UPM Institutional Repository

Artificial neural network for water quality assessment and land use pattern recognition in Kinta River catchment


Gazzaz, Nabeel Mohammad (2012) Artificial neural network for water quality assessment and land use pattern recognition in Kinta River catchment. Doctoral thesis, Universiti Putra Malaysia.


Kinta River (the State of Perak, Malaysia) was classified during the period 1997-2006 with an average class III water quality (WQ) and a water quality index (WQI) in the range of 51.9–76.5. Frequency analysis of the water quality class (WQC) revealed that over this time period Kinta River water assumed the WQCs I, II, III, IV, and V for 5.9%, 31.0%, 54.9%, 7.8%, and 0.4% of the time, respectively. From 2004 until recently, the yearly WQ reports of the Department of Environment (Malaysia) classify Kinta River as a slightly polluted river. Therefore, there is growing interest in identifying the main factors, processes, WQVs, and LU classes responsible for this deteriorated WQ and determining how the desired WQ can be secured on a sustainable basis. Therefore, the objectives of this study were to (i) assess the water quality of Kinta River during the period 1997-2006; (ii) recognize patterns and analyze trends in LUs in Kinta River catchment as well as in the river’s WQ; (iii) model these patterns and trends; and (iv) model the relative impacts of eight LU categories (agriculture, animal husbandry, forest, logging, mining, oil palm, rubber, and urban areas) on the WQ of Kinta River using data mining techniques including one major chemometric method (principal factor analysis (PFA)) and non-linear, data-driven (artificial neural network (ANN)) methods. The research adopted a comprehensive approach whereby the study scale encompassed eight sub-catchments in Kinta River basin coinciding with one WQ monitoring station, each, and investigations were carried out at six spatial scales: the whole river basin (WRB) and the 0-500 m, 0-1000 m, 0-1500 m, 0-2000 m, and 0-2500 m buffer zones (BZs)). On the other hand, the temporal scale of this study was ten years; 1997-2006. The PFA highlighted that the latent structure of the WQ data is interpreted in terms of 23 WQVs that sorted in seven factors explaining 74.2% of the total variability in this data. These factors helped in identifying the pollutant origins and the major pollution sources in the basin. A radial basis function (RBF) classification model was created to allow for prediction of the WQC from these 23 WQVs. The optimum classifier obtained had correct classification ratios (CCRs) for the WQCs I-IV ranging from 60.0%-86.7%. The self-organizing feature map (SOFM) illustrated that the spatial patterns in the WQ of Kinta River can be described by two clusters. In terms of the WQI, the first cluster generally had better WQ than the second cluster. The WQVs and LU classes characteristic of each cluster indicate that the first cluster is more influenced by surface runoff, erosion, and organic and microbial sources of pollution than the second cluster which is mainly impacted by agricultural runoff and mineral dissolution. The optimal RBF model developed for classifying the monitoring stations within these clusters had CCRs of 92.5% and 97.7%, respectively, for the first and second clusters. In addition, the SOFM showed that the temporal trends in the WQ of Kinta River are represented by two groups of similar WQ properties. The first cluster had higher mean values of the majority of the studied WQVs than the second cluster. The RBF classifier backed up this result and generated a classification model with a CCR of 78.0%. Furthermore, the SOFM showed that the spatial patterns in LUs within Kinta River catchment can be described by three clusters. This finding was reinforced by a RBF classifier which additionally provided a classification model with a CCR of 99.3%. Six ANN models were developed to compute and forecast values of the WQI using LU areas as predictors; one model for each of the six spatial scales of interest. The WQI-LU modeling efforts elucidated that the ANN was capable of rendering non-linear models of appreciably high prediction capacities, ranging from 91.2% (0-1000 m BZ) to 97.8% (0-2500 m BZ). These models were utilized in generating forecasts of the WQ status of Kinta River in the year 2020 based on the land use change designed for the area by the MBI 2020 Development Plan. The WQI and WQC forecasts spotlight that the WRB approach is more representative to the WQ of Kinta River and its degree of pollution than any of the studied buffer strips. The study results contribute to the LU planning and urban development programs and to river management plans and allow for optimization of the WQ monitoring network. For example, the approaches and results described in this study will help the WQ monitoring authorities in modifying the current monitoring scheme for Kinta River by selecting the most representative monitoring sites; months; and pollution indicators as the study identified the priority WQVs to monitor and the optimum monitoring stations and months for sample collection such that highly reliable and cost-effective WQ data is still secured. As a result, sharp reduction in monitoring time and costs will be achieved. The various modeling approaches established and presented in this study can be applied to river basins in other urban settings provided the necessary data and expertise are available.

Download File

FPAS 2012 23 IR.pdf

Download (1MB) | Preview

Additional Metadata

Item Type: Thesis (Doctoral)
Subject: Neural networks (Computer science)
Subject: Land use
Subject: Water quality - Standards
Call Number: FPAS 2012 23
Chairman Supervisor: Associate Professor Mohd Kamil Yusoff, PhD
Divisions: Faculty of Environmental Studies
Depositing User: Mas Norain Hashim
Date Deposited: 26 Feb 2019 06:39
Last Modified: 26 Feb 2019 06:39
URI: http://psasir.upm.edu.my/id/eprint/67265
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item