A Hierarchical Multiscale Super-Pixel-Based Classification Method for Extracting Urban Impervious Surface Using Deep Residual Network From WorldView-2 and LiDAR Data

• Main Authors: Mengfan Wu, Xiangwei Zhao, Zhongchang Sun, Huadong Guo
• Format: Article
• Language: English
• Published: IEEE 2019-01-01
• Series: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
• Subjects: Deep residual network; impervious surface; light detection and ranging (LiDAR); multiscale; Spatial Pyramid Pooling (SPP-net); super pixel
• Online Access: https://ieeexplore.ieee.org/document/8600363/

High-resolution optical imagery can provide detailed information of urban land objects for impervious surface extraction, while airborne light detection and ranging (LiDAR) data can provide height features of land objects. Therefore, synergistic use of high-resolution imagery and LiDAR data is considered as an effective method to improve impervious surfaces extraction. In this paper, a novel hierarchical multiscale super-pixel-based classification method is proposed and applied to the urban impervious surfaces extraction from WorldView-2 and normalized digital surface model (nDSM) images derived from airborne LiDAR data. Three subsets in rural, rural-urban, and urban subsets are selected as the study areas. First, we split nonground and ground objects based on nDSM thresholds. Second, a hierarchical multiresolution segmentation method is used to generate nonground and ground super pixels. Then, we determine the multiscale input images based on the size of super pixels. Third, we construct optimal deep residual network (ResNet) and Spatial Pyramid Pooling (SPP-net) to train the model using multiscale input images. Finally, we use our deep models to predict hierarchically total super pixels in three subsets and generate the classification and impervious surfaces results. Our proposed method adopts hierarchical classification based on LiDAR nDSM height, which significantly improves the impervious surfaces extraction accuracies. Then, the deep residual network is applied further on multispectral and height fused data to extract urban impervious surfaces. Moreover, we propose an adaptive method to determine multiscale input images based on the segmentation of super pixels, which are inputs into the ResNet+SPP-net to train the deep model. Our proposed method reduces the uncertainty of multiscale input images and extracts better multiscale features. The results of the experiment show that our proposed method has a significant superiority to traditional pixel-based method and single scale method for urban impervious surfaces extraction.