BUILDING ROOFTOOP EXTRACTION ON SATELLITE IMAGERY USING SEMANTIC SEGMENTATION AND EDGE DETECTION TO PREDICT THE NUMBER OF BUILDINGS IN THE LOCAL ENVIRONMENTAL UNIT

• Main Author: Aizatin, Anisa
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69937
• Institution: Institut Teknologi Bandung

Buildings are one of the objects in spatial planning that can be utilized in many fields such as population estimation, environmental planning, and in the field of statistics, it is used to create the master frame. A survey or census is one step in the field of statistics for data collection. In carrying out a survey or census, a master frame, which contains the number of buildings in the local environmental unit, is required as the basis for planning and carrying out a survey or census. Making the main frame in the form of census blocks and the number of buildings is done in a long time because it needs human resources and costs a lot. Satellite imagery is the result of remote sensing which depicts the earth's surface and the objects in it at a certain scale which is taken periodically. Therefore, object extraction from satellite imagery, especially buildings, can be used as an alternative for the formation of a master frame. This research proposes a hybrid segmentation method to improve the segmentation results of building extraction on satellite imagery. The proposed hybrid segmentation method utilizes UNet+ResNet101 as a deep learning method for semantic segmentation and Holistically Nested Edge Detection (HED) and Canny edge detection as edge-based segmentation. In addition, this research forms a data set in the form of images taken from satellite images of several regions in Indonesia by considering the diversity of depictions. The proposed hybrid segmentation method can improve segmentation results more than 100% in areas with dense and overlapping distances between buildings. However, accuracy and Intersection of Union values smaller than UNet+ResNet101 1% difference. The use of data sets created by considering the diversity of building descriptions can improve performance compared to using public data sets with different building characteristics from buildings in Indonesia.