The potential of empirical mode decomposition to evaluate hydrological model simulations

We explore the potential of the empirical mode decomposition (EMD), a signal-processing method, to evaluate hydrological model simulations. Usually, hydrological models are assessed in the time domain observing and comparing residuals on a point-to-point basis. An additional model evaluation in the frequency domain might provide useful and complementary insights about the model’s capability to reproduce dynamic system behaviour. EMD separates a signal (e.g. a soil moisture time series) into fast and slow oscillations based on a sifting process, in which subtracting the signals moving average from itself reveals the highest frequency oscillation. This allows for instance to analyse phase shifts of different signature modes (e.g. daily fluctuations) in different depths and by that to make assumptions on soil hydraulic properties such as the conductivity. Naturally, a model will always miss high-frequency components of the “real” signal as measurement devices used as model input already act as a filter of such. However, the ability to capture the lower frequency remains interesting as they include relevant hydrological processes. Advantages of EMD over traditional methods like Fourier or wavelet transform are that no prior assumptions are needed and that it works well for nonlinear or non-stationary signals.

We test the EMD method on soil moisture and matric potential time series of observations and a process-based hydrological model extracted for the same site and compare the phase shifts and spectral components. We want to test whether metrics such as the RMSE of frequency spectra help to further compare and elucidate different signals. First results underpin the potential of including EMD as a tool to quantify models from a different perspective. We observe difference in observation and model frequencies of soil water time series and can related certain intrinsic modes to hydrological processes.