DEVELOPMENT OF AN APPROACH FOR BUILDING EXTRACTION USING SATELLITE IMAGES

Abstract

The importance of increasing role of remote sensing and GIS technology in development of infrastructure has always posed varied challenges to the research community. Time and again, solutions have been provided by researchers for various matrices of physical built environment in different geographies. Amongst the several outcomes of the review of researches, one drawback in Indian context was that only limited studies had been carried out for fully automatic building extraction methods. It was also realized that established researches related to fully automatic building extraction methods were not tested for various types of buildings, but locally applied on construction materials. The identified gaps instigated to frame the objectivity of this research to develop a fully automated building extraction method confirming its applicability to Indian types of settlements. Based on the objectivity and target areas of these researches, each solution had its own strength and weaknesses and always left a scope for future research. In the present research, four building extraction methods were developed for which the required broad parameters were espoused from earlier researches. In first method, a threshold value was calculated for identification of buildings based on areas. The second method involved shadows and corner information for the extraction of buildings. The third extraction method was based on the texture to find the highest mean cluster value, to identify the cluster containing buildings. The fourth method involved the combination of threshold method and texture method. For verification of this method, the shadow mask was applied to detect true buildings and remove false positives. All the four developed methods accept raw images and as such no preprocessing was required before using the image in the developed program. However, in all the methods, post processing has been done to fill the small holes present in the extracted building areas, to smoothen the edges and to remove very small artifacts extracted as buildings. To test the developed methods for varied types and grouping of building in Indian settlement, the city of Jaipur in the state of Rajasthan, India, was chosen as the study area. Testing the proposed approach for complex variations in built environment and settlement formed the basis for choice of study area. The accuracy assessment of the results of developed methods was done three folds. Firstly, it was done on the basis of visual interpretation. Secondly, by comparing the ground truth produced by manually delineating the building boundaries in ArcGIS environment and the ii buildings extracted by developed methods. And thirdly, by comparing the OAP obtained from the supervised classification with the ones attained from the extracted buildings. The best texture combination was selected on the basis of ‘time taken’ and the ‘% error rate’ of area extracted. Analyzing the output results, it was observed that the method based on threshold value successfully extracted all types of buildings, but the shapes of the buildings were not retained. Also, the buildings close to each other were grouped together and extracted as one building. This method also resulted in some ‘False Positives’ and ‘False Negatives’, depending on the spectral reflectance values. The method based on shadow and corner information was not able to extract the buildings having complex shapes, and also the buildings without shadow associated with them. The method based on texture successfully extracted all types of buildings but resulted in high % error rate of area. As the number of texture methods increased in the combination, the % error decreased but the time taken for completing the extraction process also increased. However, using combination of Laws, Wavelet and GLCM, texture methods produced comparatively less % error and took less time to complete the extraction process. The method based on threshold, texture and shadow extracted all types of buildings successfully. All texture combinations used in this method gave higher accuracy than the accuracy obtained from supervised classification. Also, the % error rate of the area extracted was very less for all texture combinations. On comparison of four developed methods, it was found that the method based on threshold, texture and shadow produced best results for all types of building irrespective of their shapes, sizes and Orientation. This method also gave best results for slum buildings over other methods and extracted most of the slum buildings. The output of the research can be summarized that a fully automatic building extraction method has been developed for the complex settlement tested for an Indian case. The strength of this method is that no skills of remote sensing, Geographical Information System (GIS) or software development are required for its application, thus confirming it to be cost effective, time efficient and user friendly. Since, it is an initial attempt to develop a fully automatic building extraction method in Indian context, future researches are advised to test its application in various geographies and suggest the improvements to improve accuracy of extraction of ground features.