Characteristics of Search Spaces for Identifying Optimum Thresholds in Change Detection Studies

Abstract

This study explores the characteristics of spectral change search spaces with a range of factors that may influence binary change detection performance—type of change-enhanced features, number of change-enhanced features, sample size, and land cover change information. An automated calibration model based on an exhaustive search technique was used to create search spaces (i.e., Kappa-threshold surfaces) using single or multiple change-enhanced images. The major characteristics of the search spaces found in this research were: (1) the Kappa-threshold surfaces using single change-enhanced images were unimodal in form but contained small "pits"; (2) the search spaces using multiple change-enhanced images were a combination of "hills," and the optimum thresholds were found either where the tops of hills met or in the middle of the top of one hill; (3) the range in ideal thresholds was typically small compared to the domain of threshold values; (4) the surface was generally skewed toward the direction of no change; (5) at least 200 samples were required to produce stable Kappa-threshold surfaces; and (6) the from-to change information yielding the lowest accuracy was generally critical to identify optimum thresholds in the calibration using all change classes. These findings will help guide the development of automated and efficient search algorithms for identifying the optimum thresholds in binary change detection.