A Technical Approach of Image Segmentation in ENVI GIS to Identify Thematic Clusters for Visualization of Urban Transformations

The technical study objective is to apply segmentation techniques for clustering image into thematic areas. The data used in this research included Landsat TM and ETM+ multi-band imagery covering chosen research area. The image processing was performed using supervised classification in GIS software. The technical aim of the research is image classification which consists in automatic assignation of all pixels on an image into land cover classes that are typical for this study area. The logical algorithmic approach of clustering segmentation was applied to identify clusters for thematic mapping of land cover types in the selected study area. While using data for spatial modelling and mapping, specific study objectives should always be evaluated, hence the model may have certain limitations and dependancies on the scale and resolution. Methodology is based on Landsat TM image processing. The satellite scenes of Landsat TM with middle-sized resolution and open distribution have been taken for the research as available and appropriate images for current objective. The data pre-processing include image contrast stretching, which is useful as by default, ENVI displays images with a 2% linear contrast stretch. For better contrast the histogram equalization contrast stretch was applied to the image in order to enhance the visual quality. The Landsat TM scenes are presented by the seven spectral bands of the image taken at one time with a spatial resolution of 30 meters for Bands 1 to 5 and 7 and 120 meters for the thermal infrared 6th band. Though the original scenes are represented as grey-scale colored images, the application of the "Color Mapping" function enables to present bands in colored palette. To visualize images, the "Creating RGB Image" was applied using RGB Color button from the menu Available Bands List. The Red, Green and Blue fields correspond to the Band 7, Band 4, and Band 1, respectively. Simulation results based even using the largest scale and resolution may not always provide the best results due to the different approaches of the models. Classification was done on the basis of the multispectral data, spectral pattern, or signatures, of the pixels that represent each land cover class. Different land cover types and landscape features are detected using individual properties of digital numbers (DNs) of the pixels. The DNs showed values of the spectral reflectance of the land cover features, and individual properties of the objects. The used algorithm principle consists in merging pixels on the images into clusters, which is based on the assessment of their homogeneity and distinguishability from the neighboring pixel elements. The number of clusters was assigned to 15, which responds to the selected land cover types in the study area. These cluster centers were then located within the study area. During clustering procedure, each digital pixel on the image is categorized to the respecting cluster, to which the mean DN value of the given pixel is the closest. Upon classification of all pixels in such a way, the revised mean vectors for each of the clusters were computed. The process is repeated in an iterative way until optimal values of the class groups are received and pixels are assigned to the corresponding classes. The results show changes in the urban landscape patterns based on the comparative analysis of the image scenes.