Statistically based attribution and prediction of the frequency of flood events across the U.S. Midwest

There is growing empirical evidence that many river basins across the U.S. Midwest have been experiencing an increase in the frequency of flood events over the most recent decades. Albeit these detected changes are important to understand what happened in our recent past, they cannot be directly extrapolated to obtain information about possible future changes in the frequency of flood events. To address this problem, I adopt a methodology which consists in first establishing relationships between the frequency of flood events and its drivers, and then using predictions of the identified drivers to obtain the prediction of the quantity of interest. Focusing on 287 U.S. Geological Survey gauging stations across the U.S. Midwest, I create the time series of the seasonal number of flood events by applying a peak-over-threshold approach to the daily discharge records. I use Poisson regression statistical models to attribute seasonal changes in the frequency of flood events to seasonal changes in five covariates: precipitation, temperature, antecedent wetness conditions (as antecedent cumulated precipitation), agriculture and population (as a proxy for urbanization). Results indicate that precipitation and antecedent wetness conditions are the strongest predictors, with the role of the latter increasing as the threshold for the event identification is lowered. Temperature is an important predictor only in the northern Great Plains during spring, where snow-related processes are most relevant. Population and agriculture are less important compared to the climate predictors. Building on these outcomes, I use predictions of precipitation and temperature at different time horizons as input to the statistical models to predict the frequency of flood events. With respect to the sub- seasonal and seasonal lead time, I find that the prediction skill of the frequency of flood events is overall low. The prediction skill slightly improves when considering the shortest lead times (i.e., from 0.5 to 2.5 months) during spring in the Northern Great Plains, where antecedent wetness conditions play an important role in influencing the generation of floods. Decadal predictions also do not perform well, but downscaling and bias correcting the input variables improves the predictions. I find that the low accuracy of the predictions from the seasonal to the decadal time horizon is attributed to the input of the models rather than to the models themselves. Relatively to the long term projection analysis, I find that the frequency of flood events during the 21st century is projected to increase during spring at most of the analyzed gauging stations, with larger changes in the center area of the U.S. Midwest and regardless of the flood threshold value. My findings also point to a projected increasing number of flood events during the winter, especially in the stations in the southern and western part of the domain (Iowa, Missouri, Illinois, Ohio, Indiana and Michigan). A marked change in the frequency of flood events is not projected for the summer and fall. The framework presented in this thesis is very useful to understand which are the main climatic and land use/land cover variables regulating the frequency of flood events in gauged water basins, and it can be further improved by including other drivers. If forced with reliable climate predictions, these statistical models may provide important insights on the future conditions of flood events both at the local and regional scale, representing a fundamental tool for flood mitigation plans.