Quantitative flood forecasting using multisensor data and neural networks

Accurate quantitative forecasting of rainfall for basins with a short response time is essential to predict streamflow and flash floods. Previously, neural networks were used to develop a Quantitative Precipitation Forecasting model that highly improved forecasting skill at specific locations in Pennsylvania, using both Numerical Weather Prediction output and rainfall and radiosonde data. The objective of this study was to modify the existing artificial neural network model to include the evolving structure and frequency of intense weather systems in the mid-Atlantic region of the United States for improved flood forecasting. Besides using radiosonde and rainfall data, the model also used the satellite-derived characteristics of storm systems such as tropical cyclones, mesoscale convective complex systems, and convective cloud clusters as input. The convective classification and automated tracking system was used to identify and quantify storm properties such as lift: time, area, eccentricity, and track. As in standard expert prediction systems, the fundamental structure of the neural network model was learned from the hydroclimatology of the relationships among weather systems, rainfall production and streamflow response in the study area. Here, we present results from the application of the quantitative flood forecasting model in four watersheds on the leeward side of the Appalachian mountains in the mid-Atlantic region. The areal extent of the watersheds ranges from 750 to 8700 km(2). The reduction in the mean-squared error of the peak streamflow with respect to persistence was up to 60% for the 24 h lead-time forecasts. For the 18 h lead-time forecasts, the number of successful forecasts for streamflow peaks in the upper 5% percentile was consistently above 60%, and close to 80-90%. (C) 2001 Elsevier Science B.V. All rights reserved.