This thesis presents a method for calibrating regional scale hydrologic models using the upper South Saskatchewan River watershed as a case study. Regional scale hydrologic models can be very difficult to calibrate due to the spatial diversity of their land types. To deal with this diversity, both a manual calibration method and a multi-basin automated calibration method were applied to a WATFLOOD hydrologic model of the watershed. Manual calibration was used to determine the effect of each model parameter on modeling results. A parameter set that heavily influenced modeling results was selected. Each influential parameter was also assigned an initial value and a parameter range to be used during automated calibration. This manual calibration approach was found to be very effective for improving modeling results over the entire watershed. Automated calibration was performed using a weighted multi-basin objective function based on the average streamflow from six sub-basins. The initial parameter set and ranges found during manual calibration were subjected to the optimization search algorithm DDS to automatically calibrate the model. Sub-basin results not involved in the objective function were considered for validation purposes. Automatic calibration was deemed successful in providing watershed-wide modeling improvements. The calibrated model was then used as a basis for determining the effect of altering rain gauge density on model outputs for both a local (sub-basin) and global (watershed) scale. Four de-clustered precipitation data sets were used as input to the model and automated calibration was performed using the multi-basin objective function. It was found that more accurate results were obtained from models with higher rain gauge density. Adding a rain gauge did not necessarily improve modeled results over the entire watershed, but typically improved predictions in the sub-basin in which the gauge was located.

This thesis presents a method for calibrating regional scale hydrologic models using the upper South Saskatchewan River watershed as a case study. Regional scale hydrologic models can be very difficult to calibrate due to the spatial diversity of their land types. To deal with this diversity, both a manual calibration method and a multi-basin automated calibration method were applied to a WATFLOOD hydrologic model of the watershed. Manual calibration was used to determine the effect of each model parameter on modeling results. A parameter set that heavily influenced modeling results was selected. Each influential parameter was also assigned an initial value and a parameter range to be used during automated calibration. This manual calibration approach was found to be very effective for improving modeling results over the entire watershed. Automated calibration was performed using a weighted multi-basin objective function based on the average streamflow from six sub-basins. The initial parameter set and ranges found during manual calibration were subjected to the optimization search algorithm DDS to automatically calibrate the model. Sub-basin results not involved in the objective function were considered for validation purposes. Automatic calibration was deemed successful in providing watershed-wide modeling improvements. The calibrated model was then used as a basis for determining the effect of altering rain gauge density on model outputs for both a local (sub-basin) and global (watershed) scale. Four de-clustered precipitation data sets were used as input to the model and automated calibration was performed using the multi-basin objective function. It was found that more accurate results were obtained from models with higher rain gauge density. Adding a rain gauge did not necessarily improve modeled results over the entire watershed, but typically improved predictions in the sub-basin in which the gauge was located.

Weather radar has long been recognized as a potentially powerful tool for hydrological modelling. A single radar station is able to provide detailed precipitation information over entire watersheds. The operational use of radar in water resources applications, however, has been limited. Interpretation of raw radar data requires several rigorous analytical steps and a solid understanding of the technology. In general, hydrologists’ lack of meteorological background and the persistence of systematic errors within the data, has led to a common mistrust of radar-estimated precipitation values. As part of the Enhanced Nowcasting of Extreme Weather project, researchers at McGill University’s J.S. Marshall Radar Observatory in Montreal have been working to improve real-time quantitative precipitation estimates (QPEs). The aim is to create real-time radar precipitation products for the water resource community that are reliable and properly validated. The validation of QPEs is traditionally based on how well observed measurements agree with data from a precipitation gauge network. Comparisons between radar and precipitation gauge quantities, however, can be misleading. Data from a precipitation gauge network represents a series of single-point observations taken near ground surface. Radar, however, estimates the average rate of precipitation over a given area (i.e. a 1-km grid cell) based on the intensity of reflected microwaves at altitudes exceeding 1 km. Additionally, both measurement techniques are susceptible to a number of sources of error that further confound efforts to compare the two. One of the greatest challenges facing radar meteorologists is the variation in the vertical profile of reflectivity (VPR). A radar unit creates a volumetric scan of the atmosphere by emitting microwave beams at several elevation angles. As a beam travels away from the radar, its distance from ground surface increases. Different precipitation types are sampled at a number of heights (i.e. snow above the 0º C elevation and rain below it) that vary with range. The difficulty lies in estimating the intensity of precipitation at the Earth’s surface, based on measurements taken aloft. Scientists at McGill University have incorporated VPR correction techniques into algorithms used to automatically convert raw radar data into quantitative hydrological products. This thesis evaluates three real-time radar precipitation products from McGill University’s J.S. Marshall Radar Observatory in the context of hydrological modelling. The C0 radar product consists of radar precipitation estimates that are filtered for erroneous data, such as ground clutter and anomalous precipitation. The C2 and C3 radar products use different VPR correction techniques to improve upon the C0 product. The WATFLOOD hydrological model is used to assess the ability of each radar product to estimate precipitation over several watersheds within the McGill radar domain. It is proposed that using a watershed as sample area can reduce the error associated with sampling differences between radar and precipitation gauges and allow for the evaluation of a precipitation product over space and time. The WATFLOOD model is run continuously over a four-year period, using each radar product as precipitation input. Streamflow hydrographs are generated for 39 gauging stations within the radar domain, which includes parts of eastern Ontario, south-western Quebec and northern New York and Vermont, and compared to observed measurements. Streamflows are also modelled using distributed precipitation gauge data from 44 meteorological stations concentrated around the Montreal region. Analysis of select streamflow events reveals that despite the non-ideal placement of precipitation gauges throughout the study area, distributed precipitation gauge data are able to reproduce hydrological events with greater accuracy and consistency than any of the provided radar products. Precipitation estimates within the McGill radar domain are found to only be useful in areas within the Doppler range (120-km) where the radar beam is unobstructed by physiographic or man-made features. Among radar products, the C2 VPR-corrected product performed best during the greatest number of the flood events throughout the study area.

Weather radar has long been recognized as a potentially powerful tool for hydrological modelling. A single radar station is able to provide detailed precipitation information over entire watersheds. The operational use of radar in water resources applications, however, has been limited. Interpretation of raw radar data requires several rigorous analytical steps and a solid understanding of the technology. In general, hydrologists’ lack of meteorological background and the persistence of systematic errors within the data, has led to a common mistrust of radar-estimated precipitation values. As part of the Enhanced Nowcasting of Extreme Weather project, researchers at McGill University’s J.S. Marshall Radar Observatory in Montreal have been working to improve real-time quantitative precipitation estimates (QPEs). The aim is to create real-time radar precipitation products for the water resource community that are reliable and properly validated. The validation of QPEs is traditionally based on how well observed measurements agree with data from a precipitation gauge network. Comparisons between radar and precipitation gauge quantities, however, can be misleading. Data from a precipitation gauge network represents a series of single-point observations taken near ground surface. Radar, however, estimates the average rate of precipitation over a given area (i.e. a 1-km grid cell) based on the intensity of reflected microwaves at altitudes exceeding 1 km. Additionally, both measurement techniques are susceptible to a number of sources of error that further confound efforts to compare the two. One of the greatest challenges facing radar meteorologists is the variation in the vertical profile of reflectivity (VPR). A radar unit creates a volumetric scan of the atmosphere by emitting microwave beams at several elevation angles. As a beam travels away from the radar, its distance from ground surface increases. Different precipitation types are sampled at a number of heights (i.e. snow above the 0º C elevation and rain below it) that vary with range. The difficulty lies in estimating the intensity of precipitation at the Earth’s surface, based on measurements taken aloft. Scientists at McGill University have incorporated VPR correction techniques into algorithms used to automatically convert raw radar data into quantitative hydrological products. This thesis evaluates three real-time radar precipitation products from McGill University’s J.S. Marshall Radar Observatory in the context of hydrological modelling. The C0 radar product consists of radar precipitation estimates that are filtered for erroneous data, such as ground clutter and anomalous precipitation. The C2 and C3 radar products use different VPR correction techniques to improve upon the C0 product. The WATFLOOD hydrological model is used to assess the ability of each radar product to estimate precipitation over several watersheds within the McGill radar domain. It is proposed that using a watershed as sample area can reduce the error associated with sampling differences between radar and precipitation gauges and allow for the evaluation of a precipitation product over space and time. The WATFLOOD model is run continuously over a four-year period, using each radar product as precipitation input. Streamflow hydrographs are generated for 39 gauging stations within the radar domain, which includes parts of eastern Ontario, south-western Quebec and northern New York and Vermont, and compared to observed measurements. Streamflows are also modelled using distributed precipitation gauge data from 44 meteorological stations concentrated around the Montreal region. Analysis of select streamflow events reveals that despite the non-ideal placement of precipitation gauges throughout the study area, distributed precipitation gauge data are able to reproduce hydrological events with greater accuracy and consistency than any of the provided radar products. Precipitation estimates within the McGill radar domain are found to only be useful in areas within the Doppler range (120-km) where the radar beam is unobstructed by physiographic or man-made features. Among radar products, the C2 VPR-corrected product performed best during the greatest number of the flood events throughout the study area.

The Canadian Precipitation Analysis (CaPA) system has been developed by Environment and Climate Change Canada (ECCC) to optimally combine different sources of information to estimate precipitation accumulation across Canada. The system combines observations from different networks of weather stations and radar measurements with the background information generated by ECCC's Regional Deterministic Prediction System (RDPS), derived from the Global Environmental Multiscale (GEM) model. The main scope of this study is to assess the importance of weather stations when combined with the background information for hydrological modeling. A new approach to meteorological network design, considered to be a stochastic hydro-geostatistical scheme, is proposed and investigated which is particularly useful for augmenting data-sparse networks. The approach stands out from similar approaches of its kind in that it is comprised of a data assimilation component included based on the paradigm of an Observing System Simulation Experiment (OSSE), a technique used to simulate data assimilation systems in order to evaluate the sensitivity of the analysis to new observation network. The proposed OSSE-based algorithm develops gridded stochastic precipitation and temperature models to generate synthetic time-series assumed to represent the 'reference' atmosphere over the basin. The precipitation realizations are used to simulate synthetic observations, associated with hypothetical station networks of various densities, and synthetic background data, which in turn are assimilated in CaPA to realize various pseudo-analyses. The reference atmosphere and the pseudo-analyses are then compared through hydrological modeling in WATFLOOD. By comparing the flow rates, the relative performance of each pseudo-analysis associated with a specific network density is assessed. The simulations show that as the network density increases, the accuracy of the hydrological signature of the CaPA precipitation products improves hyperbolically to a certain limit beyond which adding more stations to the network does not result in further accuracy. This study identifies an observation network density that can satisfy the hydrological criteria as well as the threshold at which assimilated products outperforms numerical weather prediction outputs. It also underlines the importance of augmenting observation networks in small river basins to better resolve mesoscale weather patterns and thus improve the predictive accuracy of streamflow simulation. / May 2017