The accuracy of these forecasts generally decreases very rapidly during the first 30 min because of the very short lifetime of individual convective pixels. A number of observational studies have shown that individual convective cells have mean lifetimes of about 20 min, with the best performance associated with a lead-time of 10 min. Numerical simulation studies have contributed significantly to the understanding of storm composition and duration; this is just beginning to be recognized in currents nowcasting systems. The nowcasting technique developed in this work is a special kind of nonlinear model with stochastic and deterministic components. The rainfall forecasts obtained using the considered method is then routed through a rainfall runoff model Vflo. Thus, implementing a coupled rainfall-runoff forecasting procedure for a watershed in western Puerto Rico. The prediction results with lead-times of 10, 20 and 30 min were analyzed and compared using statistical methods. The forecast result with lead-time of 10 min is the alternative with least percent of error. It was used in the hydrological model Vflo to compare the estimated hydrograph with the observed hydrograph from USGS stations. Furthermore, it was used in the flooding model Inundation Animator to show the extent of flooding superimposed onto a land map.
The results obtained here, represents the first time that a Dual-Pol and Doppler X-Band radar technology has been used for hydrologic analyses and specifically for rainfall forecasting in Puerto Rico. Results from the nowcasting model at spatial and temporal scales demonstrated the capability of the model to reproduce observed rainfall, for each nowcasting lead-time with relatively good agreement. The best statistical results were found in the rainfall nowcasting model with a lead-time of 10 min, as expected. It is well known that prediction of sudden storms using rainfall nowcasting models represent the category that are the most difficult to predict, and consequently, providing accurate flash flood warnings from these types of storms is a major challenge.
A significant contribution of this research is the development of the model considering the spatial and temporal variation of rainfall rates. Several parameter estimations were developed at each spatial and temporal domain, and the stochastic behavior of rainfall intensity was represented by an exponential time and spatial lag model, which is an approximation of a stochastic transfer function. The algorithm searches for contiguous rain pixels and identifies rain cells in the last two radar images to estimate the cloud motion vector. This newly developed rainfall nowcasting algorithm was validated with ten (10) storms and results comparing the algorithm with observed data as well as the hydrological results showed that the nowcasting model is a suitable tool for predicting the most likely areas to become inundated.
Comparisons between rain gauges, X-band radar and NEXRAD demonstrated that the high resolution radar system provides a higher degree of accuracy in rainfall estimation compared to NEXRAD. The RMSE increased for heavy rain conditions, nevertheless in all cases (light, moderate and heavy rain), the X-band radar consistently yielded the smallest error as compared with rain gauges while NEXRAD produced the largest errors. This was the first attempt to evaluate a rainfall prediction in the western Puerto Rico area. The most hydrological sensitive parameter in the basin area is the initial saturation.