Improving high-resolution quantitative precipitation estimation via fusion of multiple radar-based precipitation products

Abstract

Summary For monitoring and prediction of water-related hazards in urban areas such as flash flooding, high-resolution hydrologic and hydraulic modeling is necessary. Because of large sensitivity and scale dependence of rainfall–runoff models to errors in quantitative precipitation estimates (QPE), it is very important that the accuracy of QPE be improved in high-resolution hydrologic modeling to the greatest extent possible. With the availability of multiple radar-based precipitation products in many areas, one may now consider fusing them to produce more accurate high-resolution QPE for a wide spectrum of applications. In this work, we formulate and comparatively evaluate four relatively simple procedures for such fusion based on Fisher estimation and its conditional bias-penalized variant: Direct Estimation (DE), Bias Correction (BC), Reduced-Dimension Bias Correction (RBC) and Simple Estimation (SE). They are applied to fuse the Multisensor Precipitation Estimator (MPE) and radar-only Next Generation QPE (Q2) products at the 15-min 1-km resolution (Experiment 1), and the MPE and Collaborative Adaptive Sensing of the Atmosphere (CASA) QPE products at the 15-min 500-m resolution (Experiment 2). The resulting fused estimates are evaluated using the 15-min rain gauge observations from the City of Grand Prairie in the Dallas–Fort Worth Metroplex (DFW) in north Texas. The main criterion used for evaluation is that the fused QPE improves over the ingredient QPEs at their native spatial resolutions, and that, at the higher resolution, the fused QPE improves not only over the ingredient higher-resolution QPE but also over the ingredient lower-resolution QPE trivially disaggregated using the ingredient high-resolution QPE. All four procedures assume that the ingredient QPEs are unbiased, which is not likely to hold true in reality even if real-time bias correction is in operation. To test robustness under more realistic conditions, the fusion procedures were evaluated with and without post hoc bias correction of the ingredient QPEs. The results show that only SE passes the evaluation criterion consistently. The performance of DE and BC are generally comparable; while DE is more attractive for computational economy, BC is more attractive for reducing occurrences of negative estimates. The performance of RBC is poor as it does not account for magnitude-dependent biases in the QPE products. SE assumes that the higher-resolution QPE product is skillful in capturing spatiotemporal variability of precipitation at its native resolution, and that the lower-resolution QPE product provides skill at its native resolution. While the above assumptions may not always be met, the simplicity and robustness observed in this work make SE an extremely attractive choice as a simple post-processor to the QPE process. Also, unlike the other procedures considered in this work, it is extremely easy to update the statistical parameters of SE in real time, similarly to the real-time bias correction currently used in MPE, for improved performance via self-learning.