Building zone detection and roof type classification from satellite images using convolutional neural networks

Geographic Information Systems (GIS) analyzes many aspects of a building model, including efficiency, planning, damage assessment and more. One of the key components required by GIS applications is the roof type and geometry of a building. Therefore, the identification and classification of roofs is an important topic explored in GIS. With developing technology, obtaining high-resolution satellite images of large areas can be easy and inexpensive. Large areas can be analyzed quickly with the classification method that can be done on satellite images. In this way, significant savings in time, cost and labor can be achieved. Some object detection methods can now determine the positions of building roofs in satellite images. Deep learning is one ofbthese methods usedbfor object detection. Although the classical machine learningbmethod has achieved successful results inbsolving thisbproblem,brecent research on deepblearning has shownbthat the success in accuracy ratesbcan reach higher levels. In this study, besides the regional determination of the roof locations of the building regions by using deep learning methods, the related roof types are also classified. To achieve this, the performance of the CNN method is tested using medium resolution images. When the test results of the satellite images obtained from different regions are examined, it is seen that AlexNet produced successful results in region detection with 92.30% accuracy and GoogleNet produced 87.18% accuracy. In addition, it was concluded that AlexNet was the most suita lebmodel for the classification of building roof types among the three architectures examined.