STUDY OF CELLULAR AUTOMATA TECHNIQUES FOR URBAN GROWTH SIMULATION IN GIS DOMAIN

Abstract

Land transformation is one of the foremost fields of human-induced environmental change with an extensive history dating back to antiquity. The process of land transformation has not abated, but rather accelerated and diversified with the onset of industrial revolution and the expansion of population and technological capacity. Settlement refers to the occupation of land for human living. Monitoring and evaluating the growth of urban settlements is essential in order to avoid environmental problems such as, depletion of natural resources, increased pollution levels, loss of green cover etc., especially in developing countries where cities are experiencing a rapid growth. It is of vital importance for urban planners to not only understand the past and present urban growth patterns but also be able to predict the future growth patterns. This is where spatial models of urban growth become useful. These models not only provide an understanding of urban dynamics, but also provide realizations of the numerous potential scenarios that an urban system may take. In recent years, Cellular Automata (CA) techniques have evolved as possible alternatives for urban growth simulation due to their potential for dynamic spatial simulation capability and affinity towards GIS and remote sensing. However, the CA based models highly depend on formation of transition rules, which is often subjective, as these are based on expert's opinions. Secondly in Indian context, very few attempts have been made to develop CA based models for assessing the urban growth. The present research aims to apply CA based models to simulate urban growth in two typical Indian cities having markedly different growth patterns and to examine the efficacy of Artificial neural networks (ANN) in formation of transition rules for CA based modeling and its comparison vis a vis multi-criteria evaluation technique (MCE). Besides, the effect of different neighbourhoods viz Von Neumann and Moore neighbourhood in CA modelling has also been investigated. The simulated urban growth has been evaluated, based on cell by cell match using Percent correct match (PCM) and spatially using Moran Index and Entropy. Finally using ANN, an urban growth zonation map depicting zones having different growth potential at an ordinal scale has also been generated and evaluated. The proposed CA based models have been implemented in two Indian cities namely Dehradun with geographical extents 30°15' N to 30°25' N and 77°55' E to 78°10' E. The region is experiencing a fast urban growth, which is taking place in a dispersed manner mainly along the roads and Saharanpur with geographical extents 29°55' N to 30°0' N and 77°30' E to 77°35' E. The city is expanding onto the nearby fertile agricultural land in a compact manner, mainly along the roads. The following datasets have been used to generate various thematic data layers for the two study area: a) Remote sensing data: IRS-1C LISS-III and IRS-P6 LISS IVmultispectral, PAN, and aerial photographs b) Survey of India toposheets at 1:50,000 scale (53J/3,53F/15 and 53G/9) c) Guide map at 1:20,000 scale d) Master plans at 1:20,000 scale The ERDAS Imagine image processing software has been used for processing and analysis of remote sensing data. The GIS analysis has been carried out using ArcGIS software while ANN processing has been done in Neural Network Tool Box ofMATLAB software. The following GIS data layers have been generated from remote sensing data and other data sources, i) Land cover maps showing built-up/non built-up areas for years 1997, 2001, 2005 from digital image classification of IRS LISS III images of Dehradun city. For Saharanpur city, built-up/non built-up areas maps for years 1993 and 2001 * through visual interpretation of aerial photographs (1:10,000 scale) and IRS PAN image ii) Road network maps from LISS III and PAN images, master plan and guide maps iii) City core map after consultation with local planning authorities. In these GIS layers, the boundary of reserved forests, restricted areas and public lands have been masked out. The proposed CA models take into consideration only the physical factors affecting urban growth. Social and economic factors have not been considered due to non availability of accurate data pertaining to these factors. The three physical factors considered are, i) Distance to city core ii) Accessibility to infrastructural facilities iii) Distance to road network Corresponding to these three factors, following four raster maps have been created in GIS, i) Map showing Euclidian distance of each cell from the nearest road ii) Map showing Euclidean distance of a cell from the nearest built-up iii) Map showing Euclidian distance of each cell from the city core iv) Map showing amount of built-up cells in neighbourhood In the MCE based CA model (MCE-CA), the urban growth suitability is first generated using the MCE technique. The weights are assigned to the three factors using Analytical Hierarchical Process (AHP) technique. Taking the MCE generated suitability map as an input, the MCE-CA model is run iteratively for user-defined number of iterations and the amount of built-up in the neighourhood is estimated using Von Neumann and Moore neighbourhoods of sizes varying from 3x3 cells to 39x39 cells. Using different combinations of the neighborhood and model iterations, the MCE-CA model has been executed several iii times for each of the study areas, so as to determine the optimum values of the parameters. For the Dehradun city, the model has been calibrated for the period 1997-2001. Since for this city, the actual growth for the period 2001-2005 is available, the model is validated for year 2005. For Saharanpur city, the model has been calibrated for the period 1993-2001. Using the calibrated model, future urban growth simulation for year 2011 has been carried out for both the study areas. In the ANN based CA model (ANN-CA), a multilayered feed forward ANN with one input layer, one or two hidden layers and one output layer has been designed and trained using backpropagation algorithm. The input layer consists of 4 neurons corresponding to the four variables. The output layer consists of 1 neuron corresponding to whether a cell location changed from non-built-up to built-up (l=change, 0=no change). The number of neurons in the hidden layer has been finalized in two ways: i) based on literature driven thumb rules and ii) by trial and error. The ANN, producing the highest accuracy has been selected for simulation. The output from ANN is a map showing the development potential of cells. All the cells have not transitioned immediately to built-up. Only the cells that have a development potential above a certain threshold are changed. The size ofneighbourhoods has been fixed as 5x5 cells for Dehradun city and 13x13 cells for Saharanpur city, as identified from MCE-CA model. The accuracy of the simulated urban growth has been determined using two measures: i) Percent correct match (PCM) for cell by cell assessment and ii) Moran Index (I), a spatial statistical indicator to assess the pattern of growth. An urban growth zonation map has also been generated based on the ANN outputs. The ANN output shows the development potential of each cell, based on which the study area has been categorized into three zones (i.e. high, medium and low) showing the urban growth potential on an ordinal scale. Since the ANN outputs are not normally distributed, a logit iv transformation has been applied to make the data normally distributed. The transformed data has then been categorized into three classes as low potential zone < (u-a), (u-a) < medium potential zone < (u+ a), (u+ a) < high potential zone, where u is mean and a is the standard deviation. These urban growth zonation maps have been validated by overlaying them with the actual urban growth maps for the respective years, to find the spatial distribution of actual urban growth in each zone. Further, the simulated growth patterns for both study areas have also been evaluated using relative entropy. It is a structural measurement index that assesses the goodness of fit according to the spatial domain of interest, which in this case is the distribution of urban growth with respect to distance from roads and distance from the city core. The results show that for Dehradun city which has a dispersed growth pattern, Von Neumann neighbourhood of small size produces the highest accuracy, in terms of pattern and location of urban growth. While for Saharanpur city which has a compact growth pattern, large neighbourhoods, produces the most optimum results irrespective of the neighbourhood. It has also been observed that large number of model iterations does not increase the model accuracy, as they have resulted in an increasingly compact patterns as compared to the actual growth. This may be due to unplanned and stochastic behavior of urban growth process in Indian cities, which the CA models have not been able to simulate completely. The ANN-CA model also produces comparable results as obtained from the MCE-CA model. This shows that the ANN are able to define the CA transition results directly from the database without human intervention, which proves the usefulness of ANN in urban growth simulation. In ANN-CA model, the ANN architecture based on literature driven thumb rules produces better or comparable results than those obtained from the optimal network from trail and error. The urban growth zonation maps, obtained from ANN outputs show that most of the simulated growth has taken place in the high potential zone followed by the medium and low potential zone. Thus, the delineated urban growth zones matched with the actual growth pattern. These results demonstrate that ANN can be used effectively in reducing the subjectivity involved in the urban zonation process. The simulation results have also been evaluated at a macro level using relative entropy. The evaluation of the simulation results using relative entropy, indicate that the model has been able to simulate the distribution of urban growth with respect to roads and city core accurately. The study also demonstrates the usefulness of PCM and Moran Index as simple indicators for evaluating the simulation results on a cell by cell basis and at spatial level respectively. Thus, the proposed CA based models and the urban growth zonation approach developed in this research can be very fruitful for the urban planners in planning and regulating the future growth in Indian cities.