Automatic Building Extraction using Kernel-based Deep Learning Approach from VHR Imagery

Abstract

Monitoring and analyzing the rapidly changing and growing cities in terms of buildings has become an important demand today. Deep learning approach has been widely used recently in the automatic extraction of buildings, which are important inputs for smart city systems. The recent studies demonstrate that the deep learning approaches  greatly improves the accuracy of building extraction from the high-resolution images. The purpose of the study is to investigate the performance of K-Net architecture for building extraction from VHR imagery. In this context, The Wuhan University (WHU) Aerial Building Dataset was used for training, validation and testing. The outcomes of the study demonstrate that the extraction of buildings based on deep learning architectures provides sufficient results with 98.17 % Accuracy, 92.29 % Precision, 91.20 % Recall, 84.74 % IoU and 91.74 % F1-Score.