Validation of NEXRAD multisensor precipitation estimates using an experimental dense rain gauge network in south Louisiana

SummaryThis study presents a validation analysis of a radar-based multisensor precipitation estimation product (MPE) focusing on small temporal and spatial scales that are of hydrological importance. The 4 × 4 km2 hourly MPE estimates are produced at the National Weather Service (NWS) regional River Forecast Centers and mosaicked as a national product known as Stage IV. The validation analysis was conducted during a 3-year period (2004–2006) using a high quality experimental rain gauge network in south Louisiana, United States, that was not included in the development of the MPE product. The dense arrangement of rain gauges within two MPE 4 × 4 km2 pixels provided a reasonably accurate approximation of area-average surface rainfall and avoided limitations of near-point gauge observations that are typically encountered in validation of radar-rainfall products.The overall bias between MPE and surface rainfall is rather small when evaluated over an annual basis; however, on an event-scale basis, the bias reaches up to ±25% of the event total rainfall depth during for half of the events and exceeds 50% for 10% of the events. Negative bias (underestimation) is more dominant (65% of events), which is likely caused by range-related effects such as beam overshooting and spreading over the study site (∼120 km from the closest radar site). A clear conditional bias was observed as the MPE estimates tend to overestimate small rain rates (conditional bias of 60–90% for rates lower than 0.5 mm/h) and underestimate large rain rates (up to −20% for rates higher than 10 mm/h). False detections and lack of detection problems contributed to the MPE bias, but were negligible enough to not result in significant false detection or underestimation of rainfall volumes. A significant scatter was observed between MPE and surface rainfall, especially at small intensities where the standard deviation of differences was in the order of 200–400% and the correlation coefficient was rather poor. However, the same statistics showed a much better agreement at medium to high rainfall rates. The MPE product was also successful in reproducing the underlying spatial and temporal organization of surface rainfall as reflected in the assessment of rainfall self-correlations and the extreme tail of the hourly rainfall marginal distribution. The quantitative results of this analysis emphasized the need for multiple gauges within MPE pixels as a prerequisite for validation studies. Using a single gauge within an MPE pixel as a reference representation of surface rainfall resulted in an unrealistic inflation of the actual MPE estimation error by 120–180%, especially during high-variability rainfall cases. This and the enhanced quality of the reference gauge dataset, explain the improved performance by MPE as compared to other previous studies. Compared to previous studies, the current analysis shows a significant improvement in the MPE performance. This is attributed to two main factors: continuous MPE algorithmic improvements and increased experience by its users at the NWS forecast centers, and the use of high-quality and dense rain gauge observations as a validation reference dataset. The later factor ensured that gauge-related errors are not wrongly assigned to radar estimation uncertainties.