National Weather Service

Multisensor Precipitation Estimation and Nowcasting for Flash Floods

A Radar Nowcasting Demonstration Project
to Improve Flash Flood Forecast and Warning Services
of the National Weather Service

Introduction

Quantitative precipitation estimation and nowcasting are important components of National Weather Service (NWS) flash flood warning services. They refer to the estimation of rainfall up to the current time using multiple sensors (WSR-88D, rain gauges, satellite estimates) and the forecasting of rainfall out to 1-3 hours in the future based heavily on current observed data. It is in this near-term forecast period when numerical weather prediction models currently have lesser precipitation forecast skill than extrapolation of current radar rainfall observations.

The High-Resolution Multisensor Precipitation Estimator (HPE) is a new prototype algorithm developed by the Hydrology Laboratory (HL) based on the existing operational Multisensor Precipitation Estimator (MPE) that is running at most River Forecast Centers and Weather Forecast Offices. However it has the advantage of higher spatial and temporal resolution than the current MPE, a factor that is necessary if the products are to be useful for flash flood monitoring and warning purposes. Instead of one-hour multisensor rainfall estimates at a nominal 4-km (HRAP) grid scale with updates once per hour as with the current MPE, the HPE is more flexible and generates multi-duration rainfall products on a 1-km grid (1/4th HRAP) with updates as often as every 5-15 minutes based on what the user chooses. Details of the current MPE algorithm can be found under the "MPE Training Workshop" link at http://www.nws.noaa.gov/oh/hrl/papers/papers.htm#wsr88d.

The High-Resolution Precipitation Nowcaster (HPN) algorithm is a prototype rainfall nowcasting algorithm that produces regional, gridded, one-hour rainfall nowcasts using input data generated by the HPE. The MPN is an enhancement of the Flash Flood Potential (FFP) algorithm, also developed at the HL, that uses current and recent-past WSR-88D radar data to estimate the future location of storms, their associated rainfall, and flash flood threat up to one hour into the future.

These HPE and HPN products can be used as input to distributed hydrologic forecast models or other flash flood monitoring tools at the Weather Forecast Offices. Short-term rainfall nowcasts can provide forecasters with additional guidance in evaluating flash flooding threat and additional lead time in issuing warnings to the public. The HPE and HPN algorithms have been running operationally in the mid-Atlantic state region since early 2004 in a testing and demonstration mode at the Hydrology Laboratory with experimental products made available for evaluation through this web site.

This web page is intended to be a graphical interface to real-time products generated by these two algorithms for pre-deployment evaluation by the scientific developers and forecast office staff. This demonstration project is one component of the collaborative Baltimore Flash Flood Forecasting Project at the NWS Hydrology Laboratory to demonstrate improved operational hydrologic services focused on the short time and small space scales typically associated with flash flooding.

Why a Web-based Demonstration Project?

This web page has been developed to provide a limited group of NWS developers and forecasters with real-time, 24 hours-a-day access to these experimental products for critical evaluation prior to the actual deployment of the algorithm within existing NWS operational computer systems. The World Wide Web represents a cheap, efficient forum to test new scientific algorithms and display their products to the field forecasters prior to full-scale deployment on computer systems at the 122 forecast offices. It allows them to use the products as needed during operational shifts, to learn the algorithm and its strengths and weaknesses, and to provide feedback to the developers regarding the added value and any necessary enhancements so that when the functionality is ultimately deployed nationwide it will better serve their needs from the start. This is typically called "alpha testing" by the NWS.

Users

The HE and HPN algorithms are targeted for general use by hydrologic forecasters at Weather Forecast Offices (WFO) whose primary responsibility it is to monitor flash flooding potential and issue appropriate warnings to the public. In particular, since we are using mid-Atlantic WSR-88D radar data for this web page, it is targeted for use and evaluation by the Baltimore/Washington WFO forecasters for which this region covers their county warning area. It is also targeted for use and evaluation by the scientific developers of the algorithm at the Hydrology Laboratory.

Quick Description of the Algorithm

HPN produces rainfall nowcasts up to one hour in the future on a 4-km HRAP grid by extrapolating current storm movement (as represented by consecutive pairs of instantaneous rain rate fields from HP) into the future and then estimating the amount of rainfall that would fall. These rainrate mosaics have been adjusted in real-time using mean field bias adjustment factors computed based on automated rain gauge data for each radar, and they are supplied to HPN at an interval of about 15-20 minutes. To generate these mosaics, HPE ues the Digital Hybrid Scan Reflectivity (DHR) and Digital Storm-total Precipitation (DSP) products from the WSR-88D's Precipitation Processing System (PPS). The former are used to derive the rainrate mosaics via Z-R conversion, and the latter are used to derive incremental rainfall accumulations through differencing.

The algorithm uses the rainrate mosaics to compute local storm motion using an extrapolative, pattern-matching technique and then assumes the storms will continue to move in that direction in the future. This storm motion estimation is done at a grid spacing of about 20 km, and therefore it permits a spatially-variable grid of storm motion vectors to be estimated. This is important since flash flood-producing storms often move at slower speeds and/or different directions compared to neighboring storms in the region, for example because they might be topographically forced. Options exist within the HPN to estimate and incorporate future growth or decay of the storms based on recent-past storm intensity changes. The one-hour rainfall nowcasts are updated every 5-15 minutes based on user specification (every 15 minutes for this web page). The algorithm will reinitialize all the output products to zero when there has been a 3 hour period without any detectable rain in the analysis domain.

Although not a main focus of this demonstration project, the HPN also compares both the observed and forecasted gridded rainfall with the latest gridded Flash Flood Guidance (FFG) products at 1-hr, 3-hr, and 6-hr durations from the River Forecast Centers to estimate flash flood threat within the next hour in a probabilistic sense in a manner similar though somewhat different from the existing operational AWIPS Flash Flood Monitoring and Prediction (FFMP) algorithm. Currently all such HPN comparisons are done on the 4-km grid and not within hydrologic basins, although the extension of the functionality to perform basin averages of flash flood threat within would be straightforward in the context of FFMP. Both the HPE and HPN products could be used within FFMP instead of the single-radar DHR products to provide forecasters with value-added, mosaicked, gauge-adjusted rainfall estimates and short-term nowcasts which are not possible with the current FFMP.

Additional details on the legacy FFP algorithm and processing logic can be found in Fulton and Seo (2000), references cited therein, and the associated AMS 15th Hydrology Conference Corel slide show. The "algorithm enunciation language" for the FFP's Projection and Assessment subalgorithms describes the processing logic in greater detail. Recent HPN forecast verification results are available in Guan, Ding, Fulton, and Kitzmiller (2005) from the AMS 32nd Conference on Radar Meteorology.

Description of the Image Products

The JPEG image products shown on this web site are separated into the following categories: Observations, Forecasts, and Forecast Verification. For details on the various gauge-adjusted radar rainfall products, see the MPE training documentation referenced previously. Please note that the images or only updated when rainfall is occurring in the region. If you see image dates that are old, it's probably because the weather is currently dry.

MFB Rainrate - The observed or forecasted rainrate (mm/hr) as computed using the reflectivity factor field converted to rainrate using the current Z-R relationship as defined for each WSR-88D and stored in the DHR product header. The mean field bias has been removed for each radar using rain gauges The 1-hour rainrate verification image compares the current observed rainrate field with the 1-hour rainrate forecast from one hour ago.

Radar Rainfall - A running rainfall accumulation (mm) for the past 15 or 60 minutes based on unadjusted radar data alone.

MFB Rainfall - A running rainfall accumulation (mm) for which the mean field bias has been removed for each radar using rain gauges. The 1-hour rainfall verification image compares the current 1-hour observed mean field bias-adjusted radar rainfall field with the corresponding 1-hour forecast from one hour previously.

MS Rainfall - The multisensor rainfall mosaic (mm).

Storm Motion - Quality-controlled local storm motion vectors (with length proportional to speed) overlaid on a "reliability" field. The reliability field is a relative measure of the algorithm’s ability to identify storms in the rainrate field and thus compute their velocity. You can place greater confidence in the computed storm motion vectors where the reliability values are larger. The current hourly-averaged storm motion vector over the whole radar umbrella is plotted in the lower left corner.

Storm-Relative Motion - Same as above except that the storm motion vectors are relative to the current mean hourly-averaged, regionally-averaged storm motion vector (as plotted in the lower left corner of the image). Non-zero storm-relative motion vectors, particularly those that point opposite to the mean motion vector, identify potentially dangerous storms that are back-building or stationary and therefore capable of producing flash floods.

Storm Growth/Decay - The local, lagrangian rate of change of intensity of the storms (mm/hr/hr)over the last ~15 minutes. This information is used to adjust the forecasted rainrates locally if the user desires.

Observed Probability of FFG Exceedance - The probability that the observed mean field bias-adjusted rainfall up to the current time will exceed FFG. This is computed as the maximum value for the associated 1-hour, 3-hour, and 6-hour durations of FFG.

Forecasted Probability of FFG Exceedance - The probability that the observed+forecasted mean field bias adjusted rainfall will exceed FFG. This is computed as the maximum value for the associated 1-hour, 3-hour, and 6-hour durations of FFG.

Observed Critical Rainfall Probability - The probability that the observed mean field bias-adjusted rainfall at some time during the rainfall event has exceeded the FFG. This is computed as the maximum value for the 1-hour, 3-hour, and 6-hour durations of FFG. It will always monitonically increase throughout a rain event for a given location.

Forecasted Critical Rainfall Probability - The probability that the observed+forecasted mean field bias adjusted rainfall at some time during the rainfall event has exceeded or will exceed the Flash Flood Guidance in the next hour. This is computed as the maximum value for the 1-hour, 3-hour, and 6-hour durations of FFG. It will usually monitonically increase throughout a rain event for a given location.

Flash Flood Guidance (FFG) - The RFC model-estimated basin-averaged rainfall over either 1, 3, or 6 hours duration that would cause small streams to reach bankfull.

There are certain situations when a rainfall forecast is not attempted, for example due to nonexistent or weak radar echoes. In these cases, "No Projection Attempted" is displayed on the forecast images.

Some Limitations

The MPN algorithm has been developed and optimized primarily for warm season convection. It does not always perform optimally for stratiform rainfall events where the spatial gradients of reflectivity are weak. This is because the algorithm’s ability to track storms depends on its ability to successfully perform pattern matching of consecutive, recent-past, radar rainrate fields. If it cannot identify well-defined storm features in the consecutive radar images, it will intentionally not estimate their motion and will therefore not produce a rainfall nowcast. This is a built-in quality-control feature that helps prevent the algorithm from issuing anomalous and/or questionable rainfall forecasts when the data does not adequately support it.

Future Plans

There will be on-going enhancements to the algorithm and its web presence to improve the scientific integrity of the products and the web-based user interface. There will be scientific enhancements to the algorithm itself that are briefly described below.

1) Some nowcast algorithms (e.g., NCAR, MDL, MIT/LL) utilize additional data besides radar reflectivity data, including satellite and numerical weather prediction model data as well as radar doppler velocity data. Incorporation of this technology into the existing algorithm may add value and probably improve performance, particularly as the forecast period increases beyond one hour. In particular, this data may aid in forecasting the growth or decay of new or existing storms in the future which is an active area of research currently.

2) Evaluate HPN performance for forecast periods beyond one hour. For rainfall systems with greater temporal consistency (e.g., synoptically driven systems), one could justifiably extend the nowcasts to several hours probably (with an expected increase in error with lead time). At the lead time increases, it will be necessary to utilize atmospheric model guidance by merging the current extrapolation-based forecasts with numerical model forecasts especially as lead time increases beyond an hour as is currently done at the UK Met Office.

3) Display of time series of real-time nowcast verification statistics on the web page such as forecast root mean square error, forecast bias, correlation, probability of detection, false alarm ratio, and critical success index.

4) Higher resolution (e.g., 1 km) nowcast products may be justified. This will be possible when using the FAA’s Terminal Doppler Weather Radar located near Baltimore-Washington International Airport. We are investigating methods to use this data.

5) Currently the HPN uses River Forecast Center model-generated Flash Flood Guidance (FFG) as a surrogate measure of antecedent, basin-average soil moisture for basins typically about 100 square miles in size. In the near future, radar-derived rainfall and nowcasts will be input directly into high resolution distributed hydrologic forecast models with explicit soil moisture accounting models. This will then obviate the need for comparisons of radar rainfall with surrogate measures such as FFG with its well-known deficiencies. We will produce explicit point forecasts of streamflow and depth at ungauged locations as well as high resolution flood inundation maps. See Reed, Fulton, Zhang, and Guan (2006) from the AMS 20th Hydrology Conference for more information on our activities.

Additional information on our future plans for improving multisensor rainfall estimation and nowcasting algorithms can be found at http://www.nws.noaa.gov/oh/hrl/papers/papers.htm#wsr88d.

The Product Description Document, as required by the NWS Office of Climate, Water, and Weather Services, is also available. Click Here.

Comments or Questions?

Send them to david.kitzmiller@noaa.gov.