Semi-automated segment generation for geographic novelty detection using edge and area metrics

Abstract

An approach to generating accurate image segments for land-cover mapping applications is to model the process as an optimisation problem. Area-based empirical discrepancy metrics are used to evaluate instances of generated segments in the search process. An edge metric, called the pixel correspondence metric (PCM), is evaluated in this approach as a fitness function for segmentation algorithm free-parameter tuning. The edge metric is able to converge to user-provided reference segments in an earth observation mapping problem when adequate training data are available. Two common metaheuristic search functions were tested, namely particle swarm optimisation (PSO) and differential evolution (DE). The edge metric is compared with an area-based metric, regarding classification results of the land-cover elements of interests for an arbitrary problem. The results show the potential of using edge metrics, as opposed to area metrics, for evaluating segments in an optimisation-based segmentation algorithm parameter-tuning approach.