Bridging the gap between satellite precipitation measurements and NWP forecasts

Global precipitation measurements by satellites, seamlessly extended by forecasts from numerical weather prediction (NWP) models, would be critical for global real-time flood monitoring and advance warning. In this study, we evaluated both satellite-based precipitation estimates and NWP-based forecasts, illustrated the systematic differences between the two categories, and explored techniques to consistently merge them for this purpose.

Precipitation forecasts with lead time from 1 to 5 days from ECMWF, NOAA's GFS and NASA GSFC's GEOS5 systems were studied extensively. They were evaluated against the satellite-based, gauge-corrected precipitation estimates, TMPA 3B42, over the land surface. To gain a better perspective, we also evaluated several other satellite-based precipitation products, including GPCP 1DD, TMPA 3B42RT, CMORPH and PERSIANN, against TMPA 3B42.

GEOS5, among other NWP models such as ECMWF and GFS, is shown to have the lowest biases in precipitation. Still, GEOS5 systematically over-estimate global precipitation by approximately 50%. This positive bias does not change much with lead time. In contrast, the satellite-based estimates (GPCP, TMPA, 3B42RT, CMORPH and PERSIANN) have biases mostly less than 20%. In addition, the RMS errors increase with the lead time in NWP forecasts, and in particular for GEOS5, the most increase in RMS errors takes place when the lead time goes from 1 day to 2 days. The RMS errors in the NWP products are also about twice as high as those of the satellite-based products. Moreover, GEOS5 has high probability of detection (POD) of 70-80%, but also high false alarm rate (FAR) of 50-60%. Finally, GEOS5 has much more light rain events (< 20 mm/day), but less strong events than 3B42, thus too little intermittency.

To reduce the errors in NWP precipitation and their impacts, we employed a CDF-matching method on GEOS5 forecasts. This method corrects the intensity distribution of the global precipitation field, to greatly reduce its bias and enhance its intermittency. The correction scheme also removes dubious light rain events, thus reducing FAR at the cost of reduced POD too. Evaluation of the corrected precipitation on global flood monitoring and advance warning will be presented.