85 CHAPTER 5 OBJECT ORIENTED IMAGE ANALYSIS 5.1 GENERAL Urban feature mapping is one of the important component for the planning, managing and monitoring the rapid urbanized growth. The present conventional method of mapping includes EDM and Theodolite or Total station which requires a long time, high manpower and cost component. So, the present available higher spatial resolution remote sensing data such as IKONOS and QUICKBIRD visually produce clear identification of urban features like building, road vegetation and open spaces etc. Normally there are two classification methods are followed for the classification, such as visual interpretation and digital image classification (includes pixel based image classification and object oriented image analysis). Analysis of this study comprises of four aspects such as preparation of data set, segmentation of image, classification of image and an accuracy assessment. 5.2 DATASET 5.2.1 Higher Resolution Satellite Data There are two types of data format available in QUICKBIRD satellite data such as PAN and MSS. PAN has a higher spatial resolution of 0.63m with single band and MSS has 2.6m of spatial with 4 spectral bands. Therefore to get use of both spatial and spectral advantages, the above two dataset were merged using widely accepted and accurate method called PCA.
86 5.2.2 Creation of Digital Elevation Model The elevation information of the building is one of unique data set which is given input for the segmentation and classification process. The DEM is created from Aerial photograph using Digital Photogrammetry tool following Interior Orientation, Relative Orientation and Exterior Orientation. 5.2.3 Urban Landuse Classes The National Urban Information System of India has urban landuse classes classification for the development of urban information system. For this analysis required landuse classes were taken out from the NUIS table. 5.3 SEGMENTATION The segmentation analysis is carried out using the well known object oriented image analysis software called e-cognition. The merged data product and Digital Elevation Model were given as input in the software. The required segmentation parameter were defined for the urban features such as building, road, vegetation and open area based on the guidelines, which is as follows; a. Employing possible largest scale parameter while ensuring that different classes are not merged b. To produce visually convincing results higher weightage should be given for spectral information c. Utilizing high smoothness weight to produce object with smooth borders while preserving the capacity to produce noncompact segments. d. Using a large compactness weight factor to extract compact object.
87 5.3.1 Segmentation for Buildings The building segmentation analysis is carried out using the above guidelines and the parameter values were selected such as scale parameter (20 to 110), color ( 0.8 to 0.1), shape (0.2 to 0.9), compactness of 0.5 to 0.9) and smoothness of (0.5 to 0.1). The above set of parameters were given in 5 levels and the final range were arrived based on the optimized segmentation of object. 5.3.2 Segmentation for Road The linear pattern of road got segmented by 4 level by varying the scale parameter 20 to 90, color 0.8 to 0.3, shape 0.2 to 0.7, compactness 0.5 to 0.9 and smoothness 0.5 to 0.1. The segmentation guidelines were followed for selection of the above parameter. 5.3.3 Segmentation for Open Area The open areas in the study area includes the open space and playgrounds. The above features were segmented upto 5 th level by varying the scale parameter (20 to 110), color (0.8 to 0.1) shape (0.2 to 0.9), Compactness (0.5 to 0.9) and smoothness 0.5 to 0.1) 5.3.4 Segmentation for Vegetation The vegetation were segmented upto 4 levels because of small cluster of many sub- clusters forms the meaning full object. In order to segment the values were ranged for each parameter such as Scale (20 to 90), Color (0.8 to 0.3), Shape (0.2 to 0.7), Compactness (0.5 to 0.9) and smoothness (0.5 to 0.1).
88 5.4 VALIDATION OF SEGMENTATION PARAMETER In order to ensure the performance of the developed method, the validation is one of the important criteria. Therefore, the developed method was validated by taking one of the derived segmentation parameter (scale parameter (20 to 110), color, shape) and given as input to the similar kind of image set of other area. For which the Adyar area of Chennai is taken as validation area. The merged data product and the digital elevation model is shown in Figures 5.1 and Figure 5.2. The segmentation was carried out for different level and fuzzy membership grade were obtained. Thus obtained fuzzy membership function was compared with the established membership grade. The above was procedure was adopted for the all other features such as building, road, open area and vegetation. Figure 5.1 FCC of Quickbird data of Adyar area
89 Figure 5.2 Digital Surface Model of the Adyar study area The Digital Surface Model of the study area is shown in Figure 5.19. In general, the terrain has flat to gentle slope in nature. 5.4.1 Building The segmentation parameters are selected for each class as same as that of guidelines followed in sec 5.3 of the Anna University study area. Five levels of segmentation was carried out to get the optimal segmented building and its ranges from scale parameter of (20 to 95), color (0.8 to 0.3), Shape (0.2 to 0.7), Compactness (0.5 to 0.9) and smoothness (0.5 to 0.1). 5.4.2 Road The road feature in the validated study area is segmented for the scale parameter of (20 to 110), color (0.8 to 0.1), Shape (0.2 to 0.9), Compactness (0.5 to 0.9) and smoothness (0.5 to 0.1).
90 5.4.3 Open Area The open area feature in the validated area is segmented by varying the scale parameter of (20 to 95), color (0.8 to 0.3), Shape (0.2 to 0.7), Compactness (0.5 to 0.9) and smoothness (0.5 to 0.1). 5.4.4 Vegetation The vegetation feature is segmented with its Scale parameter (20 to 95), color (0.8 to 0.3), Shape (0.2 to 0.7), Compactness (0.5 to 0.9) and smoothness (0.5 to 0.1). 5.5 OBJECT CLASSIFICATION Classification means assigning the number of object to a certain class through by the typical properties or conditions of the classes have. Then the object becomes assigned (classified) to whether they have or they have not met these properties. 5.5.1 Knowledge Base Creation Rule Base It is created using the interpretation key element such as size, shape, tone, color, texture, pattern and contextual information or otherwise it is called as spectral, spatial, textural and relationship to neighbor. Interpretation key can be organized in two ways such as selective key and elimination key (Lillesand et al 2004). In the first step, the features used for the interpretation has to be determined but in other method it can be derived out. Here, the selective key is used to select the feature and the class hierarchy is developed. 5.5.2 Class Hierarchy The significant of interpretation key elements is identified for the selected urban landuse classes of the study area, which is given in sec 4.2.1.3.
91 The classes includes such as building (includes concrete old and new, Tiled and asbestos roof), open area (vacant), vegetation, road (transportation) shadows and water bodies. The analysis were carriedout for the above landuse classes and it is discussed below. 5.5.2.1 Building There are three sub-classes of buildings is available in the study area according to the roof type and roof floor pattern such as building with concrete structure (New), concrete building covered with tar, with or without asbestos / tiled houses. The study area buildings are well separated and bigger in size (Institutional area) as well as height (>G+1 to <G+3 floors) compared to residential buildings. It get optimized in the level 5 of the segmentation process. And it is observed that the following elements / parameters in the Table 5.1 were significantly used to identify the building features. Table 5.1 Interpretation key elements and parameter for building feature Sl.No Classes / object Elements Significant Features Range Context Mean Elevation value 80 to 178 1 Building (Cement Concrete) Tone Texture Green band (Band 2) 300 to 700 Brightness 85 to 175 GLCM homogeneity 0.3778 to band2 (All direction) 0.8345 2 Building (Old tiles) Mean value Band (4) -65 to 20 of Band (4)
92 5.5.2.2 Road The road is made up of bituminous tar coated. Therefore in the study area, only the main entrance road is visible, and most of the all other inner roads are covered by the trees / vegetation. Segmentation gets optimized in the 4 th level. It is observed that the following Table 5.2 shows the significant parameters used to identify the road features. Table 5.2 Interpretation key elements and parameter for road class Sl. No Classes / object Elements Significant Features Range Membership function Spectral the band 4 (IR) 72 to 150 Full range /Tone Band 2 (Green) 190 to 270 Full range 1 Road Spatial Length / Width ratio 9 to 17 Full range GLCM Textural homogeneity 0.31 to 0.34 Full range (4)(All direction) 5.5.2.3 Open Area The selected area, has a big vacant (playfield) area of presence of sand and part of the ground is under construction of football pitch (different tone) and few of the open spaces is presented in the other part also. The open area is gets optimized in the level 5 of the segmentation process. It is inferred that the following Table 5.3 shows the significant the parameters used to identify the building features.
93 Table 5.3 Interpretation key element and parameter for open area class Sl. No Classes / object Significant Features Range Membership function Tone Mean value Red band (2) 248 to 600 Full range 1 Open area Texture GLCM homogeneity (2) (All direction) 0.4520 to 0.5289 Full range Context Not mean elevation value 87 to 140 Full range 5.5.2.4 Vegetation Vegetation class covers grass, scrub and trees. In the selected study area vegetation coves 60 % of the total area. In all the vegetation, three type of features can be discriminating. Table 5.4 shows the significant the parameters used to identify the vegetation features. Table 5.4 Interpretation key elements and its parameter for vegetation class Sl. No Classes / object Significant Features Range Membership function 1 Vegetation dark Band(4) Band 3 (Red) 220 to 350 Trapezoidal 0 to 150 Full range 2 Vegetation Light tone band 3 350 to 740 Full range Band 3 (Red) 0 to 150 Full range 3 Vegetation Grass band 3 500 to 740 Full range Band 3 (Red) 0 to 150 Full range
94 5.6 CONVENTIONAL (PIXEL BASED) METHOD OF IMAGE CLASSIFICATION In the process of satellite image classification, the image pixels are automatically assigned (by the classification software) to one of a number of output classes. Two methods of image classification were utilized, such as supervised and unsupervised. Supervised classification, as the name suggests, involves the user defining, by delineating training sites, the output classes to which pixels will be assigned. On the other hand, unsupervised classification or image clustering requires no pre-classification input from the user and pixels are split into a number of groups (to be specified by the user), based on their spectral similarity. Supervised classification formed the core of the classification work, while an unsupervised classification was performed to assist in the selection of training sites. 5.6.1 Accuracy Assessment The selected two methods of classification such as pixel based and object based methods accuracy were assessed and compared. The accuracy of the above work was carried out using the ERDAS software. In the reference sites, known surface features from field surveys and examination of other secondary data sources are compared with the classified image, often in the form of contingency table or error matrix. In this study the reference sites are derived using stratified random sampling of sites in the classification. The samplings are stratified, so that classes the samples with adequate representation into the total sample. The stratified random sampling procedure is performed using ERDAS software.