Journal of Geo-information Science ›› 2019, Vol. 21 ›› Issue (9): 1430-1443.

### Comparing Supervised and Unsupervised Segmentation Evaluation Methods for Extracting Specific Land Cover from High-Resolution Remote Sensing Imagery

ZHANG Yindan1,2,WANG Miaomiao1,LU Haixia1,LIU Yong1,*()

1. 1. College of Earth and Environmental Sciences, Lanzhou University, Lanzhou 730000, China
2. Center for Geospatial Analytics, North Carolina State University, Raleigh NC 27606, USA
• Received:2018-12-04 Revised:2019-05-21 Online:2019-09-25 Published:2019-09-24
• Contact: LIU Yong E-mail:liuy@lzu.edu.cn
• Supported by:
National Natural Science Foundation of China(41271360);Fundamental Research Funds for the Central Universities(LZUJBKY-2016-248)

Abstract:

Geographic Object-Based Image Analysis (GEOBIA), as a new paradigm, can achieve higher accuracy than pixel-based image analysis for high-resolution remote sensing imagery. Image segmentation plays an important role throughout the course of the GEOBIA. Multi-resolution segmentation (MRS) algorithm is widely used to segment an image into meaningful objects. It is a bottom-up region merging process integrated into eCognition software, which includes three key parameters (i.e. scale, shape, and compactness) to determine the size and boundary of the image objects. As a key step of GEOBIA, how to select the appropriate segmentation parameter values in MRS remains a challenge, and has an important influence on the subsequent segmentation and classification results. Previous studies focused on segmentation parameters optimization using either unsupervised or supervised methods. However, which method (i.e. unsupervised and supervised) is more suitable for analyzing specific land covers of high-resolution remote sensing imagery is still underexplored. To close the gap, we compared the optimal segmentation and classification results based on unsupervised and supervised methods. Meanwhile, we tested with three land cover types (i.e., farmland, residential area, and pond), and choose two representative segmentation parameter optimization methods for unsupervised and supervised methods, including Estimation of Scale Parameter 2 (ESP2) and Euclidean Distance 2 (ED2). The multi-source high-resolution remote sensing data (i.e., Quickbird, Worldview-2, and ALOS) were used to validate the robustness and generalizability of the unsupervised and supervised methods. We found that for a certain land cover category, the boundary of segments obtained by the supervised method seemed more consistent with the geo-objects in real world, while the optimized parameters were too large to extract the small area of geo-objects for the unsupervised method, leading to the lower classification accuracy. The supervised method performed better in analyzing the segmentation parameters optimization of the geo-object using the referenced data of land cover categories, at the same time which can break out the effect of different landscape and image resolution via reference dataset optimization, while the unsupervised method depended on image features, the artificial visual interpretation and had lower recognition accuracy due to higher subjectivity and uncertainty in land cover category identification. Although the overall classification accuracy is still above 90.08%, the omission rate is 1.43~4.65 times to the supervised method. Comparing the two methods, the supervised method obtained the optimal segmentation results with higher efficiency and accuracy using fewer segment datasets in both segmentation and classification results than the unsupervised method. Our findings suggest that the supervised method is more suitable for mapping specific land covers with high-resolution remote sensing imagery.