QUANTITATIVE APPROACHES FOR OBJECT BASED IMAGE ANALYSIS OF SATELLITE DATA

Abstract

Remote sensing has made enormous progress in recent years. In order to extract useful information from remote sensing data, typically digital image classification is applied. The land use-land cover classes on the earth's surface have different physical characteristics. A land use land cover classification can be performed in a pixel-based environment or an object-based based environment. Pixel-based classification techniques are appropriate for producing land use land cover classification from medium and low spatial resolution remote sensing data. In case of high-resolution images, pixels are smaller than the object under consideration, spectral information within an image is spatially heterogeneous. The object under consideration is composed of several pixels thus within class spectral variation increases. The pixel-based image classification is also affected by ‘salt and pepper’ noise thereby producing relatively low classification accuracy. Object-based image analysis (OBIA) produces an efficient solution to this problem, which involves three basic steps. In a first step, a segmentation process is performed by forming the homogeneous regions in order to delineate objects under consideration. In the second step, an attribute selection process may be performed to reduce the number of attributes. In the last step, a classification process applies to these segments using the selected attributes representing spectral and spatial (shape, contextual and textural) characteristics of objects. Additionally, the OBIA has been found useful in extraction of individual objects and targets. For example, urban feature extraction, military target detection etc. In this research, the object based image analysis (OBIA) has been investigated for extraction of objects from high to very high-resolution remote sensing data. In each of three basic steps, novel approaches have been proposed to improve the quality of object extraction. Two types of remote sensing datasets, a very high-resolution Quick-bird image and the high-resolution LISS-IV image have been considered for the experiments. The very high-resolution belongs to Quick-bird Pan-sharpened image acquired over Chandigarh region due to its well planned structure and separable classes. The high-resolution LISS-IV image of the typical Indian city (Delhi region) has been considered as another experimental dataset, where the classes are not easily Quantitative Approaches for Object Based Image Analysis of Satellite Data [ii] separable. Total eight classes belong to three categories; linear shape, compact regular shape and compact irregular shape features have been considered from Quick-bird image, while twelve classes have been considered for the LISS-IV image. A number of segmentation parameters are required to be set for segmenting the image via a segmentation algorithm. The selection of improper values of segmentation parameters may result into over and under segmentation. The lower value of segmentation parameters causes ‘over segmentation’ whereas higher value causes ‘under segmentation’. Thus, it is required to select proper segmentation parameters value to find critical segmentation at which meaningful objects can be formed. In this research work, semi-automatic fitness functions based on the internal properties of the segments have been proposed to fix the variations in the values of three parameters namely; scale, shape-factor/spectral, and the compactness/smoothness of the most widely used multi-segmentation algorithm. The quality of the segmentation has generally been evaluated through visual interpretation of segmented images. The quantitative evaluation of segmentation quality may therefore be appropriate. There are two ways for quantitative evaluation, the goodness based approach in which reference image is not required and the quality is evaluated using the intra-segment characteristics while the other one is discrepancy based approach, and the quality is measured by calculating the diversity in a segment from the reference object. In this study, quality measures based on the size and shape differences under the domain of discrepancy based approach have been proposed. The proposed quality assessment approach takes into the errors of omission and commission to calculate the discrepancy between the segment and the reference objects. The combined usage of size and shape differences results into a realistic estimate of the quality of segmentation. The segmented image carries a number of attributes grouped into spectral, shape, contextual and textural categories. Out of these attributes, the values of some attributes may be similar for more than one class segments. While few attributes appear exclusively in a particular segment and represent different characteristics from the segments of other classes. Working with all the generated Abstract [iii] attributes is very time consuming and also requires more memory space for storage and sometimes may create confusion, thus it is efficient to work with only important attributes. In this work, a combined decision tree-ROC curve approach has been proposed for selecting the required attributes. Significant reduction has been observed in the number of attributes. The attribute set reduces between 4 and 20 from 100 for Pan-sharpened image, from 43 to 10 and 24 for Pan image, and from 85 attributes to 14 and 25 in case of LISS-IV image. The decision tree has further been used for image classification and individual object extraction. In case of image classification, a tree has been generated and pruned using pessimistic error pruning (PEP). The image is classified using both pruned and unpruned trees. Whereas, in case of individual object extraction, one tree per class has been generated and an object has been extracted at the farthest leaf node by considering all other classes as background. The object extraction has also been done using selected attributes and also using all attributes. Further, the classification and extraction accuracies have been assessed using error matrix and ROC curve respectively. A significant improvement has been observed in both classification and extraction quality with reduced number of attributes in comparison to that when using all attributes. Highest overall classification accuracy and Kappa coefficient obtained with pruned trees are 91.03%, 0.893, 62.5%, 0.572 and 78.91%, 0.76 respectively for Quick-bird Pan-sharpened, Quick-bird Pan and LISS-IV images, respectively. The highest object extraction quality has been obtained for Quick-bird Pan-sharpened image with TPR of 1 and FAR of 0.03. The corresponding values of TPR and FAR for Quick-bird PAN image are 0.65 and 0.04 respectively and for LISS-IV multispectral image are; TPR: 0.94 and FAR: 0.021. Finally, the classified image using classification outputs and the binary image using the outputs of individual object extraction have been produced. In summary, the work has been done in three parts and compiles to form a thesis. The first part covers the image segmentation and its quality assessment, a new fitness functions have been proposed for fixation of values of segmentation parameters to create properly segmented image. To assess the quality of segmented image a set of quality indices that take into account errors of omission and commission have also been proposed. The second part of the thesis is related to the selection of the attributes, a quantitative method of attribute selection has also been suggested that Quantitative Approaches for Object Based Image Analysis of Satellite Data [iv] reduces the dimensionality of the segments attributes by selecting important attributes. The object-based image classification and individual object extraction using decision tree has comes under third part of the thesis.