IMAGE SEGMENTATION EVALUATION SOFTWARE FOR OBJECT BASED IMAGE CLASSIFICATION

Abstract

With improvements and inventions in technology today, high spatial resolution images, say < 5m, are available. These images, no doubt, make visual interpretation comfy, less tedious and precise but also invite a lot of troubles while classifying using traditional approaches. Moreover, the traditional pixel based classification approaches have limitations in taking into consideration the contextual attributes, i.e. shapes and sizes of land use features such as roads, buildings etc. into classification process. Thus, to produce quality land use maps from high resolution data, the concept of OBIA was recently introduced. The strength of OBIA is good quality image segmentation, which may not have a unique solution. There are no means to measure the goodness of segmentation embedded in software apart from visual interpretation. It varies with the perception of one analyst to the other. Hence, quantitative evaluation methods need to be developed. This study provides a review of OBIA and various segmentation techniques proposed by researchers since years. Also, a study of some approaches for segmentation evaluation is discussed. Using the best suitable approaches for evaluation, segmentation evaluation software is developed in IDL 6.2. The salient features of the software comprise 1. ' Evaluation using three approaches namely: i) discrepancy based on feature values of segments, ii) Discrepancy based on probability of error & iii) Beta index test. 2. Experimental investigation of evaluation measures on segmentation of synthetic and remote sensing datasets. 3. Object based image classification. The segmented images are implemented on the evaluation approaches. Two best segmented images are thus chosen and classified. Object based classifier is developed in IDL 6.2 using the traditional min-dist algorithm but considering an object as a pixel. For improvements in classification results, attributes of segments like, gray value mean, texture, compactness and length-width ratio are taken into consideration. The classified output clearly suggests that the image that gives better results during evaluation also gives better classification accuracy. Hence the strength of the evaluation measures is validated.