Presentation of a Method for Detecting Urban Growth using Spectral- Spatial Variation Indicators and Remote Sensing Data

Urban growth and its monitoring are important items of interest for urbans and municipalities. Remote sensing and related technologies are new tools that recently used for this purpose. Change detection techniques generally divided into two groups: object based and pixel based techniques. Pixel based techniques included algebra, transformation, fuzzy, hybrid, multi temporal direct comparison and classification based methods. In classification comparison is useful. In these methods quality and quantity of training samples are vital for producing good results. According to research conducted, accuracy of unsupervised classification techniques is less than supervised classification techniques. On the other, supervised techniques need training samples that obtaining these samples is time consuming. As first goal in this research, presentation a supervised method for obtaining training samples automatically. Second goal is using spectral- spatial variation indicators in change detection. Landsat 5 TM images in 1990, 2000 and 2011 were used for change detection of Shiraz city. In this research reference land use map was prepared by IRS, IKONOS and Google earth images.
At first, images were preprocessed. In preprocessing, geometric and radiometric calibration were done. Images were georeferenced by polynomial method and RMSE calculated less than 1 pixel. After preprocessing, spectral- spatial variation indicators were calculated by 3d wavelet decomposition. These indicators describe spectral and spatial simultaneously. Results showed that spatial variations of natural features are less than spectral variations. So these features are brighter than other features in spectral variation images. Then data were clustered into four classes urban, baresoil, road and vegetation by fuzzy c means method. Overall accuracy of fcm calculated 81.6, 86.5 and 87% in 1990, 2000 and 2011 with consideration variation indicators. Train data were selected for maximum likelihood classification automatically. Overall accuracy of maximum likelihood classification calculated 87.36, 89.5 and 89.7% in 1990, 2000 and 2011 with consideration variation indicators. Results showed variation indicators and automatic selection of training data improved overall accuracy and separation of classes. Classified images used for change detection. Some of samples were selected for assessment accuracy of final product. Confusion matrix found using reference data. Results showed spectral- spatial variation indicators improved accuracy of change detection. At the end support vector machine used in post classification comparison. Produce and user accuracies of this method were obtained 76.5, 65% for first interval and 77.84, 78.23% for second interval under consideration variation indicators. In the event that accuracies of proposed method were obtained 84.63, 78.3% for first interval and 85.89, 88.42% for second interval. So accuracies of change detection with proposed method is higher than support vector machine.