Effect of rainfall-derived inflow and infiltration on dissolved organic matter in urban sanitary sewers using chemometric and machine learning approaches
Article information
Abstract
This study investigated changes in dissolved organic matter (DOM) properties within urban sanitary sewers influenced by groundwater infiltration and excessive rainfall-derived inflow and infiltration (RDII). It employed optical indices and fluorescence excitation-emission matrix-parallel factor analysis (PARAFAC) coupled with self-organizing map (SOM) to compare DOM characteristics during wet weather flows (WWFs) and dry-weather flows (DWFs). Sampling sites impacted by RDII were identified based on flowrate. Optical indices and PARAFAC components (C1–C4) were used to differentiate DOM properties between DWFs and WWFs. In WWFs, E2/E3 and S350-400 increased, while spectral ratio (SR) decreased, indicating a shift towards smaller organic matters. Reduced fluorescence and humidification indices suggested the input of fresher/terrestrial organic matters. C3 and C4 exhibited significant distinctions, showing increased C3 and decreased C4 levels. The PARAFAC-SOM modeling further illustrated that water samples in the urban sewer system could be categorized based on the dominance of DOM properties. Principal component analysis revealed separation between DWF and WWF samples in principal component 1 (PC1), associated with molecular size. PC2 was linked to microbial activity in WWFs. Notably, DWF samples from the NY-11 site shifted to the positive side of the PC1 axis, while their corresponding WWF samples moved to the negative side.
Abstract
Graphical Abstract
1. Introduction
The flow of wastewater in a sanitary sewer system is comprised of three components, namely the base sanitary flow, groundwater infiltration, and rainfall-derived inflow and infiltration (RDII) during wet-weather flow (WWF). In most sewer systems, some inflow and infiltration (I/I) occur during dry weather flow (DWF); while small amounts of I/I in a sewer system can be tolerated, excessive I/I into capacity-constrained sewer systems may cause sanitary sewer overflows (SSOs) or bypasses [1]. RDII is stormwater entering the sanitary sewer systems in the form of inflow as well as rainfall-derived infiltration. Various sources contribute to the inflow, including direct connections (e.g., roof drains illegally connected to the sanitary sewers), surface runoff through broken manhole covers, or cross-connections between stormwater and sewer pipes [2]. Infiltration refers to groundwater that enters the sanitary sewer system through cracks or leaks in pipe sections, defective joints, and damaged manhole walls. Further, increased volumes of inflow and infiltration can dilute wastewater and change the characteristics of wastewater, which directly decreases pollutant removal efficiency of wastewater treatment plant (WWTP) operations designed for specific influent conditions. Thus, proper maintenance of wastewater collection and conveyance systems is important for effective treatment in WWTPs, prior to discharge into the environment.
Dissolved organic matter (DOM) is found almost in every type of water on earth and plays a key role in a variety of physicochemical processes and functions in aquatic environments [3–4]. For example, it serves as an energy and nutrient source for heterotrophic bacteria and some algae [5]. DOM is involved in photolytic degradation of organic pollutants as well as in pH control of aquatic systems [6]. Speciation, solubility and complexation of trace metals, and transport and fate of nanoparticles and colloids are also affected by DOM [7–8]. The quantitative and qualitative properties of DOMs vary depending on climatologic or hydrologic conditions they are exposed to, as well as their sources [9–11]. Due to the extremely complex nature of DOM, many studies have been performed to elucidate DOM properties in quantitative and qualitative ways, using bulk characterization through fine/elemental scale methods from chromatographic analyses (size exclusion, high-resolution mass spectrometry, etc.) [12–14] to multivariate data analysis (e.g., principal component analysis (PCA), parallel factor analysis (PARAFAC)) [15–17]. Among advanced techniques, fluorescence excitation-emission matrices (EEMs) coupled with PARAFAC modeling allows us to quantitatively track the variations/dynamics in DOM as well as qualitatively differentiate its sources. The unique fluorescence properties of DOM have also been used to assess the effects of anthropogenic inputs (including surface runoff of stormwater or flow into urban sewer systems or urban streams, and studies have shown that fluorescence EEM-PARAFAC could be used to understand the influence of flows and water quality based on source-differentiated DOM [18–19].
Another promising computational method for analyzing and visualizing muti-dimensional data is the self-organizing maps (SOM) [20], which is a two-layered artificial neural network (ANN) consisting of input layer and output layer. SOM, an unsupervised machine learning technique, creates a low dimensional clustering and classification map (also called Kohonen’s map) from high dimensional data [19–20].
In the present study, we coupled fluorescence EEMs obtained from multiple urban sanitary sewer points undergoing I/I with PARAFAC models to understand DOM quality under DWF and WWF conditions. The objective of this study was to assess the changes in DOM properties occurring due to RDII in urban sewer networks using EEM-PARAFAC modeling and unsupervised machine learning self-organizing map.
2. Materials and Methods
2.1. Sampling Site
Sampling sites used in the present study are in Yangju city of Kyunggi province, South Korea, and cover a catchment area of 36.80 km2 with the service population of ~28,000. In this area, sewage generated is transported to the downstream WWTP with a capacity of 13,000 m3/day and treated via the DeNiPho processes. Sewage drainage in these regions was into combined sewers until 2006, when the sewers were completely replaced with separate sewer systems (with a total length of 30,038 m for sanitary sewers) to minimize I/I under dry-weather conditions. However, recent sewer inspection and flow monitoring revealed several major problems including defective pipes, false connections in pipe joints, improper manhole covers, cracks in deteriorated pipes, causing the amount of I/I to gradually increase leading to sanitary sewer overflows (SSOs). We chose a total of fifteen sampling points (Fig. S1 and Table S1), and the flow and water quality were monitored in both DWF and WWF conditions. As shown in Table S1, the studied sites have mixed land use, with a combination of residential, commercial, and industrial development.
2.2. Sewage Flowrate Measurement and Sampling
Flow rates at each point were measured once every 10 min. Data processing was performed for missing values and outliers, and the validated data were used for estimating the amount of I/I as well as average daily sewage. Fig. S2 shows the diurnal flowrate variations on the sampling dates, which were moving averaged (window size of 5) to filter noises out. At fifteen sampling points, grab samplings were conducted once every two hours for 24 h in November 2017 for dry-weather conditions and similarly in early March, 2018 for wet-weather conditions, resulting in a total of 387 samples. On the day of wet-weather flow, the rainfall intensity at the study site was 17 mm/h and the rainfall lasted for 12 h. At each sampling point, samplers collected two liters of water manually from the manhole connected to the sewers into prewashed plastic containers. Samples were kept in an ice-chest box (4°C) and transferred to the laboratory. In the laboratory, the samples were immediately filtered using a 0.45-μm membrane for including UV/Vis scanning and measurement of dissolved organic carbon (DOC) and fluorescence further analyses.
2.3. Optical Indices
To investigate the properties of DOM, several indicators from the UV/Vis scanning spectra and EEMs have been developed. Using the UV/Vis spectrum, the specific UV absorbance at 254 nm (SUVA254) has been determined as the ratio of the absorbance at 254 nm (UV254) to the DOC concentration and is reported to have a positive correlation with aromaticity and molecular weight of DOM [21]. Based on the UV/Vis spectrum, absorption coefficients were calculated using the following Eq. (1).
where a(λ) represents the Napierian absorption coefficient (m−1) at wavelength λ, Abs(λ) is absorbance value at wavelength λ, and L is the path length of a cuvette (0.01 m).
The absorption ratio at 250 and 365 nm (E2/E3) was used as an indicator of relative molecular size of DOM, with a decreasing ratio indicating increasing molecular size, because light absorption by high molecular weight OM at longer wavelengths (i.e., at 365 nm) becomes stronger as molecular size increases [13,22]. The magnitude of spectral slopes, including S275–295 and S350–400, have been found to be associated with DOM modifications such as molecular weight and DOM aromaticity [22–23]. Typically, the higher values of S275–295 and S350–400 indicate low molecular weight (LMW) material and/or decreasing aromaticity [22,24]. Spectral slope ratio (SR) [22] was calculated from the ratio of the spectral slopes of the absorbance between 275–295 nm (S275–295) and 350–400 nm (S350–400), using nonlinear regression in MATLAB 2018b (Mathworks, USA).
Fluorescence index (FI) was obtained from the ratio of fluorescence intensity at an emission wavelength of 450 nm with that at an emission wavelength of 500 nm, at an excitation wavelength of 370 nm [25]. Humidification index (HIX) indicates the degree of maturation of DOM [26]; i.e., humidification, which is positively associated with increases in the C/H ratio and the degree of aromaticity, is reflected in fluorescence intensities at longer emission wavelengths [26]. HIX was calculated as the ratio of two spectral regions at emission wavelengths 435–480 nm and 300–345 nm, at an excitation wavelength of 254 nm [26]. Biological/autochthonous index (BIX), also known as the β:α ratio, is used for estimating the degree of biological degradation of DOM. BIX was determined as the ratio of fluorescence intensity at an emission wavelength 380 nm (β peak) with the maximum intensity between the emission wavelengths 420 nm and 435 nm, at an excitation wavelength of 310 nm (α peak) [24,27].
2.4. PARAFAC and SOM Modeling
PARAFAC modeling, which is a multi-way chemometric method applicable to large-scale data organized in third- or higher-order arrays [15]. PARAFAC modeling of a three-way dataset decomposes the data signal into a set of tri-linear terms and a residual array as Eq. (2).
where i=1, …, I; j=1, …, J; k=1, …, K. aif (first mode) is the object score (magnitude of the fluorophore), and bjf (second mode) and ckf (third mode) are the excitation loading and the emission loading, respectively; eijk is the residual and contains the variation not explained by the PARAFAC model [28]. The model components have a direct chemical interpretation in a valid model. In Eq. (2), the parameter aif is directly proportional to the concentration of fth fluorophore in sample i; the vectors bjf and ckf are scaled estimates of the emission and excitation spectra of the fth fluorophore [28].
EEMs of the samples were measured using a Hitachi F-7000 fluorometer (Hitachi, Japan) with a 1-cm quartz cuvette at a constant temperature (20°C). Prior to the measurements, samples were filtered through 0.45-μm membranes, and DOC in the samples was diluted down to ~1 mgC/L to minimize inner filter effects. EEMs were obtained by scanning over the range of 230 to 450 nm excitation wavelength at 5-nm intervals and 250 to 500 nm emission wavelength at 2-nm intervals. Thus, the dimension of each EEM consisted of 126 × 26 fluorescence intensity. Corrected EEMs were obtained by subtracting the EEM of ultrapure water (resistivity ≥ 18.2 MΩ-cm), followed by normalizing the EEM spectra with the area under the Raman scatter peak at the excitation wavelength of 350 nm measured on the same day as the day of sample measurement.
PARAFAC modeling was performed using the DOMFluor N-way v3.00 toolbox in MATLAB 2018b (MathWorks, USA). A total of 384 samples, with entire dataset consisting of 374 samples × 45 excitations × 126 emissions, was used for the modeling, and a two to six component PARAFAC model was used to evaluate the data to ensure that the optimum number of components was selected. All models were built using non-negativity constraints [15,28]. During PARAFAC modeling, three samples with high leverages were removed from the dataset. The optimum number of components was determined and validated using split-half analysis, core consistency and Tucker’s congruence coefficient [28]. The fractional contribution (i.e., %) of each component was calculated based on the Fmax (Raman units) of each component [15, 29].
SOM modeling was performed for combined data from PARAFAC components and optical indices in MATLAB 2022a (MathWorks, USA) using the SOM Toolbox (http://www.cis.hut.fi/projects/somtoolbox/) [31]. Prior to SOM analysis, the Fmax data of the PARAFAC components has been converted to the percentage of each component. All data were normalized to values in the range of 0 to 1 with unit variance to reduce the concentration effect. According to the minimum mean quantitation error (mqe) and topographic error (tge), the optimum topology of the map was selected to 13ˊ8 (104 nodes) which showed the least values for both mqe and tqe (Fig. S3) [19,31].
2.5. Statistical Analyses
Statistical analyses including two-sample (independent sample) t-test, and PCA were carried out using SPSS Version 26 (IBM Inc., USA). For PCA, the data were standardized by transforming variables to z-scores to minimize perversion due to different scales among variables. The rotation method was Varimax, and components with an eigenvalue >1 were used for data interpretation. The PCA results, referred to as loadings and scores, were obtained using a correlation matrix.
3. Results and Discussion
3.1. Assessment of Inflow and Infiltration
Patterns in diurnal flowrates were evaluated based on flow monitoring data collected from August 2017 to September 2018 to estimate the quantity of inflow and infiltration at each monitoring station. In dry weather, diurnal flowrates exhibited site-specific patterns, which were mirrored in land usage and service population, and were consistent with neighborhood development (e.g., high flowrates in large service areas, and vice versa) (Figs. S2 and S4). Flowrates at most of the monitoring stations near residential areas (except NY-2 and NY-11) showed evident patterns of low flows before daybreak and higher or peak flows during human activity hours (two peak flows at 6–9 AM and 7–10 PM), as shown in Fig. S2. The NY-2 site, located near City Hall and a dense commercial district, displayed relatively erratic patterns of sewage flowrate throughout the day, with increasing flows even at night. The NY-11 site, which had previously been shown to have particularly impaired sewer conditions by CCTV inspection [32], revealed no discernible trends in diurnal flowrates, and therefore ongoing groundwater penetration, depending on the water level was likely to offset flowrate changes.
Based on our previous report [32], infiltrations during the DWFs were estimated to be 2.7% to 18.3% of daily average sewage depending on the sites receiving unknown discharges including groundwater infiltration. The NY-8 and -9 sites showed the two least infiltration percentages (2.7% and 4.4%, respectively), followed by the NY-12 and -10 sites with 5.3% and 5.6% infiltration, respectively. The NY-11 site revealed the greatest infiltration (18.3% of average sewage), and the other remaining sites (NY-1~7, -13, and -14) were also found to have infiltrations of more than 11% of daily sewages. In fact, the contribution of NY-11 to the influent of the downstream WWTP ranged from negligible to ~2.5%; however, this monitoring point was suitable for distinguishing characteristic changes in organic matter due to relatively high infiltration to sewage, while the NY-8 and -9 sites served as kind of control sites for organic matter in urban sewage, because these points were less impacted by infiltration.
Based on flowrate monitoring of WWFs, the RDII volumes were calculated using the RTK method, which is the primary RDII method proposed by the United States Environmental Protection Agency’s (US EPA) stormwater management model (SWMM) [33–34]. The R parameters (the fraction of rainfall volume entering the sewer system as RDII) [1] and the RDII per unit sewer length (m3/m) at each site were also estimated to evaluate the probable impact of RDII at the sites (Table 1 and Figs. S5a−c). As seen from the data, the greatest RDII volume was found in NY-12 followed by that in NY-10 and NY-7, and the least RDII volumes were found in NY-13, NY-14, and NY-4. In terms of the R-value and RDII volume per unit sewer length (m3/m), NY-11 showed the greatest estimates for both calculations, indicating that sewage at the NY-11 was greatly affected by RDII throughout the entire sewer length even though the sewage volume at this site was not significant. The NY-12 showed the second greatest R-value, and a medium RDII value per sewer length (Fig. S5). Because the pipe at NY-12 was the sewer main with the largest capacity, volume increases due to RDII would be expected to significantly impact downstream WWTP performance. Thus, in terms of prioritization of sewer maintenance, NY-12 and NY-11 would need to be repaired as a matter of priority. The NY-8 and -9 sites, where the estimated infiltration % was the least in DWF conditions as mentioned above, were also affected by RDII to a relatively large extent (0.103 and 0.064 m3 of RDII per m of sewer at NY-8 and -9, respectively). In terms of RDII per unit sewer length, NY-3 showed the second largest value (0.27 m3/min), slightly lesser than NY-11 which showed 0.274 m3/m. Overall, the sewershed in the present study was found to receive substantial amounts of RDII depending on the severity of sewer conditions.
Estimated infiltration in DWF and RDII in WWF
3.2. Changes of Optical Indices
In Fig. 1a, the changes of the E2/E3 ratio, an indicator of DOM molecular size were compared in DWFs and WWFs. In DWF conditions, the mean values of E2/E3 varied in the range of 2.998–5.446, and the highest E2/E3 was observed at the NY-11 site which showed the highest influence due to infiltration. This suggests that infiltration at this point contained higher fractions of LMW DOM, thereby causing a shift to smaller DOM molecular sizes. Meanwhile, in WWF conditions, E2/E3 values showed an increasing trend at most sampling sites when compared to the DWF conditions, falling in a range of 4.126–5.255. The highest value was found at NY-5 followed by NY-11 (4.985). These results imply that the RDII changed the spectrum of molecular size distribution in sewage towards LMW substances. In addition, linking to the results from PARAFAC modeling, the C3 may have a strong association with these LMW organic constituents, because the relative fraction of C3 showed a significant increase due to RDII. Both S275–295 and S350–400 values in DWF samples were significantly higher than those in WWF samples, with S275–295 values (p < 0.00) in the range of 0.0116–0.0210 nm−1 and 0.0066–0.0181 nm−1 in DWF samples and WWF samples, respectively (Figs. 1b–c). The S350–400 values (p < 0.00) were in the range of 0.0069–0.0153 nm−1 and 0.0065–0.0185 nm−1 in DWF samples and WWF samples, respectively. Overall, the mean S275–295 and S350–400 values were 0.0152 nm−1 and 0.0107 nm−1 in DWF samples, and 0.0142 nm−1 and 0.0138 nm−1 in WWF samples.
Comparisons of (a) E2/E3, (b)–(c) spectral slopes (S275–295 and S350–400), and (d) the ratio of spectral slope (SR) in DWF and WWF conditions; red and blue boxes are DWF and WWF, respectively.
The spectral ratio SR (Fig. 1d) has been also linked to shifts in DOM molecular weight, specifically, showing a negative correlation; thus, a higher SR is an indicator of LMW [22]. In the present study, the SR ranged between 0.949 and 1.910 (mean: 1.412), and between 0.410 and 1.920 (mean: 1.078) in DWF and WWF samples, respectively. Lower SR in the WWF samples implies relatively increased presence of high molecular weight DOM compared to DWF samples, and this may be attributable to the substantial input of high molecular weight DOM of terrestrial origin, such as humic substances, due to the RDII [16].
The FI has been used as one of indicators to distinguish the autochthonous (microbially-derived) DOM from allochthonous (terrestrially sourced) DOM [25,30]. Higher FIs indicate increased autochthonous DOM while lower values mean dominance of allochthonous DOM. In previous study, for the standard reference NOMs (e.g., Suwanee River humic, fulvic acid, and NOM) derived from terrestrial sources, FIs showed values < 1, and as for the microbially-derived DOMs such as algogenic OMs or soluble microbial products, the values were > 1.4 [35]. High FIs exceeding 1.4 were observed in leachates in cyanobacterial intracellular organic matter (IOM) [36] and wastewater effluent organic matter (EfOM) [37–38]. FIs in natural waters were reported to be in the range of 1.2–1.8 [38–41]. In the present study, FIs varied in the range of 1.13–1.80 (mean: 1.63) and 1.19–1.64 (mean: 1.47) in DWF and WWF samples, respectively, as shown in Fig. 2a. Decreased FIs in the WWF samples are attributed to the RDII, resulting in a shift of DOM in sewage, especially towards predominance of allochthonous OM. This difference between DWF and WWF samples was also statistically significant by the two-sample t-tests (p < 0.00). The greatest change in the value between DWF and WWF conditions was found at the NY-11 site where the flow impacts due to groundwater infiltration and RDII were observed to be the greatest among all the sampling sites.
Comparisons of optical indices ((a) FI, (b) HIX, and (c) BIX) extracted from fluorescence EEMs in DWF and WWF conditions; red and blue boxes are DWF and WWF, respectively.
The HIX, an indicator of DOM humidification, was compared (Fig. 2b). The mean values of HIX were found to be in the range of 0.56–1.04 and 0.32–0.72 for DWF and WWF samples, respectively. In every sampling point, the HIXs in WWF samples showed a decreasing trend compared to DWF samples, indicating that organic matter in the WWF samples was composed of relatively fresh and less mature constituents with fluorophores than that in the DWF samples. Statistical test also supported that flow conditions of DWF and WWFs contributed to the differences of HIXs (p < 0.05). As for the BIX values (Fig. 2c), WWF samples showed higher values than DWF samples at the majority of sampling sites (except NY-11 and NY-14), with the means of 0.74–0.87 and 0.65–0.92 for DWF and WWF samples, respectively. In correlation analysis with other parameters, its value showed moderately positive relations with FI and S350-400 (r > 0.6, p < 0.01), especially in WWF samples (data not shown). It was reported that higher BIX values suggest that the presence of autochthonous or fresh organic matter [10] and the hydrologic conditions (e.g., the flowrates for storm periods) can be also a driver for DOM status. In this study, even though the effects on its values due to increased flowrates during the WWF event were not observed, it was assumed that increased inputs of organic matters with low molecular size and autochthonous production would be attributed to the changed BIX values.
3.3. EEM-PARAFAC Analysis of the DOM
Using PARAFAC modeling, four components were identified from the dataset consisting of DWF and WWF samples (Table 2 and Fig. 3). All the components identified have also been widely observed in other studies. One interesting observation was that the contours and spectral patterns of four components revealed multiple excitation wavelengths and one emission wavelength. Ideally, one single organic fluorophore would have one excitation and emission maxima; however, due to the complexity of natural organic matter (NOM) or DOM influenced by structures, constituents, and functional groups of humic substances and amino acids (free and bound in proteins), it is more likely to comprise a group of fluorophores [40]. The other noticeable thing in the four PARAFAC components was that all components had their excitation wavelength maxima within or lower than the UV-B (280–315 nm) range, suggesting that the DOM in the sewage in the present study consisted of relatively light-resistant/refractory constituents [43].
Spectral characteristics of the four components identified using PARAFAC and comparisons to previously reported findings.
Contours of the four components identified, and their excitation and emission loadings.
Component 1 (C1) had excitation (Ex) maxima at 230 and 275 nm and an emission (Em) wavelength of 326 nm. The spectral location and shape of C1 resembled that of the amino acid tryptophan [44]. Component 2 (C2) showed its maxima at Ex/Em ratios of 230/426 and 315/426 nm, and its spectral pattern resembled those of humic-like substances in terrestrial sources. Components analogous to C2 have also been frequently reported by previous studies [15,45]. Component 3 (C3) and component 4 (C4) had their maxima peaks in similar regions, showing Ex/Em maxima at <230 and 275/354 nm for C3 and at <230/360 nm for C4, respectively. However, split-half analysis and component number validation clearly separated them into different components. C3 resembled amino acids from autochthonous sources [46–47] and the spectral locations of its Ex/Em were similar to those of bovine serum albumin, SMP, and EfOM isolates [44]. C4 is also thought to be originated from autochthonous fluorophores, showing spectral patterns similar to those of fulvic acids associated with microbial production and/or reworking [47].
3.4. Component Variation by Flow Conditions
Fig. 4a shows fractional changes in the four components in dry- and wet weather conditions. Fig. 4b shows the fractional differences at each site, determined by subtracting the fractions of each component in WWFs from those in DWFs; thus, the plus values indicate decreases in fractional values in WWFs compared to those in DWFs, and the negative values indicate vice versa. Between the two flow conditions, the most noticeable changes were observed in C3 and C4, with increased fractions of the former (range: 12.2–34.2%), and decreased fractions of the latter in WWFs (range: 11.7–36.5%). This suggests that C3 is the representative fluorophore input due to RDII, and that C4 is one of the fluorophores typically found in sewage. C1, except at a few sites (NY-4, -8, and -15), showed slight increases at most of the sampling sites, and C2 showed decreased fractions in WWFs at all sites over a range of 0.9–10.1%. Two-sample t-tests of the mean values showed that fractional changes in the four components were statistically significant in the entire sample set (p < 0.05) and at every site (p < 0.05), implying that flow conditions influenced changes in DOM.
(a) Fractional changes in the four components identified in dry- and wet-weather conditions; (left) DWF, (right) WWF and (b) Fractional differences between DWF and WWF. The error bar indicates standard deviation.
3.5. PARAFAC-SOM Analysis
SOM analysis using the fractional compositions of PARAFAC components was performed to visually analyze the variances or similarities of sample distribution in the neuron generated by RDII event. The unified distance matrix (U-matrix), an output of SOM analysis trained on the 384-sample data (Fig. 5), depicts the distance between prototypes of fractional shares of neighboring neurons using color map units. As seen in Fig. 5a, uneven distances between neighbors were observed, indicating considerable dissimilarity in compositional pattern. The lower-middle portion of the map was darker and bluer than the upper portion, showing a significant difference between the upper and lower portions. Strong yellowish colors at the upper-left and upper-right corners indicate that the samples in these regions differed greatly from their neighbors. On the other hand, samples in the lower section of the U-matrix displayed relatively dark blue over a wide range, implying a high degree of similarity among samples in that region. The U-matrix can be used to assume the clustering of samples based on their gathering with similar color. However, the boundaries between clusters were not clearly displayed in the U-matrix, thus, clustering based on the Euclidean distance were identified for the entire samples into five different groups (clusters I, II, III, IV, and V) (Fig. 5b, and a dendrogram with node numbers in Fig. S6). As shown in Figs. 5b and c, by assigning the sample name to its closest output neuron (i.e., best-matching unit, BMU) on the map (13 × 8 topology), each sample in the original data was partitioned to the cluster. Clusters I–III were largely represented by samples of WWFs drawn in blue, with sample names beginning with the letter ‘w’, whereas clusters IV and V featured examples of DWFs beginning with the letter ‘d’, depicted in red. The dendrogram (Fig. S6), a hierarchical tree of clusters with the node number of the SOM, revealed that clusters III and IV had the greatest distance (i.e., the least similarity) between clusters whereas clusters II and III had the smallest distance (i.e., the highest similarity). The numbers marked and the size of the colored hexagons filled in the neurons (Fig. 5c) mean the number (i.e., hit) of samples falling into the winning neuron, thus, the neurons with high hits would reflect more representative features of the overall clusters. For example, in cluster I, node 3 with twelve hits showed the dominance of C1 while node 11 with nine hits presented substantially increased fraction of C4. Thus, the vertical direction from top to bottom is likely to differentiate by decreased fraction of C1 and increased fraction of C4 from top to bottom. Meanwhile, the optical characteristics extracted from nodes 97 and 9, representing clusters III and IV, respectively, with the greatest dissimilarity, were compared. The most distinguishable discrepancies between nodes 97 and 9 were SR and HIX, showing both optical indices in node 9 were much greater than those in node 97 (Fig. S7). High SR is related to DOM with LMWs, thus lower SR in samples of WWFs may indicate the inflow of DOM with high MWs. HIX is an indicator of DOM humidification. Higher HIX in node 97 than node 9 implies that DOM in WWFs is young and fresh, labile. As a result, the horizontal orientation from left to right may reflect DOM’s aging features, i.e., more aged (or old) from left to right. Meanwhile, the vertical direction of the SOM is anticipated to be dominated by shifts in major DOM components from C1 to C4, with C4 increasing and C1 decreasing in the upper to lower direction.
U-Matrix (a), Best-matching unit (BMU) with clustering (b), hit histogram (c) for the SOM using fractional compositions of PARAFAC components for DWFs and WWFs samples.
3.6. Principal Component Analysis
To identify factors responsible for changes of DOM quality in sewage due to excessive groundwater infiltration and RDII, PCA was applied using %Fmax of four components (fraction of each component in sample) along with six DOM indices (S275-290, S350-400, SR, FI, HIX, and BIX). PCA was conducted on a correlation matrix of the z-score variables (i.e., mean=0, variance=1), and three principal components (PCs) (76.67% of the explaining variance of the original data) were extracted based on the criteria above an eigenvalue of 1. However, for simplicity of interpretation, only the first two PCs (PC1 and PC2) were used, accounting for 62.93% of the variance. PCA loading and score plots presented in Figs. 6a and b show that the samples were separated due to the combined effects of PC1 and PC2, which explained 41.70% and 21.23% of the entire dataset, respectively. As shown in Fig. 6a, SR, FI, HIX and %C4 showed relatively high/positive loadings in PC1 with weak loadings of S275–295, %C1, and %C2, while %C3 and S350–400 had negative loadings in PC1. SR and S275-295 were found to be negatively related to DOM molecular size. FIs positively reflected microbial-derived organic matter. Thus, PC1 is likely to be associated with LMW-OM regardless of source. On the other hand, PC2 showed the highest positive loading in %C1, followed by HIX, BIX, %C3, and SR. %C2, %C4, S275–295, and S350–400 had negative loadings in PC2. C2 represented humic-like substances of terrestrial origin, and typically presented high aromaticity. C1 and BIX were the parameters showing biological association. Therefore, PC2 may be positively associated with pollutant loading with microbial activity. Fig. 6b shows the score scatter plot representing all the DWF and WWF samples. Samples taken in DWFs were clearly separated from those in WWFs by PC1, with DWFs showing high PC1 scores. This was consistent with high loadings of FIs, %C4, and DOM in DWFs, and although influenced by GWI, there were dominated by protein-like OM. WWFs samples showed scattered scores depending on RDII gradients. In particular, WWF samples at NY-11, showing negative loadings and scores in both PC1 and PC2, were grouped at the left-bottom, which was apparently shifted from the right-bottom. This indicates that RDII changed the overall DOM properties in sewage at the NY-11 site towards DOM absorbing light of lower energy (i.e., longer wavelength).
PCA results. (a) Loadings of DOM indices and (b) scores of DWF and WWF samples in two principal components (PC1 and PC2).
4. Conclusions
In the present study, changes in DOM properties due to RDII were assessed using several optical indices, fluorescence EEMs coupled with PARAFAC and SOM modeling. The key findings are summarized as follows:
Analysis of diurnal flow rates under both DWF and WWF conditions revealed a distinctive flow pattern at the study site (NY-11), indicating significant influence from groundwater infiltration and RDII during both flow conditions.
Indices derived from UV scanning of samples, including E2/E3, S350–400, and SR, demonstrated that RDII induces a shift in DOM constituents in sewage towards an increased fraction of LMW substances. This shift is evidenced by elevated E2/E3 and S350–400 values and a decrease in SR.
Under WWF conditions, FI and HIX values decreased, while changes in BIX were site-dependent.
The PARAFAC analysis identified four characteristic fluorophores within DOMs. Notably, C3 was identified as a potential representative influenced by RDII, whereas C4 originated from sewage. The PARAFAC-SOM modeling further demonstrated that water samples in the urban sewer system could be categorized based on the dominance of specific DOM properties.
The PCA results, utilizing both PARAFAC components and optical indices, revealed distinct separation of samples into DWF and WWF conditions. This observation underscores significant alterations in DOM properties induced by RDII. PCA showed that PC1 was associated with molecular size, while PC2 was linked to microbial activity in WWF samples.
The study’s findings indicate that the intrusion of external waters (infiltration and inflow) into the sewer networks can have a negative impact on both the quantity and quality of wastewater. Increases in sewage volume induced by RDII have been connected to poor performance of receiving wastewater treatment plants (WWTPs) as well as reductions in sewers’ effective hydraulic capacity. Deteriorated WWTP performance, which also leads to increased operation costs for utilities, can be attributed to either dilution of raw sewage or distinctive changes in organic constituents that deviate from the WWTP’s initial design parameters.
These findings enhance our understanding of how RDII affects DOM characteristics in urban sanitary sewers, providing valuable insights for water quality management and infrastructure planning. The integration of advanced chemometric and machine learning approaches, such as PARAFAC and SOM modeling, proved to be effective in characterizing and distinguishing DOM properties under varying flow conditions.
Supplementary Information
Acknowledgements
This work was supported by the Chung-Ang University research grant in 2020, and this research was also supported by the Chung-Ang University Graduate Research Scholarship (2018).
Notes
Author Contributions
S.-N.N. (Assistant Professor) conceptualized the methodology, performed the data analysis on PARAFAC modeling, and wrote the manuscript. S.L. (M.S. student) performed the experiments and analyzed the samples. J.O. (Professor) acquired the funding and finalized the submitted version.
Conflict-of-Interest Statement
The authors declare that they have no conflict of interest.