Acta Geodaetica et Cartographica Sinica ›› 2024, Vol. 53 ›› Issue (4): 700-711.doi: 10.11947/j.AGCS.2024.20220389

• Photogrammetry and Remote Sensing • Previous Articles     Next Articles

Urban impervious surface extraction based on the deep features of high-resolution remote sensing image and ensemble learning

Xuetao LI(), Pancheng WANG, Yongnian ZENG()   

  1. Center for Geomatics and Regional Sustainable Development Research, Central South University, Changsha 410083, China
  • Received:2022-06-17 Revised:2024-01-04 Published:2024-05-13
  • Contact: Yongnian ZENG E-mail:778978421@qq.com;ynzeng@csu.edu.cn
  • About author:LI Xuetao (1993—), male, master, majors in remote sensing of urban environment and application. E-mail: 778978421@qq.com
  • Supported by:
    The National Natural Science Foundation of China(42171364)

Abstract:

The effective extraction of urban impervious surface is important on the application of high-resolution remote sensing. Focusing on existing issues, an urban impervious surface extraction method is proposed based on U-Net combining with ensemble machine learning. The impervious surface areas with different density are selected as the experimental areas. Firstly, the deep features of high-resolution images are extracted by U-Net with the GF-2 multispectral data. Then, the urban impervious surface is extracted by using the ensemble learning with stacking mechanism. The experimental results show that the ensemble learning based on deep features of high-resolution remote sensing image can obtain high accuracy of urban impervious surface extraction. In the experimental areas with different density of urban impervious surface, the overall accuracy is not less than 91.66%, and Kappa is not less than 0.83. The commission error is 7.83%~9.39%; the omission error is 7.22%~14.88%. Compared with the ensemble learning, random forest and support vector machine based on image spectral features, the overall accuracy and Kappa are increased in experimental areas with relatively sparse, medium dense, dense and complex distribution of impervious surface. The commission and omission errors are significantly reduced. This indicates that the deep features can effectively improve the mapping accuracy and user accuracy of integrated learning to extract impervious surface. Compared with U-Net and SegNet, the overall accuracy and Kappa are increased by in the four experimental areas with relatively sparse, medium dense, dense and complex distribution of impervious surface. The commission errors are significantly reduced. The integration ensemble learning with deep learning can effectively improve the mapping accuracy and user accuracy of impervious surface extraction. In general, the ensemble learning based on the deep features of high-resolution remote sensing image can obtain higher accuracy of urban impervious surface extraction, which has application prospects in urban land use/cover classification.

Key words: U-Net network, ensemble learning, deep features, urban impervious surface, high spatial resolution, GF-2

CLC Number: