As one of the most destructive disasters in the world, earthquake causes death, injuries, destruction and enormous damage to the affected area. It is significant to detect buildings after an earthquake in response to reconstruction and damage evaluation. In this research, we proposed an automatic rooftop detection method based on the convolutional neural network (CNN) to extract buildings in the city of Christchurch and tuned hyperparameters to detect small detached houses from the aerial image. The experiment result shows that our approach can effectively and accurately detect and segment buildings and has competitive performance.