A novel deep neural network for multi-scale building extraction from remotely-sensed images

Building extraction is one of the most crucial requirements of urban planning. Due to their availability and affordable cast, high-resolution remotely sensed images are often used for building extraction. Owing to their impressive performances, Deep learning techniques have attracted the attention of researchers for building extraction from high-resolution images. Nevertheless, most existing models perform poorly in recovering spatial details and discriminating buildings with various sizes and shapes. Hence, this paper proposes an improvement module to address the problems associated with multi-scale building extraction. The proposed module uses dilated convolutions to increase the receiving information area to reduce the discontinuities in the results of large buildings. Extracting large buildings using the proposed module and small buildings using the main architecture of the network has turned the proposed network into an effective method for building extraction. The results of the experiments showed that the proposed module with the IoU of 0.6495 and 0.8572 for Massachusetts and WHU data sets outperformed FCN, U-Net, USSP, and DeepLab V3+. The performance analysis of the proposed module also showed that this module was able to improve the performance of building extraction considering the IoU metric by 0.1077.