Building Detection: Testing a New Object-Based Approach Against Neural Networks

Automated identification of HRBs (High-rise Buildings) on satellite images is challenging when densely populated areas are concerned. Factors that increase complexity are, among others, roads and both the azimuth and elevation angle of the sensor. In this study, two different effective HRB detection techniques are proposed. The first method is using CNNs (Convolutional Neural Networks), an extensively used tool for pattern recognition in the field of machine learning. However, domain movement considerably reduces the CNN's performance on the test data in other domains, making it difficult to generalize. Besides, obtaining the dense annotations on the remote sensing images is expensive and time-consuming. Therefore, a new object-based approach is proposed that includes multi-resolution segmentation and relief displacement by azimuth angles of the sensor. Both methods were tested using images from four regions in South Korea using VHR (Very High Resolution) satellite imagery from the KOMPSAT-3 and WorldView-3. The results show that the performance of both methods heavily depends on factors such as building size and density as well as on external factors such as the position, shape, and size of HRBs. It can be concluded that our proposed method using the relationship between the azimuth angle of the sensor and the relief displacement of the building has several distinct advantages over the CNN-based approach. E.g. the CNN performance considerably relies on the availability of a large number of training data. In addition, quantitative evaluation showed an accuracy improvement rate of at least 30% in intersection over union and F1 score compared to the object-based benchmark models. Eventually, our proposed method allows to evaluate the performance of each image individually, which helps to identify the scenarios where a certain method works best.