A comparative study of segmentation quality for multi-resolution segmentation and watershed transform

Abstract

Object-Based Image Analysis (OBIA) has gained swift popularity in remote sensing area mainly due to the increasing availability of very high resolution imagery. Image segmentation is a major step within OBIA process. Image segmentation quality remarkably influences the subsequent image classification accuracy. It is necessary to implement advanced and robust methods to increase image segmentation quality that is generally measured by several accuracy metrics including Area Fit Index (AFI and Quality Rate (Qr). In this study, two widely-used segmentation algorithms, namely region-based multi-resolution segmentation and edge-based watershed transform were applied to a very high resolution imagery acquired by VorldView-2 sensor to evaluate and compare their performance in terms of segmentation quality metrics. Totally five segmentation goodness metrics, namely under-segmentation, over-segmentation, root means square, AFI and Qr were applied through the manually digitized reference objects available on the imagery. ENVI and eCognition Developer software packages were used to perform watershed transform and multi-resolution segmentation algorithms, respectively. Nearest neighbor classification method was applied and related accuracy assessment was conducted in two software platforms. Results showed that multi-resolution segmentation was superior (about 18% higher in terms of AFI) compared to watershed transform in the delineation of segments of reference objects. Also, higher classification accuracies (about 5%) were achieved by the use of multi-resolution segmentation.