A Genetic Algorithm Approach to Moving Threshold Optimization for Binary Change Detection

Abstract

This study investigated the use of a genetic algorithm (GA) approach, a widely used optimization method, to identify optimum thresholds for remote sensing-based binary change detection. Automated GA-based calibration models using a moving threshold window (MTW) were developed and tested using a case study. Two sets of the bi-temporal QuickBird imagery were used to evaluate the new optimization models. The GA-based models using MTW were free from the assumption of symmetry of thresholds for difference- or ratio-type of change-enhanced images, unlike traditional binary change detection methods, allowing more flexibility and efficiency in selecting optimum thresholds. Exhaustive search techniques using symmetric threshold window (STW) and MTW were evaluated for comparison. The stability of the GA-based models in terms of accuracy variation was also examined. The GA-based calibration models successfully identified optimum thresholds without a significant decrease in accuracy. The GA-based models using MTW outperformed the GA-based model using STW in both calibration and validation, revealing that optimum thresholds tended to be asymmetric. Multiple change-enhanced images generally resulted in better performance than single change-enhanced images based on the GA-based models.