A comparison of unsupervised segmentation parameter optimization approaches using moderate- and high-resolution imagery
Unsupervised segmentation optimization methods have been proposed to aid in selecting an “optimal” set of scale parameters quickly and objectively for object-based image analysis. The goal of this study was to qualitatively assess three unsupervised approaches using both moderate-resolution Landsat and high-resolution Ikonos imagery from two study sites with different landscape characteristics to demonstrate the continued need for analyst intervention during the segmentation process. The results demonstrate that these methods selected parameters that were optimal for the scene which varied with method, image type, and site complexity. Several takeaways from this exercise are as follows: (1) some methods do not work as intended, (2) single-scale unsupervised optimization procedures cannot be expected to properly segment all the features of interest in the image every time, and (3) many multi-scale approaches require subjectively chosen weights or thresholds or additional testing to determine those values that meet the objective. Visual inspection of segmentation results is still required in order to assess over and under-segmentation as no method can be expected to select the best parameters for land cover classifications every time. These approaches should instead be used to narrow down parameter values in order to save time.
Natural Resources and the Environment
GIScience and Remote Sensing
Taylor & Francis
Digital Object Identifier (DOI)
Grybas, Heather, Lindsay Melendy, and Russell G. Congalton. 2017. A comparison of unsupervised segmentation optimization approaches using moderate- and high-resolution imagery. GIScience and Remote Sensing. DOI: 10.1080/15481603.2017.1287238