Environmental informatics - long-lead flood forecasting using Bayesian neural networks

Neural networks (NNs) are especially useful in exploratory data analysis to uncover and, or elucidate empirical relationships among data. Parameter estimation, the so-called "training" of neural networks is a variation of standard maximum likelihood estimation, whereby the optimal set of model parameters (the NN weights) maximizes the fit to the calibration (training) data set. In our previous applications of neural networks in hydrometeorology, we focused on the development of complex architectures of neural networks adapted to the characteristics of the available data (multisensor, multiresolution mix of ground-based and satellite observations). These architectures consist of large structures of simpler networks built to embody clearly defined hypothesis of functional relationships that are consistent with the underlying physical processes (rainfall and flood forecasting, wind, temperature and moisture profiles in the atmosphere, temporal evolution of cloud and storm morphologies). One challenge we have not addressed previously is how to quantify the uncertainty in NN-based forecasts or estimates. We begin to address this question through the use of Bayesian neural networks (BNNs) for long-lead flood forecasting (18-hours).