Building Detection Using a Hybrid Methodology Applied to High Spatial Resolution Aerial Photographs

<p>Automatic building detection using high-spatial resolution satellite and digital aerial photography is an important aspect of geographic information science research. Accurate methodologies can provide scientists with the building information that can be broadly applied in various applications, including land cover change analysis, urban planning regulation, economic analysis, disaster assessment, etc.</p> <p>This thesis focuses on applying a hybrid methodology to automatically identify the buildings on digital aerial photographs with high spatial resolution in Bluffton, South Carolina. There are five components in this hybrid methodology, including: Scale-Invariant Feature Transform (SIFT) feature extraction, K-Means (KMs) clustering, Bag-of-Visual-Words (BOVW) representation, Support Vector Machine (SVM), and Efficient Subwindow Search (ESS) techniques. This method was originally applied in object detection on natural images by Lampert et.al (2008). In this research, I proposed a new way to improve feature keypoint extraction by preserving the location information of each SIFT keypoint and replacing the original 128-dimensional keypoint descriptor with a 3-dimensioanl spectral vector. An experiment was designed to compare the performances of these two different methods. The results suggest that the newly proposed keypoints extraction is superior to the original method, achieving an correct detection rate of 71.34%, which was reasonable and useful for future exploration. It was also more computationally efficient. In the second part of the experiment, the traditional sliding window search with a fixed set of parameters was compared to the improved hybrid methodology. This exhibited advantages when finding the global maximal quality scores to locate the presence of targeted buildings. In the third part of the experiment, the performance of a traditional pixel-based, minimum distance classification was compared with the improved hybrid building detection methodology. The results indicate that the improved hybrid method performed better in terms of achieving higher accuracies and requiring less manual work.</p> <p>The results confirmed that the improved hybrid methodology can be applied to high spatial resolution aerial photographs for building detection. In order to achieve even higher detection rates, future work will focus on: 1) replacing the SIFT keypoints with spectral information with other local descriptors; and 2) replacing rectangle shape of the bounding box yielded by ESS with a freely shaped bounding box.</p>

Name thesis:degree_name

M.S.

Level thesis:degree_level

Campus Access Thesis

Discipline thesis:degree_discipline

Geography

Year

2010

Author dc:creator

• Wu, Jiang

Contributors dc:contributor

• John R Jensen

• Geography
• Social and Behavioral Sciences
• BOVW
• Building Detection
• ESS
• Remote Sensing
• SIFT
• SVM

Statement dc:rights

• © 2010, Jiang Wu

Repository record dc:identifier

https://scholarcommons.sc.edu/etd/1310

OAI identifier oai:identifier

oai:scholarcommons.sc.edu:etd-2311