Nonstationary flood-frequency analysis to assess effects of harvest and cover type conversion on peak flows at the Marcell Experimental Forest, Minnesota, USA

Forest harvest, climate change, and their interaction can alter catchment peak discharges, which have ramifications for channel geomorphology and water quality. Catchments in the boreal-temperate transition zone may be especially vulnerable to these factors. We developed a new approach to peak flow analysis using two long-term, paired catchment experiments in that landscape at the Marcell Experimental Forest (MEF) in north-central Minnesota. We investigated how forest cover change affects the magnitude, occurrence probability, event decoupling, and seasonal decoupling of annual maximum peak flows. Event decoupling is when the annual maximum flows on control and treatment catchments occur in response to different events within a given year. Seasonal decoupling is when the annual maximum flows on control and treatment catchments occur in different seasons within a given year. Commonly used statistical methods (e.g., ANCOVA) support inferences only for effects on peak flow magnitude, disregarding effects on event occurrence probability. Statistical stationarity (i.e., time-invariant parameters) is generally assumed within the ANCOVA analysis, despite temporal changes in climatic and land cover controls. Further, use of these methods requires event coupling, which is not always valid for annual maximum flows in catchments with mixed precipitation regimes. To address these limitations, we developed new nonstationary flood-frequency analysis methods with Bayesian parameter estimation to analyze two harvesting experiments, including conversion from deciduous to coniferous species. We compared results from these models to traditional least-squares based methods (e.g., ANCOVA). We assessed event and seasonal decoupling using logistic regression with Bayesian parameter estimation. There was no effect of harvesting on the annual maximum flow according to ANCOVA and generalized least squares models, but these methods were not reliable due to the presence of outliers, and both event and seasonal decoupling. Flood-frequency results from one clearcutting experiment showed increases in the annual maximum flow across nearly all return intervals, with 80–85% confidence for peak discharges between the 10 and 50-year peak flow. The 50-year return interval peak increased from an expected 17 cfs to 34 cfs after harvest. Harvesting induced substantial event decoupling in the catchment pairs, and seasonal decoupling in one catchment pair. As forest cover regenerated, decoupling probability decreased. Our coupling analysis results indicated that control and treatment catchments generated and processed flood peaks differently after harvest, and explained the poor fit of the linear models for which we assumed event coupling. Our study calls into question several key assumptions of traditional linear models in the analysis of paired-catchments given certain conditions. For example, where forest harvesting can affect the season of annual maximum peaks, and thus the phase of the generating precipitation event (snowmelt versus rainfall), it changes the relationship between the catchment pair in a way not accounted for in traditional linear models. We advocate the use of probabilistic methods that can incorporate nonstationarity and are robust to different mechanisms of catchment hydrologic change in response to forest harvest.