A Multiscale Segmentation Framework for Uncompleted Building Footprint Extraction from Remote Sensing Imagery

Building extraction from aerial and satellite images has been playing a significant role in urban development. The deep neural networks' automatic feature extraction capability provides the ease to infer building footprint from remote sensing imagery with greater accuracy. However, designing a classifier that can infer salient features such as the building category remains a challenging task. This article proposes a parameter- efficient, multiscale segmentation network for uncompleted building structure extraction. The proposed network was designed based on the architectural framework of the inception module that allows feature learning at multiscale level. Our proposed framework consists of three types of modules known as the subnets that form the encoder, the decoder, and the bottleneck of the network that allow multiscale semantic learning for segmentation application. The experimental result indicates that our proposed network required less training time to attain the best accuracy than state-of-the-art networks. We also present an approach to determine the precise geographical coordinates of the uncompleted building segment's using the georeferencing technique.