Building extraction from remotely sensed images with a novel deep learning architecture

The building layer, which is one of the basic elements of the map, has various architectures and is used in many mapping activities such as urban planning, real estate management, illegal building detection and disaster analysis. For this reason, it is important to access building object information quickly and automatically. In this thesis study, automatic building extraction from high resolution digital aerial images was aimed using the Massachusetts building dataset and a new deep learning architecture was designed for this purpose. In addition, building boundary information was added to the building representations in the masks in the data set by applying morphological operations, and the contribution of the building boundary information to the deep learning network's learning of the building object was investigated. The performance of the designed architecture on test data; for Overall Accuracy, Correctness, Completeness, F1-Score, Intersection over Union accuracy metrics was calculated as 91%, 80%, 73%, 76% and 62%, respectively. It has been observed that the proposed architecture has equal performance for the Overall Accuracy metric and an average of 2% lower performance for other metrics than the existing U-Net architecture in the literature. Despite this situation, the designed architecture has a significant advantage in completing the training process in 400 minutes less than the U-Net architecture. As a result, it is estimated that the designed architecture will approach the U-Net architectural performance, which is accepted in the literature, and will provide speed and convenience in terms of usability due to its advantage in training time.