Journal of Geo-information Science ›› 2020, Vol. 22 ›› Issue (10): 2051-2061.doi: 10.12082/dqxxkx.2020.200001

Previous Articles     Next Articles

Land Type Interpretation Authenticity Check of Vector Patch Supported by Deep Learning and Remote Sensing Image

GUO Zihui1(), LIU Wei1,2,*()   

  1. 1. School of Geographic Mapping and Urban Rural Planning, Jiangsu Normal University, Xuzhou 221116, China
    2. State Key Laboratory of Resources and Environmental Information Systems, Institute of Geographic Sciencesand Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China
  • Received:2020-01-01 Revised:2020-04-01 Online:2020-10-25 Published:2020-12-25
  • Contact: LIU Wei E-mail:guozihui@jsnu.edu.cn;liuw@jsnu.edu.cn
  • Supported by:
    Postgraduate Research & Practice Innovation Program of Jiangsu Province(KYCX20_2373);A Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions;Xuzhou Land and Resources Science and Technology Project(XZGTKJ2018001);Jiangsu Province Land and Resources Science and Technology Plan Project(2018054);Open Fund of National Key Laboratory of Resource and Environment Information System

Abstract:

Spatial data quality check is important to guarantee the accuracy and reliability of remote sensing data. Current spatial data quality check mainly focuses on topological relationship, attribute consistency, and data correlation, ignoring the authenticity of land type interpretation in vector patches. Therefore, this paper proposes a method of checking the authenticity of land type interpretation in vector patches using deep learning and high-resolution remote sensing images, which is designed to meet the urgent need of authenticity check for vector patch categories in major national and industrial projects. In this paper, a method of automatic sample labeling and purification is proposed, which uses the geographic location information of the original map spot to realize processing unit labels automatically and uses two different classifiers combined with cross validation to achieve automatic sample purification, to obtain a large number of high-quality labeled samples for deep learning model training. The classic deep learning model (i.e., Inception_v3) is used for transfer learning to classify segmented images into automatic scene classifications. The overall classification accuracy of the deep learning model in the study area reaches 0.934, the average precision is 0.937, and the average recall is 0.928. Based on the scene classification results using high-resolution remote sensing images, the original vector patches are overlapped and analyzed. By connecting related fields, the segmentation units with inconsistent category information are automatically identified as suspicious patches. We further verified the authenticity check method of land type interpretation of vector patches in Qingshan Quan Town and Dawu Town, Jiawang District, and Xuzhou City. The experimental results show that the precision and recall of our method for these study areas are as high as 0.925 and 0.817, respectively. The method proposed in this paper provides a useful technical support for the check of the authenticity of the land type interpretation of vector patches.

Key words: deep learning, Inception_v3, high resolution remote sensing image, spatial data quality check, vector patch, transfer learning, scene classification, authenticity