NASA GPCP V1B satellite-gauge precipitation |
George J. Huffman
SSAI and Laboratory for Atmospheres,
NASA Goddard Space Flight Center
March 8, 1999
CONTENTS 1. DATA SET NAMES AND GENERAL CONTENT 2. RELATED PROJECTS, DATA NETWORKS, AND DATA SETS 5. DEFINITIONS AND DEFINING ALGORITHMS 6. TEMPORAL AND SPATIAL COVERAGE AND RESOLUTION 7. PRODUCTION AND UPDATES 8. SENSORS 9. ERROR DETECTION AND CORRECTION 10. MISSING VALUE ESTIMATION AND CODES 11. QUALITY AND CONFIDENCE ESTIMATES 12. DATA ARCHIVES 13. DOCUMENTATION 14. INVENTORIES 15. HOW TO ORDER AND OBTAIN INFORMATION ABOUT THE DATA KEYWORDS absolute error variable accuracy AGPI coefficients with missing data AGPI precipitation product algorithm intercomparison projects archive and distribution sites data file access technique data set data set archive data set creators data set inventory data set revisions date documentation creator documentation revision history estimate missing values GPCP GPCP components GPI number of samples product GPI precipitation product grid intercomparison results IR IR data correction known errors missing months multi-satellite precipitation product number of samples variable obtaining data originating machine pentads period of record precipitation variable production and updates products quality index rain gauge rain gauge number of samples product rain gauge precipitation product rain gauge quality control read a month of a product read the header record references satellite-gauge precipitation product similar data sets source variable spatial coverage spatial resolution SSM/I SSM/I composite number of samples product SSM/I composite precipitation product SSM/I emission number of samples product SSM/I emission precipitation product SSM/I error detection/correction SSM/I scattering number of samples product SSM/I scattering precipitation product standard missing value technique temporal resolution variable 1. DATA SET NAMES AND GENERAL CONTENT The *data set* is formally referred to as the "GPCP Version 1b Combined Precipitation Data Set." It is also referred to as the "GPCP Combined Data Set" or the "Version 1b Data Set." The data set currently contains a suite of 19 products providing monthly, global gridded values of precipitation totals and supporting information for the 11-year period July 1987 through September 1998. The products include precipitation estimates based on several "single-source" data sets (IR, SSM/I, and rain gauge), precipitation estimates based on the combination of multiple satellite data sets, and precipitation estimates based on the combination of the satellites with the rain gauge analysis. The main refereed citation for the data set is Huffman et al. (1997) (all references are listed in section 13), which also appears in Huffman 1997b. ........................................................................... 2. RELATED PROJECTS, DATA NETWORKS, AND DATA SETS The *data set creators* are G.J. Huffman and R.F. Adler, working in the Laboratory for Atmospheres, NASA Goddard Space Flight Center, Code 912, Greenbelt, Maryland, 20771 USA, as the GPCP Merge Development Centre. ........................................................................... The work is being carried out as part of the Global Precipitation Climatology Project (*GPCP*), an international project of the WMO/WCRP/GEWEX designed to provide improved long-record estimates of precipitation over the globe. The GPCP home page is located at http://orbit-net.nesdis.noaa.gov/arad/gpcp/ ........................................................................... The Version 1b Data Set contains data from several *GPCP components*: 1. GPCP Polar Satellite Precipitation Data Centre - Emission (SSM/I emission estimates), 2. GPCP Polar Satellite Precipitation Data Centre - Scattering (SSM/I scattering estimates), 3. GPCP Geostationary Satellite Precipitation Data Centre (GPI estimates), and 4. GPCP Global Precipitation Climatology Centre (rain gauge analyses). These single-source data sets, as well as combinations based on them are contained in the Version 1b Data Set for the period of SSM/I data availability. Some single-source data sets extend beyone the SSM/I period in their original archival locations. ........................................................................... The GPCP has sponsored several *algorithm intercomparison projects* (referred to as AIP-1, AIP-2, and AIP-3) for the purpose of evaluating and intercomparing a variety of satellite precipitation estimation techniques. As well, the NASA Wetnet Project has sponsored several such projects (referred to as Precipitation Intercomparison Projects, and labeled PIP-1, PIP-2, and PIP-3). One use of these projects has been to identify competitive techniques for use in the GPCP combined data set. ........................................................................... Only a few *similar data sets* are available. The earlier combined precipitation data set produced by the GPCC, the original "Version 1 Data Set" produced at NASA/GSFC, and the Version 1a Data Set produced at GMDC are all superseded by the Version 1b Data Set. The combination data set by Xie and Arkin (1996) uses similar input data and has similar temporal and spatial coverage, but is carried out with a much different technique. Numerous single-source data sets exist that provide quasi-global coverage; several are included in this release and are described in section 5. ........................................................................... 5. DEFINITIONS AND DEFINING ALGORITHMS The GPI estimates originally reported on a 2.5x2.5-deg lat/lon grid (2.5-deg GPI) are provided as accumulations over *pentads*, which are 5-day periods starting Jan. 1 of each year. That is, pentad 1 covers Jan. 1-5, pentad 2 covers Jan. 6-10, and pentad 73 covers Dec. 27-31. Leap Day (Feb. 29) is included in pentad 12, which then covers 6 days. The pentad accumulation period prevents an exact computation of monthly average for the 2.5-deg GPI and subsequent products. We assume that a pentad crossing a month boundary contributes to the statistics in proportion to the fraction of the pentad in the month. For example, a pentad with 40 images that starts the last day of the month is assumed to contribute 8 images (one-fifth of the full pentad) of rainfall information. The 1x1-deg GPI estimates are reported as individual 3-hrly images, and all other input single-source data fields are provided to GPCP in monthly form. ........................................................................... The data set contains 19 *products*, each of which is named by concatenating a technique name with a variable name. As shown in Table 1, here are eight precipitation estimation techniques and four variables, but only 19 of the 32 possible products are considered useful and archived. Besides product availability, Table 1 displays the abbreviations used for coding the technique and variable in the file names, and the units of the various products. --> NOTE: In general, users wishing to use the "final" combined <-- --> product should use the "psg" data files (satellite-gauge <-- --> combined precipitation product). <-- Table 1. GPCP Version 1b Combined Precipitation Data Set Product List, where X denotes an available product, [] gives the abbreviation used for coding the technique or variable in the file names, and () gives the units of the various products, except Number of Samples, whose units are displayed in the last column. \ Variable | Precip | Absolute | | \ | Rate [p] | Error [e] | Source | Number of Samples Technique \ | (mm/d) | (mm/d) | [s] | [n] | (Units) ---------------------+----------+-----------+--------+--------------------- | | | | | SSMI Emission [se] | X | | | X | 55 km images | | | | | SSMI Scattering [ss] | X | | | X | overpass days | | | | | SSMI Composite [sc] | X | X | X | X | 55 km images | | | | | GPI [gp] | X | | | X | 2.5 deg images | | | | | AGPI [ag] | X | X | | | | | | | | Multi-Satellite [ms] | X | X | | | | | | | | Rain Gauge [ga] | X | X | | X | gauges | | | | | Satellite-Gauge [sg] | X | X | | | For example, the absolute error variable for the multi-satellite technique may be found in files with "ems" in the name, but there is no product giving the number-of-samples variable for the multi-satellite technique. ........................................................................... The *technique* name tells what algorithm was used to generate the product. There are eight such techniques in the Version 1b Data Set: SSMI Emission, SSMI Scattering, SSMI Composite, GPI, AGPI, Multi-Satellite, Rain Gauge, and Satellite-Gauge. ........................................................................... The *variable* name tells what parameter is in the product. There are four such variables in the Version 1b Data Set: Precipitation Rate, Absolute Error, Source, and Number of Samples. ........................................................................... The *precipitation variable* is computed as described under the individual product headings. All precipitation products have been converted from their original units of mm/mo or mm/hr to mm/d. .......................................................................... The *SSM/I emission precipitation product* is produced by the Polar Satellite Precipitation Data Centre - Emission of the GPCP under the direction of A. Chang, located in the Laboratory for Hydrospheric Processes, NASA Goddard Space Flight Center, Code 971, Greenbelt, Maryland, 20771 USA. The SSM/I (Special Sensor Microwave/Imager) data are recorded by selected Defense Meteorological Satellite Program satellites, and are provided in packed form by Remote Sensing Systems (Santa Clara, CA) for 1987-1998 and National Climatic Data Center (Asheville, NC) starting in 1999. The algorithm applied is the Wilheit et al. (1991) iterative histogram approach to retrieving precipitation from emission signals in the 19-GHz SSM/I channel. It assumes a log-normal precipitation histogram and estimates the freezing level from the 19- and 22-GHz channels. The fit is applied to the full month of data. Individual estimates on the 2.5x2.5-deg grid occasionally fail to converge. In that case the corresponding estimate on the 5x5-deg grid is substituted for precipitation, and the number of samples is estimated by smooth-filling from surrounding boxes with data. In the current version the SSM/I emission products for 1987-1995 are generated for "months" that are rounded to the nearest pentad boundary. As a result, the period and number of samples will fluctuate compared to the other estimates. Starting with 1996 the products are generated for calendar months, in agreement with the other estimates. In a future release all emission estimates will be produced on calendar months. ........................................................................... The *SSM/I scattering precipitation product* is produced by the GPCP Polar Satellite Precipitation Data Centre - Scattering under the direction of R. Ferraro, located in the Office of Research and Application of the NOAA National Environmental Satellite Data and Information Service (NESDIS), Washington, DC, 20233 USA. The SSM/I (Special Sensor Microwave/Imager) data are recorded by selected Defense Meteorological Satellite Program satellites, and are transmitted to NESDIS through theShared Processing System. The algorithm applied is based on the Grody (1991) Scattering Index (SI), supplemented by the Weng and Grody (1994) emission technique in oceanic areas. A similar fall-back approach was used during the period when the 85.5-GHz channels were unusable. Pixel-by-pixel retrievals are accumulated onto separate daily ascending and descending 0.333x0.333 deg lat/long grids, then all the grids are accumulated for the month on the 2.5 deg grid. ........................................................................... The *SSM/I composite precipitation product* is produced as part of the GPCP Version 1b Combined Precipitation Data Set by the GPCP Merge Development Centre (see section 2). The concept is to take the SSM/I emission estimate over water and the SSM/I scattering estimate over land. Since the emission technique eliminates land-contaminated pixels individually, a weighted transition between the two results is computed in the coastal zone. The merger may be expressed as | R(emiss) ; N(emiss) >= 0.75 * N(scat) | | N(emiss) * R(emiss) + ( N(scat) - N(emiss) ) * R(scat) R(composite) = | ------------------------------------------------------ ; (1) | N(scat) | N(emiss) < 0.75 * N(scat) where R is the precipitation rate; N is the number of samples; composite, emiss, and scat denote composite, emission, and scattering, respectively; and the 0.75 threshold allows for fluctuations in the methods of counting samples in the emission and scattering techniques. Note that the second expression reduces to R(scat) when N(emiss) is zero. Important Note: The emission and scattering fields used in this merger have been edited to remove known and suspected artifacts, such as high values in polar regions. These edited fields may be approximated by using the source variable to mask the emission and scattering fields contained in this data set. That is, the user may infer that editing must have occurred for points where the source variable indicates that the scattering or emission (or both) are not used, but the scattering or emission (or both) values are non-missing. ............................................................................ The *GPI precipitation product* is produced by the Geostationary Satellite Precipitation Data Centre of the GPCP under the direction of J. Janowiak, located in the Climate Prediction Center, NOAA National Centers for Environmental Prediction, Washington, DC, 20233 USA. Each cooperating geostationary satellite operator (the Geosynchronous Operational Environmental Satellites, or GOES, United States; the Geosynchronous Meteorological Satellite, or GMS, Japan; and the Meteorological Satellite, or Meteosat, European Community) accumulates three-hourly infrared (IR) imagery which are forwarded to GSPDC. The global IR rainfall estimates are then generated from a merger of these data using the GPI (GOES Precipitation Index; Arkin and Meisner, 1987) technique, which relates cold cloud-top area to rain rate. For the period 1986-March 1998 the GPI data are accumulated on a 2.5x2.5-deg lat/lon grid for pentads (5-day periods), preventing an exact computation of the monthly average. We assume that a pentad crossing a month boundary contributes to the statistics in proportion to the fraction of the pentad in the month. For example, given a pentad that starts the last day of the month, 0.2 (one-fifth) of its samples are assigned to the month in question and and 0.8 (four-fifths) of its samples are assigned to the following month. Starting with October 1996 the GPI data are accumulated on a 1x1-deg lat/lon grid for individual 3-hrly images. In this case monthly totals are computed as the sum of all available hours in the month. In both data sets gaps in geo-IR are filled with low-earth-orbit IR (leo-IR) data from the NOAA series of polar orbiting meteorological satellites. However, the 2.5x2.5-deg data only contain the leo-IR used for fill-in, while the 1x1-deg data contain the full leo-IR. The latter allows a more accurate AGPI (see "AGPI precipitation product"). The GPI product is based on the 2.5x2.5-deg data for the period 1987-96, and the 1x1-deg beginning in 1997. The boundary is set at January 1997 to avoid the 1997-1998 ENSO event. ........................................................................... The *AGPI precipitation product* is produced as part of the GPCP Version 1b Combined Precipitation Data Set by the GPCP Merge Development Centre (see section 2). The technique follows the Adjusted GPI (AGPI) of Adler et al. (1994). Separate monthly averages of approximately coincident GPI and SSM/I precipitation estimates are formed by taking cut-outs of the 3-hourly GPI values that correspond most closely in time to the local overpass time of the DMSP platform. The ratio of SSM/I to GPI averages is computed, controlled to prevent unstable answers, and smoothly filled in regions where the SSM/I is missing but the GPI is available. In regions of light precipitation an additive adjustment is computed as the difference between smoothed SSM/I and ratio-adjusted GPI values when the SSM/I is greater, and zero otherwise. The spatially varying arrays of adjustment coefficients are then applied to the full set of GPI estimates. In regions lacking geo-IR data, leo-GPI data are calibrated to SSM/I, then these SSM/I-calibrated leo-GPI are calibrated to the geo-AGPI. This two-step process tries to mimic the information contained in the AGPI, namely the local bias of the SSM/I and possible diurnal cycle biases in the geo-AGPI. The second step can only be done in regions with both geo- and leo-IR data, and then smooth-filled across the leo-IR fill-in. In the case of the 2.5x2.5-deg IR, which lacks leo-IR in geo-IR regions, the missing SSM/I-calibrated leo-GPI is approximated by smoothed SSM/I for doing the calibration to geo-AGPI. ........................................................................... The *multi-satellite precipitation product* is produced as part of the GPCP Version 1b Combined Precipitation Data Set by the GPCP Merge Development Centre (see section 2) following Huffman et al. (1995). Geo-AGPI estimates are taken where available (latitudes 40 deg N-S), the weighted combination of the SSM/I combined estimate and the leo-AGPI elsewhere in the 40 deg N-S belt, and the SSM/I combined outside of that zone. The combination weights are the inverse (estimated) error variances of the respective estimates. Such weighted combination of microwave and leo-AGPI is done because the leo-IR lacks the sampling to support the full AGPI adjustment scheme. ........................................................................... The *rain gauge precipitation product* is produced by the Global Precipitation Climatology Centre (GPCC) under the direction of B. Rudolf, located in the Deutscher Wetterdienst, Offenbach a.M., Germany (Rudolf 1993). Rain gauge reports are archived from about 6700 stations around the globe, both from Global Telecommunications Network reports, and from other world-wide or national data collections. An extensive quality-control system is run, featuring an automated step and then a manual step designed to retain legitimate extreme events that characterize precipitation. A variant of the SPHEREMAP spatial interpolation routine (Willmott et al. 1985) is used to analyze station values to area averages. The analyzed values have been corrected for climatological estimates of systematic error due to wind effects, side-wetting, evaporation, etc., following Legates (1987). ........................................................................... The *satellite-gauge precipitation product* is produced as part of the GPCP Version 1b Combined Precipitation Data Set by the GPCP Merge Development Centre (see section 2) in two steps (Huffman et al. 1995). First, the multi-satellite estimate is adjusted toward the large-scale gauge average for each grid box over land. That is, the multi-satellite value is multiplied by the ratio of the large-scale (5x5 grid-box) average gauge analysis to the large-scale average of the multi-satellite estimate. Alternatively, in low-precipitation areas the difference in the large-scale averages is added to the multi-satellite value when the averaged gauge exceeds the averaged multi-satellite. In the second step, the gauge-adjusted multi-satellite estimate and the gauge analysis are combined in a weighted average, where the weights are the inverse (estimated) error variance of the respective estimates. ........................................................................... The *absolute error variable* is produced as part of the GPCP Version 1b Combined Precipitation Data Set by the GPCP Merge Development Centre (see section 2). Following Huffman (1997a), bias error is neglected compared to random error (both physical and algorithmic), then simple theoretical and practical considerations lead to the functional form H * ( rbar + S) * [ 720 + 268 * SQRT ( rbar ) ] VAR = ----------------------------------------------- (2) Ni for absolute error, where VAR is the estimated error variance of an average over a finite set of observations, H is taken as constant (actually slightly dependent on the shape of the precipitation rate histogram), rbar is the average precipitation rate in mm/mo, S is taken as constant (approximately SQRT(VAR) for rbar=0), Ni is the number of INDEPENDENT samples in the set of observations, and the expression in square brackets is a parameterization of the conditional precipitation rate based on work with the Goddard Scattering Algorithm, Version 2 (Adler et al. 1994) and fitting of (2) to the Surface Reference Data Center analyses (McNab 1995). The "constants" H and S are set for each of the data sets for which error estimates are required by comparison of the data set against the SRDC and GPCC analyses and tropical Pacific atoll gauge data (Morrissey and Green 1991). The computed value of H actually accounts for multiplicative errors in Ni and the conditional rainrate parameterization (the [] term), in addition to H itself. Table 2 shows the numerical values of H and S. All absolute error fields have been converted from their original units of mm/mo to mm/d. Table 2. Numerical values of H and S constants used to estimate absolute error for various precipitation estimates. | S | Technique | (mm/mo) | H ---------------------+---------+----------------------- | | SSMI Emission [se] | 30 | 3.25 (55 km images) | | SSMI Scattering [ss] | 30 | 4.5 (55 km images) | | AGPI [ag] | 20 | 0.6 (2.5 deg images) | | Rain Gauge [ga] | 6 | 0.005 (gauges) For the indenpendent data sets rbar is taken to be the independent estimate of rain itself. However, when these errors are used in the combination, theory and tests show that the result is a low bias. Rbar needs to have the same value in all the error estimates; so we estimate it as the simple average of all rainfall values contributing to the combination. Note that this scheme is only used in computing error. ........................................................................... The *source variable* is produced as part of the GPCP Version 1b Combined Precipitation Data Set by the GPCP Merge Development Centre (see section 2). It is only available for the SSM/I composite technique and gives the fractional contribution to the composite by the SSM/I scattering estimate. Referring to (1) in the "SSM/I composite precipitation product" description, the source SOURCE may be expressed as | 0 ; N(emiss) >= 0.75 * N(scat) | SOURCE = | ( N(scat) - N(emiss) ) (3) | ---------------------- ; N(emiss) < 0.75 * N(scat) | N(scat) where N is the number of samples, emiss and scat denote emission and scattering, respectively, and the 0.75 threshold allows for fluctuations in the methods of counting samples in the emission and scattering techniques. Note that the second expression reduces to 1 when N(emiss) is zero. ........................................................................... The *number of samples variable* is produced in a variety of units as described under the individual product headings. ........................................................................... The *SSM/I emission number of samples product* is provided to the GPCP as the number of pixels contributing to the grid box average for the month (i.e., the number of "good" pixels). As part of the Version 1b Data Set processing, this number is converted to the number of 55x55 km boxes that the number of pixels can evenly and completely cover. This conversion provides a very approximate (over)estimate of the number of independent samples contributing to the average. ........................................................................... The *SSM/I scattering number of samples product* is provided to the GPCP as the number of "overpass days," the count of days in the month that had at least one ascending pass plus days that had at least one descending pass. As part of the Version 1b Data Set processing, this number is converted to the number of 55x55 km boxes that the number of pixels can evenly and completely cover. This conversion provides a very approximate (over)estimate of the number of independent samples contributing to the average. ........................................................................... The *SSM/I composite number of samples product* is produced as part of the GPCP Version 1b Combined Precipitation Data Set by the GPCP Merge Development Centre (see section 2). Due to the different units for the SSM/I emission and scattering numbers of samples, it is necessary to convert at least one before doing the merger. We have chosen to convert overpass days (SSM/I scattering estimates) to an estimate of complete 55x55 km boxes (our modified units for the SSM/I emission). In the latitude belt 60 deg N-S, orbits in the same direction don't overlap on a single day, and there is an approximate linear relationship between overpass days and 55 km boxes. Outside that belt the overlaps cause non-linearity, but we ignore it because the general lack of reliable SSM/I at higher latitudes overwhelms details about the numbers of samples. The separate numbers of samples for each technique, measured in 55 km boxes, are merged according to the same formula as the rainfall: | N(emiss) ; N(emiss) >= 0.75 * N(scat) | | N(emiss) * N(emiss) + ( N(scat) - N(emiss) ) * N(scat) N(composite) = | ------------------------------------------------------ ; (4) | N(scat) | N(emiss) < 0.75 * N(scat) where N is the number of samples; composite, emiss, and scat denote composite, emission, and scattering, respectively; and the 0.75 threshold allows for fluctuations in the methods of counting samples in the emission and scattering techniques. Note that the second expression reduces to N(scat) when N(emiss) is zero. ........................................................................... The *GPI number of samples product* is provided to the GPCP as the number of IR images that contribute to the 2.5x2.5-deg grid box. For the 2.5x2.5-deg IR data it is provided as the number of images per pentad (5-day period), while for the 1x1-deg IR data each 3-hrly image is a separate dataset. For the 2.5x2.5-deg IR data the contribution by pentads that cross month boundaries are taken to be proportional to the fraction of the pentad in the month.to the fraction of the pentad in the month. For example, given a pentad that starts the last day of the month, 0.2 (one-fifth) of its samples are assigned to the month in question and and 0.8 (four-fifths) of its samples are assigned to the following month. .......................................................................... The *rain gauge number of samples product* is provided to the GPCP as the number of stations providing gauge reports for the month in the 2.5x2.5 deg grid box. .......................................................................... 6. TEMPORAL AND SPATIAL COVERAGE AND RESOLUTION The *date* for a file is the year in which the months it contains occurred. The date for a grid is the year/month over which the observations were accumulated to form the averages and estimates. All dates are UTC. ........................................................................... The *temporal resolution* of the products is one calendar month. The temporal resolution of the original single-source data sets is also one month, except the GPI data source has pentad (five-day) or 3-hrly temporal resolution for the 2.5x2.5-deg and 1x1-deg IR data sets, respectively. Some of the single-source data sets are available from other archives at a finer resolution. ........................................................................... The *period of record* for the GPCP Version 1b Combined Precipitation is July 1987 through September 1998, without December 1987. The start and gap are based on the availability of SSM/I data. The end is based on the availability of input analyses, and will be extended in future releases. Some of the single-source data sets have longer periods of record in their original archival sites. ........................................................................... The *grid* on which each field of values is presented is a 2.5x2.5 deg latitude--longitude (Cylindrical Equal Distance) global array of points. It is size 144x72, with X (longitude) incrementing most rapidly West to East from the Prime Meridian, and then Y (latitude) incrementing North to South. Whole- and half-degree values are at grid edges: First point center = (88.75N,1.25E) Second point center = (88.75N,3.75E) Last point center = (88.75S,1.25W) ........................................................................... The *spatial resolution* of the products is 2.5x2.5 deg lat/long, as it was for the original single-source data sets, except the 1x1-deg IR (used starting January 1997). Some of the single-source data sets are available from other archives at a finer resolution. ........................................................................... The *spatial coverage* of the products is global in the sense that they are provided on a global grid. However, all of the products have meaningful values only on a subset of the grid points. The single-source products have the largest holes, and the combination products cover successively more of the globe. See the sensor descriptions (section 8) for additional discussion of coverage by the single-source products. ........................................................................... 7. PRODUCTION AND UPDATES The GPCP is responsible for managing *production and updates* of the GPCP Combined Precipitation Data Set (WCRP 1986). Version 1b is produced by the GPCP Merge Development Centre (GMDC), located at NASA Goddard Space Flight Center in the Laboratory for Atmospheres. Various groups in the international science community are given the tasks of preparing precipitation estimates from individual data sources, then the GMDC is charged with combining these into a "best" global product. This activity takes place after real time, at a pace governed by agreements about forwarding data to the individual centers and activities designed to ensure the quality in each processing step. The techniques used to compute the individual and combination estimates are described in section 5. Updates will be released to (1) extend the data record, (2) take advantage of improved combination techniques, or (3) correct errors. Updates resulting from the last two cases will be given new version numbers. At the present the GPCP is working to approve Version 2, which will feature extension back to Jauary 1986 (before the beginning of SSM/I coverage, and filling the December 1987 gap in SSM/I), as well as extension of the observation-only product to the polar regions. A second, more approximate product is being developed to provide globally complete observation-only monthly estimates for 1979-1985. In parallel, the GPCP is working to approve a Daily 1-Deg. product that will provide daily estimates on a global 1x1-deg lat/lon grid for the period January 1997 through the present (with a delay). It is based on SSM/I and geo-IR data with calibration by the monthly GPCP Satellite- Gauge Combination product. Experimental versions of both are available from the GMDC, with formal release possible in Summer 1999. .......................................................................... A number of *data set revisions* have been integrated into the GPCP Version 1b Combined Precipitation, and additional enhancements are planned for Version 2. In 1996 the GPCP started operational experiments with collecting geosynchronous-orbit IR (geo-IR) data on a 1x1-deg lat/lon grid in parallel with the operational 2.5x2.5-deg gridded data. In 1998 the GPCP decided that the 1x1-deg data had sufficient similarity with the 2.5x2.5-deg data that only the 1x1-deg data would be collected. Combinations computed with the new dataset revealed certain processing problems with the old dataset. None of the problems was considered so severe that immediate reprocessing of the entire dataset was demanded. However, unless reprocessing occurred (or unless the new data were processed with the same mistakes) the combinations based on the 1x1-deg IR would show systematic differences in climatology from the preceeding years of data based on the 2.5x2.5-deg IR. 1. ISSUE: In Version 1a intersatellite calibration was applied to 2.5x2.5-deg pentad IR, but not 2.5x2.5-deg 3-hrly pentad IR, so AGPI adjustment coefficients were computed without the calibration, but applied to calibrated data. Since the AGPI scheme acts to calibrate all the IR to the SSM/I, the AGPI thus produced actually contained approximately two times the actual intersatellite calibration. FIX: Histograms without intersatellite calibration have been used for Version 1b. Tests show the expected approximately neutral difference between 2.5x2.5- and 1x1-deg results. Done. The old intersatellite calibration scheme depended on ISCCP data which ceased in 1994, so this needed work anyway. The Geostationary Satellite Precipitation Data Centre (GSPDC) will pursue the intersatellite calibration issue for Version 2. Pending. NOTE: Having intersatellite calibration would provide a better GPI and at second order refine the AGPI at satellite data boundaries. 2. ISSUE: In Version 1a the choice of satellite source is strictly by the number of images in the 2.5x2.5-deg 3-hrly pentad IR (used to compute adjustment coefficients), but in the 2.5x2.5-deg pentad IR the distance to the satellite is also considered (used to compute the AGPI). So, at some locations nearly equidistant between the two satellites the AGPI is derived for one satellite, but applied to the other. FIX: GSPDC will produce 2.5x2.5-deg 3-hrly pentad IR for Version 2. Pending. NOTE: In the 1x1-deg 3-hrly GPI it is possible for the two satellites to cut in and out on successive hours. As long as the relative contribution of each is in the same proportion for both the SSM/I- matched subset and the full data set this is not too important. Using intersatellite calibrated data would overcome this issue, although it is likely a second-order effect. 3. ISSUE: The 1x1-deg IR dataset provides comprehensive leo-IR data while the 2.5x2.5-deg IR only provides leo-IR in regions lacking geo-IR. The additional data in the 1x1-deg IR allows more accuracy in estimating the calibration of the SSM/I-calibrated leo-GPI to the geo-AGPI, causing biases between the 1x1- and 2.5x2.5-deg AGPI in leo regions (the Indian Ocean being the prime case) of up to 15%. FIX: First, smoothed SSM/I is used in Version 1b to estimate the SSM/I-calibrated leo-GPI in regions where the 2.5x2.5-deg lacks leo-IR data, rather than the Version 1a scheme of trying to estimate the missing data from gradients in the available data. Second, the leo-GPI is now not smoothed (in either the 1x1- or 2.5x2.5-deg case) before calibration against the (smoothed) SSM/I. The GPI is already pretty smooth, and this prevents differences between the 1x1-deg leo-GPI and the smoothed SSM/I from bleeding into the leo-only regions due to smoothing. Done. NOTE: Alternatively, a whole different 2.5x2.5-deg pentad low-orbit GPI dataset could be generated, and then integrated into the system. The improvement over the FIX should be only second-order. 4. ISSUE: The Version 1a floor of 0.2 on the SSM/I--IR calibration ratio is too high to control cold surface (Himalayas, Rockies) and non-raining cirrus (Sahara) in very dry regions, causing the AGPI to unphysically exceed the calibrating SSM/I values in those areas. FIX: The ratio floor is reduced to 0.02. Done. NOTE: This ameliorates a long-standing artifact in the Himalayas. 5. ISSUE: Regions that lack SSM/I data, but have IR were treated differently in the initial versions of the Version 1b 1x1- and 2.5x2.5-deg AGPI. Smoothfilling the coefficients leaves residual artifacts, while smoothfilling the SSM/I underestimates precipitation and provides better continuity between the 1x1- and 2.5x2.5-deg AGPI. The third choice is to leave the no-SSM/I region blank as in Version 1a. FIX: Since coverage is important, but we want to avoid noticeable artifacts, smoothfilling the SSM/I is preferred. Done. 6. ISSUE: The use of individual average rainrates in the Version 1a random error estimates causes the inverse error variance combination scheme to tend toward the lower rainrate, causing the final combined value to be systematically lower than expected. FIX: The inverse error variances used for the combination in a gridbox should be computed with a single average rainrate; in Version 1b this has been implemented using the simple average of all rainrates that contribute to the combination for the gridbox. Done. NOTE: This only affects combination values over land and in the leo-IR fill-ins. 7. ISSUE: The data boundary between 1x1-deg and 2.5x2.5-deg input can be put any time between Jan. 1997 and April 1998. An early start ensures that the entire 1997-98 ENSO event is covered with homogeneous data, while the latest date would allow the longest possible homogeneous record (of 2.5x2.5-deg input). FIX: Continuity through the ENSO seems most important, so the Jan. 1997 start is used. Done. 8. ISSUE: During the lifespan of Version 1a, incremental upgrades were made to the GPCC analysis system. Similarly, incremental upgrades were made to the software for controlling artifacts in the SSM/I data, principally in coastal areas susceptible to ice-contamination artifacts. The revisions were considered too minor to recompute the entire data record. FIX: Version 1b is a complete recomputation, so the latest GPCC analysis has been pulled and the SSM/I artifact control is applied throughout. One additional issue has no action at this time: 9. ISSUE: The GMS 2.5x2.5-deg histograms were collected with temperature bin boundaries at half-degree values, but the 1x1-deg histograms are being collected on whole-degree boundaries; this causes GPI differences in excess of 10% at 30-40 deg latitude, and everywhere the 1x1-deg GPI is smaller. FIX: The AGPI largely calibrates out this problem. Done. NOTE: The fix leaves the GPI somewhat damaged. If the GPI itself needs to be consistent, one could split the 235K class in the 1x1-deg histograms. ........................................................................... 8. SENSORS The Special Sensor Microwave/Imager (*SSM/I*) is a multi-channel passive microwave radiometer that has flown on selected Defense Meteorological Satellite Program (DMSP) platforms since mid-1987. The DMSP is placed in a sun-synchronous polar orbit with a period of about 102 min. The SSM/I provides vertical and horizontal polarization values for 19, 22, 37, and 85.5 GHz frequencies (except only vertical at 22) with conical scanning. Pixels and scans are spaced 25 km apart at the suborbital point, except the 85.5-GHz channels are collected at 12.5 km spacing. Every other high-frequency pixel is co-located with the low-frequency pixels, starting with the first pixel in the scan and the first scan in a pair of scans. The channels have resolutions that vary from 12.5x15 km for the 85.5 GHz (oval due to the slanted viewing angle) to 60x75 km for the 19 GHz. The polar orbit provides nominal coverage over the latitudes 85 deg N-S, although limitations in retrieval techniques prevent useful precipitation estimates in cases of cold land (scattering), land (emission), or sea ice (both scattering and emission). The SSM/I is an operational sensor, so the data record suffers the usual gaps in the record due to processing errors, down time on receivers, etc. Over time the coverage has improved as the operational system has matured. As well, the first 85.5 GHz sensor to fly degraded quickly due to inadequate solar shielding. After launch in mid-1987, the 85.5 GHz vertical- and horizontal-polarization channels became unusable in 1989 and 1990, respectively. Further details are available in Hollinger et al. (1990). The SSM/I emission estimates are based on data from the F8 instrument from mid-1987 through 1991, with the F11 being used for 1992 through 1995, and the F13 thereafter. In contrast, the SSM/I scattering estimates are based on F13 starting in May 1995. In a future release the emission estimates will also be based on F13 starting in May 1995. ........................................................................... The infrared (*IR*) data are collected from a variety of sensors. The primary source of IR data is the international constellation of geosynchronous-orbit meteorological satellites -- the Geosynchronous Operational Environmental Satellites (GOES, United States), the Geosynchronous Meteorological Satellite (GMS, Japan), and the Meteorological Satellite (Meteosat, European Community). There are usually two GOES platforms active, GOES-EAST and -WEST, which cover the eastern and western United States, respectively. Gaps in geosynchronous coverage (most notably over the Indian Ocean) are filled with IR data from the NOAA-series polar-orbiting meteorological satellites. The geosynchronous data are collected by scanning (parts of) the earth's disk, while the polar-orbit data are collected by cross-track scanning. The data are accumulated for processing from full-resolution (4x8 km) images. For the period 1986-March 1998 the GPI data are accumulated on a 2.5x2.5-deg lat/lon grid for pentads (5-day periods). Starting with October 1996 the GPI data are accumulated on a 1x1-deg lat/lon grid for individual 3-hrly images. In both data sets gaps in geo-IR are filled with low earth orbit IR (leo-IR) data from the NOAA series of polar orbiting meteorological satellites. However, the 2.5x2.5-deg data only contain the leo-IR used for fill-in, while the 1x1-deg data contain the full leo-IR. The GPI product is based on the 2.5x2.5-deg data for the period 1987-196, and the 1x1-deg beginning in 1997. The boundary is set at January 1997 to avoid placing the boundary during the 1997-1998 ENSO event. The combination of IR satellites provides near-global coverage, but limitations in retrieval techniques prevent useful precipitation estimates poleward of about latitude 40 deg in the summer hemisphere, and about latitude 30 deg in the winter hemisphere. The various IR instruments are operational sensors, so the data record suffers the usual gaps in the record due to processing errors, down time on receivers, sensor failures, etc. Most notably, the GOES series experienced successive failures and replacement over the whole period of record. Further details are available in Janowiak and Arkin (1991). ........................................................................... The *rain gauge* data are quite heterogeneous. Unlike the fairly uniform preparation of satellite data sets, gauge data sources and qualities are extremely variable. Choice of instrumentation, including wind-shielding (if any), siting, observing practices, error detection/correction, and data transmission techniques are all governed by national or regional rules. Typical rain-gauge models include simple 8-inch cylinders (read manually), weighing (ink trace on graph paper), or tipping bucket (digital or analog record) devices located in an open area. Reports are usually generated manually and transmitted to a central regional or national site. Most of the rain gauge reports contributing to the GPCP Version 1b Combined Precipitation Data Set were transmitted as SYNOP or CLIMAT reports on the Global Telecommunications System, although these were supplemented by national and regional collections retrieved after real time. There are about 6700 stations in the current data set, mostly in land areas and concentrated in developed countries. Version 1b uses the January 1999 version of the GPCC "monitoring analysis" for 1986- September 1998, together with real-time pulls of analyses for subsequent months. Further details are available in Rudolf (1993). ........................................................................... 9. ERROR DETECTION AND CORRECTION *SSM/I error detection/correction* has several parts. Built-in hot- and cold-load calibration checks are used to convert counts to Antenna Temperature (Ta). An algorithm has been developed to convert Ta to Brightness Temperature (Tb) for the various channels (eliminating cross-channel leakage). As well, systematic navigation corrections are performed. All pixels with non-physical Tb and local calibration errors are deleted. Accuracies in the Tb's are within the uncertainties of the precipitation estimation techniques. For the most part, tests show only small differences among the SSM/I sensors flying on different platforms. ........................................................................... The dominant *IR data correction* is for slanted paths through the atmosphere. Referred to as "limb darkening correction" in polar-orbit data, or "zenith-angle correction" in geosynchronous-orbit data, this correction accounts for the fact that a slanted path through the atmosphere increases the chances that (cold) cloud sides will be viewed, rather than (warm) surface, and raises the altitude dominating the atmospheric emission signal (almost always lowering the equivalent Tb). In addition, the various sensors have a variety of sensitivities to the IR spectrum, usually including the 10-11 micron band. Inter-satellite calibration differences are documented, but they are not implemented in the current version. They are planned for a future release. The AGPI largely corrects intersatellite calibration, except for small effects at boundaries between satellites. The satellite operators are responsible for detecting and eliminating navigation and telemetry errors. ........................................................................... The *rain gauge quality control* scheme is discussed in Rudolf (1993) and section 13. For the most part, quality-control errors are deleted. The largest correctable error for individual reports is the systematic bias. The use of the Legates (1987) climatological correction is only an approximate solution, since the correction ought to be applied to the gauges before averaging. The availability of rain-gauge reports is extremely variable in space and time, and within a box the coverage by gauges is often not uniform. As a result, even the "ground truth" of rain gauge data has non-trivial errors. ........................................................................... Seven types of *known errors* are contained in part or all of the current data set, and will be corrected in a future general re-run. They have been uncovered by visual inspection of the combined data fields over several years of production, but are considered too minor or insufficiently understood to provoke an immediate reprocessing. 1. Limit checks on sea ice contamination in the SSM/I emission estimates continue to be refined as additional cases are uncovered. 2. The climatological bias correction to the gauge data have artifacts in a few areas, particularly in Antarctica and Siberia. 3. Exact-zero values in marginally snowy land regions (from the SSM/I scattering field) are probably not reliable, and should simply be "small." 4. Isolated exact-zero values surrounded by significantly non-zero values (i.e., >30 mm/mo) in oceanic regions are not reliable and are replaced with the average of the surrounding points. 5. The emission estimates are not fully converged, resulting in a 13% over-estimate, on average. 6. The AGPI calibration coefficients for the 2.5x2.5-deg IR input (1987- 1996) are derived on one choice of satellites in regions of overlap between geo satellites, and applied to another. 7. There is no inter-satellite calibration applied to the GPI. ........................................................................... 10. MISSING VALUE ESTIMATION AND CODES There is generally no effort to *estimate missing values* in the single-source data sets, although a few missing days of gauge data are tolerated in computing monthly values. ........................................................................... We must handle two cases of computing the *AGPI coefficients with missing data*. First, when SSM/I data are missing in a region, but GPI data exist, the SSM/I data are smoothly filled across the blank for the purpose of computing the AGPI. Second, when leo-GPI data are used to fill holes in the geo-GPI, the leo-GPI data are used to estimate a leo-AGPI. That is, the calibration of the AGPI and SSM/I-calibrated leo-GPI is computed around the edge of the hole, the calibration coefficients are smoothly filled across the hole, and applied to the SSM/I-calibrated leo-GPI in the hole. Because the 2.5x2.5-deg IR lacks leo-GPI in the geo-GPI region, smoothed SSM/I is used to estimate SSM/I-calibrated leo-GPI in the geo-GPI region. ........................................................................... All products in the GPCP Version 1b Data Set use the *standard missing value* '-99999.' Some of the single-source data sets possess coded missing values in other archives of the data set. ........................................................................... Within a year file, *missing months* are filled entirely with the standard missing value, so that the month number and the position of the month in the file always agree. ........................................................................... 11. QUALITY AND CONFIDENCE ESTIMATES The *accuracy* of the precipitation products can be broken into systematic departures from the true answer (bias) and random fluctuations about the true answer (sampling), as discussed in Huffman (1997a). The former are the biggest problem for climatological averages, since they will not average out. However, on the monthly time scale the low number of samples tends to present a more serious problem. That is, for most of the data sets the sampling is spotty enough that the collection of values over one month is not yet representative of the true distribution of precipitation. Accordingly, the "random error" is assumed to be dominant, and estimates are computed as discussed for the "absolute error variable" (section 5). Note that the rain gauge analysis' random error is just as real as that of the satellite data, even if somewhat smaller. Random error cannot be corrected. The "bias error" is not corrected in the SSM/I emission, SSM/I scattering, SSM/I composite, and GPI precipitation estimates. In the AGPI the GPI is adjusted to the large-scale bias of the SSM/I, which is assumed lower than the GPI's. As noted in the "satellite-gauge precipitation product" discussion (section 5), the Multi-Satellite product is adjusted to the large-scale bias of the Gauge analysis before the combination is computed. It continues to be the case that biases over ocean are not corrected by gauges in the Multi-Satellite and Satellite-Gauge products. ........................................................................... All four of the single-source estimates have shown strong *intercomparison results* in various intercomparison projects (section 2), which was the basis for selection for this project. The concept of combination is relatively new, so there is no strong comparison available. An early validation against the Surface Reference Data Center analysis yields the statistics in Table 3. Overall, the combination appears to be working as expected. Table 3. Summary statistics for all cells and months comparing the SSM/I composite, Multi-satellite, Gauge, and Satellite-gauge products to the SRDC analysis for July 1987 -- December 1991. | Bias | Avg. Diff. | RMS Error Product | (mm/mo) | (mm/mo) | (mm/mo) ----------------+---------+------------+---------- | | | SSM/I composite | 4.03 | 60.10 | 88.05 | | | Multi-satellite | -5.80 | 44.20 | 62.47 | | | Gauge (GPCC) | 6.77 | 18.85 | 35.11 | | | Satellite-gauge | 3.70 | 20.29 | 32.98 .......................................................................... The *quality index* variable has recently been proposed by Huffman et al. (1997) and developed in Huffman (1997a) as a way of comparing the errors computed for different techniques. Absolute error tends to zero as the average precipitation tends to zero, while relative error tends to infinity. According to (2), the dependence is approximately SQRT(rbar) and 1/SQRT(rbar), respectively. Thus, it is hard to illustrate overall dependence on sample size with either representation. However, if one inverts (2) it is possible to get an expression for a number of samples as a function of precipitation rate and the estimated error variance: Hg * ( rbarx + Sg) * [ 720 + 268 * SQRT ( rbarx ) ] Neg = --------------------------------------------------- (5) VARx where rbarx and VARx are the precipitation rate and estimated error variance for technique X, Hg and Sg are the values of H and S for the gauge analysis, and Neg is the number of "equivalent gauges," an estimate of the number of gauges that corresponds to this case. Tests show that Neg is well-behaved over the range of rbar, largely reflecting the sampling that provided rbarx and VARx, but also showing differences in the functional form of absolute error over the range of rbar for different techniques. Qualitatively, higher Neg denotes more confident answers. Values above 10 are relatively good. The SSM/I composite estimates tend to have Neg around 1 or 2, while the AGPI has Neg around 3 or 4. The rain gauge analysis runs the whole range from 0 to a few grid boxes in excess of 40. .......................................................................... 12. DATA ARCHIVES The major *archive and distribution sites* for the GPCP Version 1b Combined Precipitation Data Set are as follows: Dr. Alan McNab World Data Center A (WDC-A) National Climatic Data Center (NCDC) Rm 514 151 Patton Ave. Asheville, NC 28801-5001 USA Phone: 704-271-4592 Fax: 704-271-4328 Internet: amcnab@ncdc.noaa.gov WDC-A Home Page: http://www.ncdc.noaa.gov/wdcamet.html#GPCP Dr. Bruno Rudolf Global Precipitation Climatology Centre (GPCC) Deutscher Wetterdienst (DWD) Postfach 10 04 65 D-63004 Offenbach a.M., Germany Phone: +49-69-8062-2765 Fax: +49-69-8062-2880 Internet: brudolf@dwd.d400.de GPCC Home Page: http://www.dwd.de/research/gpcc Independent archive and distribution sites exist for the single-source data sets, and a current list may be obtained by contacting Dr. McNab at NCDC. .......................................................................... 13. DOCUMENTATION The *documentation creator* is: Dr. George J. Huffman Code 912 NASA Goddard Space Flight Center Greenbelt, MD 20771 USA Phone: 301-614-6308 Fax: 301-614-5492 Internet: huffman@agnes.gsfc.nasa.gov .......................................................................... The *documentation revision history* is: April 24, 1996 Draft 1 by GJH May 2, 1996 Draft 2 by GJH May 3, 1996 Draft 3 by GJH May 8, 1996 Draft 4 by GJH May 10, 1996 Version 1 by GJH June 5, 1996 Version 1a by GJH August 13, 1996 Version 1a Rev. 1 by GJH September 11, 1996 Version 1a Rev. 2 by GJH June 20, 1997 Version 1a Rev. 3 by GJH November 11, 1997 Version 1a Rev. 4 by GJH January 18, 1998 Version 1a Rev. 5 by GJH April 28, 1998 Version 1a Rev. 6 by GJH May 5, 1998 Version 1a Rev. 7 by GJH June 15, 1998 Version 1a Rev. 8 by GJH August 10, 1998 Version 1a Rev. 9 by GJH January 26, 1999 Version 1a Rev. 10 by GJH March 8, 1999 Version 1b by GJH The latest version has a new version number, 1b, and features new combination errors, consistent lack of inter-satellite calibration, a revised leo-AGPI scheme, and use of 1x1-deg IR data starting January 1997. .......................................................................... The list of *references* used in this documentation is: Adler, R.F., G.J. Huffman, and P.R. Keehn 1994: Global rain estimates from microwave-adjusted geosynchronous IR data. Remote Sens. Rev., 11, 125-152. Arkin, P.A., and B. N. Meisner, 1987: The relationship between large-scale convective rainfall and cold cloud over the Western Hemisphere during 1982-1984. Mon. Wea. Rev., 115, 51-74. Grody, N.C., 1991: Classification of snow cover and precipitation using the Special Sensor Microwave/Imager (SSM/I). J. Geophys. Res., 96, 7423-7435. Huffman, G.J., 1997a: Estimates of root-mean-square random error contained in finite sets of estimated precipitation. J. Appl. Meteor., 36, 1191-1201. __________, ed., 1997b: The Global Precipitation Climatology Project monthly mean precipitation data set. WMO/TD No. 808, WMO, Geneva, Switzerland. 37pp. __________, R.F. Adler, B. Rudolf, U. Schneider, and P.R. Keehn, 1995: Global precipitation estimates based on a technique for combining satellite-based estimates, rain gauge analysis, and NWP model precipitation information. J. Climate, 8, 1284-1295. __________, __________, P.A. Arkin, A. Chang, R. Ferraro, A. Gruber, J. Janowiak, R.J. Joyce, A. McNab, B. Rudolf, U. Schneider, and P. Xie, 1997: The Global Precipitation Climatology Project (GPCP) Combined Precipitation Data Set. Bull. Amer. Meteor. Soc., 78, 5-20. Janowiak, J.E., and P.A. Arkin, 1991: Rainfall variations in the tropics during 1986-1989. J. Geophys. Res., 96, 3359-3373. Legates, D.R, 1987: A climatology of global precipitation. Pub. in Climatol., 40, U. of Delaware. McNab, A., 1995: Surface Reference Data Center Product Guide. National Climatic Data Center, Asheville,NC, 10 pp. Morrissey, M.L., and J. S. Green, 1991: The Pacific Atoll Raingauge Data Set. Planetary Geosci. Div. Contrib. 648, Univ. of Hawaii, Honolulu, HI, 45 pp. Rudolf, B., 1993: Management and analysis of precipitation on a routine basis. Proc. Internat. WMO/IAHS/ETH SYMP. on Precip. and Evap., Slovak Hydromet. Inst., Bratislava, Sept. 1993, 1, 69-76. Weng, F., and N.C. Grody, 1994: Retrieval of cloud liquid water using the Special Sensor Microwave Imager (SSM/I). J. Geophys. Res., 99, 25535-25551. Wilheit, T., A. Chang and L. Chiu, 1991: Retrieval of monthly rainfall indices from microwave radiometric measurements using probability distribution function. J. Atmos. Ocean. Tech., 8, 118-136. Willmott, C.J., C.M. Rowe, and W.D. Philpot, 1985: Small-scale climate maps: A sensitivity analysis of some common assumptions associated with grid-point interpolation and contouring. Amer. Cartographer, 12, 5-16. WCRP, 1986: Report of the workshop on global large scale precipitation data sets for the World Climate Research Programme. WCP-111, WMO/TD - No. 94, WMO, Geneva, 45 pp. Xie, P., and P.A. Arkin, 1996: Analysis of global monthly precipitation using gauge observations, satellite estimates, and numerical model predictions. J. Climate, 9, 840-858. .......................................................................... 14. INVENTORIES The *data set inventory* may be obtained by accessing the home pages or contacting the representatives listed in section 12. .......................................................................... 15. HOW TO ORDER DATA AND OBTAIN INFORMATION ABOUT THE DATA Users interested in *obtaining data* should access the home pages or contact the representatives listed in section 12. ..........................................................................