Open Access
Add to BibSonomy
Abstract
The sustainable use of water resources is inseparable from water pollution detection. The sensing of toxic ammonia nitrogen in water currently requires auxiliary reagents, which may cause secondary pollution. Benefiting from the ability of substances to change light characteristics, this work proposes polarimetry-inspired feature fusion spectroscopy (PIFFS) to detect ammonia. The PIFFS system mainly includes a light source, a quarter-wave plate (QWP), a linear polarizer (LP) and a fiber spectrometer. The target light containing substance information is polarization modulated by adjusting the QWP and LP angles. Then, the Stokes parameters of target light can be calculated by appropriate modulations. The feasibility of PIFFS method to detect ammonia nitrogen is verified by experiments on both standard water samples and environmental water samples. Experimental results show that inspired by the first Stokes parameter, the fused features provide superiority in classifying ammonia concentration. The results also demonstrate the effectiveness of support vector machine-based concentration classification and random forests-based spectral selection. The interaction between light and substances ensures that the proposed PIFFS method has the potential to detect other pollutants.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Water is not only the origin of life, but also the necessity of life. Whereas, water scarcity is an increasingly serious environmental issue worldwide [1]. It is well recognized that water scarcity is mainly caused by water pollution. Unsurprisingly, water pollution usually comes from our daily lives, including but not limited to dumpsites and non-sanitary landfills [2], swine wastewater [3], agriculture and mining activities [4]. These aware behaviors and other unaware behaviors are therefore bringing various pollutants to the water. Nevertheless, more and more researches have been focusing on some major pollutants such as microplastics [5], heavy metals [6], petroleum hydrocarbons [7] and antibiotics [8]. In terms of chemical composition, ammonia nitrogen has attracted continuous attention worldwide due to its ubiquity in surface water and high toxicity [9]. The oxidation of ammonia nitrogen can reduce the concentration of dissolved oxygen in water to blacken and odorize the water. The deterioration of water quality further affects the survival of aquatic animals and plants [10]. That is, the water supporting a ship can also upset it.
In order to improve water quality and save lives, it is critical and urgent to detect pollutants in the water for further treatment. Recently, the detection of ammonia nitrogen mainly includes optical, electrochemical and biological enzyme methods [11]. In terms of optical methods focusing on the optical properties of ammonia nitrogen, sensing strategies based on spectrophotometry [12], fluorescence [13] and fiber-optic [14] have been widely developed and applied worldwide.
Initially, spectrophotometry adopts Berthelot’s reagent, an alkaline solution of phenol and hypochlorite, to detect ammonia with a mixed green indicator [15]. The Berthelot reaction was then improved to optimize reaction conditions mainly by replacing toxic phenol with salicylate [16], altering the substituents of salicylate [17], and evaporating dissolved ammonia [18]. Furthermore, ammonia can react with fluorescent reagents to generate fluorescent compounds, thereby absorbing external energy to emit light with a certain wavelength and concentration-related intensity [19–21]. However, both spectrophotometry and fluorescence require auxiliary reagents to quantify ammonia, which complicates the operation, restricts reaction conditions, and may cause secondary pollution.
As a novel method, fiber-optic detection utilizes optical fibers to transmit light that is sensitive to the external environment. Generally, ammonia reacts with the coating material of optical fiber to first change the refractive index and then modulate the transmitted light [22–24]. The modulated light is detected and further demodulated to quantify the ammonia. However, the changed refractive index or the poor adsorption of the coating material may reduce the detection sensitivity at low concentrations. Meanwhile, these current optical methods require ammonia to react with designated substances, which limits their application in actual environment and long-term real-time monitoring.
To overcome these current defects, this paper proposes polarimetry-inspired feature fusion spectroscopy (PIFFS) to detect ammonia nitrogen without any auxiliary reagents. In the proposed PIFFS, the light emitted by the light source is first changed by the water sample, then modulated by the polarization optical elements, and finally received by the fiber spectrometer. Obviously, the critical ammonia concentration is reflected in the light changed by the water sample. The changed light is modulated multiple times to further calculate four Stokes parameters in the polarization domain. Then, the modulated or calculated spectra of ammonia samples with various concentrations can be classified by machine learning algorithms. In this work, the random forests (RF) [25], k-nearest neighbor (k-NN) [26] and support vector machine (SVM) [27] are comprehensively compared in the ammonia classification. Inspired by the first Stokes parameter, the experimental results show that the classification accuracy can be improved by fusing the features received after multiple modulations. In addition, the classified spectra can be sorted and selected based on RF algorithm to reduce computational and time costs without loss of accuracy. This work paves a new way for optical detection of ammonia nitrogen by combining polarimetry and spectroscopy.
2. Ammonia sensing principle and PIFFS system
2.1 Ammonia sensing principle
The PIFFS method is proposed to detect ammonia nitrogen in water mainly based on reflection principle and fluorescence effect. The reflectance spectrum is the main component. Ammonia nitrogen is very sensitive to the light from 634 nm to 688 nm [28], resulting in the formation of spectral peaks in this spectral band. Moreover, the spectral response is closely related to the ammonia concentration, so the reflectance spectra under different ammonia concentrations are significantly different in the spectral band from 634 nm to 688 nm.
In addition, the fluorescence effect cannot be ignored. Although the fluorescence effect lasts only tens to hundreds of nanoseconds after illumination, continuous light sources can enhance the fluorescence effect. Meanwhile, ammonia nitrogen molecules keep moving in the water. The motion changes the polarization state of the fluorescence relative to the light source, which is called fluorescence depolarization [29]. Exactly, this change in polarization state can be detected by the PIFFS system. Therefore, fluorescence polarization detection further increases the spectral differences from 634 nm to 688 nm under different ammonia concentrations.
In conclusion, the polarization spectroscopy introduced in this paper can maximize the spectral difference to a certain extent, thereby reducing the detection limit of ammonia nitrogen. Moreover, most of other regular indices in water, such as chemical oxygen demand (COD) and turbidity, will not affect ammonia detection due to different characteristic wavelengths [30].
2.2 PIFFS system
As shown in Fig. 1, the proposed PIFFS system mainly comprises of a light source, a reflector, a quarter-wave plate (QWP), a linear polarizer (LP), a fiber lens, an optical fiber and a fiber spectrometer. In our system established on the testbed, the manufacturer, model and spectral range of each component are listed in Table 1. The spectral range refers to the most efficient wavelengths transmitted by each component. Therefore, the established system works efficiently at visible wavelengths from 510 nm to 700 nm. Nevertheless, other wavelengths can also be transmitted, but with slightly lower efficiency.
Table 1. The manufacturer, model and spectral range of each component.
Herein, QWP and LP perform polarization modulation in the system. Let $\boldsymbol{P} = {\left ( {{S_0},\;{S_1},\;{S_2},\;{S_3}} \right )^{\textrm{T}}}$ represent the polarization parameters of the light containing the target information at any wavelength. Then, the polarization parameters are sequentially modulated by the Mueller matrix of QWP and LP, which can be expressed as
The matrix represented by Eq. (4) can also be regarded as the polarization modulation matrix of the PIFFS system. Adjusting the QWP angle and the LP angle can modulate the polarization information of the target light as required. Furthermore, the polarization parameters of the target light can be calculated from the received light intensities modulated by designing required matrices.
3. PIFFS methodology
3.1 Spectropolarimetric designing
The QWP and LP angles of four polarization modulations are listed in Table 2. Under the four polarization modulations, the combined Mueller matrices of QWP and LP described in Eq. (4) are as follows:
Table 2. The QWP and LP angles of the four polarization modulations.
The spectrometer only receives the light intensity information after polarization modulation, which is modulated by the first row of the Mueller matrix. Therefore, the four Stokes parameters and the degree of polarization (DoP) representing the target light can be calculated by
According to the combined Mueller matrix of QWP and LP, the Stokes parameters can be determined from light intensities under several modulation angles. The selected four angles in Table 2 are a convenient combination to calculate the four Stokes parameters, though other angles could also be used. Each modulation angle corresponds to a combined Mueller matrix. As long as the Mueller matrices determined by the four modulation angles satisfy a certain condition, i.e., the $4 \times 4$ matrix formed by the first row of four Mueller matrices is invertible, the four Stokes parameters can be solved by matrix operations.
3.2 Spectropolarimetric feature fusion
Inspired by the first Stokes parameter, the fusion of spectral features is proposed to improve sensing accuracy. The feature fusion strategy and index are listed in Table 3. The feature fusion strategies can be divided into two categories, one based on measured light intensities and the other based on calculated polarization parameters. Each feature fusion refers to the information superposition on all spectral bands. The 33 feature fusions are indexed to better visualize and analyze the results below.
Table 3. The feature fusion strategy and index.
4. PIFFS procedure for ammonia sensing
Figure 2 shows the PIFFS procedure for ammonia sensing. First, we establish the PIFFS system in the laboratory, as shown in Fig. 3. Table 1 lists the main components in the established system. In addition, the sample stage consists of a linear translation stage (Thorlabs, LTS300/M), a YZ translation stage (Thorlabs, LX20YZ/M) and a right-angle bracket (Thorlabs, MT402). The sample stage provides free movement in three-dimensional space for the sample. The light source illuminates the sample through a fiber ring illuminator (Thorlabs, FRI61F50). The reflector is mounted on a right-angle kinematic mirror mount (Thorlabs, KCB2/M). The QWP and LP are installed on a motorized rotation stage (Thorlabs, KPRM1E/M) and a cage rotation mount (Thorlabs, CRM1/M), respectively.
Then, water samples containing different ammonia concentrations are prepared. Section 3.1 introduces the polarization spectroscopy and Stokes calculation, and Section 3.2 introduces the feature fusion strategy. Finally, the polarization spectra of ammonia concentrations can be classified based on machine learning algorithms.
5. Standard water samples and experiments
5.1 Standard water samples
According to China’s national standards [31], ammonia nitrogen in drinking water is limited to no more than 0.5 mg/L. In addition, ammonia nitrogen is limited to no more than 5 mg/L in toilet flushing and vehicle washing, and no more than 8 mg/L in urban landscaping, street sweeping, fire protection, construction site and concrete production [32]. In order to satisfy practical applications, these limitations and their neighboring values are therefore included in the concentrations selected for ammonia nitrogen in this work.
Table 4 lists the detailed distribution of the 29 concentrations of ammonia nitrogen standards. The ammonia concentration ranges from 0 mg/L to 100 mg/L at different intervals. As listed in Table 4, there are 11 concentrations from 0 mg/L to 1 mg/L at 0.1 mg/L intervals, 10 concentrations from 1 mg/L to 10 mg/L at 1 mg/L intervals, and 10 concentrations from 10 mg/L to 100 mg/L at 10 mg/L intervals. Noted that the 1 mg/L that belongs to the two concentration ranges is counted only once, as is the case for 10 mg/L, hence there are totally 29 concentrations from 0 mg/L to 100 mg/L.
Table 4. The ammonia concentrations in standard water samples.
The ammonia nitrogen standards are prepared by Aoke Reference Materials of China, with ammonium chloride as the solute and ultrapure water as the solvent. A standard substance certificate is issued only if the tested concentration is qualified. The standard water sample for each ammonia concentration is 50 mL. For each of our experimental measurements, we use a pipette (Thermo, 100-1000 $\mathrm{\mu}$L) to take 5 mL water sample.
Then, the overall PIFFS procedure for ammonia sensing in standard water samples is summarized as Fig. 4. Two sets of ammonia samples with 29 concentrations are polarization modulated and measured to further calculate the polarization parameters, as described in Section 3.1. Furthermore, data preprocessing includes experimental data mixing, feature fusion, concentration range division, and training-test spectra distribution. Finally, we classify all spectral bands based on machine learning algorithms in Section 5.3.1, and select spectral bands for classification based on machine learning algorithms in Section 5.3.2.
5.2 Polarization spectroscopy
For each ammonia concentration in standard water samples, 100 spectra are collected under one polarization modulation. Then, the QWP and LP are rotated to successively collect the spectra under the four polarization modulations. After replacing the standard samples with 29 concentrations, the spectra are measured again under the above four polarization modulations. Finally, we can measure two sets of experimental spectra under the same experimental conditions and different samples with the same concentrations.
The average measured spectra of the two experiments are shown in Fig. 5. Each experiment contains the spectra under the four polarization modulations listed in Table 2. Obviously, the spectral distribution is similar overall in the two experiments, although there are subtle fluctuations caused by water samples or system stability. In the same experiment, the spectral peaks decrease slightly in the order of the third, second, fourth and first polarization modulations. In the same polarization modulation, there is no significant difference in the spectra of ammonia nitrogen standards with 29 concentrations. Therefore, appropriate algorithms are then required to classify these spectra.
5.3 Experimental results and discussion
5.3.1 Classification results based on feature fusion
To verify the reliability of the proposed method, the classification is performed based on not only the first set of experimental spectra, but also the mixed spectra of the two sets of experiments. Furthermore, the ratio of the number of spectra between the training set and the test set generally affects the classification results. Table 5 lists the number of spectra of the training set and the test set in different experimental data. It can be known from Section 5.2 that in each set of experiments, 100 spectra are collected for each polarization modulation at each concentration. Thus, a total of 200 spectra are collected in the two experiments under the same concentration and polarization modulation. Then, the 100 spectra of the first experiment and the 200 spectra of the mixture of two experiments are divided into the training set and the test set at a ratio of 9:1, 3:1 and 1:1, respectively.
Table 5. The number of spectra of the training set and test set in different experimental data.
Then, the RF [25], k-NN [26] and SVM [27] algorithms are used to classify all spectral bands based on each feature in Table 3, each training-test ratio in Table 5, and each concentration range in Table 4. Integrating these variable factors, the classification results based on the first set of experimental data are shown in Fig. 6. Coincidentally, the classification results based on the mixture of two sets of experimental data are shown in Fig. 7.
Fig. 6. Classification results of all spectral bands under different features based on the first set of experimental data.
Fig. 7. Classification results of all spectral bands under different features based on the mixture of two sets of experimental data.
The experimental settings for different algorithms used in this paper keep the same. In RF, the number of trees is 1000, and the number of randomly-selected features for each tree is the square root of the total number of features. In k-NN, the Euclidean distance is used, and the number of nearest neighbors in the predictors is 1. In SVM, we choose the linear kernel function with the penalty coefficient as 1. The calculations are conducted on a computer with an Intel Core i7-10700K CPU, Windows 10 system and 64 GB RAM.
It can be seen from Fig. 6 that the classification accuracy in the concentration range from 0 mg/L to 1 mg/L is generally higher than that in other concentration ranges. The classification accuracy decreases as the ammonia concentration grows, largely due to the fact that the ammonia detection is based on the sensitivity of ammonia to light. The sensitivity may vary more significantly at lower concentrations. The stronger sensitivity is reflected in the form of larger spectral differences and hence higher classification accuracy. In addition, the accuracy difference may also be caused by the experimental spectra acquired from low to high concentrations. The fluctuations of light source and spectrometer slightly increase after a long-time experiment, which affects the acquired spectra and thus reduces the classification accuracy. In the same concentration range, the classification accuracy increases as the training-test ratio increases. Even if the training-test ratio is as low as 1:1, the classification accuracy can be higher than 98% in each concentration range. In terms of the 33 classified features, the classification accuracy of the last 11 features is slightly higher than that of the middle 18 features and significantly higher than that of the first 4 features. In other words, the first Stokes parameter can effectively improve the classification accuracy. Expanding the concept of the first Stokes parameter, the fusion of several light intensities can further improve the classification accuracy. Obviously, the four measured light intensities are fused to obtain the classification accuracy closest to 100% in each concentration range and each training-test ratio. In the classification algorithms, SVM shows superiority in both classification accuracy and classification time. This is because SVM is a model trained by maximizing the distance interval, and it performs better on the classification problem of small sample sets. However, k-NN cannot form a model. The test set is identified by the distance calculation of training set, which is too dependent on the training set, and the model is prone to overfitting. RF is an integrated model, and multiple sub-models are obtained through sampling. However, for the small sample sets, the number of features is much larger than the number of samples, which also leads to overfitting of the model.
Following the above analysis on Fig. 6, similar conclusions can be drawn directly from Fig. 7. Based on the mixed data, the classification accuracy of SVM can reach 100% on the training set and higher than 95% on the test set in each concentration range and each training-test ratio. The classification time mentioned here refers to the time required to complete the entire process, including loading data, training model and predicting label. In terms of test time, different models vary slightly. For example, to identify 29 spectra (one spectrum from each concentration), SVM, k-NN and RF need 0.669s, 0.092s and 0.021s, respectively. Therefore, the proposed method can be applied to the high-precision classification of ammonia nitrogen standards. Due to the small concentration interval, this classification can be regarded as an approximate quantitative detection.
5.3.2 Classification results based on spectral selection
All spectral bands can be used for classification to achieve high accuracy, but contain some redundant information. Meanwhile, using all spectral bands to classify consumes a lot of computational cost and time cost. In order to reduce these costs, the spectral bands of the first Stokes parameter are first selected by RF, and then classified by SVM. The number distribution of selected spectral bands is listed in Table 6. It can be seen that the number of selected bands ranges from 20 to 3600 with varying intervals. The classification results based on the first set of data and the mixed data are shown in Fig. 8 and Fig. 9, respectively. Moreover, Fig. 8 and Fig. 9 are marked with the classification accuracy using all spectral bands, and the minimum number of selected bands whose classification results are closest to the marked accuracy.
Fig. 8. Classification results after selecting the spectral bands of the first Stokes parameter based on the first set of experimental data.
Fig. 9. Classification results after selecting the spectral bands of the first Stokes parameter based on the mixture of two sets of experimental data.
Table 6. The number distribution of selected spectral bands.
Although the classified data are different, Fig. 8 and Fig. 9 reflect similarities in the overall trend of classification accuracy. The classification accuracy increases as the number of selected bands increases in each concentration range and each training-test ratio. In the same concentration range, increasing the training-test ratio can reduce the minimum number of selected bands. In the same training-test ratio, the minimum number of selected bands in the low concentration range is smaller than that in the high concentration range. As shown in Fig. 8, only 40 bands are required to achieve 100% classification accuracy in the concentration range from 0 mg/L to 1 mg/L and the training-test ratio with 9:1. The results show that the band selection effectively eliminates a large number of redundant bands without loss of classification accuracy. Whereas, the minimum number of selected bands required for the mixed data is generally greater than that required for the first set of data. This is because the two sets of data with the same concentration and different samples are inherently different, resulting in slight deviations in their band selection results. Furthermore, more selected bands are required to classify the mixed data. Nevertheless, the band selection and high-precision classification can be implemented simultaneously.
6. Environmental water samples and experiments
6.1 Environmental water samples
In order to further verify the feasibility and effectiveness of the PIFFS method, experiments are conducted on environmental water samples. Different from standard samples, environmental samples contain a variety of substances other than a certain research object. This paper is devoted to the detection of ammonia nitrogen in water.
First, Nongfu Spring drinking water, which is taken from natural water sources, is selected as the solvent. The characteristic indices of the 5-liter Nongfu Spring we used are listed in Table 7. It can be seen that Nongfu Spring is rich in a variety of natural mineral elements required by the human body, such as metasilicic acid, potassium, sodium, magnesium, calcium, etc.
Table 7. The characteristic indices of the 5-liter Nongfu Spring.
Then, ammonium chloride is used as the solute, and Nongfu Spring is used as the solvent to prepare the ammonia nitrogen solution. Since the ammonia nitrogen in drinking water is limited to less than 0.5 mg/L [31], the ammonia concentration is prepared from 0 mg/L to 1 mg/L with an interval of 0.1 mg/L. In addition, based on the ammonia nitrogen solution prepared by Nongfu Spring, we further add several regular indices in drinking water, such as COD, nitrate, chromaticity and turbidity. The COD, nitrate, chromaticity and turbidity in drinking water are limited to 3 mg/L, 10 mg/L, 15 Hazen and 1 NTU, respectively [31]. Therefore, the indices of the prepared environmental water samples are listed in Table 8. In preparing the six groups of samples, we consider not only the influence of a single regular index, but also the comprehensive influence of the four regular indices.
Table 8. The indices of the prepared environmental water samples.
Similarly, Nongfu Spring-based ammonia water samples (50 mL) from 0 mg/L to 1 mg/L at 0.1 mg/L intervals are prepared by Aoke Reference Materials of China, which also produces COD standard (50 mL, 30 mg/L), nitrate standard (50 mL, 100 mg/L), chromaticity standard (20 mL, 150 Hazen) and turbidity standard (50 mL, 10 NTU). Based on these water samples, we use a pipette (Thermo, 100-1000 $\mathrm{\mu}$L) to prepare the six groups of samples listed in Table 8. The water sample measured in each experiment is also 5 mL.
6.2 Polarization spectroscopy
The polarization spectra of environmental water samples are measured based on the system and method described in Section 4. It can be seen from Table 8 that each group of samples has 11 ammonia concentrations. For each ammonia concentration, we measure four sets of polarization spectra. Meanwhile, each set of polarization spectra contains 100 spectra. The 100 spectra are averaged to obtain the spectral curve of each ammonia concentration in each group of samples, as shown in Fig. 10.
Obviously, for the same group of samples, each set of polarization spectra varies among the 11 ammonia concentrations with similar trends. Moreover, the spectral changes are most pronounced in the first group of samples. This phenomenon suggests that other regular indices have a slight effect on the polarization spectrum. In addition, the third and sixth groups of samples show the smallest spectral changes. The third group may be because that nitrate contains the same nitrogen element as ammonia nitrogen. The sixth group is due to the comprehensive effect of several regular indices.
6.3 Experimental results and discussion
It is critical to classify the polarization spectra of various ammonia concentrations in complex environmental water samples. We therefore adopt the SVM algorithm that performs the best in Section 5.3.1. Meanwhile, among the 33 classified features, we select the 5th, 23rd and 33rd features whose accuracy increases sequentially in Section 5.3.1. Similarly, we still consider three training-test ratios of 9:1, 3:1 and 1:1. Based on the above parameters, the classification results of 11 ammonia concentrations in environmental water samples are listed in Table 9.
Table 9. The classification results of ammonia concentrations in environmental water samples.
Overall, the classification results of environmental water samples are satisfactory. As listed in Table 9, the classification accuracy of samples 1, 2, 4 and 5 are all 100%. This is mainly because that the environmental samples are classified based on the algorithm and features that perform better in classifying standard samples. Moreover, the experimental conditions of environmental samples are relatively improved in terms of shading and stabilization. In addition, there are individual classification accuracies slightly below 100% in samples 3 and 6. This phenomenon is consistent with the spectral distribution in Fig. 10, where the spectra of samples 3 and 6 vary relatively little between different ammonia concentrations. Nevertheless, 100% classification accuracy can be achieved by increasing the training-test ratio (9:1 or 3:1) or by fusing the four polarization spectra (feature index 33). Therefore, the PIFFS method can effectively classify the ammonia concentrations in environmental water samples.
7. Conclusions
The PIFFS method is proposed in this work to simplify ammonia detection and avoid secondary pollution. The PIFFS system mainly consists of a light source, a QWP, an LP and a fiber spectrometer. In the spectral domain, the established system can effectively acquire the visible spectrum. In the polarization domain, the target light is modulated four times by adjusting QWP and LP angles to calculate its four Stokes parameters. Furthermore, the modulated light intensities and calculated Stokes parameters are respectively superimposed to obtain 33 fused features. Referring to the national standards on the limitation of ammonia nitrogen in water, 29 concentrations with different intervals are selected from 0 mg/L to 100 mg/L for classifying ammonia in standard water samples. Various experimental results fully show that the higher classification accuracy can be obtained by superimposing the modulated light intensities inspired by the first Stokes parameter. Meanwhile, compared to k-NN and RF, SVM shows overall superiority in classification accuracy and time. Regardless of the training-test ratio, the overall classification accuracy of 29 concentrations can always reach 100% for the first experimental data and more than 98% for the mixed data of the two experiments. The experimental results also show that a few spectra selected by RF can generally achieve the same high classification accuracy as all spectra. Significantly, the PIFFS method can successfully detect ammonia as low as 0.1 mg/L in both standard water samples and environmental water samples. Due to the exceptional reaction between light and substance, the PIFFS method can be further applied to detect other water pollution.
Funding
National Key Scientific Instrument and Equipment Development Projects of China (61527802).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. C. He, Z. Liu, J. Wu, X. Pan, Z. Fang, J. Li, and B. A. Bryan, “Future global urban water scarcity and potential solutions,” Nat. Commun. 12(1), 4667 (2021). [CrossRef]
2. A. K. Morita, C. Ibelli-Bianco, J. A. Anache, J. V. Coutinho, N. S. Pelinson, J. Nobrega, L. M. Rosalem, C. M. Leite, L. M. Niviadonski, C. Manastella, and E. Wendland, “Pollution threat to water and soil quality by dumpsites and non-sanitary landfills in Brazil: A review,” Waste Manage. 131, 163–176 (2021). [CrossRef]
3. D. Cheng, H. H. Ngo, W. Guo, S. W. Chang, D. D. Nguyen, Y. Liu, Q. Wei, and D. Wei, “A critical review on antibiotics and hormones in swine wastewater: Water pollution problems and control approaches,” J. Hazard. Mater. 387, 121682 (2020). [CrossRef]
4. Y. Liu, P. Wang, B. Gojenko, J. Yu, L. Wei, D. Luo, and T. Xiao, “A review of water pollution arising from agriculture and mining activities in Central Asia: Facts, causes and effects,” Environ. Pollut. 291, 118209 (2021). [CrossRef]
5. L. Beichert, Y. Binhammer, J. R. C. Andrade, R. Mevert, A.-K. Kniggendorf, B. Roth, and U. Morgner, “Real-time stimulated Raman spectroscopy with a non-collinear optical parametric oscillator,” Opt. Express 29(20), 31499–31507 (2021). [CrossRef]
6. P. Gupta, C. E. Rahm, D. Jiang, V. K. Gupta, W. R. Heineman, G. Justin, and N. T. Alvarez, “Parts per trillion detection of heavy metals in as-is tap water using carbon nanotube microelectrodes,” Anal. Chim. Acta 1155, 338353 (2021). [CrossRef]
7. S. Uddin, S. W. Fowler, T. Saeed, B. Jupp, and M. Faizuddin, “Petroleum hydrocarbon pollution in sediments from the Gulf and Omani waters: Status and review,” Mar. Pollut. Bull. 173, 112913 (2021). [CrossRef]
8. M. Gros, N. Catalán, J. Mas-Pla, M. Celic, M. Petrovic, and M. J. Farré, “Groundwater antibiotic pollution and its relationship with dissolved organic matter: Identification and environmental implications,” Environ. Pollut. 289, 117927 (2021). [CrossRef]
9. G. Yang, L. Liu, T. Wang, L. Fan, X. Huang, D. Tian, L. Jiang, J.-F. Silvain, and Y. Lu, “Laser-induced breakdown spectroscopy of ammonia gas with resonant vibrational excitation,” Opt. Express 28(2), 1197–1205 (2020). [CrossRef]
10. T. A. Neubauer, T. Hauffe, D. Silvestro, J. Schauer, D. Kadolsky, F. P. Wesselingh, M. Harzhauser, and T. Wilke, “Current extinction rate in European freshwater gastropods greatly exceeds that of the late Cretaceous mass extinction,” Commun. Earth Environ. 2(1), 97 (2021). [CrossRef]
11. D. Li, X. Xu, Z. Li, T. Wang, and C. Wang, “Detection methods of ammonia nitrogen in water: A review,” Trends Anal. Chem. 127, 115890 (2020). [CrossRef]
12. A. Jain, S. Soni, and K. K. Verma, “Combined liquid phase microextraction and fiber-optics-based cuvetteless micro-spectrophotometry for sensitive determination of ammonia in water and food samples by the indophenol reaction,” Food Chem. 340, 128156 (2021). [CrossRef]
13. L. Placer, L. Estévez, I. Lavilla, F. Pena-Pereira, and C. Bendicho, “Assessing citric acid-derived luminescent probes for pH and ammonia sensing: A comprehensive experimental and theoretical study,” Anal. Chim. Acta 1186, 339125 (2021). [CrossRef]
14. F. Liu, H. Bao, H. L. Ho, W. Jin, S. Gao, and Y. Wang, “Multicomponent trace gas detection with hollow-core fiber photothermal interferometry and time-division multiplexing,” Opt. Express 29(26), 43445–43453 (2021). [CrossRef]
15. J. J. Giner-Sanz, G. M. Leverick, V. Pérez-Herranz, and Y. Shao-Horn, “Salicylate method for ammonia quantification in nitrogen electroreduction experiments: The correction of iron III interference,” J. Electrochem. Soc. 167(13), 134519 (2020). [CrossRef]
16. J. J. Giner-Sanz, G. Leverick, V. Pérez-Herranz, and Y. Shao-Horn, “Optimization of the salicylate method for ammonia quantification from nitrogen electroreduction,” J. Electroanal. Chem. 896, 115250 (2021). [CrossRef]
17. H. Yu, G. Zhang, Y. Cai, and F. Dong, “Altering the substituents of salicylic acid to improve Berthelot reaction for ultrasensitive colorimetric detection of ammonium and atmospheric ammonia,” Anal. Bioanal. Chem. 413(23), 5695–5702 (2021). [CrossRef]
18. Y. B. Cho, S. H. Jeong, H. Chun, and Y. S. Kim, “Selective colorimetric detection of dissolved ammonia in water via modified Berthelot’s reaction on porous paper,” Sens. Actuators, B 256, 167–175 (2018). [CrossRef]
19. X. Chen, T. Xiong, J. Xu, Y. Li, M. Zhang, and Y. Liang, “Determination of ammonium in natural water using a quinoline-based o-dialdehyde fluorescent reagent with visible excitation wavelength,” Anal. Methods 13(43), 5231–5239 (2021). [CrossRef]
20. J. Zhang, Z. Xu, C. Shi, and X. Yang, “A fluorescence method based on N, S-doped carbon dots for detection of ammonia in aquaculture water and freshness of fish,” Sustainability 13(15), 8255 (2021). [CrossRef]
21. Z. Li, T. Wang, X. Xu, C. Wang, and D. Li, “An “on–off” fluorescent probe based on cucurbit[7] uril for highly sensitive determination of ammonia nitrogen in aquaculture water,” Anal. Methods 13(36), 4090–4098 (2021). [CrossRef]
22. P. Ma, N. Hu, J. Ruan, H. Song, and X. Chen, “In-situ measurement of ammonium in wastewater using a tilted fiber grating sensor,” J. Lightwave Technol. 39(12), 4055–4061 (2021). [CrossRef]
23. S. H. Girei, H. N. Lim, M. Z. Ahmad, M. A. Mahdi, A. R. M. Zain, and M. H. Yaacob, “High sensitivity microfiber interferometer sensor in aqueous solution,” Sensors 20(17), 4713 (2020). [CrossRef]
24. D. Rong, G. Meng, X. Fang, L. You, and Z. Deng, “Delafossite AgAlO2 modified long-period grating for highly-sensitive ammonia sensor,” Opt. Express 29(25), 42005–42019 (2021). [CrossRef]
25. L. Breiman, “Random forests,” Machine Learning 45(1), 5–32 (2001). [CrossRef]
26. G. M. White and P. J. Fong, “k-nearest-neighbor decision rule performance in a speech recognition system,” IEEE Trans. Syst., Man, Cybern. SMC-5(3), 389 (1975). [CrossRef]
27. C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector machines,” ACM Trans. Intell. Syst. Technol. 2(3), 1–27 (2011). [CrossRef]
28. M. Wu, X. Li, and J. Sha, “Spectral inversion models for prediction of red soil total nitrogen content in subtropical region (Fuzhou),” Spectrosc. Spectr. Analysis 33(11), 3111–3115 (2013). [CrossRef]
29. H. G. Weber, F. Bylicki, F. König, H. Burkhardt, and J. Wieland, “A new molecular fluorescence depolarization effect: The coherence decay time of NO2,” Z. Phys. A: At. Nucl. 309(4), 363–364 (1983). [CrossRef]
30. X. Chu, “Molecular spectroscopy analytical technology combined with chemometrics and its applications,” (Chemical Industry, 2011), pp. 154–155.
31. “Standards for drinking water quality,” GB 5749–2006.
32. “The reuse of urban recycling water—water quality standard for urban miscellaneous use,” GB/T 18920-2020.
