Object-based land surface segmentation scale optimisation: An ill-structured problem

Abstract

The segmentation of continuous land surfaces into morphologically representative objects has received increasing attention in recent years. Multi-resolution segmentation (MRS) has been shown to delimit morphological boundaries accurately and the use of unsupervised data-driven local variance (LV) based methods for detecting characteristic levels of scale parameter (SP) in land surfaces has been established. However, whether the detected SPs accurately delimit target morphological features is unclear. This study illustrates that multi-resolution land surface segmentation SP optimisation is an ill-structured problem that can be divided into subsets of well-structured problems by defining “conceptual” morphometric primitive (henceforth referred to as morphometric primitive) conditions. A new methodology is proposed where an ensemble of unsupervised data-driven LV-based SP optimisation techniques are implemented to evaluate objects against each of the morphometric primitive conditions. To construct an ensemble of SP optimisation techniques, an established method, estimation of scale parameter 2 (ESP 2), is reviewed and existing LV concepts expanded to include two new SP optimisation techniques, namely: object boundary local variance (OBLV) and local variance ratio (LVR). Agreement between the different SP optimisation techniques are indicative of SPs where the probability that morphometric primitives are delimited is higher than when applying SPs selected by single SP optimisation approaches.