Improved change monitoring using an ensemble of time series algorithms

Abstract

Listen

Translate article

An ensemble of time series algorithms improves land change monitoring. The methodology combines the Continuous Change Detection and Classification (CCDC; Zhu & Woodcock, 2014) and Cumulative Sum of Residuals (CUSUM) algorithms for break detection and the Chow Test (Chow, 1960) for removing false positives (or breaks in time series not representing land change). The algorithms included are based on fundamentally different approaches to change detection and therefore offer unique advantages. The ensemble, or the combination of the three algorithms, was applied to 3 Landsat scenes in the United States and the results were assessed based on their ability to correctly discern structural breaks from stable time periods. The CUSUM test was shown to detect significant breaks 84.18% of the time and the Chow Test correctly removed breaks in 87.4% of the breaks analyzed. The ensemble produced results with lower frequency of errors of omission and commission (Type-I and Type-II errors) than a single algorithm approach. These results indicate that using a combination of break detection algorithms can be an improvement over typical approaches that utilize only one algorithm.