Omid Ashkriz; Babak Mirbagheri; Ali Akbar Matkan; Alireza Shakiba
Introduction: Urban growth has accelerated in recent decades, therefore, predicting the future growth pattern of the city is very important to prevent environmental, economic, and social problems. The city of Tabriz has also experienced rapid growth of urban lands due to significant demographic changes, ...
Introduction: Urban growth has accelerated in recent decades, therefore, predicting the future growth pattern of the city is very important to prevent environmental, economic, and social problems. The city of Tabriz has also experienced rapid growth of urban lands due to significant demographic changes, which requires accurate simulation of urban growth to prevent negative environmental and economic consequences. The purpose of this study was to evaluate the performance accuracy of the proposed machine learning algorithms by spatial cross-validation method in combination with the cellular automata model to simulate urban growth.Material and methods: In this study, to analyze urban land-use changes, Landsat satellite images related to the years 1997, 2006, and 2015 were classified using the support vector machine algorithm. In the next step, change potential maps of non-urban to urban areas were produced using random forest algorithms, support vector machine, and multilayer perceptron neural network for two periods of calibration (1997 and 2006) and validation (2006 and 2015) based on distance from the main roads, distance from the city center, distance from built-up areas, distance from the rivers and railways, as well as slope, elevation, and two-class (agricultural/barren) land use layer as effective factors in the growth of the city. Finally, using the cellular automata model, the growth simulation of Tabriz city based on land use and change potential maps obtained from machine learning algorithms for the mentioned periods was performed. To prevent over-fitting of algorithms to training samples and to obtain optimistic results, in the process of extracting optimal parameters of machine learning algorithms, the spatial cross-validation method was used to reduce the spatial correlation between training and test data.Results and discussion: The results showed that the random forest algorithm with the area under the ROC curve of 0.9228 compared to the support vector machine and multilayer perceptron neural network algorithms with 0.8951 and 0.8726, respectively, had a better performance in estimating the change potential of non-urban to urban areas. Furthermore, in comparison with others, the random forest also clearly showed local variations in potential change. Finally, the growth of Tabriz city was simulated using the cellular automata model based on the obtained change potential maps. Comparison of the prediction map in the validation period with the current situation of urban areas in 2015 showed that the accuracy of an urban growth simulation model based on random forest with a Figure of Merit index of 0.3569 compared to models based on support vector machine and artificial neural network was more accurate in allocating non-urban to urban lands with 0.3496 and 0.3434, respectively.Conclusion: As machine learning algorithms such as artificial neural networks, support vector machines, and random forest are capable of solving non-linear problems, using them is strongly recommended for urban growth simulation. Also, among the algorithms used in this research, the random forest algorithm based on ensemble learning has a higher advantage than the two-support vector machine and the artificial neural network algorithms.