samples were collected in bearing
model for COD index is established, which provides ideas for rapid containers made of quartz cuvette, glass cuvette, polyethylene Petri dish
testing of COD index for wastewater. and glass Petri dish respectively. The spectral curves of the five samples
collected are shown in Fig. 2.
2. Experimental materials and methods It can be seen from the analysis of the experimental results of the four
containers that samples with the same COD concentration collected by
This paper takes textile desizing wastewater as the research object. four containers have differences in peak values, trough values and trends
of spectral reflectance curve of hyperspectral data and differences in

2
D. Huang et al. Journal of Cleaner Production 385 (2023) 135681

Fig. 2. Spectral data of wastewater collected with different bearing containers.

Fig. 3. The curve of correlation coefficient between Continuum Removal and COD.

trends of concentration gradient curve. We compared the changes of ensure the scattering consistency of spectral curves of samples
curve with that of sample concentration and experimental results with with the same concentration (Le et al., 2022);
the best correlation was chosen for as the best collection method. it was (3) Remove the envelopes of spectral curves (Continuum Removal,
found that the spectral curves detected with quartz cuvette were clear CR) to highlight spectral features, and obtain new smooth spec­
and the reflectance difference between samples with different concen­ tral curves with prominent features (Yousefi et al., 2018).
trations was obvious. Therefore, the quartz cuvette was finally used as
the bearing containers in this study to collect hyperspectral data of 2.2. Detection method
wastewater samples.
The hyperspectral data processing was divided into the following 2.2.1. Selection of detection model
steps: The purpose of COD detection method proposed in this paper is to
establish a regression detection model by taking the hyperspectral data
(1) Extract the spectral reflectance by using whiteboard and dark of wastewater samples with different concentrations as the independent
current normalization and eliminate the noise caused by ambient variable and the standard COD value as the target variable. Pearson
light on spectral reflectance curves using polynomial convolution correlation coefficient between spectral data and COD standard value
smoothing algorithm (Savitzky-Golay Smoothing, SG)(Shi et al., (Adler and Parmryd, 2010) is calculated to analyze the correlation
2021); characteristics of independent variable and target variable, so as to
(2) Eliminate data offset between different pixels of the same sample select an appropriate regression model.
through Multiple Scattering Correction (MSC) algorithm to According to the results shown in Fig. 3, the maximum absolute
value of correlation coefficient between wastewater COD index and

Fig. 4. SPA algorithm screening feature bands of COD.

3
D. Huang et al. Journal of Cleaner Production 385 (2023) 135681

Table 2 Table 3
Feature bands of COD. Training effect of SVR model.
Algorithm Bands Feature bands（nm） REMS Model R2c R2p RMSE RPD

SPA VNIR 431.09、436.34、442.92、444.23、446.87、 90.2449 VNIR-SVR 0.8165 0.8352 313.6159 2.1691

457.40、491.75、560.92、638.81、939.06 VNIR-SPA- SVR 0.9826 0.9776 207.4437 5.8323
NIR 973.65、1032.6、1056.9、1091.7、1119.6、 53.9046 VNIR-GA- SVR 0.9990 0.9989 276.0959 21.8627
1133.5、1147.5、1150.9、1168.4、1178.9、 NIR- SVR 0.9695 0.9375 102.2259 4.0691
1206.9、1224.4、1238.4、1248.9、1266.4、 NIR-SPA- SVR 0.9911 0.9795 82.4563 8.5015
1364.9、1393.1、1488.6、1563.2、1577.4、 NIR-GA- SVR 0.9870 0.9702 83.7335 9.2162
1620.2、1645.2、1677.3、1680.9
GA VNIR 424.52、432.40、435.03、457.40、460.04、 136.7055
462.68、466.64、475.88、481.17、497.05、 experiment, SPXY (Sample Set Partitioning based on Joint X–Y Distance)
505.01、520.94、803.78
NIR 977.11、1060.4、1067.3、1074.3、1119.6、 34.9141
algorithm was used to process data (Espel et al., 2020) to ensure the
1133.5、1140.5、1168.4、1245.4、1277、 difference and representativeness between samples and obtain a dataset
1347.3、1442.6、1478、1492.2、1513.5、 that can improve the model stability.
1577.4、1655.9、1670.1、1673.7、1677.3
3.2.1. Evaluation method of detection model
In the paper, the following three indexes were used for comprehen­
spectral band is 0.21, indicating that the linear correlation between
sive evaluation of the detection performance of the selected regression
them is low. Based on correlation analysis result and the hyperspectral
models in order to explore the optimal detection model of wastewater
data characteristics of wastewater, the experiment adopted SVR, 1DCNN
COD index. The determination coefficient (R2) was used to evaluate the
and RF models. The optimal regression detection model suitable for
fitting degree and stability of the regression model. The closer R2 is to 1,
wastewater COD index was found through the comprehensive evalua­
the better the stability of the model is. The model determination coef­
tion of the training and verification results of the regression model.
ficient (R2c ) and prediction determination coefficient (R2p ) are used for
2.2.2. Dimension reduction method of wastewater hyperspectral data comprehensive evaluation of the stability of the model. The Root Mean
In this paper, two algorithms, the Successive Projections Algorithm Square Error (RMSE) was used to evaluate the detection accuracy of the
(SPA)(Li and Guo, 2021) based on the principle of geometric distance regression model. The smaller RMSE is, the higher the detection accu­
separability and the Genetic Algorithm (GA)(Peng et al., 2021) based on racy of regression models is. The Residual Predictive Deviation (RPD)
the principle of probability density separability were used to explore the was used to evaluate the predictive ability of the regression model of the
optimal dimensionality reduction method of the spectral characteristic data, and the larger the RPD is, the stronger the predictive ability of the
band of wastewater COD index. During compression, to prevent the loss model is (Cozzolino et al., 2004; Li et al. n.d.). The evaluation process is
of important features, PRESS (Prediction Error of Square Sum) criterion divided into two stages. In the first stage, we use the above three in­
was added to monitor the accuracy of feature band selection. Using this dicators obtained from the training results to evaluate and get the
criterion, whether verification of the Root Mean Square Error (REMS) pre-selection model. In the second stage, we use new samples to test the
reaches the minimum is taken as the standard to judge the accuracy of model, and use RMSE index to evaluate the accuracy of the model. In the
COD index feature band extraction. second stage, we mainly verify the generalization ability of the model to
ensure that the pre-selected model is adaptive to the samples.
3. Results and discussion
3.2.2. The experiment
3.1. Analysis of spectral characteristics
(1) Support Vector Regression prediction model
SPA algorithm and GA algorithm were adapted to extract the char­
acteristic bands of COD. Fig. 4a) shows the feature bands of COD index In the experiment of SVR model, radial basis function (RBF) was used
extracted by SPA algorithm in NIR band, which show multi-point as the kernel function. The center of RBF was selected, and weight pa­
concentrated distribution at 1200 nm and 1600 nm, and discrete dis­ rameters were updated by gradient descent method. The training results
tribution in other bands. Fig. 4b) shows that under PRESS criterion of the model are shown in Table 3.
verification, the REMS of verification results is minimum when the The data in Table 3 shows that, among the training effects of SVR
number of feature bands is 24. model, NIR-SPA-SVR model has the best stability and prediction ability.
Finally, the results of COD characteristic wavelength selected by the Additionally, the SVR detection model was tested with the test set to
two algorithms are shown in Table 2. As can be seen from the data in ensure the generalization ability, the results are shown in Fig. 5. The
Table 2, the SPA algorithm effectively reduces the characteristic optimal SVR detection model is NIR-SPA-SVR model, and its test REMS
dimension of hyperspectral data and the characteristic wavelengths are is 82.06 mg/L.
mainly concentrated in the range of NIR band. In the inspection and
verification, the final RMSE of NIR data was also smaller than that of (2) One-dimensional convolution regression prediction model
VNIR band, which indicated that the spectral response characteristics of
COD were mainly concentrated in NIR band. The 1DCNN model only performed effective convolution in the
process of convolution, and did not process boundary data. The con­
volutional and dense layers were activated using exponential linear
3.2. Comparison and evaluation of models
units (ELU), and the convolutional kernel weights and bias vectors were
initialized using Glorot_Uniform and Zeros methods respectively. The
The experiment was implemented in PyCharm 2020 professional
loss function adopted mean square error (MSE), and the training effect of
edition. Three regression models (SVR, 1DCNN and RF) were built using
1DCNN model is shown in Table 4.
Tensorflow2.0 environment, evaluated and tested. In order to verify the
The data in Table 4 show that the CNN model with the best training
effectiveness of COD spectral feature extraction in eliminating Hughes
effect is NIR-SPA-CNN, which R2 is above 0.98. RMSE and RPD are
phenomenon, three kinds of data were used as model datasets, including
78.0901 and 24.8758 respectively. The test set was used to verify the
the full band spectrum after removing envelopes and the spectral feature
prediction and generalization ability of CNN’s detection model and the
bands extracted by SPA and GA algorithm respectively. Before the

4
D. Huang et al. Journal of Cleaner Production 385 (2023) 135681

Fig. 5. Test effect of SVR model.

models, to some degrees, are all able to detect COD index of industrial
Table 4 wastewater. After comprehensive evaluation, NIR-GA-RF model has the
Training effect of CNN model.
highest detection accuracy and its REMS is 58 mg/L. In other words, RF
Model R2c R2p RMSE RPD regression model, which was built from hyperspectral data of textile
VNIR-CNN 0.8695 0.8775 301.2259 3.0691
desizing wastewater collected by FX17 camera, dealt by data processing
VNIR-SPA-CNN 0.9055 0.9846 207.5262 7.7548 and extracted characteristic bands by GA algorithm, is the best detection
VNIR-GA-CNN 0.9032 0.8760 327.9245 2.4299 scheme for COD index.
NIR-CNN 0.9695 0.9375 162.2259 4.0691
NIR-SPA-CNN 0.9950 0.9861 78.0901 24.8758
NIR-GA-CNN 0.9302 0.9889 70.4489 39.7872 3.3. Stability test

In order to verify the stability of the detection model proposed in this

test results are shown in Fig. 6.
paper, test equipment was set up for cyclic sampling test in the 2SSSAB-
The test results show that the optimal CNN detection model is NIR-
AFB anaerobic reactor. The structure of the test equipment is shown in
GA-CNN and the test REMS is 75.13 mg/L.
Fig. 8a).
In order to avoid the influence of ambient light on the accuracy of
(3) Random forest regression prediction model
hyperspectral data, the data acquisition module was installed in a dark
box with a calibration plate, halogen lamp and hyperspectral camera
Firstly, the decision tree was built by random sampling with
above. The quartz pipe was used to connect the sampling port (with a
replacement and completely splitting. Then the decision tree set split
water pump), and the pipe was equipped with a liquid level sensor at the
attribute nodes in the randomly selected attributes through information
end. When the sample volume of the pumped wastewater reached the
gain and other methods so that the decision tree branched from top to
liquid level sensor, the liquid level sensor would transmit the signal to
bottom and iterated through optimization criteria until the leaf nodes
the control module. After data collection, the air pump and electronic
could not split. The above steps were repeated to build the decision tree
valve at the outlet end of the collection system would pump out the
to form the forest. Average value of the prediction results of all decision
trees were took as the regression result of the model. RF model was used
to establish the inversion prediction model of COD index of desizing Table 5
Training effect of RF model.
wastewater, and the evaluation results are shown in Table 5.
According to the training effect of RF model in Table 5, the training Model Tree-Number R2c R2p RMSE RPD
results of the six network models all show good fitting degree and pre­ VNIR-RF 58 0.8695 0.8775 205.1616 3.0691
diction ability. The generalization ability of RF model was tested by VNIR-SPA-RF 75 0.9873 0.9728 195.3248 6.8391
using the test set, and the test results are shown in Fig. 7. The test results VNIR-GA-RF 86 0.8961 0.9618 153.2259 2.5439
show that the optimal RF detection model is NIR-GA-RF model and the NIR-RF 55 0.9695 0.9375 75.8526 4.0691
NIR-SPA-RF 61 0.9880 0.9627 68.0626 3.8700
REMS is 58.35 mg/L. NIR-GA-RF 45 0.9958 0.9971 54.1740 9.3728
The experimental results show that the three nonlinear regression

Fig. 6. Test effect of CNN model.

5
D. Huang et al. Journal of Cleaner Production 385 (2023) 135681

Fig. 7. Test effect of RF model.

Fig. 8. Hyperspectral dynamic wastewater detection system.

Table 6
Model stability test results.
Detect method Index SSSAB I(mg/L) SSSAB II(mg/L) AFB(mg/L)

Concentration ranges RMSE Concentration range RMSE Concentration range RMSE

PDM COD 900–4000 34.76 700–1200 38.56 200–700 58.46

NIR-GA-RF COD 900–4000 38.54 700–1200 40.68 200–700 63.46

wastewater sample in the pipeline. CRediT authorship contribution statement

In the process of changing the wastewater COD index concentration
decreased from 4000 mg/L to 200 mg/L. There were 80 times of peri­ Danping Huang: Conceptualization, Methodology. Ying Tian:
odically sampling and the NIR-GA-RF model was used to detect the COD Investigation, Software, Writing – original draft, Writing – review &
index of samples. The results are shown in Table 6. editing. Shaodong Yu: Formal analysis. Xiaomei Wen: Writing – re­
Comparing the RMSE in test results, it is found that the detection view & editing. Siyu Chen: Investigation. Xiang Gao: Supervision.
accuracy of NIR-GA-RF proposed in this study is close to that of di­ Luotong Ren: Validation. Jia Zhen: Investigation. Xiaoguang Chen:
chromate method, which can meet the detection requirements of textile Resources, Data curation.
desizing wastewater.
Declaration of competing interest
4. Conclusion
The authors declare that they have no known competing financial
This study verified that among SPA and GA algorithms, the charac­ interests or personal relationships that could have appeared to influence
teristic bands of wastewater COD index extracted by GA algorithm are the work reported in this paper.
more conducive to improve the accuracy of the detection model. By
comparing the three regression models of SVR, CNN and RF, the best Data availability
detection model of COD index of textile desizing wastewater is obtained.
Finally, through comprehensive evaluation, the optimal detection model The authors do not have permission to share data.
of COD index is determined as NIL-GA-RF. NIL-GA-RF has been verified
repeatedly in the testing equipment and its detection accuracy can meet Acknowledgements
the testing standard of COD index in the industrial wastewater industry.
It provides a reliable solution for rapid and non-contact detection of This work was sponsored by Enterprises and institutions entrust
COD index in industrial wastewater. scientific and technological projects(CXY2019ZR006); Liquor Making
Bio-Technology & Application of Key Laboratory of Sichuan Province
(NJ 2018-05).

6
D. Huang et al. Journal of Cleaner Production 385 (2023) 135681

References Retrieval of Water Quality, 2007. Retrieval of Water Quality in the Pearl River Estuary
Using Hyperspetral Technique. A Thesis Submitted for the Degree of Doctorate at the
Graduate School of the Chinese Academy of Sciences.
Adler, J., Parmryd, I., 2010. Quantifying colocalization by correlation: the pearson
Shi, X., Yao, L., Pan, T., 2021. Visible and near-infrared spectroscopy with multi-
correlation coefficient is superior to the Mander’s overlap coefficient. Cytometry 77,
parameters optimization of savitzky-golay smoothing applied to rapid analysis of soil
733–742. https://doi.org/10.1002/cyto.a.20896.
Cr content of pearl river delta. J. Geosci. Environ. Protect. 9, 75–83. https://doi.org/
Chander, S., Gujrati, A., Abdul Hakeem, K., Garg, V., Issac, A.M., Dhote, P.R., Kumar, V.,
10.4236/gep.2021.93006.
Sahay, A., 2019. Water quality assessment of River Ganga and Chilika lagoon using
Song, Q., Chen, X., Tang, L., Zhou, W., 2021a. Bioresource Technology Treatment of
AVIRIS-NG hyperspectral data. Curr. Sci. 116, 1172–1181. https://doi.org/
polyvinyl alcohol containing wastewater in two stage spiral symmetrical stream
10.18520/cs/v116/i7/1172-1181.
anaerobic bioreactors coupled a sequencing batch reactor. Bioresour. Technol. 340,
Cozzolino, D., Kwiatkowski, M.J., Parker, M., Cynkar, W.U., Dambergs, R.G., Gishen, M.,
125702 https://doi.org/10.1016/j.biortech.2021.125702.
Herderich, M.J., 2004. Prediction of phenolic compounds in red wine fermentations
Song, Q., Sun, Z., Chang, Y., Zhang, W., Lv, Y., Wang, J., Sun, F., Ma, Y., Li, Y., Wang, F.,
by visible and near infrared spectroscopy. Anal. Chim. Acta 513, 73–80. https://doi.
Chen, X., 2021b. Efficient degradation of polyacrylate containing wastewater by
org/10.1016/j.aca.2003.08.066.
combined anaerobic–aerobic fluidized bed bioreactors. Bioresour. Technol. 332
Dai, R., Chen, X., Xiang, X., Huang, D., Lin, H., Xu, M., 2016. Dispersion characteristics of
https://doi.org/10.1016/j.biortech.2021.125108.
a spiral symmetry stream anaerobic bio-reactor. Biochem. Eng. J. 110 https://doi.
Tu, J., Yang, X., Chen, C., Gao, S., Wang, J., Sun, C., 2019. Water quality prediction
org/10.1016/j.bej.2016.02.005.
model based on GRU hybrid network. In: Proceedings - 2019 Chinese Automation
Deng, C., Zhang, L., Cen, Y., 2019. Retrieval of chemical oxygen demand through
Congress. CAC 2019, pp. 1893–1898. https://doi.org/10.1109/
modified capsule network based on hyperspectral data. Appl. Sci. 9 https://doi.org/
CAC48633.2019.8996847.
10.3390/app9214620.
Xing, Z., Chen, J., Zhao, X., Li, Y., Li, X., Zhang, Z., Lao, C., Wang, H., 2019. Quantitative
Espel, D., Courty, S., Auda, Y., Sheeren, D., Elger, A., 2020. Submerged macrophyte
estimation of wastewater quality parameters by hyperspectral band screening using
assessment in rivers: an automatic mapping method using Pléiades imagery. Water
GC, VIP and SPA. PeerJ. https://doi.org/10.7717/peerj.8255, 2019.
Res. 186 https://doi.org/10.1016/j.watres.2020.116353.
Yang, W.C., Choe, C.M., Kim, J.S., Om, M.S., Kim, U.H., 2022. Materials selection
Fernandez-Beltran, R., Plaza, A., Plaza, J., Pla, F., 2018. Hyperspectral unmixing based
method using improved TOPSIS without rank reversal based on linear max-min
on dual-depth sparse probabilistic latent semantic analysis. IEEE Trans. Geosci. Rem.
normalization with absolute maximum and minimum values. Mater. Res. Express 9.
Sens. 56, 6344–6360. https://doi.org/10.1109/TGRS.2018.2837150.
https://doi.org/10.1088/2053-1591/ac2d6b.
Huang, D., Ye, J., Yu, S., Tian, Y., Wen, X., Wang, Y., Ren, L., Chen, X., 2022. Study on a
Yousefi, B., Sojasi, S., Ibarra Castanedo, C., Maldague, X.P.V., Beaudoin, G.,
fast non-contact detection method for key parameters of refractory organic
Chamberland, M., 2018. Continuum removal for ground-based LWIR hyperspectral
wastewater treatment. Biochem. Eng. J. 177, 1–30. https://doi.org/10.1016/j.
infrared imagery applying non-negative matrix factorization. Appl. Opt. 57, 6219.
bej.2021.108269.
https://doi.org/10.1364/ao.57.006219.
Le, C.T., Phan, N.L., Vu, D.D., Ngo, C., Le, V.H., 2022. Effect of multiple rescatterings on
Zhang, J., Chen, X., Liu, J., Huang, B., Xu, M., 2018. Structural characteristics of a spiral
continuum harmonics from asymmetric molecules in multicycle lasers. Phys. Chem.
symmetry stream anaerobic bioreactor based on CFD. Biochem. Eng. J. 137 https://
Chem. Phys. 24, 6053–6063. https://doi.org/10.1039/d2cp00245k.
doi.org/10.1016/j.bej.2018.05.016.
Li, L., Guo, S., 2021. A wavelength selection model based on successive projections
Zhang, Y., Wu, L., Deng, L., Ouyang, B., 2021. Retrieval of water quality parameters from
algorithm for pH detection of water by VIS-NIR spectroscopy. J. Phys. Conf. Ser.
hyperspectral images using a hybrid feedback deep factorization machine model.
1813 https://doi.org/10.1088/1742-6596/1813/1/012002.
Water Res. 204, 117618 https://doi.org/10.1016/j.watres.2021.117618.
Li, Z., Cheng, Y., Zhang, X., Zhang, Y., 2022. A Novel INS/ADS Integrated Navigation
Zhou, W., Chen, X., Ismail, M., Wei, L., Hu, B., 2021. Simulating the synergy of electron
Method Based on INS Error Model-Aided.
donors and different redox mediators on the anaerobic decolorization of azo dyes:
Peng, D., Tan, G., Fang, K., Chen, L., Agyeman, P.K., Zhang, Y., 2021. Multiobjective
can AQDS-chitosan globules replace the traditional redox mediators? Chemosphere
optimization of an off-road vehicle suspension parameter through a genetic
275. https://doi.org/10.1016/j.chemosphere.2021.130025.
algorithm based on the particle swarm optimization. Math. Probl Eng. 2021 https://
doi.org/10.1155/2021/9640928.