The presence of significant non-stationarity in a hydrologic time series cannot be ignored when estimating design values for future time horizons. This paper introduces a second-order non-stationary approach to pooled flood frequency analysis. The proposed approach separates the non-stationary pooled quantile function into a local time-dependent component, comprising the location and scale distribution parameters, and a regional component that can be regarded as time-independent under the assumption of second-order non-stationarity. A local trend analysis is used for identification, local significance assessment and estimation of the changes in the time-dependent components. A regional trend analysis based on a regional bootstrap-resampling algorithm is then applied for assessment of the changes at a regional scale. A Monte-Carlo experiment is used for evaluating the performance of the method in the estimation of trend magnitudes in the location and scale parameters of a non-stationary series. The model is then applied on a study catchment from a homogeneous region. The results show that ignoring even a weakly significant non-stationarity in the data series may seriously bias the quantile predicted for time horizons as near as 0-20 years in the future.