Elsevier

Water Research

Volume 45, Issue 14, August 2011, Pages 4183-4197
Water Research

Advancing assessment and design of stormwater monitoring programs using a self-organizing map: Characterization of trace metal concentration profiles in stormwater runoff

https://doi.org/10.1016/j.watres.2011.05.021Get rights and content

Abstract

Stormwater runoff poses a great challenge to the scientific assessment of the effects of diffuse pollution sources on receiving waters. In this study, a self-organizing map (SOM), a research tool for analyzing specific patterns in a large array of data, was applied to the monitoring data obtained from a stormwater monitoring survey to acquire new insights into stream water quality profiles under different rainfall conditions. The components of the input data vectors used by the SOM included concentrations of 10 metal elements, river discharge, and rainfall amount which were collected at the inlet and endpoint of an urban segment of the Yeongsan River, Korea. From the study, it was found that the SOM displayed significant variability in trace metal concentrations for different monitoring sites and rainfall events, with a greater impact of stormwater runoff on stream water quality at the upstream site than at the downstream site, except under low rainfall conditions (≤4 mm). In addition, the SOM clearly determined the water quality characteristics for “non-storm” and “storm” data, where the parameters nickel and arsenic and the parameters chromium, cadmium, and lead played an important role in reflecting the spatial and temporal water quality, respectively. When the SOM was used to examine the efficacy of stormwater quality monitoring programs, between 34 and 64% of the sample size in the current data set was shown to be sufficient for estimating the stormwater pollutant loads. The observed errors were small, generally being below 10, 6, and 20% for load estimation, map resolution, and clustering accuracy, respectively. Thus, the method recommended may be used to minimize monitoring costs if both the efficiency and accuracy are further determined by examining a large existing data set.

Highlights

► We apply a self-organizing map (SOM) to characterize stormwater quality profiles. ► SOM clearly addresses concentration profiles over space and time. ► The SOM indicates that 34–64% of the samples can be cut from the data set. ► The actual pollutant loads agree well with those estimated from the reduced data set. ► Thus, the SOM may be used to better assess and design stormwater monitoring programs.

Introduction

In recent years, wet weather pollutant discharge (i.e., stormwater pollution) has been given particular emphasis in surface water quality management, due mainly to its various impacts on environmental degradation, including (inland and coastal) water quality impairment, an increase of public health risks, and alterations or destruction of ecosystems (U.S.EPA, 1994, Surbeck et al., 2006, Kim et al., 2007a, Kim et al., 2007b, Lee et al., 2007, Kang et al., 2009, Chun et al., 2010). Primary sources of pollution include all types of pollutants (e.g., suspended solids, nutrients, organic compounds, bacteria, metals, etc.; U.S.EPA, 1994, Grout et al., 1999, Gobel et al., 2007, Kayhanian et al., 2007, Chun et al., 2010), which are diffused by surface runoff from many potential hot spots, particularly in mining zones, highway roads, agricultural lands, and urban areas (Lee et al., 2004, Kim et al., 2007a, Kim et al., 2007b). Depending on the properties of various pollutants, they typically experience—once distributed—multiple fate and transport processes in the environment (e.g., sorption, decay, deposition, etc.), resulting in a wide range of water quality variability in space and time that current stormwater quality monitoring programs or regulations may not fully address (Surbeck et al., 2006, Lee et al., 2007).

The dynamic nature of stormwater quality has been ascribed to a number of factors that may contribute to water quality variation in different ways. Previous studies have shown that the observed pollutant concentration matrix of stormwater was highly sensitive to local environmental conditions, including: i) season (e.g., seasonal first flush, rainfall pattern, antecedent dry day, etc.; Brezonik and Stadelmann, 2002, Lee et al., 2004), ii) location (e.g., land use activities, drainage area, etc.; Kim et al., 2005, Gobel et al., 2007, Kayhanian et al., 2007), iii) water quality conditions (e.g., pH, salinity, suspended solids, etc.; Guećguen and Dominik, 2003, Nguyen et al., 2005), and iv) pollutants of interest (e.g., mass loading or decay rates, composition or morphology of colloids, etc.; Grout et al., 1999, Surbeck et al., 2006). Along with the intrinsic randomness of nature, most of the above studies also showed that many complex interactions among these factors occurred in various dimensions (Kim et al., 2005). In such conditions, while conventional statistical methods, i.e., multivariate (e.g., principal component analysis and multiple linear regression) and uni/bivariate approaches (e.g., correlation and ANOVA), have been routinely employed to describe the relationship between water quality and various influence factors, the results were often prone to various sources of statistical bias unless all the requirements in the analysis (e.g., parametric assumptions, multicollinearity, interaction terms, model selection criteria, etc.; Ge and Frick, 2007, Cho et al., 2009) were rigorously addressed or adjusted. Accordingly, there appears to be an inherent difficulty in assessing the variability of stormwater quality, because all these issues currently hinder a complete understanding of water quality changes and their various effects on water resources in response to storm events.

Differences in the accuracy of estimating stormwater pollutant loads are another hurdle to overcome (Hicks et al., 1997, Leecaster et al., 2002, Lee et al., 2007, Ackerman et al., 2010). An indicator popularly used in computing mass emissions is the event mean concentration (EMC), which allows for a comparison of the contribution of pollutant loads in different local areas over any time period (Kim et al., 2005, Lee et al., 2007). However, the EMCs estimated from different stormwater monitoring programs may not be directly comparable, because various sampling schemes (e.g., grab, flow-weighted, and time-weighted composite samples) significantly affect the variability and uncertainty of the load estimations (Lee et al., 2007). For this reason, the use of EMC in assessing the efficacy of the monitoring programs from different sampling schemes is not desirable under various environmental conditions (e.g., seasonal variation in pollution sources under different antecedent conditions), possibly introducing much error in a variable pollutograph over time (Ackerman et al., 2010). In addition, there is less scientific consensus on substitution of analytical detection limits (DLs) for data evaluation (Hicks et al., 1997, Clarke, 1998), which again lessens the accuracy of EMCs. In the absence of reliable data, a deterministic modeling approach seems to be an excellent candidate for identifying the effect of sampling methods (Ackerman et al., 2010). However, the quality of such an assessment is not always certain, as the performance of the model again depends on the quality of field data, available resources, and the assumptions made. Therefore, designing effective stormwater monitoring programs may be challenging yet greatly rewarding.

Compared to previous research, the present study describes a novel method for analyzing stormwater monitoring data based on the characteristics involved in water quality profiles, thereby providing an advanced understanding of the dynamic nature in stormwater quality. For this study, a self-organizing map (SOM), which was shown to have a wide range of applications in the field of data analysis in complex dimensions (Vesanto, 2002, Bieroza et al., 2009, Bieroza et al., 2010), was applied to a data set that included 10 trace metal elements in an attempt to characterize concentration profiles in stormwater runoff. Specifically, this study used the SOM: i) to investigate relationship between elements and their variability in relation to different locations and rainfall amounts, ii) to identify dominant parameters reflecting spatial and temporal stream water quality, and iii) to develop a tool for assessing and designing stormwater monitoring programs with special needs in reducing non-beneficial samples from the data set. Note that although a few case studies in stormwater monitoring survey are introduced in this study, the results drawn are sufficient for exploring such ideas in question; it is recognized that the effectiveness and accuracy of the recommended methodology can be greatly enhanced by a subsequent analysis of a large existing data set (as a future study). In addition, the study raises a question on basic functions and algorithms implemented in the SOM (e.g., the role of active/inactive neurons, map quality, and clustering accuracy), showing how such tools can be used to address the research questions. From this study, we expect that the findings not only provide valuable information on developing enforceable policies for diffuse source control, but also strongly broaden the potential applications of the SOM in the field of water resource engineering.

Section snippets

Field site description

The Yeongsan River, located in the southwest region of the Korean Peninsula, is home to 1.7 million people and associated ecosystems in the South Jeolla Province. The river stretches over 135 km along the western sector of the province, and the drainage basin includes the territories of three cities and fifteen districts, with a total area of about 3500 km2. In the basin, agriculture and farming mainly dominate the land use, along with forest coverage, and there are many artificial structures

Overview of trace metal concentration

Table 2 shows the mean and coefficient of variation (CV) values for the concentrations of the 10 trace metal elements in dissolved phase and the river discharge at UC and GS during the five monitoring events. In the table, a large CV value was typically observed during precipitation for each metal element, reflecting a great variability in stream water quality for different sampling sites and rainfall amounts. In particular, the water quality at UC during storm events was more variable than at

Conclusions

In the present study, a SOM was applied to a set of monitoring data collected during non-storm and storm events in an attempt to gain new insights into stormwater quality, i.e., its dynamic patterns across time and space and their impacts on description of stormwater pollutant loads. The main conclusions drawn from this study are as follows.

  • Stomwater quality is highly variable depending on the monitoring site and rainfall amount, although only a few storm events were examined in this study.

Acknowledgments

This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (No. 2010-0011822). We also acknowledge the high quality public services provided by the Ministry of Land, Transport, and Maritime Affairs and Meteorological Administration in Korea in supplying the monitoring data.

References (30)

  • H. Lee et al.

    Design of stormwater monitoring programs

    Water Research

    (2007)
  • M.K. Leecaster et al.

    Assessment of efficient sampling designs for urban stormwater monitoring

    Water Research

    (2002)
  • H.L. Nguyen et al.

    Correlations, partitioning and bioaccumulation of heavy metals between different compartments of Lake Balaton

    Science of the Total Environment

    (2005)
  • D. Ackerman et al.

    Evaluating performance of stormwater sampling approaches using a dynamic watershed model

    Environmental Monitoring and Assessment

    (2010)
  • M. Bieroza et al.

    Exploratory analysis of excitation-emission matrix fluorescence spectra with self-organizing maps as a basis for determination of organic matter removal efficiency at water treatment works

    Journal of Geophysical Research-Biogeosciences

    (2009)
  • Cited by (42)

    • A novel approach for extracting functional brain networks involved in mesial temporal lobe epilepsy based on self organizing maps

      2022, Informatics in Medicine Unlocked
      Citation Excerpt :

      Since there is not any rigid boundary between the neurons that define the clusters by SOM, the subset of neurons that are related to the set of clusters cannot be determined [61]. We used the k-means method for clustering SOM results proven to be efficient in such applications [62,63]. There are several methods used for evaluating clusters in previous studies, such as Davies–Bouldin index (DBI) [64], Silhouette [65] and Elbow [66].

    • Assessment of extrinsic and intrinsic influences on water quality variation in subtropical agricultural multipond systems

      2021, Environmental Pollution
      Citation Excerpt :

      Specifically, self-organizing maps (SOM), which belongs to the category of competitive learning networks in artificial neural networks, are effective for classifying and visualizing data patterns in a highly variable environment (Vesanto and Alhoniemi, 2000; Céréghino and Park, 2009). SOM provides a feasible but rarely tested solution to parse MPS water quality variations for associated source attributions with promising applications in correlating agricultural practices and water chemistry (Romić et al., 2020), selecting biological indicators to reflect environmental degradation (Milošević et al., 2016), and optimizing storm water monitoring programs (Ki et al., 2011). Over the past several decades, NFWs, such as MPSs, have faced a severe decline in number and in ecosystem service provisioning owing to expanding urbanization, agricultural intensification, and climatic extremes (Marton et al., 2015; Lischeid et al., 2018).

    View all citing articles on Scopus
    View full text