Exponentially sampling scale parameters for the efficient segmentation of remote-sensing images

Abstract

ABSTRACTScale parameter(s) of multi-scale hierarchical segmentation (MSHS), which groups pixels as objects in different size and hierarchically organizes them in multiple levels, such as the multiresolution segmentation (MRS) embedded into the eCognition software, directly determines the average size of segmented objects and has significant influences on following geographic object-based image analysis. Recently, some studies have provided solutions to search the optimal scale parameter(s) by supervised strategies (with reference data) or unsupervised strategies (without reference data). They focused on designing metrics indicating better scale parameter(s) but neglected the influences of the linear sampling method of the scale parameter they used as default. Indeed, the linear sampling method not only requires a proper increment and a proper range to balance the accuracy and the efficiency by supervised strategies, but also performs badly in the selection of multiple key scales for the MSHS of complex landscapes by unsupervised strategies. Against these drawbacks, we propose an exponential sampling method. It was based on our finding that the logarithm of the segment count and the logarithm of the scale parameter are linearly dependent, which had been extensively validated on different landscapes in this study. The scale parameters sampled by the exponential sampling method and the linear sampling method with increments 2, 5, 10, 25, and 100 that most former studies used were evaluated and compared by two supervised strategies and an unsupervised strategy. Results indicated that, when searching by the supervised strategies, the exponential sampling method achieved both high accuracy and efficiency where the linear sampling method had to balance them through the experiences of an expert; and when searching by the unsupervised strategy, multiple key scale parameters in MSHS of complex landscapes could be identified among the exponential sampling results, while the linear sampling results hardly achieved this. Considering these two merits, we recommend the exponential sampling method to replace the linear sampling method when searching the optimal scale parameter(s) of MRS.