Urban Growth Simulation Using Remote Sensing Imagery and Neural Networks

This paper presents a methodology for simulating urban growth phenomenon through utilizing remote sensing imagery and neural network (NN) algorithms. Historical satellite images of Indianapolis city, IN were used. All images were rectified and registered to Universal Transverse Mercator (UTM) NAD83 zone 16N. Supervised classification was used to classify the images to different land use categories. Seven classes were identified: water, road, residential, commercial, forest, pasture grasses and row crops. Image fusion was tested to examine its effect on the classification results. Overall, the classification accuracy using original images was better than the results from the fused ones. To implement NN algorithms to simulate the urban growth; focus was directed to the residential and commercial classes and their growth. The boundaries of these areas were extracted for each of the growth years. Radial extent of the boundary at specified angles was measured using different city centres. Radial distances and growth years were used as inputs to train the neural network algorithms. Two NN algorithms were used to simulate the urban growth: simple linear NN and back propagation (BP). Each algorithm was trained using the available data as well as interpolated data produced through NN function approximation algorithm. Short and long term urban boundary predictions were performed. Results showed that both algorithms after increasing the volume of the dataset succeeded in simulating the growth trends with better results achieved using the simple linear NN. Visual check and similarity of simulated and real growths were tested.