Rainfall Estimation from WSR-88D Reflectivities Using Artificial Neural Networks

Spatially explicit rainfall estimates are vital drivers for a number of important research problems (e.g. flash flooding, crop yield modeling, water availability). Due to the complex atmospheric and spatial properties of the precipitation process, accurate rainfall estimation is not easily accomplished. This is especially true for high intensity short duration (HISD) events. The National Weather Service WSR-88D provides spatially explicit reflectivity measurements at various levels of the atmosphere. Although these reflectivities provide useful information for rainfall estimation, the exact relationships between radar reflectivity and rainfall are not well established. Static Z-R relations have been used in the past but do not account for the temporal variations caused by variant atmospheric conditions, nor do they account for the inherent spatial dependencies present in the precipitation field. Systematic differences in WSR-88D grid cell sizes with range and altitude also are not accounted for in traditional Z-R relations. In this paper, I describe a new method to translate WSR-88D reflectivities into rainfall rates using artificial neural networks (ANNs) for the Albuquerque, NM radar (KABX). The ANNs use the spatially explicit WSR-88D reflectivity information in combination with other valuable atmospheric and ground-level information to estimate rainfall rates for HISD events. The results of the new method are compared to the performance of Z-R relations using new time series econometric tests for out-of-sample predictive accuracy. The benefits of the new method are further illustrated through its designed connection with a distributed hydrological model for flash flood prediction.