• Arkin, P. A., and B. N. Meisner. 1987. The relationship between large-scale convective rainfall and cold cloud over the Western Hemisphere during 1982–84. Mon. Wea. Rev. 115:5174.

    • Search Google Scholar
    • Export Citation
  • Atlas, D., , D. Rosenfeld, , and D. B. Wolff. 1990. Climatologically tuned reflectivity–rain rate relations and links to area–time integrals. J. Appl. Meteor. 29:11201135.

    • Search Google Scholar
    • Export Citation
  • Austin, G. L., and A. Bellon. 1982. Very-short-range forecasting of precipitation by the objective extrapolation of radar and satellite data. Nowcasting, K. A. Browning, Ed., Academic Press, 177–190.

    • Search Google Scholar
    • Export Citation
  • Bellon, A., and G. L. Austin. 1978. The evaluation of two years of real-time operation of a short-term precipitation forecasting procedure (SHARP). J. Appl. Meteor. 17:17781787.

    • Search Google Scholar
    • Export Citation
  • Bellon, A., , S. Lovejoy, , and G. L. Austin. 1980. Combining satellite and radar data for the short-range forecasting of precipitation. Mon. Wea. Rev. 108:15541566.

    • Search Google Scholar
    • Export Citation
  • Browning, K. A. 1979. The FRONTIERS plan: A strategy for using radar and satellite imagery for very short-range precipitation forecasting. Meteor. Mag. 108:161184.

    • Search Google Scholar
    • Export Citation
  • Browning, K. A. . Ed.,. . 1982. Nowcasting. Academic Press, 256 pp.

  • Chandler, D. G., and J. L. Collins Jr.. 1994. The D-Day Encyclopedia. Simon and Schuster, 665 pp.

  • Cooper, H. J., and E. A. Smith. 1993. The importance of short-term forecasting of thunderstorms to launch operations at Cape Canaveral. Bull. Amer. Meteor. Soc. 74:8186.

    • Search Google Scholar
    • Export Citation
  • Endlich, R. M., , D. E. Wolf, , D. J. Hall, , and A. E. Brain. 1971. Use of a pattern recognition technique for determining cloud motions from sequences of satellite photographs. J. Appl. Meteor. 10:105117.

    • Search Google Scholar
    • Export Citation
  • Ferraro, R. R. 1997. Special Sensor Microwave Imager derived global rainfall estimates for climatological applications. J. Geophys. Res. 102:1671516735.

    • Search Google Scholar
    • Export Citation
  • Ferraro, R. R., , E. A. Smith, , W. Berg, , and G. J. Huffman. 1998. A screening methodology for passive microwave precipitation retrieval algorithms. J. Atmos. Sci. 55:15831600.

    • Search Google Scholar
    • Export Citation
  • Gandin, L. S., and A. H. Murphy. 1992. Equitable skill scores for categorical forecasts. Mon. Wea. Rev. 120:361370.

  • Goerss, J. S., , C. S. Velden, , and J. D. Hawkins. 1998. The impact of multispectral GOES-8 wind information on Atlantic tropical cyclone track forecasts in 1995. Part II: NOGAPS forecasts. Mon. Wea. Rev. 126:12191227.

    • Search Google Scholar
    • Export Citation
  • Grody, N. C. 1991. Classification of snow cover and precipitation using the Special Sensor Microwave Imager. J. Geophys. Res. 96:74237435.

    • Search Google Scholar
    • Export Citation
  • Hasler, A. F., , K. Palaniappan, , C. Kambhammeta, , P. Black, , E. Uhlhorn, , and D. Chesters. 1998. High-resolution wind fields within the inner core and eye of a mature tropical cyclone from GOES 1-minute images. Bull. Amer. Meteor. Soc. 79:24832496.

    • Search Google Scholar
    • Export Citation
  • Huffman, G. J. Coauthors,. 1997. The Global Precipitation Climatology Project (GPCP) combined precipitation dataset. Bull. Amer. Meteor. Soc. 78:520.

    • Search Google Scholar
    • Export Citation
  • Kedem, B., , L. S. Chiu, , and G. R. North. 1990. Estimation of mean rain rate: Application to satellite observations. J. Geophys. Res. 95:19651972.

    • Search Google Scholar
    • Export Citation
  • Kessler, E. 1966. Computer program for calculating average lengths of weather radar echoes and pattern bandedness. J. Atmos. Sci. 23:569574.

    • Search Google Scholar
    • Export Citation
  • Laurent, H. 1993. Wind extraction from Meteosat water vapor channel image data. J. Appl. Meteor. 32:11241133.

  • Lazzara, M. A. Coauthors. 1999. The Man computer Interactive Data Access System: 25 years of interactive processing. Bull. Amer. Meteor. Soc. 80:271284.

    • Search Google Scholar
    • Export Citation
  • Leese, J. A., , C. S. Novak, , and B. B. Clark. 1971. An automated technique for obtaining cloud motion from geosynchronous satellite data using cross correlation. J. Appl. Meteor. 10:118132.

    • Search Google Scholar
    • Export Citation
  • Menzel, W. P., , F. C. Holt, , T. J. Schmit, , R. M. Aune, , A. J. Schreiner, , G. S. Wade, , and D. G. Gray. 1998. Application of GOES-8/9 soundings to weather forecasting and nowcasting. Bull. Amer. Meteor. Soc. 79:20592077.

    • Search Google Scholar
    • Export Citation
  • Merrill, R. T., , W. P. Menzel, , W. Baker, , J. Lynch, , and E. Legg. 1991. A report on the recent demonstration of NOAA's upgraded capability to derive cloud motion satellite winds. Bull. Amer. Meteor. Soc. 72:372376.

    • Search Google Scholar
    • Export Citation
  • Nieman, S. J., , W. P. Menzel, , C. M. Hayden, , D. Gray, , S. T. Wanzong, , C. S. Velden, , and J. Daniels. 1997. Fully automated cloud drift winds in NESDIS operations. Bull. Amer. Meteor. Soc. 78:11211133.

    • Search Google Scholar
    • Export Citation
  • Purdom, J. F. W. 1976. Some uses of high resolution GOES imagery in the mesoscale forecasting of convection and its behavior. Mon. Wea. Rev. 104:14741483.

    • Search Google Scholar
    • Export Citation
  • Reynolds, D. W., and E. A. Smith. 1979. Detailed analysis of composited digital radar and satellite data. Bull. Amer. Meteor. Soc. 60:10241037.

    • Search Google Scholar
    • Export Citation
  • Rossow, W. B., and L. C. Garder. 1993. Cloud detection using satellite measurements of infrared and visible radiances for ISCCP. J. Climate 6:23412369.

    • Search Google Scholar
    • Export Citation
  • Schmetz, J., , K. Holmlund, , J. Hoffman, , B. Strauss, , B. Mason, , V. Gaertner, , A. Koch, , and L. van de Berg. 1993. Operational cloud-motion winds from Meteosat infrared images. J. Appl. Meteor. 32:12061225.

    • Search Google Scholar
    • Export Citation
  • Shin, K-S., , P. E. Riba, , and G. R. North. 1990. Estimation of area-averaged rainfall over tropical oceans from microwave radiometry: A single channel approach. J. Appl. Meteor. 29:10311042.

    • Search Google Scholar
    • Export Citation
  • Smith, E. A. 1975. The McIDAS System. IEEE Trans. Geosci. Electron. GE-13:123136.

  • Smith, E. A., and D. R. Phillips. 1972. Automated cloud tracking using precisely aligned digital ATS pictures. IEEE Trans. Comput. C17:715729.

    • Search Google Scholar
    • Export Citation
  • Smith, E. A. Coauthors,. 1998. Results of WetNet PIP-2 Project. J. Atmos. Sci. 55:14831536.

  • Spencer, R. W. 1993. Global oceanic precipitation from the MSU during 1979–91 and comparisons to other climatologies. J. Climate 6:13011326.

    • Search Google Scholar
    • Export Citation
  • Taylor, B. C., and K. A. Browning. 1974. Towards an automated weather radar network. Weather 29:202216.

  • Tsonis, A. A., and G. L. Austin. 1981. An evaluation of extrapolation techniques for the short-term prediction of rain amounts. Atmos. Ocean 19:5456.

    • Search Google Scholar
    • Export Citation
  • Turk, F. J., , G. Rohaly, , J. D. Hawkins, , E. A. Smith, , A. Grose, , F. S. Marzano, , and A. Mugnai. 2000. Analysis and assimilation of rainfall from blended SSM/I, TRMM and geostationary satellite data. Preprints, 10th Conf. on Satellite Meteorology and Oceanography, Long Beach, CA, Amer. Meteor. Soc., 66–69.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , C. M. Hayden, , S. J. Nieman, , W. P. Menzel, , S. Wanzong, , and S. Goerss. 1997. Upper-tropospheric winds derived from geostationary satellite water vapor observations. Bull. Amer. Meteor. Soc. 78:173195.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , T. L. Olander, , and S. Wanzong. 1998. The impact of multispectral GOES-8 wind information on Atlantic tropical cyclone track forecasts in 1995. Part I: Dataset methodology, description, and case analysis. Mon. Wea. Rev. 126:12021218.

    • Search Google Scholar
    • Export Citation
  • Wetzel, M. A., , R. D. Borys, , and L. E. Xu. 1996. Satellite microphysical retrievals for land-based fog with validation by balloon profiling. J. Appl. Meteor. 35:810829.

    • Search Google Scholar
    • Export Citation
  • Wilheit, T. T., , A. T. C. Chang, , and L. S. Chiu. 1991. Retrieval of monthly rainfall indices from microwave radiometric measurements using probability distribution functions. J. Atmos. Oceanic Technol. 8:118136.

    • Search Google Scholar
    • Export Citation
  • Wilks, D. S. 1995. Statistical Methods in the Atmospheric Sciences. Academic Press, 464 pp.

  • View in gallery

    Map of South Korea indicating locations of 395-site automated weather station network

  • View in gallery

    (top) Three grid systems used for nowcasting within domain of full-resolution IR image sectors. Larger boxes outlined with thick lines represent 21 × 21 pixel target subgrids making up the “target grid system.” Smaller boxes outlined with thin dashed lines (inside target grid system) represent 7 × 7 pixel filtered subgrids making up the “filtered grid system.” Expanded small box domain makes up the “forecast grid system.” Edge domain of 4-pixel width is necessitated by ±25 lag dimensions used in seeking match subgrids. (bottom) Relationship between centers of target subgrid and match subgrid found in 71 × 71 search area by minimum difference method defines CMW vector between consecutive pair of IR images

  • View in gallery

    Examples of (top) initial CMW field and (bottom) filtered CMW field for Amazon basin on 31 Jan 1999 extracted from 2245–2345 UTC IR image pair. Filtered field is created by applying a priori Monte Carlo averaging operator using 50% vector selection with 10 realizations, followed by two-dimensional linear interpolation onto high-resolution grid, and nine-point binomial smoother. Vector directions are indicated with arrowheads; vector magnitudes with grayscale. Every third gridpoint vector has been selected from high-resolution field for display in bottom panel

  • View in gallery

    Examples of raw histogram, smooth histogram, and gamma function PDFs used for EBBTs in probability matching process: (top) ocean case and (bottom) land case from original GMS-5–SSM/I paired pixel dataset over Korean peninsula region

  • View in gallery

    Example of lognormal PDF fit to RR histogram from GMS-5–SSM/I paired pixel ocean dataset over Korean peninsula region: (top) comparison for light rain rates (0.5–5.0 mm h−1) and (bottom) comparison for medium–heavy rain rates (5.0–35.0 mm h−1)

  • View in gallery

    Normalized frequency distributions for (left) ocean and (right) land using RHM, SHM, and FM algorithms applied to 0715 UTC GOES-8 IR image for Caribbean basin on 1 Apr 1998. Bin interval is 2.5 mm h−1

  • View in gallery

    Final EBBT–RR mapping rules for ocean and land based on SHM algorithm

  • View in gallery

    (a) Cloud–rain areas created from (top) 1015 and (bottom) 1045 UTC IR images on 1 Apr 1998 over Caribbean basin. (middle) Filtered CMW field and EBBT tendencies derived from the two IR images. (middle) Arrows indicate CMW vector directions; sizes of attached circles indicate CMW vector magnitudes; and cross, open, or filled patterns within circles indicate negative, near-zero, or positive EBBT tendencies. Every 12th grid point in east–west and every 35th grid point in north–south are selected for display. (b) The (left) 1-, 2-, and 3-h cloud–rain area forecasts from conditions shown in (a) and (right) verification images at 1145, 1245, and 1345 UTC. RESS, FAR, and RISS skill scores are indicated numerically between pairs of forecast and verification panels

  • View in gallery

    (Continued)

  • View in gallery

    (a) Same as Fig. 8a but for Amazon basin on 1 Feb 1999 starting at 0745 UTC, in which every 10th grid point in both east–west and north–south is selected for display (middle). (b) Same as Fig. 8b but for Amazon basin on 1 Feb 1999 starting at 0945 UTC

  • View in gallery

    (Continued)

  • View in gallery

    (a) Same as Fig. 8a but for Korean peninsula on 1 May 1999 starting at 1530 UTC, in which every 10th grid point in both east–west and north–south is selected for display (middle). (b) Same as Fig. 8b but for Korean peninsula on 1 May 1999 starting at 1730 UTC

  • View in gallery

    (Continued)

  • View in gallery

    Ensemble-averaged RESS, FAR, and RISS skill score comparisons out to 3-h forecasts used to evaluate four advection techniques. Comparisons presented for (left) Caribbean basin, (middle) Amazon basin, and (right) Korean peninsula

  • View in gallery

    Ensemble-averaged RESS and FAR skill score time decay curves out to 10-h forecasts for (left) Caribbean basin and (right) Amazon basin used for determination of asymptotic limit of forecasts. Forecast ensemble counts are over 10 times as large for Amazon basin relative to Caribbean basin, explaining why former's skill score time decay curves are much smoother

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 204 204 4
PDF Downloads 65 65 1

Possibilities and Limitations for Quantitative Precipitation Forecasts Using Nowcasting Methods with Infrared Geosynchronous Satellite Imagery

View More View Less
  • a Department of Meteorology, The Florida State University, Tallahassee, Florida
  • | b NASA Goddard Space Flight Center, Greenbelt, Maryland
  • | c Meteorological Research Institute, Korean Meteorological Administration, Seoul, Korea
  • | d School of Earth and Environmental Sciences, Seoul National University, Seoul, Korea
  • | e Marine Meteorology Division, Naval Research Laboratory, Monterey, California
  • | g Department of Meteorology, The Florida State University, Tallahassee, Florida
© Get Permissions
Full access

Abstract

A rainfall nowcasting system is developed that identifies locations of raining clouds on consecutive infrared geosynchronous satellite images while predicting the movement of the rain cells for up to 10 h using cloud-motion-based winds. As part of the analysis, the strengths and weaknesses of various kinds of cloud wind filtering schemes and both steady and nonsteady storm advection techniques as forecast operators for quantitative precipitation forecasting are evaluated. The first part of the study addresses the development of a probability matching method (PMM) between histograms of equivalent blackbody temperatures (EBBTs) and Special Sensor Microwave Imager (SSM/I)–derived rain rates (RRs), which enables estimating RRs from instantaneous infrared imagery and allows for RR forecasts from the predicted EBBT fields. The second part of the study addresses the development and testing of the nowcasting system built upon the PMM capability and analyzes its success according to various skill score metrics. Key processes involved in the nowcasting system include the retrieved cloud-motion wind field, the filtered cloud-motion wind field, and the forecasting of a future rain field by storm advection and EBBT tendencies. These processes allow for the short-term forecasting of cloud and rain locations and of rain intensity, using PMM-based RRs from different datasets of infrared Geostationary Meteorological Satellite (GMS) and Geostationary Operational Environmental Satellite (GOES) imagery. For this study, three convective rain sequences from the Caribbean basin, the Amazon basin, and the Korean peninsula are analyzed. The final part of the study addresses the decay of forecast accuracy with time (i.e., the point at which the asymptotic limit on forecast skill is reached). This analysis indicates that the nowcasting system can produce useful rainfall forecast information out to approximately 6 h.

Corresponding author address: Eric A. Smith, NASA Goddard Space Flight Center, Code 912.1, Greenbelt, MD 20771. easmith@pop900.gsfc.nasa.gov

Abstract

A rainfall nowcasting system is developed that identifies locations of raining clouds on consecutive infrared geosynchronous satellite images while predicting the movement of the rain cells for up to 10 h using cloud-motion-based winds. As part of the analysis, the strengths and weaknesses of various kinds of cloud wind filtering schemes and both steady and nonsteady storm advection techniques as forecast operators for quantitative precipitation forecasting are evaluated. The first part of the study addresses the development of a probability matching method (PMM) between histograms of equivalent blackbody temperatures (EBBTs) and Special Sensor Microwave Imager (SSM/I)–derived rain rates (RRs), which enables estimating RRs from instantaneous infrared imagery and allows for RR forecasts from the predicted EBBT fields. The second part of the study addresses the development and testing of the nowcasting system built upon the PMM capability and analyzes its success according to various skill score metrics. Key processes involved in the nowcasting system include the retrieved cloud-motion wind field, the filtered cloud-motion wind field, and the forecasting of a future rain field by storm advection and EBBT tendencies. These processes allow for the short-term forecasting of cloud and rain locations and of rain intensity, using PMM-based RRs from different datasets of infrared Geostationary Meteorological Satellite (GMS) and Geostationary Operational Environmental Satellite (GOES) imagery. For this study, three convective rain sequences from the Caribbean basin, the Amazon basin, and the Korean peninsula are analyzed. The final part of the study addresses the decay of forecast accuracy with time (i.e., the point at which the asymptotic limit on forecast skill is reached). This analysis indicates that the nowcasting system can produce useful rainfall forecast information out to approximately 6 h.

Corresponding author address: Eric A. Smith, NASA Goddard Space Flight Center, Code 912.1, Greenbelt, MD 20771. easmith@pop900.gsfc.nasa.gov

Introduction

In the last few decades, nowcasting has played an increasingly important role in the field of meteorology. The significance of nowcasting weather events can be seen in matters that range from routine daily planning (e.g., Purdom 1976) to pivotal situations such as hurricane preparedness (Velden et al. 1998; Goerss et al. 1998), planning military maneuvers (Chandler and Collins 1994), or evaluating weather conditions prior to the launch of a space shuttle (Cooper and Smith 1993). In conventional terms, nowcasting incorporates meteorological observations to produce short-term (order 3 h) mesoscale weather forecasts based on extrapolation (Browning 1982), with greatest emphasis on use of radar and geosynchronous (GEO; see appendix for list of acronyms) satellite measurements (e.g., Reynolds and Smith 1979; Bellon et al. 1980; Tsonis and Austin 1981; Austin and Bellon 1982). Such data inherently contain the short-term dynamical manifestations of the ever-changing atmosphere. This study investigates the prospects of using a pattern recognition technique to produce short-term quantitative precipitation forecasts (QPFs) of convective rainfall using infrared GEO-satellite imagery.

The concept of nowcasting of rain using radar data was formally introduced in the 1960s (e.g., Kessler 1966) and was rapidly assimilated in the 1970s [e.g., see Taylor and Browning (1974), Browning (1979), Browning's (1982) compilation on the subject of nowcasting]. Radar-based nowcasting quickly developed to the use of real-time reflectivity pattern recognition techniques (e.g., Bellon and Austin 1978). The building blocks for satellite nowcasting also began in the 1970s and focused on methods to discern the wind field through quantitative estimates of cloud motion from GEO satellite imagery. For example, Endlich et al. (1971), Leese et al. (1971), Smith and Phillips (1972), and Smith (1975) described ever-improving techniques for determining cloud motion through the use of pattern recognition for sequences of visible and infrared (IR) imagery from National Aeronautics and Space Administration (NASA) Applications Technology Satellites and National Oceanic and Atmospheric Administration (NOAA) Geostationary Operational Environmental Satellites (GOES).

More recent examples of GEO imagery–based wind retrieval are given by Merrill et al. (1991), who described an updated version of the GOES-based cloud motion wind algorithm used by NOAA; Laurent (1993), who introduced a Meteosat-based water vapor feature tracking method (Meteosat is the European operational GEO satellite); Wetzel et al. (1996), who developed a GOES-based technique for forecasting fog visibility hazards; and Velden et al. (1997), who set forth a GOES-based water vapor method focused on tropical cyclone prediction. Current operational systems using GEO imagery–based wind retrieval are described in Schmetz et al. (1993) for the European Community Meteosat IR winds; Nieman et al. (1997) for the NOAA National Environmental Satellite, Data, and Information Service (NESDIS) GOES cloud winds; Hasler et al. (1998) for the NASA Goddard Space Flight Center GOES hurricane analysis winds; and Lazzara et al. (1999) for the University of Wisconsin Space Science and Engineering Center GEO mesoscale modeling support winds.

Satellite-based rainfall retrieval has become one of the more intense research topics in the discipline of satellite meteorology. The range of studies includes the creation of global annual rainfall climatological descriptions from various types of operational satellite data products (e.g., Spencer 1993; Huffman et al. 1997), quantitative performance characteristics of Special Sensor Microwave Imager (SSM/I) rainfall retrieval algorithms (e.g., Smith et al. 1998), exploration of short-term rainfall forecasting using GOES sounding products (e.g., Menzel et al. 1998), resurrecting historical satellite datasets (e.g., Shin et al. 1990), and diagnosing causes for diurnal variability of rainfall using recent Tropical Rainfall Measuring Mission results (S. Yang 2001, personal communication). The numerous studies dedicated to this topic demonstrate the complexity involved in understanding rainfall from space-based observations.

One of the inherent problems in performing rainfall studies from satellite measurements is the general dearth of adequate ground data with which to verify instantaneous satellite retrievals—particularly over oceans. In this study we have sought to address this topic by verifying with 1-min rain gauge data from a dense gauge network over the Korean peninsula.

Another problem area is distinguishing precipitating clouds from background features that may exhibit similar signals. In the centimeter–millimeter spectrum, this is relatively straightforward. For example, Grody (1991) explored the differentiation of precipitation from other atmospheric and surface features through the use of an 85-GHz “degree of scattering” index obtained from passive microwave (PMW) SSM/I brightness temperatures TB. [SSM/I instruments are carried aboard the Defense Meteorological Satellite Program near-polar orbiting satellites.] Based on this approach, Ferraro et al. (1998) implemented a hierarchical screening procedure for use with the NESDIS operational SSM/I rainfall algorithm. Smith et al. (1998) included a review of a number of SSM/I screening procedures in an intercomparison of 20 published SSM/I precipitation retrieval algorithms and how their respective inversion-screening designs influenced retrieval differences in the second WetNet Precipitation Intercomparison Project.

The screening of GEO IR images for precipitating clouds is less tractable because there is no unique separation point between rain and no rain based on cloud-top radiometric temperature [usually referred to as equivalent blackbody temperature (EBBT)]. Nevertheless, based on the use of EBBT thresholds, a number of studies have developed GEO IR rain algorithms. These are perhaps best exemplified by the GOES Precipitation Index developed by Arkin and Meisner (1987), which uses an invariant 235-K EBBT threshold to map out precipitating cloud area. Screening IR images for precipitating clouds encompasses several problems. Clouds deemed cold enough to produce rain according to an EBBT threshold may have evolved beyond the rain stage (e.g., inactive optically thick cirrus anvils); other clouds not deemed cold enough may be undergoing pre-ice-phase rain microphysics (e.g., orographic warm rain). Despite these drawbacks, GEO IR data provide important physical characteristics of precipitation-producing storms and, above all, are available over the diurnal cycle.

In this context, GEO IR data provide the foundation for the nowcasting system developed in this study. Cloud motions between sequential images, as well as changes in IR temperatures, are used to forecast the location and intensity of convective rain elements for the short term (order 3–6 h). We submit that the potential value of such a QPF method is as a nowcasting tool that could expand the domain of a radar-based nowcasting system—as originally advocated in Reynolds and Smith (1979).

Section 2 of the paper describes the nowcasting methodology, and section 3 describes the datasets used in the study. Section 4 presents the results of the study consisting of analyses of the components of the nowcasting system, along with forecast verification results for three different GEO IR datasets. Conclusions are presented in section 5, including discussion of the strengths and weaknesses of the current methodology and suggestions for its possible improvement.

Nowcasting methodology

The development of the rain nowcasting system is presented in two stages. The first addresses development and validation of a satellite rain retrieval algorithm, based on probability matching between GEO IR temperature distribution functions and SSM/I PMW-retrieved rain-rate distribution functions. The second stage addresses development and testing of the several components of a short-term QPF method. This involves estimation of the initial wind field from cloud motions and prediction of rain cell locations and intensities into the near future through steady or nonsteady advection techniques.

Rain retrieval

PMM algorithms

In the first stage of the study, an algorithm is sought that deterministically relates measured EBBTs to rain rates (RRs) at individual GEO IR pixel locations. This is accomplished using coincident Japanese Geostationary Meteorological Satellite (GMS) IR and SSM/I RR datasets (as described in detail in section 3), in conjunction with application of the rain-rate probability matching method (PMM). This method consists of assigning a mapping rule between an independent variable's probability distribution function (PDF; in this case the PDFs of IR EBBTs), to a rain-rate PDF (in this case PDFs generated from SSM/I-retrieved rain rates). The method has been successfully applied to radar reflectivities and rain gauge–measured rain rates by Atlas et al. (1990), as well as to GEO IR EBBTs and SSM/I-retrieved RRs by Turk et al. (2000) for both land and ocean locales. In the Turk et al. algorithm, EBBT and RR histograms are first created, then a cumulative distribution function (CDF) is used to assign discrete EBBT ranges to discrete RR intervals.

Part of our interest in this study is in using a PMM algorithm with EBBT and RR histograms to evaluate the use of statistical distribution functions in representing the original histograms. Prior studies, such as those of Kedem et al. (1990) and Wilheit et al. (1991), have shown that nonzero RRs over a sufficiently large time–space domain generally obey a lognormal distribution, which, following Wilks (1995), is expressed by
i1520-0450-41-7-763-e1
where x represents rain rate and μ and σ are the logarithmic mean and standard deviation of the RR distribution. This PDF has been found suitable for representing RR histograms in a number of prior studies as noted in the Kedem et al. (1990) and Wilheit et al. (1991) papers.
We also seek a suitable statistical distribution for representing EBBT histograms. With no well-known proven schemes for representing EBBT histograms in rain-rate probability matching, a number of PDF formulations were tested. From these tests, the two most suitable candidates were the lognormal distribution described above and the two-parameter gamma distribution, which, following Wilks (1995), is expressed by
i1520-0450-41-7-763-e2
in which the two parameters of the distribution are the shape parameter α and the distribution mode parameter β, with Γ(α) representing the gamma function and x representing EBBT.

Once statistical fits have been created for both the RR and EBBT distributions, finding the PMM algorithm between two fits is somewhat similar to the approach used by Turk et al. (2000) for raw histograms. The difference is that the EBBT–RR mapping rule is based on matching equivalent percentage areas under the function curves rather than matching equivalent normalized histogram frequencies. In creating the initial histograms, 1° bins between 199 and 273 K are used for EBBTs while 0.5 mm h−1 bins between 0 and 35 mm h−1 are used for RRs. As in Turk et al. (2000), the RRs are based on the NESDIS operational SSM/I rain-retrieval algorithm (see Ferraro et al. 1998). All EBBTs less than 199 K are set to 35 mm h−1 (the upper maximum allowed in the NESDIS algorithm) while all EBBTs greater than 273 K are set to zero. Note that probability matching only takes place for nonzero RRs with this algorithm, which separates it slightly from algorithms that include the histogram bin of 0 RR in the probability matching rule.

This algorithm is denoted as function matching (FM) and is tested against two other histogram-based PMM algorithms that follow the method of Turk et al. (2000) to create the EBBT–RR mapping rules. As in that algorithm, CDFs are first created from the PDFs, beginning with the warm end of the EBBT histogram and working toward the cold end, and beginning with the low-intensity end of the RR histogram (including 0 rain values) and working toward the high-intensity end. For each weighted moment of the RR CDF, the EBBT CDF is scanned from warm to cold for the matching EBBT indices. This method creates a series of EBBT–RR pairs that are used as the mapping rule in the nowcasting system to predict RR values from advected IR imagery.

The first of these algorithms is denoted as raw histogram matching (RHM). It uses the original EBBT and RR histograms to create an EBBT–RR mapping rule similar to Turk et al. (2000). The second algorithm, denoted as smooth histogram matching (SHM), uses a five-point binomial filter to create a smoothed version of the EBBT histogram in order to eliminate cloud-top surface noise prior to defining the PMM algorithm. In the following analysis, the FM, RHM, and SHM algorithms are tested to determine the most effective mapping rule for rainfall nowcasting.

PMM algorithm validation

The PMM algorithms are validated with 1-min rain gauge data from a dense gauge network deployed over South Korea. The network, illustrated in Fig. 1, consists of 395 uniformly distributed, 0.5-mm-resolution, automatically reporting tipping rain gauges continuously operated by the Korean Meteorological Administration (KMA). Given the geographic size of South Korea, this network yields a gauge density of approximately 35 per unit degree of latitude–longitude. The algorithm estimates are validated by compiling gauge–satellite RR difference statistics—but only for pixels surrounding gauge measuring sites and coincident in time with the measurements (within 1 min). This restriction eliminates the great uncertainty involved in comparing rain rates derived from instantaneous satellite retrievals with ground validation measurements, which are typically disjoint in time–space and accumulated over periods of 1 h or more.

Going to 1-min-resolution gauge measurements nearly eliminates the time-difference factor. Moreover, selecting only those satellite pixels that surround valid gauge reports nearly eliminates the space-difference factor. In creating an IR-retrieved rain rate to pair with a measured rain rate, the weighted average of a 5 × 5 pixel subgrid is used (∼25 km × 25 km region), with the center pixel selected according to its proximity to the associated gauge position. A two-dimensional binomial filter is used for the pixel subgrid weighting scheme. After compiling many pairings over two extensive rain days (some 6312 samples), the differences are aggregated to enable calculation of bias, scale factor, and correlation coefficient comparison terms. Aggregation is essential because 0.5-mm resolution for a tipping rain gauge quantizes the gauge measurements to a step size of 30 mm h−1. This is actually reduced to 10 mm h−1 by taking 3-min arithmetic gauge averages centered at the times of matchup.

Nowcasting system components

The second stage of the nowcasting system involves forecasting areas of rainfall and their associated rain rates over 3–6-h forecast periods. The system uses consecutive IR images from GOES and GMS satellites to produce the QPFs. The forecasting process consists of three steps: 1) retrieving cloud motion wind (CMW) fields, 2) creating filtered CMW fields, and 3) making rain forecasts through application of the PMM algorithms to predicted IR images generated from a CMW-based advection technique.

CMW retrieval

Prior to forecasting, an estimated wind field is created based on the movement of clouds between consecutive IR images. During this process, it is essential to ensure that only clouds and not background features enter the cloud-tracking procedure. Therefore, screening out noncloudy background pixels is of primary concern. Each IR image over a diurnal cycle used in the nowcasting process is analyzed to find a characteristic background temperature. Under the assumption that convective clouds exhibit EBBTs lower than that of the background, Rossow and Garder (1993) applied a threshold temperature method in which values of 3.5°C below ocean background and 6.5°C below land background defined the cloud EBBT thresholds. We opted to tighten the selective procedure and subtract 10°C from the background EBBT to define the cloud EBBT threshold. Afterward, any IR pixel in an image sector with an EBBT greater than this threshold is designated as background and is set to that value.

The next step in the CMW retrieval process is to define three grid systems for the sequential IR image sectors. The top panel of Fig. 2 shows how these grid systems are organized within the domain of a given IR sector. The first is the “target grid system” and contains what are called target subgrids, outlined in thick black in the top panel. These subgrids in the first of a consecutive pair of full-resolution IR images contain the individual cloud tracer targets used in conjunction with search areas in the second IR image (bottom panel of Fig. 1) to derive the CMW vectors. Each target subgrid is 21 × 21 full-resolution IR pixels in size, resulting in a target subgrid domain of about 100 km × 100 km for GMS and about 80 km × 80 km for GOES imagery.

A process similar to cross correlation is employed to detect cloud motion between an IR image pair as indicated in the bottom panel of Fig. 2. A search area of 71 × 71 pixels based on ±25 pixel lags is established around each target subgrid on the second IR image. The minimum difference method described by Smith and Phillips (1972) is then used to calculate the cloud pattern difference matrix over the lag domain, each element of which contains the accumulated pixel differences between pairs of 21 × 21 subgrids. The matrix element with the least difference identifies the 21 × 21 match subgrid, with its center and the center of the corresponding target subgrid forming the endpoints of a CMW vector and thus the estimated wind direction. The time difference between the pair of IR images determines the vector magnitude and thus the estimated wind speed. Vectors are only calculated for target subgrids with a sufficient number of cloudy pixels (>5% of the 441 possible target subgrid pixels).

This scheme creates a field of CMW vectors, but with possible gaps (voids) associated with target subgrids that contain too few cloudy pixels. Before the initial CMW field is ready for use in rain forecasting, a second grid system is introduced called the “filtered grid system.” It is represented by the grid of 7 × 7 subgrids outlined with dashed thin lines in the top panel of Fig. 2, the subgrids having a domain size of about 35 km × 35 km for GMS or about 30 km × 30 km for GOES. The filtered subgrid elements are created by a process of interpolating and smoothing the initial CMW vector field. Note the filtered grid system is located inside the domain of the target grid system and is smaller in dimension by the equivalent of two filtered subgrids. Forecasting is accomplished through the use of the filtered CMW field to which two parameters are assigned for each subgrid element: 1) the filtered CMW vector and 2) the change in EBBT (ΔEBBT) between associated EBBTs from the target and match subgrids. This latter parameter is used as an EBBT tendency for the QPFs.

The third and final grid system is designated to contain forecast RRs and is therefore denoted the “forecast grid system.” Individual elements of this grid system are equivalent in size to those of the filtered grid system but encompass almost the entire IR sector domain, as seen in the top panel of Fig. 2 (with the exception of a 4-pixel border). Therefore, this grid system is larger than the target grid system, which is confined to the interior of an image sector because of the ±25-pixel lag dimensions used for cloud tracking. This grid system is employed for each forecast time step and leads to a predicted EBBT field—which, after application of the PMM algorithm, enables assigning forecast rain areas and rain rates.

Filtered CMW field

Upon creation of an initial set of CMWs, gaps in the estimated wind field may exist. It is found that by creating a continuous, resolution-enhanced, and smoothed CMW field from the initial low-resolution CMW field containing voids, it is possible to optimize forecast skill. We call this product the filtered CMW field. Thus, after an initial CMW field is derived, a combination of interpolation and smoothing is performed to fill the voids, enhance the estimated wind field resolution, and obtain a wind field better suited for forecasting. Also, through inspection of the initial CMW fields, it is apparent that a scheme is needed to minimize the influence of the occasional “bad” CMW vectors that arise from using an automated cloud tracking algorithm. This can be done by invoking a Monte Carlo averaging operator prior to the application of the interpolation operator. In this part of the analysis, the most effective combination of Monte Carlo averaging, interpolation, and smoothing is determined.

A total of 25 filtering schemes were tested involving 1) an optional a priori Monte Carlo averaging operator to lessen bad-vector effects, 2) an interpolation operator to fill gaps in the initial 100-km (or 80 km) low-resolution CMW fields while enhancing the resolution to the 35-km (or 30 km) high-resolution fields, and 3) a smoothing operator to suppress discontinuities in the forecast fields.

A two-dimensional linear interpolation operator is used to fill voids onto the high-resolution grid and is employed either with or without the a priori Monte Carlo averaging operator designed to suppress bad-vector effects. The latter operator consists of multiple randomized selections of either 50% or 75% of the initial low-resolution CMW vectors prior to interpolation for either 10 or 100 realizations (repetitions), then averaging all realizations at low resolution before interpolation to high resolution. In addition, four two-dimensional smoothing operators are tested. The first is a 5 × 5 cusp filter with a weighting vector of 1-1-4-1-1. The remaining three are binomially weighted filters of dimensions 3 × 3 (9 point), 5 × 5 (25 point), and 7 × 7 (49 point). Table 1 summarizes the 25 schemes, and Fig. 3 provides examples of both an initial CMW field with voids and a filtered field.

Rain forecasting

Assigning estimated u, υ wind vectors at each element of the filtered grid system enables the QPF process to be invoked. Filtered CMWs are used to advect each subgrid from the second IR image into a forecast image. In doing so, the tendency term ΔEBBT is used to adjust the EBBT values of individual subgrids. Note that a negative (positive) ΔEBBT represents a radiometric cooling (warming) trend. Thus, for each forecast time step, 7 × 7 subgrids are advected according to their associated filtered CMW vector, while the individual EBBT pixels are adjusted according to their associated ΔEBBT.

After the forecast process, it is inevitable that gaps remain within the forecast grid. Thus, an interpolation scheme is used to smooth the gaps not assigned EBBTs but left containing the preset background temperature. The interpolation operator audits the forecast EBBT field for subgrids that contain preset background temperatures and, when found, considers the EBBTs of the eight surrounding subgrids. The requirement for temperature interpolation of preset background pixels is that at least five of the eight immediately surrounding subgrids must contain EBBTs colder than the cloud threshold. In this event, an EBBT replaces a subgrid's background values based on weighted averaging of the surrounding subgrid values with weights assigned according to distance proximity.

Upon completion, each full-resolution pixel within the forecast grid system contains a predicted or interpolated EBBT consistent with the movements and temperature changes of the filtered subgrid elements. The individual pixels of the forecast image fall into one of three categories: 1) cloudy with rain, 2) cloudy with no rain, and 3) clear. Pixels surmised to contain rain are assigned RR values using the appropriate EBBT–RR mapping rule.

One aspect of the QPF process that invites careful examination is the advection technique. Although the above describes use of two consecutive IR images for creating a forecast using the constant velocity advection technique, we also examine performances of steady, linear, and nonlinear acceleration advection techniques, which requires extending IR sequences beyond two members. Overall, four techniques are examined: 1) constant velocity advection (CVA), 2) constant acceleration advection (CAA), 3) linear acceleration advection (LAA), and 4) nonlinear acceleration advection (NAA).

CVA uses one CMW field created from two IR images; that is, the u and υ components remain constant (no acceleration) for all forecast time steps throughout the forecast period. CAA uses two consecutive CMW fields to define a linear trend in velocity, which requires three consecutive IR images. Changes in the u and υ components are linearly extrapolated forward in time, producing a constant acceleration. LAA uses three consecutive CMW fields with a quadratic extrapolation of velocity components, which requires four consecutive IR images. Second-order polynomials are fit between the three consecutive uυ component pairings at each interpolated grid point, thus enabling nonsteady acceleration or linear acceleration advection. NAA is a nonlinear acceleration advection process based on four CMW fields derived from five consecutive IR images. Here, cubic splines are generated between the four consecutive u and υ components at each filtered grid point and used for extrapolation. The best of these four techniques is selected on the basis of forecast skill scores in order to design the most effective nowcasting system.

Description of datasets

The dataset used to create land and ocean PMM algorithms was developed from approximately 250 000 coincident EBBT–RR pairings derived from GMS-5 and SSM/I measurements [processed by the Ferraro (1997) algorithm] over the time period from June 1997 to January 1998. Data were collected over a 10° × 10° sector encompassing the Korean peninsula (32°–42°N, 123°–133°E), which provided a sufficient area to create EBBT–RR mapping rules for both land and ocean cases.

As noted, the dataset used to validate the PMM algorithms is derived from the KMA rain gauge network situated over South Korea. The measurements are reported as accumulations in 1-min intervals from 0.5-mm-resolution tipping gauges and have been provided by the KMA. (The KMA uses the gauge network operationally in their weather forecasting division.) The dataset covers two full days, 31 July and 3 August 1998, days that had extensive rains and some flooding over the southern Korean peninsula. Hourly GMS-5 infrared imagery were collected over these two days (a total of 48 image sectors) to enable an independent validation analysis.

Three separate GOES and GMS-5 IR datasets were used to test the nowcasting system. The first of these is a set of 41 half-hourly IR images from GOES-8 on 1 April 1998. Data were collected over a 26° × 55° sector (7°–33°N, 50°–105°W) containing both the Gulf of Mexico and the Caribbean Sea, hereinafter referred to as the Caribbean basin. The second dataset is also from GOES-8 and consists of 167 hourly IR images situated over the Amazon basin collected between 29 January and 7 February 1999, covering a 30° × 30° sector (15°S–15°N, 55°–85°W). This dataset provides a long time series for testing the nowcasting scheme over the highly convective Amazon region during its wet season. The third dataset is from GMS-5, consisting of 17 hourly IR images for 1 May 1999, covering a 30° × 30° sector (20°–50°N, 110°–140°E) situated over the Korean peninsula. These three datasets enable testing the entire nowcasting system over several areas of the world, including both land and ocean regimes.

Discussion and interpretation of results

Rain retrieval

Selection of PMM algorithm

Three PMM algorithms and the associated EBBT–RR mapping rules were developed for both land and ocean cases based on the FM, RHM, and SHM methods. To ensure consistency in the comparisons, each of the three mapping rules have equivalent ranges and bin sizes for both EBBTs and RRs. The ranges and bin sizes were determined from different PMM algorithms. For RR, the FM method was used, in which the lognormal PDF is resolved at a 0.5 mm h−1 spacing starting from 0 and continuing to the maximum allowable value of 35 mm h−1. For EBBTs, the rain–no rain threshold is established by the RHM method for both land and ocean. Thresholds are 233.5 K for ocean and 228.5 K for land; these values represent the warmest raining EBBTs for each PMM algorithm. Matching of EBBTs continue through the colder end of the distributions until the minimum allowable EBBT for matching occurs at 199 K (any colder values are redefined to be 199 K), with the EBBT bins determined by the RR intervals. Figure 4 illustrates the differences between the raw histogram, smooth histogram, and two-parameter gamma PDF distributions used both for land and ocean EBBTs in the probability matching process.

As might be expected, the RHM and SHM algorithms produce similar EBBT–RR mapping rules. However, with the FM algorithm, although the gamma distribution provides a suitable fit to the raw EBBT histograms, Fig. 5 indicates problems are encountered in fitting lognormal PDFs to raw RR histograms. Although this PDF produces a good fit at the light end of the RR distribution (top panel), a significant low bias is found at the heavy end (bottom panel). Thus, a PMM algorithm based on the lognormal PDF representation of RRs and the gamma PDF representation of EBBTs will of necessity create a low bias in the transformed RRs. The source of the bias stems from fitting a lognormal PDF to RR data limited to a maximum of 35 mm h−1, as stipulated by the operational NESDIS algorithm.

Figure 6 presents normalized frequency distributions complied for the three EBBT–RR mapping rules, for both land and ocean cases, extracted from all pixels in the 0715 UTC GOES-8 IR image for the Caribbean basin. The RHM and SHM results for the ocean case show similar tapered distributions, with peaks at relatively light RR values. The ocean FM distribution exhibits even lower RRs (due to the low bias), largely restricting RR values from exceeding 3 mm h−1. The RHM and SHM results for the land case exhibit more broadened distributions with modal frequencies at light RRs, but also moderate frequencies of medium to heavy RRs. The land FM results show the bias toward light RRs with values confined to less than 10 mm h−1. Average RRs produced from each algorithm are given in Table 2. For both land and ocean cases, the FM algorithm produces average values that are less than one-half of those produced by the RHM and SHM algorithms. This highlights the rain-rate bias that would be transferred to the forecasting process were the FM algorithm selected for the nowcasting system.

Because the objective is to identify the probability matching method that is best suited for the nowcasting system, the FM algorithm must be rejected. Although the RHM and SHM algorithms produce nearly similar results, the SHM algorithm is selected for both ocean and land cases because it eliminates spurious cloud-top surface noise contained in nearly all EBBT distributions. Notably, the smoothing also accentuates a consistent feature observed in all EBBT histograms we have analyzed, that being the wave form within the distribution region of ∼225–255 K, as is apparent in Fig. 4. This represents a feature that two- or three-parameter FM algorithms could not characterize. Figure 7 illustrates the final land and ocean EBBT–RR mapping rules by the SHM algorithm, and Table 3 provides the numerical EBBT–RR pairings. Similar temperatures are paired with different RRs for land and ocean because the PMM, as described in section 2a(2), was performed over both kinds of backgrounds, creating separate EBBT–RR mapping rules.

Validation of SHM algorithm over South Korea

Each of the two rain days selected for the validation analysis over South Korea (31 July and 3 August 1998) involved heavy convective storms over different regions of South Korea and extensive light showers over much of the rest of the southern peninsula. On 31 July, there were three major convective storm systems: 1) a slow-moving concentration of morning storms that developed west of Seoul (∼37.5°N, 126.5°E) by around 0030 UTC (0930 LT), moved slowly eastward, and dissipated by around 0330 UTC; 2) a multicell evening storm that developed in the southwest of the peninsula (∼35°N, 127°E) at around 1130 UTC and moved northeast along the southern coast until dissipation at around 2030 UTC; and 3) an east-southeast-moving late-evening squall line that developed along the west coast at around 1630 UTC and moved across the peninsula until dissipation at around 2130 UTC.

On 3 August, the heavy rains were situated around the Seoul area (∼37.5°N, 127°E), with two major events occurring in the early evening and early morning at around 1430 UTC and around 2030 UTC. In addition, light to moderate shower activity took place all along the west coast throughout the day, with propagation eastward in central South Korea, with the heaviest events taking place along the central west coast and southwest coast.

Based on the 6312 gauge–satellite RR differences compiled for the 48 hourly IR images associated with the two days, bias b, scale factor sf, and modified cross-correlation r terms have been calculated. Because of the quantization effect in the gauge measurements, the root-mean-square difference has little meaning and is thus not considered. Here, b is defined as the overall satellite mean RRs minus the overall gauge mean RRg, and sf is given by (1 + b/RRg). The bias for the total sample pool is −1.16 mm h−1 with a corresponding scale factor of 0.79. The associated rain gauge and satellite means are 5.40 and 4.24 mm h−1, respectively. This means that the systematic difference between the GMS-5 IR rain-rate estimates (as dictated by the SSM/I RR retrievals through the EBBT–RR land mapping rule) and the observed rain rates from individual rain gauges is approximately 20%.

Direct correlation is poor when all samples are considered because of both the quantization effect and the numerous light IR rain rates—the resultant r is only 0.13. However, if just the larger rain rates are considered and a modified correlation coefficient is taken by correlating between the bin indices for the gauges (every 10 mm h−1 between 0 and 120 mm h−1 for 3-min gauge averages) and the averaged IR values within each bin, correlations improve from 0.15 to 0.84 as a light satellite rain-rate cutoff between 1 and 5 mm h−1 is applied. For purposes of nowcasting, we regard the above validation results as acceptable measures of validation success.

Nowcasting

Our main objective in nowcasting is to employ the selected PMM algorithms to generate QPFs out to 6 h, or as long as the forecasts produce useful results. This involves determining the most successful forecasting procedures based on the variety of options described in section 2. To achieve this aim involves four steps: 1) selecting the most effective interpolation-smoothing scheme for producing the filtered CMW fields, 2) designing and testing a forecast verification procedure, 3) selecting the most effective EBBT advection technique, and 4) assessing the decay of forecast success toward its asymptotic limit.

Selection of CMW filtering scheme

Each filtering scheme was tested on the three GEO datasets for QPFs out to 6 h using CVA as the forecasting procedure. As will be shown later, CVA proved to be the most effective advection technique. Verification of rainfall forecasts is performed by intercomparing a forecast IR image to an observed IR image at the associated verification time. This procedure is explained in detail in the next section. It was found that a small selection of the filtering schemes produced superior forecast results.

Scheme S19 (see Table 1) exhibited significant success in forecasting rain areas beyond 3 h over the Korean peninsula. Scheme S9 exhibited the most promise in forecasting rain rates, when considering all three datasets together. Although schemes S4 and S25 exhibited general skills, they were offset by a negative “overforecasting” effect in that both consistently produced too little rain area coverage. The most consistently successful scheme considering all aspects of the forecasting was S8. This scheme consists of applying the a priori Monte Carlo averaging operator using 50% vector selection with 10 realizations, followed by two-dimensional linear interpolation onto the high-resolution grid, and finishing with a nine-point binomial smoother. This is the final scheme of choice for generating the filtered CMW fields.

Design and testing of forecast verification procedure

To quantify the skill of the QPFs, forecast IR images are compared with observed IR images at the associated verification times. A grid system is first established within an observed IR image that coincides with the associated forecast image. This grid is equivalent to the forecast grid system described in section 2 (see Fig. 2). Thus, a grid of 7 × 7 pixel subgrids is defined over the IR image sector with averaged EBBTs calculated for each subgrid. If the average EBBT within a subgrid is colder than the rain–no rain threshold established by the SHM algorithm, then the box is designated as “raining” and an RR is assigned using the EBBT–RR mapping rule. If the average EBBT is warmer than the rain–no rain threshold but colder than the cloud threshold, the box is denoted as “cloudy.” Cloudy areas do not figure into rainfall verification but provide a means of visually analyzing forecast images for veracity. Areas with EBBTs warmer than the cloud threshold are designated as “clear.” Thus, rain areas and corresponding RRs serve as the primary basis for forecast verification.

Both rain area and rain rate are inspected when verifying a rainfall forecast. They are considered separately so the value of the nowcasting system can be evaluated along different lines. In considering rain area, a measure denoted as the rain-existence skill score (RESS) is defined by compiling a binary truth table assigning a “yes” or “no” considering the presence of rain in each 7 × 7 pixel subgrid in the verification and forecast images. The RESS is then the ratio of table counts in agreement concerning the presence of rain (i.e., sum of counts in yes–yes bin) to the sum of raining table counts in the verification image (i.e., sum of counts in both no–yes and yes–yes bins). This quantity scores the forecast purely on its ability to predict rain area.

A second rain area skill score, denoted as the false-alarm rate (FAR), is used to measure overforecasting. It also uses the binary truth table; however, this quantity is the fraction of 7 × 7 subgrids that are forecast to contain rain but do not contain rain in the verification image (i.e., ratio of yes–no bin count to sum of yes–no and yes–yes bin counts). When analyzing forecasts for the presence of rain in the correct location, it is desirable to obtain a high RESS in conjunction with a low FAR.

A third verification measure focuses on rain-rate magnitude and is denoted the rain-intensity skill score (RISS). The RISS is not so much a measure of the success of the advection process but rather is a measure of how well the nowcasting system utilizes the initial EBBT and ΔEBBT information, because it is the forecast EBBTs that determine the resulting RRs within each 7 × 7 pixel box. To calculate the RISS, RRs are divided into three categories: 1) light, 2) moderate, and 3) heavy. The light category contains RRs of less than 5 mm h−1, moderate RRs fall between 5 and 20 mm h−1, and heavy RRs are greater than 20 mm h−1. A 3 × 3 contingency table [as described in Wilks (1995)] is created to compare the intensity of the forecast RRs with the verification RRs. Only matches (forecast light RR with verified light RR, etc.) are considered successful in this table, that is, the sum of counts in bins along the diagonal. Therefore, the RISS measures the ratio of the successfully matched light, medium, or heavy RRs to the total number of nonzero RRs in the comparison (sum of all table counts). Because 0 RRs are not considered, a high RISS indicates that the nowcasting system successfully utilizes the initial EBBT and ΔEBBT information, without attention to skill in forecasting rain area.

Tests of the verification procedure have been conducted with all three GEO datasets and are illustrated in Figs. 8–10. As before, the CVA technique has been used in the forecasts. Figures 8a,b show an example of 1–3-h forecast results for the Caribbean basin. Figure 8a shows rain-rate conditions at 1015 and 1045 UTC 1 April 1998 (top and bottom panels), along with the filtered CMW field and ΔEBBT tendencies derived from the associated EBBT images (middle panel). Rain-rate pixels are 30-km resolution associated with the forecast grid. The salient rain feature in this panel is the large storm cluster near 60°W longitude. The six panels in Fig. 8b show the 1-, 2-, and 3-h rain forecasts along with the verification rain images and associated skill scores. Table 4 summarizes the skill scores for these forecasts and the corresponding 4–6-h forecasts and for the related Amazon basin and Korean peninsula forecasts.

The upper-left panel presents the 1-h forecast with an RESS value of 66.0% and a FAR value of 39.6%. A feature that lowers the RESS and increases the FAR is the north-northeast-to-south-southwest-oriented streak of rainfall forecast to the north of the main storm cluster that does not appear in the verification images. If one overlooks that feature, the forecast is generally successful in short-term prediction of rain location. The middle-left panel shows the 2-h forecast with an RESS of 62.1% and a FAR of 40.1%. The degradation of the RESS is due to the dispersion of the main storm cluster by the forecast. The lower-left panel shows further dispersion of the main storm cluster by hour 3. This forecast has an RESS of 48.6% and a FAR of 40.5%. RISS values for all three forecast periods are relatively high at ∼70% and do not degrade over the three hours. As seen in Table 4, the RESS, FAR, and RISS scores for the 4–6-h forecasts do degrade with forecast time (as expected) but nevertheless exhibit noticeable accuracy out to the 6-h point.

Figures 9a,b present, in a similar format, successful 3-h forecasts for the Amazon basin. RESS values (Fig. 9b and Table 4) range from 72.1% for the 1-h forecast to 68.8% for the 2-h forecast, ending with 62.6% for the 3-h forecast. The associated FAR values increase from 36.1%, through 42.1%, to 50.6%. Success with these forecasts is due to the near-linear progression of the storm in the center of the region throughout the 4-h period between 0745 and 1145 UTC on 1 February 1999. RISS values are lower than found for the Caribbean basin, degrading from 53.1% to 40.9% over the 3-h forecast period. As with the Caribbean basin, the skill scores for the 4–6-h forecasts continue to deteriorate; however, even at the 6-h point, the RESS scores are still indicating reasonable accuracy.

Although Figs. 10a,b depict a successful 3-h forecast for the Korean peninsula, the associated skill scores are lower than found for the other two cases (4–6-h forecasts were not produced for this case). RESS values (Fig. 10b and Table 4) range from 46.2% for the 1-h forecast to 37.0%, then to 30.2% for the 3-h forecast. The associated FAR values are relatively high, ranging from 65.8% to 68.1%. The explanation for the high FARs is the presence of Typhoon Leo in the southwest corner of the verification panel. Development of high, thick cirrus clouds exiting the northeast quadrant of the typhoon leads to predicted EBBTs cold enough to produce rain. This is seen in the southern third of each forecast panel, where spotty predicted rain areas do not verify. The RISS values decrease from 63.0% in the 1-h forecast to 53.4% in the 3-h forecast. This is a good example of the nowcasting system forecasting reliable rain rates, albeit overlooking the false-alarm rates.

In an additional test to assess forecast accuracy, equitable skill scores (ESSs) were calculated in the manner of Gandin and Murphy (1992) for the same forecast time sequences as presented in Table 4 and then were compared with RESS. ESS measures the accuracy of forecasts relative to procedures such as using mean climatic values (“climatology”) or persistence of current conditions. The ESS values calculated for the aforementioned time sequences were slightly lower than RESS values (but within 5%) and exhibited a similar falloff with respect to forecast time.

Note that there are discernible gaps around the borders of the forecast panels in Figs. 9b and 10b. These gaps are due to the domains of the forecast grid system, from which the forecasts emanate, being inside the actual IR image sectors (see Fig. 2). These gaps remain until cloud systems are advected to the edges of a sector domain.

All of the three cases have been subjected to synoptic and mesoscale analysis to relate the meteorological conditions and properties of the embedded convection with the magnitudes of the skill scores. In general, the highest skill scores are associated with dynamically organized, intense convective storms—in the front half of their life cycle and moving from slow to moderate speeds (3–15 m s−1). For example, high skill scores were obtained with the moderately paced storms shown in Figs. 8a and 9a. Fast-moving storms (>15 m s−1) tend to create dispersive effects in the forecast EBBTs, whereas stratiform and cirrus cloudiness produced in older storms do not track as well as convective elements and at times lead to spurious tracer motion because pattern definition becomes transitory. An example of this can be seen with the relatively low skill scores for the storm near the Korean peninsula in Fig. 10a. Weaker storms characterized more by buoyancy than by dynamic flow conditions, such as low-level speed and directional convergence, vertical shear, and upper-level potential vorticity advection (PVA), are not forecast as well for the one-hour-to-longer forecast periods.

Selection of advection technique

The next step involves selecting the optimal QPF procedure among the CVA, CAA, LAA, and NAA advection techniques. Comparisons are made on the basis of verification of 1-, 2-, and 3-h forecasts with each of the three GEO datasets. Whereas Figs. 8–10 and Table 4 have provided a look at skill scores associated with the CVA technique, here we examine all four advection techniques together. Ensembles of forecasts were created for the entire Caribbean basin, Amazon basin, and Korean peninsula datasets. Forecast ensemble counts are 31, 27, and 24 for the 1-, 2-, and 3-h forecasts for the Caribbean basin, respectively. The associated counts for the Amazon basin are 132, 121, and 114, and for the Korean peninsula are 9, 8, and 7.

The left column of three panels in Fig. 11 shows comparisons of the ensemble-averaged forecast skill scores for each advection technique for the Caribbean basin. Note these are averages of each skill score (i.e., for RESS, FAR, and RISS taken individually) from the forecast ensembles for this particular dataset for the three individual forecast periods. The average RESS values are similar for all four advection techniques for the 1-h forecasts and then diverge for the 3-h forecasts. The steady techniques show more promise, considering average RESS values, with CVA surpassing all techniques. Each set of average FAR values increases about 5%—with CVA exhibiting the best results. Although the average RISS values exhibit little variation with technique, the two steady techniques exhibit slightly higher values, with CAA producing the highest values for the 2- and 3-h forecasts. Considering all three verification skill scores, the steady techniques outperform the nonsteady techniques, with CVA exhibiting the best overall results.

The middle column of three panels in Fig. 11 presents results for the Amazon basin. Average RESS values exhibit a nearly linear degradation over the 1–3-h forecasts, with CVA and CAA scoring higher than the nonsteady acceleration techniques; that is, average RESS values for the latter two decrease more rapidly. Although the four average FAR values for the 1-h forecasts are close to one another, those for LAA and NAA undergo larger increases for the 3-h forecasts in comparison with the scores for CVA and CAA, which exhibit smaller increases. Average RISS values are fairly low for all techniques in this dataset. The steady techniques are some 10% more skillful than the nonsteady techniques for the 1-h forecasts, although they exhibit little differentiation for the 2- and 3-h forecasts—that is, with the exception of the poorer performance of the NAA technique for 3-h forecasts. As with the Caribbean basin, the steady techniques consistently outperform the nonsteady techniques.

The right column of three panels in Fig. 11 presents results for the Korean peninsula. Average RESS values are relatively low overall in comparison with the other two cases, less than 50% for the 1-h forecasts. Despite the lower average RESS scores, the steady techniques again indicate better results. Average FAR values are very high for each technique. For this case, the NAA technique exhibits the best results for 1-h forecasts but undergoes sharp increases for the 2- and 3-h forecasts, as does the LAA technique. The steady techniques exhibit smaller increases. The average RISS scores are similar for the steady techniques for all three forecasts and are consistently higher than the nonsteady techniques by some 5%–10%. As with the Caribbean basin and Amazon basin cases, the steady techniques produce overall superior results.

Based on the evidence presented in Fig. 11, the steady CVA and CAA advection techniques produce the most accurate rainfall forecasts. Moreover, CVA produces better overall performance than CAA, considering all three cases. Although all four techniques occasionally indicate similar skill scores, the similarities are largely confined to the 1-h forecasts. Therefore, CVA is selected as the optimal technique. This result suggests, albeit with caution because of the limited number of cases, that there is little memory in the motion of rain storms prior to a few hours.

Asymptotic limit of forecasts

Having established and tested the optimal nowcasting system, we conclude the analysis by seeking an objective estimate of the maximum forecast period for which useful QPFs can be made with this approach. This is accomplished by extending CVA forecasts out to 10 h (a much longer period of time than that to which nowcasting methods are generally applied) and examining the decay behavior in the skill scores. The objective here is to determine the time at which the skill scores deteriorate toward an asymptotic limit, which we will define as the time when forecasts no longer contain any useful prediction information. Because the RESS and FAR scores are most indicative of the success of the derived CMW fields and advection scheme, these are the scores analyzed to answer the question concerning the asymptotic limit. We note that RISS scores exhibit significant degradation beyond 3 h, which is why they are not considered in this analysis. We also note that because the skill score curves are not true mathematical functions, the analysis for asymptotic convergence is necessarily subjective.

The left two panels in Fig. 12 present ensemble-averaged RESS and FAR values for up-to-10-h forecasts produced from the entire Caribbean basin dataset (ensemble count for 1-h forecasts is 31, decreasing to 20 for 10-h forecasts). A persistent decline in average RESS values is seen for the first 6 h, with approximately 30% success achieved for the 6-h forecasts. Thereafter, the decline in average RESS is relatively small. The FAR values exhibit a persistent increase up through approximately 7 h of forecasting, overlooking the abrupt increase between the 9- and 10-h forecasts. Beyond that point, the rise in FAR is relatively small. An assessment of the forecast limit for the Caribbean basin is therefore approximately 6–7 h.

The right two panels of Fig. 12 present the same analysis for the entire Amazon basin dataset (ensemble count for 1-h forecasts is 132, decreasing to 80 for 10-h forecasts). Here the ensemble-averaged RESS and FAR values exhibit much smoother behavior from which to determine the asymptotic limits. This is because the forecast ensemble counts for the Amazon basin are over 10 times as large as for the Caribbean basin. The average RESS scores decline steadily out to the 6-h forecast point, reaching an ∼20% skill score—converging thereafter to an ∼10% skill score. Average FAR scores increase steadily toward the 6-h forecast, at which point nearly 80% of the forecast rain area fails to verify—converging thereafter to ∼90%. Therefore, the Amazon basin probably has a forecast limit of approximately 6 h. Based on the results presented in Fig. 12, we cautiously estimate that rainfall forecasts containing useful skill can be produced by this nowcasting system out to approximately 6 h. Beyond that point, there is little meaningful prediction information.

Summary and conclusions

A method for nowcasting rainfall through the use of sequential infrared geosynchronous satellite imagery, along with forecast skill scores obtained for three separate GMS and GOES datasets, are described. Probability matching between infrared temperature measurements and rain rates retrieved from SSM/I brightness temperatures is used to generate a rain algorithm for forecasting areas and intensity of rainfall based on cloud motions between satellite images.

A verification procedure to determine how, where, and when the nowcasting system excels and fails in relative terms is devised. Each component of the nowcasting system is analyzed to define the most successful methods. This includes designing an optimal PMM algorithm for generating the land–ocean EBBT–RR mapping rules, producing a cloud motion–based wind field, selecting a filtering scheme for generating a continuous cloud-motion wind field, selecting an advection technique for forecasting, and determining the forecast limit for which meaningful QPFs can be expected.

One of the conclusions drawn from the study addresses probability matching between EBBTs and RRs. Two of the matching methods, raw histogram matching and smooth histogram matching, produce viable results, but function matching produces less than satisfactory results. For example, matching between a lognormal PDF fit to RR data and a gamma PDF fit to EBBT data cannot provide results comparable to those from smooth histogram matching. This is because the success of the fit of a lognormal PDF to the SSM/I rain-rate data deteriorates toward the intense end of the rain-rate histogram, generating a low rain-rate bias. However, it is the maximum allowable rain rate in the NESDIS SSM/I algorithm of 35 mm h−1 that limits the effectiveness of lognormal fitting. Therefore, we do not conclude that function matching lacks merit for future studies involving probability matching between EBBTs and RRs. Instead, we conclude that function matching requires analysis of the RR algorithm to ensure that statistical distributions closely conform to retrieval histogram properties.

The probability matching method that shows most promise in this study is smooth histogram matching, similar to the raw histogram matching, differing only in its propensity to suppress noise in the raw EBBT histograms. In addition, there is a subtlety associated with this result consisting of a reoccurring and accentuated wave form that appears in the EBBT distributions over the temperature range of about 225–255 K. Upon inspection, it is evident that this feature is due to a consistent tendency of the EBBTs to exhibit bimodality within their intermediate temperature distributions. The colder of the bimodal peaks represents cirrus cloud cover associated with the stratiform stage of precipitation and could represent a useful signal in attempting to decompose rainfall into its convective and stratiform parts. Such a feature could not be fit with two- or three-parameter statistical PDF forms and would not be uniformly represented when using raw histograms.

We note that the Turk et al. (2000) study has analyzed the effectiveness of continuous updating of the EBBT–RR mapping rules for regional domains, which we did not address. Whereas continuous updating can generate a more accurate rain-rate algorithm because it compensates for the everchanging relationship in time and space between rain rates and cloud-top EBBTs, it would not be needed in a nowcasting system until it could be shown that the accuracy gain from a refined mapping rule would exceed that of the intrinsic accuracy in EBBT predictions.

The nowcasting system design that exhibits the best skill scores combines smooth histogram matching for generating the PMM algorithm, a Monte Carlo averaging operator to suppress bad CMW retrievals, a two-dimensional interpolation operator with a nine-point binomial filter to create a suitably filtered cloud-motion wind field at approximately 30–35-km resolution, a constant EBBT tendency term, and a constant velocity advection technique for creating the forecast infrared imagery. The CVA technique consistently outperformed the three other nonsteady velocity advection techniques. The differences between the CVA and CAA techniques were not found to be all that significant; however, CVA consistently exhibited the highest skill scores. By the same token, the nonsteady acceleration advection techniques consistently exhibited poor skill scores, particularly for forecasts of 2 h and longer. Even for 1-h forecasts, the nonsteady acceleration techniques were almost always outperformed by the steady velocity or acceleration techniques. These results suggest that there is little meaningful memory in storm motion, considered independently, prior to several hours.

The best success with the nowcasting system was found with storms driven by low-level convergence and upper-level PVA, during the front end of their life cycle with ground speeds that are slow to moderate. Less intense airmass thunderstorms, well-organized but short-lived storms that are not driven by lower- or upper-level dynamics, that exhibit less definitive movement and development, as well as older storms in their stratiform stage and producing excessive amounts of cirrus, produced lower skill scores. Also, skill scores appear to deteriorate as storm speeds increase, although this is difficult to confirm with only a few case studies.

The largest verification skill scores were found for storms in the Amazon basin, with RESS scores sometimes exceeding 70% for 1-h forecasts and 60% for 4-h forecasts, although with FAR scores generally only less than 50% for 1-h forecasts. The Caribbean basin forecasts also exhibited high RESS scores, exceeding 60% for 1- and 2-h forecasts, and relatively low FAR scores of about 40%. Over Korea, RESS values were smaller, typically about 45% for the 1-h forecasts, diminishing to near 30% for 3-h forecasts. The associated FAR values were typically about 65% and 75%, respectively.

The variability of forecast skill scores with the three datasets suggests that unexplored factors may affect the performance of the nowcasting procedure. For example, one of the questions concerns the characteristic differences in verification scores between the Caribbean basin and Amazon basin forecasts, given that both datasets contain persistent storms. The major difference in acquiring these two datasets is the change from half-hourly Caribbean basin imagery to hourly Amazon basin imagery. To determine the influence of the imagery time step, we ran the 1-, 2-, and 3-h Caribbean basin forecasts for the 1015–1045 UTC image pair, switching the 1015 UTC IR image with the 0945 UTC image and generating a forecast from an hourly time step. Although, the skill scores changed, the changes were small, with RESS scores actually increasing by a few percent for the 1- and 3-h forecasts, the FAR scores decreasing a few percent for the same two forecasts, and the RISS scores decreasing about 3.5% and 6.5%, respectively. This result suggests that the basic difference between Caribbean basin and Amazon basin forecasts was largely due to meteorological conditions.

In considering improvements to the nowcasting scheme, the use of visible imagery might help because of the generally higher resolutions available from shortwave channels, although any such improvement is only achievable during daylight. Water vapor imagery might also aid with the generation of the cloud-motion wind field, although the emphasis here is to develop tracer vectors associated with storm movement more so than with the true wind field, with which water vapor motions tend to agree. Last, bringing topographic information into the forecast process might aid in handling orographic convection and warm rain situations.

The skill of any radar- or satellite-based nowcasting system deteriorates with forecast time. The analysis presented here suggests an upper limit of approximately 6 h on reliable QPFs using infrared geosynchronous satellite imagery. This facet of nowcasting should be given more attention in future studies. By the same token, this draws attention to what might be the most valuable aspect of a rain nowcasting system, that being the use of short-term predicted rain fields to help to launch the early integration period in mesoscale QPF models that suffer in their early stages from sluggish spinup error when lacking realistic physical initialization information.

Acknowledgments

The authors thank Dr. Harry Cooper of The Florida State University (FSU) for helpful discussions on the technical underpinnings of nowcasting, Mr. James Merritt of FSU for technical assistance on computer graphics and software development, and the Korean Meteorological Administration for generously providing their 1-min rain gauge data. This research has been supported by National Science Foundation Grant ATM-9714361, a part of NSF's contribution to the U.S. Weather Research Program.

REFERENCES

  • Arkin, P. A., and B. N. Meisner. 1987. The relationship between large-scale convective rainfall and cold cloud over the Western Hemisphere during 1982–84. Mon. Wea. Rev. 115:5174.

    • Search Google Scholar
    • Export Citation
  • Atlas, D., , D. Rosenfeld, , and D. B. Wolff. 1990. Climatologically tuned reflectivity–rain rate relations and links to area–time integrals. J. Appl. Meteor. 29:11201135.

    • Search Google Scholar
    • Export Citation
  • Austin, G. L., and A. Bellon. 1982. Very-short-range forecasting of precipitation by the objective extrapolation of radar and satellite data. Nowcasting, K. A. Browning, Ed., Academic Press, 177–190.

    • Search Google Scholar
    • Export Citation
  • Bellon, A., and G. L. Austin. 1978. The evaluation of two years of real-time operation of a short-term precipitation forecasting procedure (SHARP). J. Appl. Meteor. 17:17781787.

    • Search Google Scholar
    • Export Citation
  • Bellon, A., , S. Lovejoy, , and G. L. Austin. 1980. Combining satellite and radar data for the short-range forecasting of precipitation. Mon. Wea. Rev. 108:15541566.

    • Search Google Scholar
    • Export Citation
  • Browning, K. A. 1979. The FRONTIERS plan: A strategy for using radar and satellite imagery for very short-range precipitation forecasting. Meteor. Mag. 108:161184.

    • Search Google Scholar
    • Export Citation
  • Browning, K. A. . Ed.,. . 1982. Nowcasting. Academic Press, 256 pp.

  • Chandler, D. G., and J. L. Collins Jr.. 1994. The D-Day Encyclopedia. Simon and Schuster, 665 pp.

  • Cooper, H. J., and E. A. Smith. 1993. The importance of short-term forecasting of thunderstorms to launch operations at Cape Canaveral. Bull. Amer. Meteor. Soc. 74:8186.

    • Search Google Scholar
    • Export Citation
  • Endlich, R. M., , D. E. Wolf, , D. J. Hall, , and A. E. Brain. 1971. Use of a pattern recognition technique for determining cloud motions from sequences of satellite photographs. J. Appl. Meteor. 10:105117.

    • Search Google Scholar
    • Export Citation
  • Ferraro, R. R. 1997. Special Sensor Microwave Imager derived global rainfall estimates for climatological applications. J. Geophys. Res. 102:1671516735.

    • Search Google Scholar
    • Export Citation
  • Ferraro, R. R., , E. A. Smith, , W. Berg, , and G. J. Huffman. 1998. A screening methodology for passive microwave precipitation retrieval algorithms. J. Atmos. Sci. 55:15831600.

    • Search Google Scholar
    • Export Citation
  • Gandin, L. S., and A. H. Murphy. 1992. Equitable skill scores for categorical forecasts. Mon. Wea. Rev. 120:361370.

  • Goerss, J. S., , C. S. Velden, , and J. D. Hawkins. 1998. The impact of multispectral GOES-8 wind information on Atlantic tropical cyclone track forecasts in 1995. Part II: NOGAPS forecasts. Mon. Wea. Rev. 126:12191227.

    • Search Google Scholar
    • Export Citation
  • Grody, N. C. 1991. Classification of snow cover and precipitation using the Special Sensor Microwave Imager. J. Geophys. Res. 96:74237435.

    • Search Google Scholar
    • Export Citation
  • Hasler, A. F., , K. Palaniappan, , C. Kambhammeta, , P. Black, , E. Uhlhorn, , and D. Chesters. 1998. High-resolution wind fields within the inner core and eye of a mature tropical cyclone from GOES 1-minute images. Bull. Amer. Meteor. Soc. 79:24832496.

    • Search Google Scholar
    • Export Citation
  • Huffman, G. J. Coauthors,. 1997. The Global Precipitation Climatology Project (GPCP) combined precipitation dataset. Bull. Amer. Meteor. Soc. 78:520.

    • Search Google Scholar
    • Export Citation
  • Kedem, B., , L. S. Chiu, , and G. R. North. 1990. Estimation of mean rain rate: Application to satellite observations. J. Geophys. Res. 95:19651972.

    • Search Google Scholar
    • Export Citation
  • Kessler, E. 1966. Computer program for calculating average lengths of weather radar echoes and pattern bandedness. J. Atmos. Sci. 23:569574.

    • Search Google Scholar
    • Export Citation
  • Laurent, H. 1993. Wind extraction from Meteosat water vapor channel image data. J. Appl. Meteor. 32:11241133.

  • Lazzara, M. A. Coauthors. 1999. The Man computer Interactive Data Access System: 25 years of interactive processing. Bull. Amer. Meteor. Soc. 80:271284.

    • Search Google Scholar
    • Export Citation
  • Leese, J. A., , C. S. Novak, , and B. B. Clark. 1971. An automated technique for obtaining cloud motion from geosynchronous satellite data using cross correlation. J. Appl. Meteor. 10:118132.

    • Search Google Scholar
    • Export Citation
  • Menzel, W. P., , F. C. Holt, , T. J. Schmit, , R. M. Aune, , A. J. Schreiner, , G. S. Wade, , and D. G. Gray. 1998. Application of GOES-8/9 soundings to weather forecasting and nowcasting. Bull. Amer. Meteor. Soc. 79:20592077.

    • Search Google Scholar
    • Export Citation
  • Merrill, R. T., , W. P. Menzel, , W. Baker, , J. Lynch, , and E. Legg. 1991. A report on the recent demonstration of NOAA's upgraded capability to derive cloud motion satellite winds. Bull. Amer. Meteor. Soc. 72:372376.

    • Search Google Scholar
    • Export Citation
  • Nieman, S. J., , W. P. Menzel, , C. M. Hayden, , D. Gray, , S. T. Wanzong, , C. S. Velden, , and J. Daniels. 1997. Fully automated cloud drift winds in NESDIS operations. Bull. Amer. Meteor. Soc. 78:11211133.

    • Search Google Scholar
    • Export Citation
  • Purdom, J. F. W. 1976. Some uses of high resolution GOES imagery in the mesoscale forecasting of convection and its behavior. Mon. Wea. Rev. 104:14741483.

    • Search Google Scholar
    • Export Citation
  • Reynolds, D. W., and E. A. Smith. 1979. Detailed analysis of composited digital radar and satellite data. Bull. Amer. Meteor. Soc. 60:10241037.

    • Search Google Scholar
    • Export Citation
  • Rossow, W. B., and L. C. Garder. 1993. Cloud detection using satellite measurements of infrared and visible radiances for ISCCP. J. Climate 6:23412369.

    • Search Google Scholar
    • Export Citation
  • Schmetz, J., , K. Holmlund, , J. Hoffman, , B. Strauss, , B. Mason, , V. Gaertner, , A. Koch, , and L. van de Berg. 1993. Operational cloud-motion winds from Meteosat infrared images. J. Appl. Meteor. 32:12061225.

    • Search Google Scholar
    • Export Citation
  • Shin, K-S., , P. E. Riba, , and G. R. North. 1990. Estimation of area-averaged rainfall over tropical oceans from microwave radiometry: A single channel approach. J. Appl. Meteor. 29:10311042.

    • Search Google Scholar
    • Export Citation
  • Smith, E. A. 1975. The McIDAS System. IEEE Trans. Geosci. Electron. GE-13:123136.

  • Smith, E. A., and D. R. Phillips. 1972. Automated cloud tracking using precisely aligned digital ATS pictures. IEEE Trans. Comput. C17:715729.

    • Search Google Scholar
    • Export Citation
  • Smith, E. A. Coauthors,. 1998. Results of WetNet PIP-2 Project. J. Atmos. Sci. 55:14831536.

  • Spencer, R. W. 1993. Global oceanic precipitation from the MSU during 1979–91 and comparisons to other climatologies. J. Climate 6:13011326.

    • Search Google Scholar
    • Export Citation
  • Taylor, B. C., and K. A. Browning. 1974. Towards an automated weather radar network. Weather 29:202216.

  • Tsonis, A. A., and G. L. Austin. 1981. An evaluation of extrapolation techniques for the short-term prediction of rain amounts. Atmos. Ocean 19:5456.

    • Search Google Scholar
    • Export Citation
  • Turk, F. J., , G. Rohaly, , J. D. Hawkins, , E. A. Smith, , A. Grose, , F. S. Marzano, , and A. Mugnai. 2000. Analysis and assimilation of rainfall from blended SSM/I, TRMM and geostationary satellite data. Preprints, 10th Conf. on Satellite Meteorology and Oceanography, Long Beach, CA, Amer. Meteor. Soc., 66–69.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , C. M. Hayden, , S. J. Nieman, , W. P. Menzel, , S. Wanzong, , and S. Goerss. 1997. Upper-tropospheric winds derived from geostationary satellite water vapor observations. Bull. Amer. Meteor. Soc. 78:173195.

    • Search Google Scholar
    • Export Citation
  • Velden, C. S., , T. L. Olander, , and S. Wanzong. 1998. The impact of multispectral GOES-8 wind information on Atlantic tropical cyclone track forecasts in 1995. Part I: Dataset methodology, description, and case analysis. Mon. Wea. Rev. 126:12021218.

    • Search Google Scholar
    • Export Citation
  • Wetzel, M. A., , R. D. Borys, , and L. E. Xu. 1996. Satellite microphysical retrievals for land-based fog with validation by balloon profiling. J. Appl. Meteor. 35:810829.

    • Search Google Scholar
    • Export Citation
  • Wilheit, T. T., , A. T. C. Chang, , and L. S. Chiu. 1991. Retrieval of monthly rainfall indices from microwave radiometric measurements using probability distribution functions. J. Atmos. Oceanic Technol. 8:118136.

    • Search Google Scholar
    • Export Citation
  • Wilks, D. S. 1995. Statistical Methods in the Atmospheric Sciences. Academic Press, 464 pp.

APPENDIX

Definitions of Acronyms

  • CAA  constant acceleration advection
  • CDF  cumulative distribution function
  • CMW  cloud-motion wind
  • CVA  constant velocity advection
  • EBBT  equivalent blackbody temperature
  • ΔEBBT  change in EBBT
  • ESS  equitable skill score
  • FAR  false-alarm rate
  • FM  function matching
  • GEO  geosynchronous
  • GMS  Geostationary Meteorological Satellite
  • GOES  Geostationary Operational Environmental Satellite
  • IR  infrared
  • KMA  Korean Meteorological Administration
  • LAA  linear acceleration advection
  • NAA  nonlinear acceleration advection
  • NASA  National Aeronautics and Space Administration
  • NESDIS  National Environmental Satellite, Data, and Information Service
  • NOAA  National Oceanic and Atmospheric Administration
  • PDF  probability distribution function
  • PMM  probability matching method
  • PMW  passive microwave
  • PVA  potential vorticity advection
  • QPF  quantitative precipitation forecast
  • RESS  rain-existence skill score
  • RHM  raw histogram matching
  • RISS  rain-intensity skill score
  • RR  rain rate
  • SHM  smooth histogram matching
  • SSM/I  Special Sensor Microwave Imager

Fig. 1.
Fig. 1.

Map of South Korea indicating locations of 395-site automated weather station network

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 2.
Fig. 2.

(top) Three grid systems used for nowcasting within domain of full-resolution IR image sectors. Larger boxes outlined with thick lines represent 21 × 21 pixel target subgrids making up the “target grid system.” Smaller boxes outlined with thin dashed lines (inside target grid system) represent 7 × 7 pixel filtered subgrids making up the “filtered grid system.” Expanded small box domain makes up the “forecast grid system.” Edge domain of 4-pixel width is necessitated by ±25 lag dimensions used in seeking match subgrids. (bottom) Relationship between centers of target subgrid and match subgrid found in 71 × 71 search area by minimum difference method defines CMW vector between consecutive pair of IR images

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 3.
Fig. 3.

Examples of (top) initial CMW field and (bottom) filtered CMW field for Amazon basin on 31 Jan 1999 extracted from 2245–2345 UTC IR image pair. Filtered field is created by applying a priori Monte Carlo averaging operator using 50% vector selection with 10 realizations, followed by two-dimensional linear interpolation onto high-resolution grid, and nine-point binomial smoother. Vector directions are indicated with arrowheads; vector magnitudes with grayscale. Every third gridpoint vector has been selected from high-resolution field for display in bottom panel

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 4.
Fig. 4.

Examples of raw histogram, smooth histogram, and gamma function PDFs used for EBBTs in probability matching process: (top) ocean case and (bottom) land case from original GMS-5–SSM/I paired pixel dataset over Korean peninsula region

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 5.
Fig. 5.

Example of lognormal PDF fit to RR histogram from GMS-5–SSM/I paired pixel ocean dataset over Korean peninsula region: (top) comparison for light rain rates (0.5–5.0 mm h−1) and (bottom) comparison for medium–heavy rain rates (5.0–35.0 mm h−1)

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 6.
Fig. 6.

Normalized frequency distributions for (left) ocean and (right) land using RHM, SHM, and FM algorithms applied to 0715 UTC GOES-8 IR image for Caribbean basin on 1 Apr 1998. Bin interval is 2.5 mm h−1

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 7.
Fig. 7.

Final EBBT–RR mapping rules for ocean and land based on SHM algorithm

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 8.
Fig. 8.

(a) Cloud–rain areas created from (top) 1015 and (bottom) 1045 UTC IR images on 1 Apr 1998 over Caribbean basin. (middle) Filtered CMW field and EBBT tendencies derived from the two IR images. (middle) Arrows indicate CMW vector directions; sizes of attached circles indicate CMW vector magnitudes; and cross, open, or filled patterns within circles indicate negative, near-zero, or positive EBBT tendencies. Every 12th grid point in east–west and every 35th grid point in north–south are selected for display. (b) The (left) 1-, 2-, and 3-h cloud–rain area forecasts from conditions shown in (a) and (right) verification images at 1145, 1245, and 1345 UTC. RESS, FAR, and RISS skill scores are indicated numerically between pairs of forecast and verification panels

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 8.
Fig. 8.

(Continued)

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 9.
Fig. 9.

(a) Same as Fig. 8a but for Amazon basin on 1 Feb 1999 starting at 0745 UTC, in which every 10th grid point in both east–west and north–south is selected for display (middle). (b) Same as Fig. 8b but for Amazon basin on 1 Feb 1999 starting at 0945 UTC

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 9.
Fig. 9.

(Continued)

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 10.
Fig. 10.

(a) Same as Fig. 8a but for Korean peninsula on 1 May 1999 starting at 1530 UTC, in which every 10th grid point in both east–west and north–south is selected for display (middle). (b) Same as Fig. 8b but for Korean peninsula on 1 May 1999 starting at 1730 UTC

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 10.
Fig. 10.

(Continued)

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 11.
Fig. 11.

Ensemble-averaged RESS, FAR, and RISS skill score comparisons out to 3-h forecasts used to evaluate four advection techniques. Comparisons presented for (left) Caribbean basin, (middle) Amazon basin, and (right) Korean peninsula

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Fig. 12.
Fig. 12.

Ensemble-averaged RESS and FAR skill score time decay curves out to 10-h forecasts for (left) Caribbean basin and (right) Amazon basin used for determination of asymptotic limit of forecasts. Forecast ensemble counts are over 10 times as large for Amazon basin relative to Caribbean basin, explaining why former's skill score time decay curves are much smoother

Citation: Journal of Applied Meteorology 41, 7; 10.1175/1520-0450(2002)041<0763:PALFQP>2.0.CO;2

Table 1. 

Summary of 25 schemes tested for creating filtered CMW fields

Table 1. 
Table 2. 

Summary of average rain rates produced by three different PMM rules for land and ocean cases within the Caribbean basin

Table 2. 
Table 3. 

Numerical EBBT–RR matching rules for optimal land and ocean SHM rain algorithms based on probability matching

Table 3. 
Table 4. 

Summary of skill scores for up-to-6-h CVA forecasts for each of three GEO datasets in testing of verification procedure

Table 4. 
Save