Ensemble Kalman filtering versus sequential self-calibration for inverse modelling of dynamic groundwater flow systems

SummaryMonte-Carlo (MC) type inverse modelling techniques like the sequential self-calibration (SSC) method, can be used as a benchmark to test other inverse modelling procedures. In comparison studies MC type inverse modelling methods outperformed other inverse parameter estimation methods, but the large amount of CPU time needed can become prohibitive to use such methods. Therefore, an interest exists to develop alternative methods that perform nearly as good and need much less CPU time. In this paper Ensemble Kalman Filtering (EnKF) is promoted for the off-line calibration of transient groundwater flow models with many nodes. An augmented state vector approach is used to calibrate parameters together with the updating of the states. EnKF and SSC are compared in two calibration experiments, for a mildly heterogeneous case (σlnT2=1.0) and a strongly heterogeneous case (σlnT2=2.7). For the mildly heterogeneous case, EnKF gives very similar results to SSC, also in terms of calibrated log-transmissivity fields. For the strongly heterogeneous case, EnKF gives still similar results as SSC, although the characterisation of the log T field improves less compared to SSC. On the other hand, EnKF needed in both cases around a factor of 80 less CPU time than SSC. In addition, the performance of EnKF and SSC was compared in two prediction experiments: (1) the prediction of groundwater flow in a different flow situation (without pumping and with a different time series of recharge rate), (2) the prediction of solute transport towards a pumping well. Both in the mildly and strongly heterogeneous cases the quality of the predictions with EnKF was as good as for SSC. Given the good performance of EnKF, the strength of EnKF to include multiple sources of uncertainty (e.g., related to external forcing) and the reduced CPU time needed compared to MC type inverse modelling, EnKF seems to be an interesting candidate for the stochastic calibration of large subsurface hydrological models.