A hierarchical scale setting strategy for improved segmentation performance using very high resolution images

Abstract

Land use/land cover (LULC) classification is a specific implementation to define terrain features to the closest real world object. Object-based image analysis (OBIA) has been proved to improve classification accuracy, particularly for very high resolution remotely sensed images. In this study, multiresolution segmentation algorithm was utilized in the image segmentation process using a pan-sharped Qickbird-2 image. Segmentation scales were determined by widely-used estimation of segmentation parameter (ESP-1) tool that produces rate of change graph (LV-RoC) in terms of local variance of the image. In this study, the LV-RoC graph of the image was evaluated to determine optimal scale values ranging from fine to coarse levels. An attempt was made to estimate optimal scale parameter for an image considering not only single-scales but also multi-scales for an image using a hierarchical scale setting strategy. Nearest neighbour classifier was used on single-scale segmented images and fuzzy classifier employing membership functions was used on multi-scale segmented image. Equal numbers of pixels for each class were randomly selected to estimate accuracy metrics (i.e. overall accuracy and kappa coefficient). The differences in classifier performances (? 6%) were statistically significant according to McNemar's test. It was found that the proposed strategy has a great potential for LULC classification using very high resolution imagery. © 2017 Elsevier B.V., All rights reserved.