The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLII-2/W7
https://doi.org/10.5194/isprs-archives-XLII-2-W7-779-2017
https://doi.org/10.5194/isprs-archives-XLII-2-W7-779-2017
13 Sep 2017
 | 13 Sep 2017

AUTOMATIC OBJECT-ORIENTED ROUNDABOUTS EXTRCTION FROM HIGH RESOLUTION MULTISPECTRAL IMAGES

X. Li and W. Zhang

Keywords: Roundabout Extraction, Remote Sensing, Object-oriented Feature Extraction, GF-2

Abstract. Road roundabouts, a typical class of road facilities to avoid collision, are generally not directed extracted in existing road extraction methods. This paper presents a novel four-step approach for automatic vegetated roundabout extractions from high resolution multispectral satellite images, which combines object-oriented extraction, Support Vector Machine (SVM) classification and spatial relationship estimation. Firstly, after proper preconditioning, the vegetated roundabouts are extracted by object-oriented extract algorithm in ENVI with rules that simultaneously taking area, roundness and vegetation index (NDVI) into consideration. After a certain number of experiments, the set of three items’ thresholds can be found, which may stand as the general rules for vegetated roundabouts extraction in similar conditions. Next, the roads are classified using Support Vector Machine (SVM) and the outputs are several band shaped polygons. Then, the holes in road polygons will be detected by examining the topological relation in ArcGIS. Lastly, since the margin of extracted roundabout and the biggest detected hole may not strictly coincide, by comparing the distance between central points of both the extracted roundabout and the hole with the threshold, convincing determination can be made. The proposed automatic approach has been proved to have very high production accuracy that all above 85 % in each case of the test set, which is good enough for automatic vegetated roundabouts extraction from high resolution remote sensing images without manual interpretation.