An Optimized Feed-Forward Neural Network and Neighborhood Filter with Different Thresholds for Urban Expansion Simulation; a Case Study: Tehran

Due to increasing of population and its impacts on urban development and to avoid creating problems, studying and predicting of urban expansion and its impacts is one of the most important issues in scientific societies and environmental academies. For a better planning for the future, a vision of the future would be needed. Prediction of urban expansion gives a good view about the future to experts and decision makers. In this regard, they can make a well plan to meet the needs of the city in the future. Tehran is the capital and one of the most growing cities of Iran. So, the major purpose of this paper is to examine and display the growth of urbanization in Tehran in recent years and forecasting it for future. Thus, the combined results of the optimized feed-forward artificial neural network with different neighboring filters were used to model this issue. Moreover, for the urban growth purpose, only the building land use class was considered as the urban land use. The Landsat images at 1994, 2004, and 2014 were used to produce the urban map with the SVM method in ENVI. To access the trend of changes, the images for the years 1994 and 2004 were used for temporal mapping. In this paper, twelve parameters were considered as the effective parameters in urban expansion for Tehran, including digital elevation model, slope, population density, distance from Building blocks, distance from a new buildings created from 1994 to 2004, distance from villages, distance from roads, distance from farm lands, distance from jungle and parks, number of buildings pixel in 3×3 neighborhood, non-proliferation areas, and existing buildings. These factors were extracted from various existing maps and remotely sensed data using the ArcGIS and ENVI software. The feed-forward artificial neural network was performed with these parameters in two phases: (i) for training the network and (ii) for prediction. In order to predict the suitability map with high precision, the feed-forward artificial neural network architecture was optimized according to the least RMSE value for each number of hidden layer neurons. Then, the predicted suitability map was combined with different neighborhood filters with different thresholds in order to produce the urban map of 2014. Then, a combined method concluded from the results of the different neighborhood filters was presented. For accuracy assessment, the output of the model was validated in two phases: (i) 92.46 % of accuracy of the suitability map was obtained using Relative Operating Characteristic (ROC), and (ii) the final urban map of 2014 resulted from the combined method was compared with the real map of 2014 with the kappa index of agreement, i.e., 82.31 %, and the overall accuracy, i.e., 92.22 %. Finally, the proposed combined method was used to predict the urban map for the year 2024. The results indicate that an uncontrolled growth is going to happen in West and Southwest of Tehran. Accordingly, the results of this paper warned the city managers and experts for better planning and having a good strategy in critical situations.