ARTICLE IN PRESS WAT E R R E S E A R C H 42 (2008) 3519 – 3527 Available at www.sciencedirect.com journal homepage: www.elsevier.com/locate/watres Assessment of the nutrient removal performance in integrated constructed wetlands with the self-organizing map Liang Zhanga,b, Miklas Scholzb,, Atif Mustafab, Rory Harringtonc a School of Resource and Environmental Science, Wuhan University, Wuhan 430079, Hubei Province, People’s Republic of China Institute for Infrastructure and Environment, School of Engineering and Electronics, William Rankine Building, The King’s Buildings, The University of Edinburgh, Edinburgh EH9 3JL, Scotland, UK c Water Services, National Parks and Wildlife, Department of Environment, Heritage and Local Government, The Quay, Waterford, Ireland b art i cle info ab st rac t Article history: The self-organizing map (SOM) model was applied to predict outflow nutrient concentra- Received 29 February 2008 tions for integrated constructed wetlands (ICWs) treating farmyard runoff. The SOM Received in revised form showed that the outflow ammonia-nitrogen concentrations were strongly correlated with 22 April 2008 water temperature and salt concentrations, indicating that ammonia-nitrogen removal is Accepted 27 April 2008 effective at low salt concentrations and comparatively high temperatures in ICWs. Soluble Available online 16 May 2008 reactive phosphorus removal was predominantly affected by salt and dissolved oxygen Keywords: Ammonia-nitrogen Farmyard runoff Modeling Self-organizing map Chloride Soluble reactive phosphorus concentrations. In addition, pH and temperature were weakly correlated with soluble reactive phosphorus removal, suggesting that soluble reactive phosphorus was easily removed within ICWs, if salt concentrations were low, and dissolved oxygen, temperature and pH values were high. The SOM model performed very well in predicting the nutrient concentrations with water quality variables such as temperature, conductivity and dissolved oxygen, which can be measured cost-effectively. The results indicate that the SOM model was an appropriate approach to monitor wastewater treatment processes in ICWs. & 2008 Elsevier Ltd. All rights reserved. 1. Introduction Constructed wetlands are often used as artificial wastewater treatment systems usually composed of one or more treatment cells, which are planted with aquatic vegetation such as macrophytes (US EPA, 2000). They are used to treat many types of wastewaters including urban runoff, municipal and industrial wastewaters, agricultural runoff and wastewater, and acid mine drainage (US EPA, 2000; Scholz, 2006). Constructed wetlands are usually efficient in reducing chemical oxygen demand, biochemical oxygen demand and suspended solids, but the corresponding removal efficiencies for nitrogen and phosphorus are often low (US EPA, 2000; Vymazal, 2007). Nitrogen and phosphorus are considered to be the most important nutrients causing water pollution. Nitrogen has an intricate biogeochemical cycle with various biotic and abiotic transformations. The important inorganic forms of nitrogen in wetlands are ammonia-nitrogen (NH+4 -N), nitrate-nitrogen (NO–3-N) and nitrite-nitrogen (NO–2-N) according to Vymazal (2007). Phosphorus occurs predominantly as phosphate in natural waters and wastewater. Phosphates are classified as ortho-phosphate, condensed (pyro-, meta- and Corresponding author. Tel.: +44 131 6506780; fax: +44 131 6506554. E-mail addresses: whu7733@yahoo.com.cn (L. Zhang), m.scholz@ed.ac.uk (M. Scholz), a.mustafa@ed.ac.uk (A. Mustafa), rory_harrington@environ.ie (R. Harrington). 0043-1354/$ - see front matter & 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.watres.2008.04.027 ARTICLE IN PRESS 3520 WAT E R R E S E A R C H 42 (2008) 3519– 3527 poly-) phosphates and organically bound phosphate (US EPA, 2000). The treatment mechanisms and processes within constructed wetlands are highly complex, and include microbial, biological, physical and chemical processes that may occur sequentially or simultaneously (Hammer and Bastian, 1989; US EPA, 2000; Scholz, 2006; Vymazal, 2007). The processes of nitrogen removal and retention during wastewater treatment in constructed wetlands are manifold, and include ammonia volatilization, nitrification, denitrification, nitrogen fixation, plant and microbial uptake, mineralization (ammonification), nitrate-ammonification, anaerobic ammonia oxidation, ammonia adsorption and burial (Vymazal, 2007). Phosphorus removal and retention mechanisms during wastewater treatment in constructed wetlands include adsorption, desorption, precipitation, plant and microbial uptake, fragmentation, leaching, mineralization, sedimentation (peat accretion) and burial (Hammer and Bastian, 1989; Vymazal, 2007). In comparison to conventional constructed wetlands, which cannot remove significant amounts of nitrogen and phosphorus, integrated constructed wetlands (ICWs) are a more effective type of constructed wetland suitable for nitrogen and phosphorus removal (US EPA, 2000; Harrington and Ryder, 2002; Harrington et al., 2005). The ICW concept promoted by the ICW Initiative of the Irish National Parks and Wildlife service has been described in detail by Scholz et al. (2007). ICWs are free surface flow constructed wetlands, which are based on the holistic use of land to control water quality and include elements of good landscape fit, and biodiversity and habitat enhancement into their design (Carroll et al., 2005). Modeling and predicting treatment processes are significant for elucidating the complex nutrient removal mechanisms, and assessing the corresponding water treatment potential of ICWs. It is necessary to model and predict the nutrient removal processes to optimize the design, operation, management and water quality monitoring strategy of an ICW. The self-organizing map (SOM) is based on an unsupervised neural network algorithm, and has been used to analyze, cluster and model various types of large databases (Kohonen et al., 1996; Lee and Scholz, 2006; Kalteh et al., 2007). Astel et al. (2007) and Scholz (2008) applied SOM models successfully for classification of large water and environmental data sets. The advantages of the SOM algorithm and its classification and visualization ability were exploited. The SOM was used as the first abstraction level in clustering. The original data set was represented using a smaller set of prototype vectors, which allowed efficient use of clustering algorithms to divide the prototypes into groups, and the 2-D grid allowed rough visual presentation and interpretation of the clusters as outlined by Vesanto and Alhoniemi (2000). The SOM model, which has not been as often implemented in water treatment process control strategies in comparison to traditional neural networks, was successfully used for the first time as a prediction tool for heavy metal removal in constructed wetland systems by Lee and Scholz (2006). However, the SOM model has never been applied to model and predict the nitrogen and phosphorus removal efficiencies within constructed wetland systems such as ICWs. The aims of this study were as follows: 1. to assess the farmyard runoff treatment potential in terms of nutrient removal with ICWs; 2. to identify relationships between nutrients and other water quality variables; and 3. to predict the nutrient concentration removal performances with the SOM model using water quality parameters, which are more cost-effective, quicker and easier to measure. 2. Methodology and software 2.1. Case study sites The ICW study presented in this paper relates to 13 ICW systems, which were constructed to treat farmyard runoff and wastewater within the Anne Valley near Waterford in Ireland. The farmyard runoff and waste entering the ICW typically consists of yard and dairy washings, and rainfall on open yard and farmyard roofed areas along with silage (usually only spillages) and manure (occasional droppings) effluents. Construction of the ICW systems began in 2000, and was followed by commissioning in February 2001. Scholz et al. (2007) describe these systems and their catchments in detail. ICWs 3, 9 and 11 were built on dairy farms operated for 50, 55 and 77 cows, respectively. The corresponding wetland sizes were 10,288, 7964 and 7676 m2. ICWs 9 and 11 had four cells while ICW 3 had five cells; all wetland cells had a linear sequential arrangement. The mean ICW size was approximately 1.7 times the size of the farmyard areas. The primary vegetation types planted in the ICW systems were emergent species (helophytes). Fig. 1 shows the representative ICW 11 in winter. 2.2. Data and variables The ICW data were collected by monitoring the inflow and outflow water qualities of all 13 ICW systems for more than six years (August 2001 to December 2007). However, this paper is based on only a fraction of the overall data set to address the aims of this paper. Only data obtained from the representative and typical ICW system sites 3, 9 and 11 (characterized by Scholz et al., 2007) were combined, and subsequently used in this paper, because these systems have linear sequential cell configurations and single influent entry points. In contrast, the other ICWs have either multiple influent entry points and/or parallel treatment cells. All three selected ICW sites are typical farm constructed wetlands (specific application of ICW to treat farmyard runoff), previously defined by Carty et al. (2008). Water samples were analyzed for ammonia-nitrogen, soluble reactive phosphorus (SRP), dissolved oxygen (DO), temperature, pH, chloride and conductivity according to standard methods (Allen, 1974; APHA, 1998). Ammonianitrogen and chloride were determined using automated colorimetry. SRP was determined as molybdate reactive phosphorus with an auto analyzer (Method 2540-D; APHA, 1998). Dissolved oxygen, temperature, pH and conductivity ARTICLE IN PRESS WAT E R R E S E A R C H 42 (2008) 3519 – 3527 3521 conductivity (mS). The corresponding expensive and timeconsuming to measure model output parameters were outflow ammonia-nitrogen (mg/L) or SRP (mg/L). 2.3. Statistical analyses All statistical analyses were performed using the standard software packages Origin 7.0, Matlab 7.0 and Econometrics Views 5.0. Significant differences (usually po0.05, if not stated otherwise) between data sets are indicated where appropriate. 2.4. Self-organizing map The SOM is a neural network model and algorithm that implements a characteristic non-linear projection from the high-dimensional space of sensory or other input signals onto a low-dimensional array of neurons, and has been widely applied for visualization of dimensional systems and data mining (Kohonen et al., 1996). The SOM is a competitive learning neural network and based on unsupervised learning, which means that no human intervention is required during the learning process and that little needs to be known about the characteristics of the input data (Alhoniemi et al., 1999). In the SOM algorithm, the topological relations and the number of the neurons or nodes are fixed from the beginning. Each neuron i is represented by an n-dimensional weight, or model vector mi ¼ [mi1,y,min] (n, dimension of the input vectors). Each neuron contains a weight vector. At the start of the model, the weight vectors are initialized to random values. During the training, the weight vectors are calculated using some distance measure such as the Euclidian distance, which is defined as vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi uX u n (1) Di ¼ t ðxij  mij Þ2 ; i ¼ 1; 2; . . . ; M j¼1 where Di ¼ Euclidian distance between the input vector and the weight vector i; xij ¼ jth element of the current input vector; mij ¼ jth element of the weight vector i; M ¼ number of the neurons in the SOM; and n ¼ dimension of the input vectors. Node c (Eq. (2)), whose weight vector is closest to the input vector, is chosen as the best matching unit (BMU). When the BMU is found, the weight vectors mi are updated. The BMU and its topological neighbors are moved closer to the input vector. The update rule of the weight vector is shown in Eq. (3). Fig. 1 – Representative integrated constructed wetland system 11 in winter 2006: (a) sedimentation tank; (b) site overview; and (c) inlet arrangement to the first ICW cell. jjx  mc jj ¼ minfjjx  mi jjg, where x ¼ input vector; m ¼ weight vector; and jj jj ¼ a distance measure. mi ðt þ 1Þ ¼ mi ðtÞ þ aðtÞhci ðtÞ½xðtÞ  mt ðtÞ, were measured in the field with portable meters. Scholz et al. (2007) provides a detailed description of the water quality analysis, which is beyond the scope of this paper. The inexpensive and easy to measure SOM input water quality variables of the outflow were DO (mg/L), temperature (1C), pH (dimensionless), chloride (mg/L) and (2) (3) where m(t) ¼ weight vector indicating the output unit’s location in the data space at time t; a(t) ¼ learning rate at time t; hci(t) ¼ neighborhood function centered in the winner unit c at time t; and x(t) ¼ input vector drawn from the input data set at time t. After this competitive learning exercise, the clusters corresponding to characteristic features can be shown on ARTICLE IN PRESS 3522 WAT E R R E S E A R C H 42 (2008) 3519– 3527 the map. The quality of the mapping is usually measured with the quantization error and the topographic error. The learning rate and neighborhood radius were set with default values. The default number of neurons was determined by the heuristic equation (4). The ratio between side lengths of the map grid was set to the square root of the ratio of the two highest eigenvalues of the data sample (Vesanto et al., 2000) pffiffiffi M  5 n, (4) where M ¼ number of neurons and n ¼ total number of data samples. A two-dimensional lattice with a map size of M ¼ 14  7 hexagonal units was used for both ammonia-nitrogen and SRP modeling. The final quantization and topographic errors were 8.852 and 0.096, and 6.541 and 0.123 for ammonianitrogen and SRP, respectively. These values were relatively low if compared to the error values with other parameter settings, indicating that the quality of the mappings was relatively good. Since the codebook vectors of the SOM represent the local mean of the input vector, the SOM can be used for the prediction of missing components of an input vector. A prediction can be made by seeking the BMU for a vector with unknown components. The predicted values can be obtained from the BMU. The application of the SOM for prediction purposes is illustrated in Fig. 2. The model is trained using the training data set, which is removed from the vector to predict a set of variables as part of an input vector. The depleted vector is subsequently presented to the SOM to identify its BMU. The values for the missing variables are then obtained by their corresponding values in the BMU (Rustum et al., 2008). Lee and Scholz (2006) had applied a SOM model to elucidate heavy metal removal mechanisms and to predict heavy metal concentrations in experimental constructed wetlands. The results demonstrated that heavy metals could be efficiently estimated by utilizing the SOM model. The SOM toolbox (version 2) for Matlab 7.0 developed by the Laboratory of Computer and Information Science at Helsinki University of Technology was used in this study. The toolbox is available online at http://www.cis.hut.fi/projects/ somtoolbox (Vesanto et al., 1999). The SOM model was applied for ammonia-nitrogen and SRP removal data to better understand the corresponding removal mechanisms in ICWs. 3. Results and discussion 3.1. Overall performance Mean inflow and outflow concentrations of ammonia-nitrogen and SRP are presented in Table 1. The reduction rates of ammonia-nitrogen and SRP were highest in 2001 because the ICW systems were young, and therefore not matured. The reduction rates were higher in the first three years compared to those in the following four years. Nevertheless, the ICW had a good treatment capacity for ammonia-nitrogen and SRP during a period of more than six years with removal efficiencies ranging between 97.4% and 99.2%, and between 82.6% and 95.8%, respectively. In comparison, the reduction rates of ammonia-nitrogen and SRP were higher compared to those of a constructed wetland with horizontal sub-surface flow (Kyambadde et al., 2005): between 45.5% and 68.6%, and between 45.2% and 73.5%, respectively. The relationships between the inflow and removed concentrations of ammonia-nitrogen and SRP are shown in Table 2. The removal performances for ammonia-nitrogen were more stable in comparison to those for SRP. This corresponded well with the higher reduction rates of ammonia-nitrogen in comparison to those for SRP. Detailed water quality results have been published by Scholz et al. (2007). Concerning the selected ICW systems 3, 9 and 11, the influent concentrations for ammonia-nitrogen, SRP, DO, temperature, pH, chloride and conductivity were 36.10 mg/L, 11.14 mg/L, 5.5 mg/L, 13.7 1C, 7.21, 107.9 mg/L and 994 mS/cm, respectively, between 2001 and 2007. The corresponding outflow concentrations were 0.59 mg/L, 1.12 mg/L, 5.8 mg/L, 13.1 1C, 7.51, 42.9 mg/L and 358 mS/cm, respectively. 3.2. Application of the SOM model to assess nutrient removal performance The SOM model was applied to identify the relationships between the outflow ammonia-nitrogen concentrations and Best matching unit search Self-organizing map Missing values Best matching unit Predictions Fig. 2 – Predicting missing components of the input vector using the self-organizing map. ARTICLE IN PRESS WAT E R R E S E A R C H 3523 42 (2008) 3519 – 3527 Table 1 – Mean reduction rates of ammonia-nitrogen and soluble reactive phosphorus of integrated constructed wetland systems 3, 9 and 11 Year 2001 2002 2003 2004 2005 2006 2007 Ammonia-nitrogen Soluble reactive phosphorus Sample number Inflow (mg/L) Outflow (mg/L) Reduction rate (%) Sample number Inflow (mg/L) Outflow (mg/L) Reduction rate (%) 12 26 32 54 80 40 70 35.508 38.654 36.482 39.845 46.169 32.766 23.621 0.293 0.404 0.554 0.551 1.192 0.595 0.565 99.2 99.0 98.5 98.6 97.4 98.2 97.6 12 27 29 57 80 32 69 9.071 14.671 11.067 13.007 12.658 9.473 8.014 0.382 0.668 0.964 1.570 1.864 1.645 0.718 95.8 95.4 91.3 87.9 85.3 82.6 91.0 Table 2 – Relationships between inflow concentrations and removed concentrations of ammonia-nitrogen and soluble reactive phosphorus in 2005 (example only) Variables Ammonia-nitrogen (mg/L) Soluble reactive phosphorus (mg/L) Regression equation R2 p value Removed ¼ 0.97  inflow Removed ¼ 0.93  inflow 0.995 0.966 o0.01 o0.01 U-matrix other water quality variables. The component planes for each variable of the SOM model are shown in Fig. 3. The unified distance matrix (U-matrix) representation of the SOM visualizes the distances between the map neurons (Vesanto et al., 1999; Lee and Scholz, 2006). The distances between the neighboring map neurons were calculated, and subsequently visualized by applying gray shade scaling between them; e.g. the lighter gray shades are associated with the high relative component value of the corresponding weight vector. This helps to identify and subsequently illustratively show the clusters in the input data. The component plane shows the value of the variable in each map unit (Lee and Scholz, 2006). The component plane helps to visualize the relationships between ammonia-nitrogen and other variables. High outflow ammonia-nitrogen concentrations (41.951 mg/L) are linked to high chloride concentrations (448 mg/L), high conductivity values (4391 mS) and low temperatures (o13.1 1C). Ammonianitrogen concentrations do not reveal an obvious association with DO concentrations and pH. Low reduction rates are apparently associated with high outflow ammonia-nitrogen concentrations as shown in Table 1. High levels of conductivity and chloride represent high salt concentrations in the runoff. The linear relationship between effluent conductivity and chloride concentration is shown in Eq. (5). Furthermore, Eqs. (6)–(8) show regression equations for ammonia-nitrogen. It can be seen that ammonia-nitrogen removal was influenced by high salt concentrations: Conductivity ¼ 3:79  chloride þ 209:7, R2 ¼ 0:44 and po0:01 Ammonia-nitrogen ¼ 0:08  chloride  2:6, 2 R ¼ 0:15 and po0:01 DO 8.6 17.6 15.1 6.5 13.1 0.8 pH Temperature 29.3 7.8 d Chloride 61 7.5 d Ammonia-nitrogen 7.2 4.4 d Conductivity 34 391 286 d 3.702 1.951 d 0.195 Fig. 3 – Abstract visualization of the relationships between outflow ammonia-nitrogen (mg/L), and outflow dissolved oxygen (DO, mg/L), temperature (1C), pH (dimensionless), chloride (mg/L) and conductivity (lS) using a self-organizing map model. Ammonia-nitrogen ¼ 0:02  conductivity  4:9, (5) 496 48 d 8.6 R2 ¼ 0:18 and po0:01 (7) Ammonia-nitrogen ¼ 0:04  chloride (6) 2 þ 0:01  conductivity  4:9, R ¼ 0:20 and po0:01 ðchloride and conductivityÞ (8) ARTICLE IN PRESS 3524 WAT E R R E S E A R C H 42 (2008) 3519– 3527 Chapanova et al. (2007) demonstrated that ammonia conversion is sensitive to the salinity of the wastewater to be treated; after adding salinity to the input wastewater, ammonia degradation was markedly reduced. However, Dincer and Kargi (1999) showed that salt concentrations 42% resulted in significant reductions in performances of both nitrification and denitrification. In contrast, the outflow temperature is negatively correlated (R ¼ 0.38) with the ammonia-nitrogen concentration, suggesting that temperature had a positive effect on the ammonia-nitrogen removal. The elevated water temperature can enhance nitrate volatilization. Relatively high temperatures (413.1 1C) are better for both nitrification and denitrification, if compared to temperatures p13.1 1C (US EPA, 2000). Chapanova et al. (2007) reported that at 5 1C the ammonianitrogen removal rate was on an average three to five times lower than at temperatures between 15 and 25 1C. Nitrification is greatly affected by temperature; nitrification rates are slow in cold compared to warm climates (Chapanova et al., 2007; Vymazal, 2007). No obvious correlation (R ¼ 0.10) between pH and ammonia-nitrogen could be identified (Fig. 3). Most outflow pH values were in the range between 7.0 and 8.0 at temperatures o17.6 1C. Ammonia-nitrogen concentrations did not reduce at this pH range. Ammonia-nitrogen may be found in the unionized form (NH3) or ionized form (NH+4 ) depending on water temperature and pH. The ionized form is predominant in wetlands; e.g. at pH 7.0 and 25 1C, the percentage of unionized ammonia is approximately 0.6% (US EPA, 2000). It was also reported that at high pH ranging between 8.0 and 8.5, the proportion of ammonia might increase to between 20 and 25% at 20 1C, if surface turbulence is high due to wind action. Significant losses of nitrogen may occur in the open water areas via ammonia gas (NH3) volatilization (US EPA, 2000; Camargo Valero and Mara, 2007). Many papers (Schaafsma et al., 1999; US EPA, 2000; Noorvee et al., 2007; Iamchaturapatr et al., 2007) indicate that DO significantly influences the removal rate of ammonia-nitrogen in constructed wetland systems. However, the DO concentrations had no obvious impact on ammonia-nitrogen removal in ICW based on the visualization of the relationship between outflow ammonia-nitrogen and DO (R ¼ 0.02). Therefore, it can be seen that ammonia-nitrogen removal was largely influenced by salt concentrations and temperature. The variables pH and DO seemed to be of less importance. The visualization of relationships between the outflow SRP concentrations and other water quality parameters of the SOM model is shown in Fig. 4. High outflow SRP concentrations (42.641 mg/L) are linked to high chloride concentrations (449 mg/L) and high conductivity values (4394 mS). Unlike the case of ammonia-nitrogen removal, SRP removal was largely influenced by the DO and salt concentrations, and correlated comparatively weakly with temperature and pH. Chloride (R ¼ 0.48) and conductivity (R ¼ 0.56) correlated positively with SRP, indicating that elevated salt concentrations had a negative impact on SRP removal. With increasing salt concentrations, the phosphorus removal rates of the tested ICW systems decreased. This was probably the case because phosphate-accumulating microorganisms were U-matrix 36.9 DO 9.7 19.7 d 13.2 3.9 d Chloride d 18.5 6.8 2.5 pH Temperature 7.8 Conductivity 7.9 64 516 7.5 49 394 7.1 d 33 d 272 SRP 4.668 2.641 d 0.621 Fig. 4 – Abstract visualization of the relationships between outflow soluble reactive phosphorus (SRP, mg/L), and outflow dissolved oxygen (DO, mg/L), temperature (1C), pH (dimensionless), chloride (mg/L) and conductivity (lS) using a self-organizing map model. sensitive to salinity (Scholz, 2006). The salt accumulation in phosphate-accumulating microorganism cells might have reached a certain threshold indicative of a significant increase in osmotic pressure in cells. Diminished phosphate accumulation capabilities subsequently result in reduced removal efficiencies as discussed previously by Carucci et al. (1997), Panswad and Anan (1999) and Wang et al. (2007). In contrast, DO is negatively correlated (R ¼ 0.46) with SRP indicating that high DO concentrations had positive effects on SRP removal. Dissolved oxygen is an important variable influencing phosphorus removal in ICWs. Case studies undertaken by Girija et al. (2007) revealed that the phosphorus concentrations decreased from 6.0 to 0.1 mg/L as DO concentrations increased from 0.1 to 8.6 mg/L. Low DO concentrations can cause the release of phosphorus from the sediment, causing an increase of SRP (Golterman, 1995; Maine et al., 2007). Furthermore, Wang et al. (2007) reported that phosphorus concentrations between 0.22 and 1.79 mg/L within a biological reactor effluent could be obtained when the corresponding influent phosphorus concentration ranged between 15 and 20 mg/L. The DO was controlled at 3.070.2 mg/L during the aerobic phase and pH was maintained at 7.070.1. Phosphorus removal of 90% was achieved in the reactor. Phosphorous might precipitate as calcium phosphate or coprecipitate with iron colloids or with calcium carbonate (Golterman, 1995). For example, the US EPA (2000) reported that phosphorus might precipitate as calcium phosphate within sediment pore water or in the water column near active phytoplankton growth at pH values 47.0. Furthermore, ARTICLE IN PRESS WAT E R R E S E A R C H 3525 42 (2008) 3519 – 3527 as pH decreases, SRP sorption to carbonates decreases while adsorption to iron increases (Golterman, 1995). Concerning the ICW study, the negative correlation (R ¼ 0.16) between pH and SRP is weak, indicating that a high pH had a small positive influence on SRP removal. Since overall pH values were comparatively low (Fig. 4), the influence of pH on SRP removal was weak. However, the chemical composition of the three ICW systems and their effluents is complex. It follows that the key precipitation processes cannot be discussed in detail within the scope of this paper. However, a detailed discussion on water quality issues with respect to an earlier directly related study has been published by Scholz et al. (2007). Pietro et al. (2006) observed that phosphorus removal was weakly correlated with water temperature in (USA) freshwater marsh located in the South of Florida. In comparison, high SRP concentrations in ICWs are also associated with low temperatures (R ¼ 0.21). However, the influence of temperature was lower for SRP removal than for ammonia-nitrogen removal. 3.3. Ammonia-nitrogen and SRP predictions with the SOM model The SOM model was applied to predict the ammonia-nitrogen and SRP removal performances of ICWs. Table 3 summarizes the results from a correlation analysis comprising the input variables DO, temperature, pH, chloride and conductivity, and the target variables ammonia-nitrogen and SRP. Findings are in agreement with Figs. 3 and 4 highlighting the key relationships revealed by the SOM. For example, it can be seen that ammonia-nitrogen concentrations were highly correlated with temperature, chloride and conductivity. In comparison, SRP concentrations were highly correlated with DO, chloride and conductivity. In general, the measurements of input variables used for prediction should be more cost-effective and time-efficient, and easier in comparison to those of the target variables. Based on this consideration, temperature (R ¼ 0.38) and conductivity (R ¼ 0.43) were selected as input variables to predict ammonia-nitrogen in the outflow of the ICW. In comparison, DO (R ¼ 0.46) and conductivity (R ¼ 0.56) were selected as input variables to predict SRP. Considering the relatively high costs and long time associated with most chloride measurement techniques, chloride was not selected for predicting both ammonia-nitrogen and SRP concentrations, even though it had comparatively strong relationships with ammonia-nitrogen (R ¼ 0.38) and SRP (R ¼ 0.48). For modeling purposes, the data set was mixed in a random order and then subdivided into two sets. The first subset was used as a training data set, and the second subset was used as a testing data set. Training and test data sets are summarized in Table 4. The model was verified with the test data set. For example, when predicting the ICW treatment performance concerning ammonia-nitrogen removal, the corresponding ammonia-nitrogen data entries were omitted form the test data set, implying that ammonia-nitrogen concentrations were in fact missing values. After running the simulation, the predicted ammonia-nitrogen concentrations were subsequently compared with the actual values. The SOM modeling performances in terms of predicting the outflow ammonianitrogen and SRP concentrations are shown in Fig. 5. Table 4 shows the summary statistics of the SOM model for the test. The SOM model has a comparatively lower mean absolute scaled error indicating its relatively high accuracy in prediction if compared to previous results (Lee and Scholz, 2006). In general, the SOM model performed very well in predicting the nutrient concentrations in representative ICW systems. 4. Conclusions 1. Representative integrated constructed wetlands (ICWs) were very efficient in removing ammonia-nitrogen and soluble reactive phosphorus. Table 4 – Summary statistics of the self-organizing map model applied for prediction purposes Statistics Ammonianitrogen prediction Soluble reactive phosphorus prediction 240 250 74 84 0.934 0.951 0.015 0.048 Number of training data sets Number of test data sets Correlation coefficient Mean absolute scaled error Table 3 – Correlation coefficients and corresponding p values (in brackets) related to a correlation analysis comprising input (column headings) and target (row headings) variables (n ¼ 314 for ammonia-nitrogen; n ¼ 334 for soluble reactive phosphorus) Variables Ammonia–nitrogen (mg/L) Soluble reactive phosphorus (mg/L) Dissolved oxygen (mg/L) Temperature (1C) pH (dimensionless) Chloride (mg/L) Conductivity (mS) 0.016 (0.779) 0.463 (o0.01) 0.376 (o0.01) 0.206 (o0.01) 0.096 (0.088) 0.163 (o0.01) 0.384 (o0.01) 0.477 (o0.01) 0.428 (o0.01) 0.562 (o0.01) ARTICLE IN PRESS 3526 WAT E R R E S E A R C H 42 (2008) 3519– 3527 Acknowledgements Ammonia-nitrogen (mg/L) 5 The authors wish to acknowledge the support received from Fuxing Gan, who is Liang Zhang’s supervisor at Wuhan University. Paul Carroll and Susan Cook (Waterford County Council) supplied most of the data. Funding was provided by the Federation of European Microbiological Societies and the Irish Department of Environment, Heritage and Local Government. Liang Zhang’s overall research at The University of Edinburgh was funded by the China Scholarship Council project ‘‘National Study-abroad Project for Postgraduates of Key Constructed High Level Universities in China in 2007’’. 4 3 Actual Predicted 2 1 0 0 10 20 30 40 50 Data number 60 70 80 8 7 Actual Predicted SRP (mg/L) 6 5 4 3 2 1 0 0 10 20 30 40 50 60 Data Number 70 80 90 Fig. 5 – Comparison of the actual and predicted outflow (a) ammonia-nitrogen and (b) soluble reactive phosphorus (SRP) concentrations. 2. The self-organizing map (SOM) showed that the ammonia-nitrogen outflow concentrations correlated with water temperature and the salt concentrations (indicated by conductivity and chloride). High ammonianitrogen removal efficiency can be achieved, if salt concentrations are low and temperatures are high. The SOM model also revealed that soluble reactive phosphorus removal was predominantly affected by salt and dissolved oxygen. Soluble reactive phosphorus can easily be removed within ICWs, if salt concentrations are low, and dissolved oxygen, temperature and pH values are high. 3. The SOM performed very well in modeling and predicting the nutrient removal in the selected ICW. Nutrients such as ammonia-nitrogen and soluble reactive phosphorus can be accurately predicted by other more costeffective, rapid and easier to measure water quality variables such as temperature, conductivity and dissolved oxygen. R E F E R E N C E S Alhoniemi, E., Hollmen, J., Simula, O., Vesanto, J., 1999. Process monitoring and modeling using the self-organizing map. Integr. Comput. Aided Eng. 6 (1), 3–14. Allen, S.E., 1974. Chemical Analysis of Ecological Materials. Blackwell, Oxford, UK. APHA, 1998. Standard methods for the examination of water and wastewater, 20th ed. American Public Health Association (APHA), American Water Works Association and Water and Environmental Federation, Washington, DC, USA. Astel, A., Tsakovski, S., Barbieri, P., Simeonov, V., 2007. Comparison of self-organizing maps classification approach with cluster and principal components analysis for large environmental data sets. Water Res. 41 (19), 4566–4578. Camargo Valero, M.A., Mara, D.D., 2007. Nitrogen removal via ammonia volatilization in maturation ponds. Water Sci. Technol. 55 (11), 87–92. Carroll, P., Harrington, R., Keohane, J., Ryder, C., 2005. Water treatment performance and environmental impact of integrated constructed wetlands in the Anne Valley watershed, Ireland. In: Dunne, E.J., Reddy, K.R., Carton, O.T. (Eds.), Nutrient Management in Agricultural Watersheds: A Wetlands Solution. Wageningen Academic Publishers, Wageningen, The Netherlands. Carty, A., Scholz, M., Heal, K., Gouriveau, F., Mustafa, A., 2008. The universal design, operation and maintenance guidelines for farm constructed wetlands (FCW) in temperate climates. Bioresour. Technol., in press, doi:10.1016/j.biortech. 2008.01.045. Carucci, A., Majone, M., Ramadori, R., Rossetti, S., 1997. Biological phosphorus removal with different organic substrates in an anaerobic/aerobic sequencing batch reactor. Water Sci. Technol. 35 (1), 161–168. Chapanova, G., Jank, M., Schlegel, S., Koeser, H., 2007. Effect of temperature and salinity on the wastewater treatment performance of aerobic submerged fixed bed biofilm reactors. Water Sci. Technol. 55 (8–9), 159–164. Dincer, A.R., Kargi, F., 1999. Salt inhibition of nitrification and denitrification in saline wastewater. Environ. Technol. 20 (11), 1147–1153. Girija, T.R., Mahanta, C., Chandramouli, V., 2007. Water quality assessment of an untreated effluent impacted urban stream: the Bharalu Tributary of the Brahmaputra River, India. Environ. Monit. Assess. 130 (1–3), 221–236. Golterman, H.L., 1995. The labyrinth of nutrient cycles and buffers in wetlands: results based on research in the Camargue (Southern France). Hydrobiologia 315 (1), 39–58. ARTICLE IN PRESS WAT E R R E S E A R C H 42 (2008) 3519 – 3527 Hammer, D.A., Bastian, R.K., 1989. Wetlands ecosystems: natural water purifiers. In: Hammer, D.A. (Ed.), Constructed Wetlands for Wastewater Treatment. Lewis Publishers, Chelsea, MI. Harrington, R., Ryder, C., 2002. The Use of Integrated Constructed Wetlands in the Management of Farmyard Runoff and Waste Water. National Hydrology Seminar on Water Resources Management Sustainable Supply and Demand. The Irish National Committees of the International Hydrological Programme and the International Commission on Irrigation and Drainage, Tullamore, Ireland. Harrington, R., Dunne, E.J., Carroll, P., Keohane, J., Ryder, C., 2005. The concept, design and performance of integrated constructed wetlands for the treatment of farmyard dirty water. In: Dunne, E.J., Reddy, K.R., Carton, O.T. (Eds.), Nutrient Management in Agricultural Watersheds: A Wetlands Solution. Wageningen Academic Publishers, Wageningen, The Netherlands. Iamchaturapatr, J., Yi, S.W., Rhee, J.S., 2007. Nutrient removals by 21 aquatic plants for vertical free surface-flow (VFS) constructed wetland. Ecol. Eng. 29 (3), 287–293. Kalteh, A.M., Hjorth, P., Berndtsson, R., 2007. Review of the selforganizing map (SOM) approach in water resources: analysis, modelling and application. Environ. Model. Softw. (in press). Available online since 19 November 2007. Kohonen, T., Oja, E., Simula, O., Visa, A., Kangas, J., 1996. Engineering applications of the self organizing map. Proc. IEEE 84 (10), 1358–1384. Kyambadde, J., Kansiime, F., Dalhammar, G., 2005. Nitrogen and phosphorus removal in substrate-free pilot constructed wetlands with horizontal surface flow in Uganda. Water Air Soil Pollut. 165 (1–4), 37–59. Lee, B.-H., Scholz, M., 2006. Application of the self-organizing map (SOM) to assess the heavy metal removal performance in experimental constructed wetlands. Water Res. 40, 3367–3374. Maine, M.A., Suñé, N., Hadad, H.R., Sánchez, G., 2007. Temporal and spatial variation of phosphate distribution in the sediment of a free surface water constructed wetland. Sci. Total Environ. 380 (1–3), 75–83. Noorvee, A., Poldvere, E., Mander, Ü., 2007. The effect of preaeration on the purification processes in the long-term performance of a horizontal subsurface flow constructed wetland. Sci. Total Environ. 380 (1–3), 229–236. 3527 Panswad, T., Anan, C., 1999. Impact of high chloride wastewater on an anaerobic/anoxic/aerobic process with and without inoculation of chloride acclimated seeds. Water Res. 33 (5), 1165–1172. Pietro, K.C., Chimney, M.J., Steinman, A.D., 2006. Phosphorus removal by the Ceratophyllum/periphyton complex in a south Florida (USA) freshwater marsh. Ecol. Eng. 27 (4), 290–300. Rustum, R., Adeloye, J.A., Scholz, M., 2008. Applying Kohonen selforganizing map as a software sensor to predict biochemical oxygen demand. Water Environ. Res. 80 (1), 32–40. Schaafsma, J.A., Baldwin, A.H., Streb, C.A., 1999. An evaluation of a constructed wetland to treat wastewater from a dairy farm in Maryland, USA. Ecol. Eng. 14 (1–2), 199–206. Scholz, M., 2006. Wetland Systems to Control Urban Runoff. Elsevier, Amsterdam, The Netherlands. Scholz, M., 2008. Classification of flood retention basins: the Kaiserstuhl case study. Environ. Eng. Geosci. 24 (2) (in press). Scholz, M., Harrington, R., Carroll, P., Mustafa, A., 2007. The integrated constructed wetlands (ICW) concept. Wetlands 27 (2), 337–354. US EPA, 2000. Constructed Wetlands Treatment of Municipal Wastewater. United States (US) Environmental Protection Agency (EPA), Office of Research and Development, Cincinnati, OH, USA. Vesanto, J., Alhoniemi, E., 2000. Clustering of the self-organizing map. IEEE Trans. Neural Networks 11 (3), 586–600. Vesanto, J., Himberg, J., Alhoniemi, E., Parhankangas, J., 1999. Selforganizing map in Matlab: the SOM toolbox. In: Proceedings of the Matlab DSP Conference, November 1999, Espoo, Finland, pp. 34–40. Software available online at /http://www.cis.hut.fi/ projects/somtoolbox/S. Vesanto, J., Himberg, J., Alhoniemi, E., Parhankangas, J., 2000. SOM Toolbox for Matlab 5 Documentation. Helsinki University of Technology, Helsinki, Finland Available online at /http:// www.cis.hut.fi/projects/somtoolbox/S. Vymazal, J., 2007. Removal of nutrients in various types of constructed wetlands. Sci. Total Environ. 380, 48–65. Wang, D., Li, X., Yang, Q., Zeng, G., Liao, D., Zhang, J., 2007. Biological phosphorus removal in sequencing batch reactor with single-stage oxic process. Bioresour. Technol. (in press). Available online since 21 December 2007.