Uncertainty quantification using the particle filter for non-stationary hydrological frequency analysis

Recent changes in climate, anthropogenic activities and land-use patterns have significantly altered the hydrological cycle and thus led to the presence of non-stationarity in hydrological data series. Existing conventional approaches for hydrological frequency analysis (HFA) have commonly overlooked non-stationarity and consequently they might produce false estimates of hydrological design events. Assessing the effect of non-stationarity through uncertainty quantification is a potentially feasible approach for HFA. This paper proposed to incorporate the particle filter (PF) into HFA (here flood frequency analysis (FFA) was exemplified) (named PF-FFA) for quantifying prediction uncertainty in flood quantile estimates. The feasibility of the PF-FFA was verified through comparison with the conventional L-moment based FFA (LM-FFA) as well as the random sampling based FFA (RS-FFA) in terms of both accuracy and precision respectively using several selected evaluation indices. Furthermore, the comparison of the use of constant and varying shape parameter in the PF-FFA demonstrated that the use of constant shape parameter would deteriorate the performance in both accuracy and precision, especially for datasets showing a high-level degree of non-stationarity. Through these elaborate investigations, the PF-FFA was shown to be an effective approach when dealing with non-stationary datasets as it successfully captured the effect of non-stationarity compared to the LM-FFA and RS-FFA.