Extracting Urban Impervious Surface from WorldView-2 and Airborne LiDAR Data Using 3D Convolutional Neural Networks

Abstract

The urban impervious surface has been recognized as a key quantifiable indicator in assessing urbanization and its environmental impacts. Adopting deep learning technologies, this study proposes an approach of three-dimensional convolutional neural networks (3D CNNs) to extract impervious surfaces from the WorldView-2 and airborne LiDAR datasets. The influences of different 3D CNN parameters on impervious surface extraction are evaluated. In an effort to reduce the limitations from single sensor data, this study also explores the synergistic use of multi-source remote sensing datasets for delineating urban impervious surfaces. Results indicate that our proposed 3D CNN approach has a great potential and better performance on impervious surface extraction, with an overall accuracy higher than 93.00% and the overall kappa value above 0.89. Compared with the commonly applied pixel-based support vector machine classifier, our proposed 3D CNN approach takes advantage not only of the pixel-level spatial and spectral information, but also of texture and feature maps through multi-scale convolutional processes, which enhance the extraction of impervious surfaces. While image analysis is facing large challenges in a rapidly developing big data era, our proposed 3D CNNs will become an effective approach for improved urban impervious surface extraction.