The movement of computational power and communications capabilities onto networks of sensors in the environment through the concept of pervasive or ubiquitous computing has initiated opportunities for the delivery of ground-based data in real-time and the development of adaptive monitoring systems. Measurements of water level taken by a network ofwireless sensors called 'FloodNet' were assimilated into a one-dimensional hydrodynamic model using an ensemble Kalman filter, to create a forecasting model. The ensemble Kalman filter led to an increase in forecast accuracy of between 50% and 70% depending on location for forecast lead times of less than 4 hours. This research then focused on methods for targeting measurements in real-time, such that the power limited but flexible resources deployed by the FloodNet project could be used optimally. Two targeting methods were developed. The first targeted measurements systematically over space and time until the forecasting model predicted that the probability of the water level exceeding a pre-defined threshold was less than 5%. The second method targeted measurements based on the expected decrease in forecasted water level error variance at a validation time and location, quickly calculated for various sets of measurements by an ensemble transform Kalman filter. Estimates of forecast error covariance from the ensemble Kalman filter and ensemble transform Kalman filter were significantly correlated, with correlations ranging between 0.979 and 0.292. Targeting measurements based on the decrease in forecast error variance was found to be more efficient than the systematic sampling method. The ensemble transform Kalman filter based targeting method was also used to estimate the 'signal variance' oftheoretical measurements at any computational node in the hydrodynamic model. Furthermore, time series data, different sensors types and measurements of floodplain stage could all be taken into account either as part of the targeting process or prior to measurement targeting.

Key uncertainties within inundation modelling of storm tide overflow were investigated for two regions. A northern Bay of Bengal LISFLOOD-FP inundation model was developed from freely available data sources, and forced with a storm surge model (IIDT) hind-cast of the 2007 cyclone Sidr flood event because no quality water-level records exist. Validation showed inundation prediction accuracy, with a Root Mean Squared Error (RMSE) on predicted water-level of...., 2 rn, which was similar in magnitude to the forcing water-level uncertainty. Indeed, when observed natural variability within five key cyclone parameters was propagated through the IID-T storm surge model, extreme water-level uncertainty was found to be very high in the Bay of Bengal, and should be considered in future work (and flood risk managers). Future flood hazard mapping uncertainty is much less in the data rich UK; however, when some key uncertainties were propagated through a North Somerset LISFLOOD•FP inundation model of the 1981 historic flood, storm tide spatial variability was found to significantly affect flood risk estimates, second only to sea level rise. A new method for prescribing the still peak water-level along a coastline was developed (Method C), which characteristics the spatial variability using a relatively short record of modelled extreme water-level events, relative to a tide gauge. Good agreement (RMSE 36 cm) was found between Method C predicted water-levels and tide gauge observations for two historic flood events in East Anglia (1953 and 2007). Furthermore, remotely sensed storm tide observations along the North Somerset coast indicated the accuracy of Method C between tide gauge observations; however, fine-scale wave and bathymetry effects need to be resolved for accurate coastal flood risk estimates in the UK. Indeed, the quantification of uncertainty, and the characterisation of natural variability, is necessary for a robust flood risk prediction.

This thesis focuses on integrating rainfall-runoff modelling with a mesoscale numerical weather prediction (NWP) model to make real-time flood forecasts at the catchment scale. Studies carried out are based on catchments in Southwest England with a main focus on the Brue catchment of an area of 135 km2 and covered by a dense network of 49 rain gauges and a C-band weather radar. The studies are composed of three main parts: Firstly, two data mining issues are investigated to enable a better calibrated rainfall-runoff model for flood forecasting. The Probability Distributed Model (PDM) is chosen which is widely used in the UK. One of the issues is the selection of appropriate data for model calibration regarding the data length and duration. It is found that the information quality of the calibration data is more important than the data length in determining the model performance after calibration. An index named the Information Cost Function (ICF) developed on the discrete wavelet decomposition is found to be efficient in identifying the most appropriate calibration data scenario. Another issue is for the impact of the temporal resolution of the model input data when using the rainfall-runoff model for real-time forecasting. Through case studies and spectral analyses, the optimal choice of the data time interval is found to have a positive relation with the forecast lead time, i.e., the longer is the lead time, the larger should the time interval be. This positive relation is also found to be more obvious in the catchment with a longer concentration time. A hypothetical curve is finally concluded to describe the general impact of data time interval in real-time forecasting. The development of the NWP model together with the weather radar allows rainfall forecasts to be made in high resolutions of time and space. In the second part of studies, numerical experiments for improving the NWP rainfall forecasts are carried out based on the newest generation mesoscale NWP model, the Weather Research & Forecasting (WRF) model. The sensitivity of the WRF performance is firstly investigated for different domain configurations and various storm types regarding the evenness of rainfall distribution in time and space. Meanwhile a two-dimensional verification scheme is developed to quantitatively evaluate the WRF performance in the temporal and spatial dimensions. Following that the WRF model is run in the cycling mode in tandem with the three-dimensional variational assimilation technique for continuous assimilation of the radar reflectivity and traditional surface/ upperair observations. The WRF model has shown its best performance in producing both rainfall simulations and improved rainfall forecasts through data assimilation for the storm events with two dimensional evenness of rainfall distribution; while for highly convective storms with rainfall concentrated in a small area and a short time period, the results are not ideal and much work remains to be done in the future. Finally, the rainfall-runoff model PDM and the rainfall forecasting results from WRF are integrated together with a real-time updating scheme, the Auto-Regressive and Moving Average (ARMA) model to constitute a flood forecasting system. The system is tested to be reliable in the small catchment such as Brue and the use of the NWP rainfall products has shown its advantages for long lead-time forecasting beyond the catchment concentration time. Keywords: rainfall-runoff modelling, numerical weather prediction, flood forecasting, real-time updating, spectral analysis, data assimilation, weather radar.

The need for reliable, easy to set up and operate, hydrological forecasting systems is an appealing challenge to researchers working in the area of flood risk management. Currently, advancements in computing technology have provided water engineering with powerful tools in modelling hydrological processes, among them, Artificial Neural Networks (ANN) and genetic algorithms (GA). These have been applied in many case studies with different level of success. Despite the large amount of work published in this field so far, it is still a challenge to use ANN models reliably in a real-time operational situation. This thesis is set to explore new ways in improving the accuracy and reliability of ANN in hydrological modelling. The study is divided into four areas: signal preprocessing, integrated GA, schematic application of weather radar data, and multiple input in flow routing. In signal preprocessing, digital filters were adopted to process the raw rainfall data before they are fed into ANN models. This novel technique demonstrated that significant improvement in modelling could be achieved. A GA, besides finding the best parameters of the ANN architecture, defined the moving average values for previous rainfall and flow data used as one of the inputs to the model. A distributed scheme was implemented to construct the model exploiting radar rainfall data. The results from weather radar rainfall were not as good as the results from raingauge estimations which were used for comparison. Multiple input has been carried out modelling a river junction with excellent results and an extraction pump with results not so promising. Two conceptual models for flow routing modelling and a transfer function model for rainfall-runoff modelling have been used to compare the ANN model's performance, which was close to the estimations generated by the conceptual models and better than the transfer function model. The flood forecasting system implemented in East Anglia by the Environment Agency, and the NERC HYREX project have been the main data sources to test the model.

Existing methods used in the prediction of flood magnitudes likely to be experienced in British rivers are examined and reasons for prediction inaccuracies as well as problems of interpretation which may arise due to the use of multiple regression techniques, are discussed. Data from over five hundred river gauging stations in the British Isles are analysed in the development and calibration of a new causeeffect oriented rainfall-runoff model, and a new parameter is identified as representing the efficiency with which a catchment discharges its flood waters. Large errors and inconsistencies as well as a serious lack of range are found in the present soil classification system being used in Britain and improvements are suggested. A major new finding is that flood magnitudes generally increase in proportion to rainfall magnitude raised to a power of 1.3 and this relationship has now given considerable insight into the regional differences which have been observed in flood magnitude related to frequency of occurrence. Relationships describing the particular effects of the presence of lakes, urban development, and chalk areas, have also been obtained, and the thesis concludes with a survey of each region of the British Isles and examines any evidence of regional trends in floods behaviour and whether any abnormally large floods have been recorded in the past.

A simple four parameter conceptual model of catchment response has been developed with the object of modelling only isolated storm events and thus providing a useful tool for flood estimation. Most previous conceptual modelling has been concerned with fitting continuous long-term records in which flood events are small in number. The isolated event model (IEM) is concerned only with such flood events. At its present stage of development, the isolated event model is more suited to steep upland gauged catchments. In essence, the model is fitted to such catchments by using an automatic computer optimisation technique to find those optimum values of the model parameters which enable the model to reproduce existing isolated events to an acceptable degree of accuracy. Design rainstorms may be routed through a fitted model to yield design flood hydrographs. After considerable development work on a sequence of model structures and strategies of optimisation, the IEM4 model has been found acceptable on the proving data used, consisting of 500 events from 21 catchments, and a two stage optimisation strategy has been adopted. A further version of the model, incorporating a number of improvements, was applied exhaustively to individual events from the River Ray catchment. An interesting idea that emerged was the concept of the augmented hydrograph. The latter half of this thesis is devoted to expanding this concept, incorporating evidence from other researchers, and making comparisons with the standard unit hydrograph method of analysis of isolated events.

Methods of estimating floods with return periods of up to one hundred years are reasonably well established, and in the main rely on extrapolation of historical flood data at the site of interest. However, extrapolating the tails of fitted probability distributions to higher return periods is very unreliable and cannot provide a satisfactory basis for extreme flood estimation. The probable maximum flood concept is an alternative approach, which is often used for critical cases such as the location of nuclear power plants, and is viewed as a consequence of a combination of a probable maximum precipitation with the worst possible prevailing catchment conditions. Return periods are not usually quoted although they are implicitly thought to be of the order of tens of thousand of years. There are many less critical situations which still justify greater flood protection than would be provided for an estimated one-hundred year flood. There is therefore a need for techniques which can be used to estimate floods with return periods of up to several thousand years. The predictive approach adopted here involves a combination of a probabilistic storm transposition technique with a physically-based distributed rainfall-runoff model. Extreme historical storms within a meteorologically homogeneous region are, conceptually, moved to the catchment of interest, and their return periods are estimated within a probabilistic framework. Known features of storms such as depth, duration, and perhaps approximate shape will, together with catchment characteristics, determine much of the runoff response. But there are other variables which also have an effect and these include the space-time distribution of rainfall within the storm, storm velocity and antecedent catchment conditions. The effects of all these variables on catchment response are explored.

A model for estimating seasonal flood risks which uses flow readings which are equally spaced in time is presented in this thesis. The model is referred to as the Spectral Model. This model can be used to estimate the probability of at least 1 exceedance of given critical levels. The model is based on the Rice distribution for peaks of a Gaussian stochastic process, whose parameters are associated with the spectral moments of the process. In the simpler form of the model, peaks are assumed independent. Simulation results obtained using realisation of Gaussian AR(1) processes indicated that the estimates of the risks using Spectral Model are less biased than those obtained from the EV1 and the POT Model, especially for higher critical levels. A modification which removes the assumption that peaks are independent using the multi-fold integrals of Gupta and Moran is also considered. Gupta's method assumes that the correlation between peaks at any lag are equal to the first autocorrelation. The Monte Carlo simulation of Moran has no such restriction on the autocorrelations but may not converge. The Model was applied to River Greta, a small catchment in County Durham in the North of England.

A stochastic rainfall-runoff model has been developed to generate synthetic series of floods, which are routed through idealised channel-flood plain configurations using a hydraulic flood routing model. Results are presented which show the effect of varying six geometrical parameters which are thought to be important in the transformation of flood frequency curves.

Two models for estimating the short term risk of a flood exceeding some critical flow, which take account of the season and prevailing conditions have recently been published. The model proposed by Ettrick (1986) is based on conditional probability distributions, while Smith and Karr (1986) relate the rate of exceedence of the critical level of interest to relevant covariates. Both models are fitted to a 1000 year synthetic data set, to compare the results with empirically derived immediate and 30 day ahead risk estimates. After some modifications to the Smith and Karr model, both models demonstrate reasonable accuracy. A second comparison is then made using summer data from a U.K. catchment. The results demonstrate the sensitivity of the risk to the prevailing conditions at the beginning of the period of interest. The assumptions, data requirements and accuracy of the models are compared and discussed. The Ettrick model is chosen for further consideration, given that this model is based on a precipitation threshold, whilst the Smith and Karr model is based on a flow threshold, and the data record is longer for the former. The Ettrick model is then applied to two other U.K. catchments to give all year flood risk estimates. These cover the immediate, 7 day and 30 day ahead time periods. For the immediate flow risk estimates, the importance of snow on the catchment to the levels of flood risk is highlighted. In the case of the 7 day ahead estimates, the significance of the snow is reduced. Given this latter result, the 30 day ahead estimates are not conditional on snow, but still highlight a strong seasonality in the flood risk. Application of the Ettrick model is shown to imply the variable source area concept of runoff production. This may not be the dominant runoff production mechanism on certain catchments, and as such, restricts applicability of the Ettrick model.