A dynamical statistical framework for seasonal streamflow forecasting in an agricultural watershed

Abstract

The state of Iowa in the US Midwest is regularly affected by major floods and has seen a notable increase in agricultural land cover over the twentieth century. We present a novel statistical-dynamical approach for probabilistic seasonal streamflow forecasting using land cover and General Circulation Model (GCM) precipitation forecasts. Low to high flows are modelled and forecast for the Raccoon River at Van Meter, a 8900 km2 catchment located in central-western Iowa. Statistical model fits for each streamflow quantile (from seasonal minimum to maximum; predictands) are based on observed basin-averaged total seasonal precipitation, annual row crop (corn and soybean) production acreage, and observed precipitation from the month preceding each season (to characterize antecedent wetness conditions) (predictors). Model fits improve when including agricultural land cover and antecedent precipitation as predictors, as opposed to just precipitation. Using the dynamically-updated relationship between predictand and predictors every year, forecasts are computed from 1 to 10 months ahead of every season based on annual row crop acreage from the previous year (persistence forecast) and the monthly precipitation forecasts from eight GCMs of the North American Multi-Model Ensemble (NMME). The skill of our forecast streamflow is assessed in deterministic and probabilistic terms for all initialization months, flow quantiles, and seasons. Overall, the system produces relatively skillful streamflow forecasts from low to high flows, but the skill does not decrease uniformly with initialization time, suggesting that improvements can be gained by using different predictors for specific seasons and flow quantiles.