Assessing Edge and Area Metrics for Image Segmentation Parameter Tuning and Evaluation

Image segmentation algorithms allow the creation of variables that influence the attributes of the resulting objects. There is currently no set way of identifying the best segmentation parameters for a specific application. It would consequently be useful to calculate the accuracy of a particular segmentation for determining how well it can delineate an object of interest. This article assesses the use of both an area and edge metric to evaluate and quantify the ability of a segmentation algorithm to delineate objects of interest based on manually collected reference data. SPOT 5 imagery was used for the segmentation of two 2x3km study areas in the Eastern-Cape of South Africa. The aims of this study were to use the area metric to identify a scale parameter for which the ratio between oversegmentation (OS) and undersegmentation (US) yields the most favourable result and, use the edge metric to determine the accuracy to which the boundary of objects of interest is delineated. The usefulness of estimating scale parameter (ESP) and segmentation parameter tuner (SPT) are also investigated to identify the suitable scale parameter. The root mean square error (RMSE) between OS and US, was used in conjunction with the edge metric results to produce an overall accuracy value between 01. The ESP tool was found useful as a guide, but failed to identify the scale parameter yielding the most accurate results, whereas, the SPT tool succeeded in both cases to identify this scale parameter. It was concluded that the edge and area metric can be used in conjunction to, not only determine the ideal scale parameter, but also calculate the overall accuracy of the segmentation algorithm. * Corresponding author