Trend, seasonality, and abrupt change detection method for land surface temperature time-series analysis: Evaluation and improvement

Long-term land surface temperature (LST) variation is vital for the study of climate change and environmental monitoring. Change detection methods provide access to recovery trajectories of trend and seasonality and detect abrupt changes in LST time series, but a comprehensive evaluation of the published methods is lacking. In this study, simulated LST data with a temporal resolution of 8 days under different scenarios were used to evaluate the performance of three commonly used methods: Detecting Breakpoints and Estimating Segments in Trend (DBEST), Breaks for Additive Seasonal and Trend (BFAST), and Bayesian Estimator of Abrupt change, Seasonal change, and Trend (BEAST). The results obtained using the simulated data indicated that BEAST was the best method for decomposing LST time series into trend and seasonality (mean RMSEs were 0.28 K and 0.27 K, respectively) and for detecting abrupt changes in these two components (mean F1 scores were 0.83 and 0.95, respectively). BFAST was less robust to high-complexity data (F1: 0.56 and 0.52, RMSE: 1.34 K and 1.46 K). 0.91 K and 1.29 K). DBEST is recommended to capture component details because it yields the least generalized output (F1 for trend: 0.37, RMSE: 0.64 K and 1.37 K). Both BFAST and DBEST exhibited reduced accuracy when the time-series data has long-lasting continuous missing data. An application using the 20-year MODIS LST time series supports the results obtained using the simulated data. BEAST exhibited the highest detection accuracy for land cover change (13 correct detections among 15 true changes), followed by DBEST (9) and BFAST (7). All three methods were ineffective for detecting low-magnitude disturbances: wildfires, heatwaves, and cold spells due to their low intensity or short duration. To reduce the non-negligible commission error of BEAST, this study proposes an improved BEAST, which eliminates the false breakpoints in BEAST using a set of thresholds. Compared with BEAST, the user accuracy of the improved BEAST was significantly increased by 13.9% in the simulated data, resulting in an F1 increase of 0.04, and 15 false breakpoints were eliminated among 53 detected disturbances in the MODIS LST time series. This study outlines commonly used change detection methods and offers guidance for choosing the optimal method to detect changes in LST time series. Furthermore, suggestions on the determination of parameters and false breakpoints elimination in the improved BEAST enable it more practical. • Three methods (i.e., DBEST, BFAST, BEAST) were evaluated for detecting LST changes. • BEAST performed best in detecting abrupt changes in LST time series. • BEAST accurately decomposed LST time series into trend and seasonality. • BEAST is practical in detecting land cover changes using LST time series. • The improved BEAST could significantly reduce the number of false breakpoints.