Improving time series anomaly detection based on exponentially weighted moving average (EWMA) of season-trend model residuals

摘要

Continuous anomaly detection in satellite image time series is important for studying spatial-temporal processes of land cover changes. In another study, we proposed a method based on Z-scores of Season-Trend model Residuals (Z-STR), which can continuously detects anomaly regions in image time series. However, Z-STR only detects anomalies with large shifts but cannot detect anomalies with small or gradual shifts. To improve the effectiveness of continuous anomaly detection, in this study, we propose a new method based on Exponentially Weighted Moving Average (EWMA) of Season-Trend model Residuals (EWMA-STR). Experiment for detecting spatial-temporal anomaly regions caused by severe flooding validated the better performance of EWMA-STR than Z-STR for time series anomaly detection. By using information of all historical time series data, EWMA-STR can detects anomalies with small or gradual shifts as well as anomalies with large or abrupt shifts.