Akademik Veri Yönetim Sistemi
Tezin Türü: Yüksek Lisans
Tezin Yürütüldüğü Kurum: Hacettepe Üniversitesi, Fen Bilimleri Enstitüsü, Geomatik Mühendisliği A.B.D., Türkiye
Tezin Onay Tarihi: 2015
Tezin Dili: İngilizce
Öğrenci: SALAR GHAFFARIAN
Danışman: Mustafa Türker
Özet:
Building extraction is an important task for detecting the changes of buildings, detecting the destroyed buildings, updating the vector maps as well as reconstructing 3D building models. Satellite and airborne images are greatly used for this purpose in various integrated forms. In this study, a novel approach is presented for automatic extraction of buildings from high resolution multispectral space imagery. The main goal of the study was to develop an automatic method for building extraction by utilizing the relationship between the cast shadows and the buildings with various types, shapes, sizes, heights and environmental scenes. If the characteristics of the shadow areas and the buildings that cast them are considered it can be seen that the geometry of the shadows highly depends on the geometry of buildings that share a border with their cast shadows. Therefore, the developed method has been developed with regard to meaningful relations between the buildings and their cast shadows.
In the beginning, the shadow regions of the buildings are extracted through a novel developed technique which operates on the features of the shadows in LAB color space. Next, to segment the buildings automatically, a novel sampling method entitled ‘buffer zone generation’ is carried out. This method utilizes the geometric relations between the buildings and their relevant cast shadows as well as the illumination direction. Then, the generated buffer zone is enlarged to generate appropriate initial contours for the Gradient Vector Flow (GVF) Snake segmentation algorithm. To enlarge the sampling region and localize the initial contours in reliable positions, the pixel-based Region-Growing segmentation algorithm is carried out to produce the required vectors to be used by the GVF Snake algorithm. The region-growing segmentation is used as an intermediate step as a connectivity-bridge between the generated buffer zone and the GVF segmentation to automate the GVF Snake algorithm. The results of Region-Growing segmentation are then converted into vector form and used as initial contours for the GVF Snake segmentation algorithm to extract buildings. This thesis study contains three main contributions. The first contribution is that it automatically extracts the cast shadows of the buildings as the essential requirements for the method. The second contribution is that it automatically collects samples from the rooftops of the buildings. The third contribution is that it segments the building areas in an automatic manner by means of the GVF snake algorithm.
The
developed approach was tested on 50 test sites selected from urban and
sub-urban areas in Ankara, the capital of Turkey. The high resolution multispectral
satellite images used for the test sites were obtained from Google Earth. For
each of the three main steps followed throughout the approach, the accuracy
values were computed using the well-known metrics Precision, Recall, and FB-score.
For each metric, the overall accuracy and the individual accuracies were
computed. The results achieved are quite satisfactory. For the shadow
extraction part, the overall FB-score result was computed to be 87.3%, while
for the final extracted buildings the overall FB-score result was computed to
be 84.9%. The results achieved in this study also demonstrated that the GVF
Snake algorithm improved the overall results of the region growing segmentation
about 10%.