Erad 2008 - the Fifth European Conference on Radar in Meteorology and Hydrology

Short-term prediction, or nowcasting, of rain rate derived from ground-based weather radar observations has become an important filed of research for several applications such as civil protection, water resource management and hydrometeorological alarming (Germann and Zawadzki, 2002). Most nowcasting techniques are based on advection approaches, solved by means of cross-correlation or optimization procedures, coupled with statistical prediction methods (e.g., Pierce et al., 2004). Within the context of space-time prediction methodologies, artificial neural networks (ANNs) may represent an appealing tool due to their capability to extrapolate and generalize input-output data relations (e.g., Elman, 1990; Hecht-Nielsen, 1991). Different architectures and learning algorithms may be envisaged for ANNs, their effectiveness depending on the problem non-linearity and dimensionality under consideration. ANNs have been successfully used in several remote sensing contexts such satellite meteorology retrieval and nowcasting (e.g., Hsu et al., 1997; Grimes et al., 2003; Marzano et al., 2007). In this work a preliminary assessment of ANNs for rainfall radar nowcasting is performed. The proposed technique, named NNowRad hereafter, uses either a direct multi-layer network or recurrent neural network whose input layer is fed by radar-derived rain rate intensities taken from a sub-set of pixels adjacent to the pixel whose intensity we want to predict at future time steps. The neural network and its different configurations are then applied to each available rainfall radar map in order to predict the rain intensity at the next time step, typically at 15 or 30 minutes here. The prediction of rain intensities for longer periods of time is beyond the scope of the present work since we are principally aimed in pursuing the reliability of neural networks for radar nowcasting purpose. In the next sections, after a brief description of available radar datasets, the procedure to select the optimal ANN approach to the specific problem is discussed. More in detail, in section 3 the performances of different network architectures for a case study are compared. In section 4 the application of the selected optimal neural architecture to several case studies is discussed in terms of overall performances by comparing different criteria for training patterns. In the conclusions, preliminary results will be summarized stressing the fact that a short time sequence of events appears to be more suitable to train the ANN tool for operational applications exploiting radar nowcasting and its benefits. 2. Radar Data sets