Journal of Geo-information Science ›› 2019, Vol. 21 ›› Issue (8): 1265-1274.doi: 10.12082/dqxxkx.2019.190071

Previous Articles     Next Articles

Comparison of Evaluation Indexes for Supervised Segmentation of Remote Sensing Imagery

LI Zeyu,MING Dongping(),FAN Yinglin,ZHAO Linfeng,LIU Simin   

  1. School of Information Engineering, China University of Geosciences (Beijing), Beijing 100083, China
  • Received:2019-02-21 Revised:2019-04-05 Online:2019-08-25 Published:2019-08-25
  • Contact: MING Dongping
  • Supported by:
    Innovation and Entrepreneurship Training Program for College Students of China University of Geosciences, Beijing(201811415118)


The improving spatial resolution of remote sensing imagery provides more abundant information for users, but also increases the difficulty of accurate and efficient extraction of information. Image segmentation is a fundamental step in target extraction from remote sensing imagery. The quality of image segmentation directly affects the accuracy of information extraction from high spatial resolution remote sensing imagery. With various segmentation algorithms, image segmentation evaluation has become one of the research focuses in remote sensing information extraction and target recognition. Aiming at the issue of typical target recognition and from an experimental perspective, this paper compared and analyzed in detail eight representative supervised segmentation indexes: Area Fitness Index (AFI), Similarity of Size (SimSize), Relative Area in Sub-Object (RAsub), Quality Rate (QR), Euclidian Distance 1 (ED1), Euclidian Distance 2 (ED2), area discrepancy index (ADI) and Distance-Based Measure (D). Firstly, we employed a series of experiments to calculate the difference between segmentation image and reference image by using different segmentation methods, then discussed the calculation results and evaluated the advantages and disadvantages of the different supervised segmentation evaluation indexes. The comparison results show that the AFI, ED1, ED2 and D could representatively and synthetically assess the segmentation quality. Further, based on the indexes analysis result, this paper proposed a comprehensive evaluation scheme for remote sensing imagery supervised segmentation evaluation by using weighted calculation of the four representative indexes. Through the experiment of comprehensive evaluation, we conclude that the effect of simple shape objects (such as the baseball field and oil tank) by using various segmentation methods is generally ideal. When the shape of objects is complex and the contour is blurred, the accuracy of image segmentation will be sensitive to the segmentation result to some extent. Meanwhile, the effect of segmentation methods (e.g., Otsu-2D, Regional Growth, and Mean Shift) are in general better than the other segmentation methods (e.g., split and merge, maximal entropy, and fuzzy threshold). In addition, the experiments also suggest that the comprehensive method is helpful for the scientific selection of segmentation methods and it can improve the efficiency of information extraction from high spatial resolution imagery. Finally, this paper systematically analyzed the experimental results from the two aspects of evaluation index and segmentation method, and pointed out the existing problems and development trends of image supervised segmentation evaluation.

Key words: high resolution remote sensing image, image segmentation quality, supervisory evaluation method, target recognition, GeOBIA, empirical discrepancy method