JIANG Yao, XU Zongxue, WANG Jing

Abstract: Long-term trend is one of major characteristics of hydrological cycle. It is a basic content to accurately detect the trend of different hydrological variables for understanding the evolution processes of hydrological cycle under changing environment. Presently, there are many trend detection methods, but they are affected by many factors. Thus, the results from different methods are often uncertain and also difficult to judge which one is accurate and reliable. Therefore, the accuracy and reliability of five widely used methods for trend detection (Cumulative anomaly (CA), Linear regression (LR), Mann-Kendall (MK), Empiric mode decomposition (EMD) and Discrete wavelet transfer (DWT)) are analyzed based on three artificially generated sequences with known components and different change characteristics in this study. Three different observed annual runoff series, which respectively showed typical trend characteristics, were collected and used to further validate the performance and applicability of five methods for hydrological time series. Case studies indicated that the performance of CA method is only applicable for preliminary trend analyses. The results from MK method are similar to those from LR method, while the performance of MK method is better than that of LR method. Both of these two methods have higher accuracy for trend detection of time series with monotonous trend and insignificant periodicity, and is thus suitable for annual runoff with simple variation and large magnitude of the trend. EMD and DWT methods can accurately identify and separate the fluctuation components at different scale contained in the sequence itself, so as to avoid the influence of the periodicity, nonlinearity of time series on the trend detection. They are thus relatively accurate for the trend detection of non-monotonous hydrological time series. Comparatively, the results of DWT method are better than that of EMD method. The conclusion can provide reference for the trend detection as well as its method selection for hydrological series.

Keywords: runoff; trend detection; EMD method; discrete wavelet transfer; MK method