GPCP (Monthly): Global Precipitation Climatology Project

The following was contributed by Angeline Pendergrass (NCAR), July, 2014:

"GPCP monthly data

 The GPCP monthly dataset is a widely-used global monthly gridded precipitation dataset.  It has global coverage on a 2.5 x 2.5 degree grid.   The dataset begins in 1979 and continues through the present, with some delay for processing.  It is used to calculate the climatology of precipitation, for comparison with climate models, and for analysis of changes over the recent past. 

GPCP combines microwave and infrared measurements, and outgoing longwave radiation retrievals from satellites and, over land, also incorporates rain gauge observations.  The available input data change over time as satellites are launched and retired and the gauge network evolves.   The microwave satellite data in particular are thought to be quite accurate.  Over land, microwave scattering data are used, while over ocean microwave emission measurements are used. The microwave retrievals are problematic over ice and at high latitudes, and are phased out starting at 40 degrees of latitude.

 All GPCP versions are multi-satellite combination datasets. The analysis method changed from version 1 (described in Huffman et al 1997) to 2 (described in Adler et al 2003). Version 2.1 incorporated a substantial update of the gauge input dataset (described in Huffman et al 2009); version 2.2 included data from the SSMIS satellite (currently described only in the documentation, Huffman and Bolvin 2012). 

GPCP “was formed to improve understanding of seasonal to inter-annual and longer term variability of the global hydrological cycle, determine the atmospheric latent heating rates needed for weather and climate prediction models, and provide an observational data set for model validation and initialization and other hydrological applications.” (Gruber and Levizzani 2008).  Some examples of studies that use the dataset follow. The climatology from GPCP is often used to validate climate models (e.g., Donner et al 2011, Giorgetta et al 2013) and reanalyses (Bosilovich et al 2011).  It is used to study variations in precipitation at global (e.g., Allan et al 2013, Trenberth 2011, John et al 2009, Su and Neelin 2003) and regional scales (e.g., Giannini et al 2008). Wentz et al (2007) use the land data along with their own analysis over ocean to examine trends in global precipitation.  It can be used to examine relationships between circulation and precipitation (e.g., Screen et al 2013, Frierson et al 2013, Loeb et al 2014, Johnson and Xie 2010).

Uncertainty is characterized via an error estimate of each datapoint available alongside the dataset.  This estimate accounts only for the random error, and depends on the mean rain rate and the number of samples used to calculate it (Huffman 1997).  Other systematic sources of error (such as the dependence on the observing system) are not quantified, though they are also present.  The most important systematic errors are over ocean, where there are few regular surface measurements to provide validation of satellite estimates. There is a at least one attempt at validation using measurements from small islands (Pfeifroth et al 2013), though their representativeness is still in question (Wang et al 2014).

 The key limitation of this dataset (like all merged-satellite precipitation products) is the indirect and complex nature of translating sparse satellite precipitation measurements into gridded precipitation estimates.  Satellites can only indirectly measure quantities related to rain rate at the surface: microwave and infrared satellites measure brightness temperature, which is then converted to rain rate indirectly, while radars measure energy reflected by cloud and rain drops throughout the depth of the column.  Then, these indirect measurements (along with direct gauge measurements over land) are used as an input to a complex algorithm that produces estimates of surface rain rate on a regular grid in time and space.  Many of the data that go into the precipitation estimates are actually measurements of clouds, rather than precipitation itself.  Also, the lack of rain gauges for ground validation over ocean introduces unknown systematic errors into this and all ocean precipitation estimates from satellites. 

There is active debate about the possibility that the global-mean amount of precipitation in GPCP is lower than is necessary to balance other components of the global energy budget  (see Stephens et al 2012, Wild et al 2012, Behrangi et al 2014, Trenberth et al 2009).  Stephens et al (2012) and Behrangi et al (2014) argue that the GPCP dataset may be missing light rain over ocean, and may suffer from a lack of ground validation in the Southern Ocean in particular.  The satellite estimates used in GPCP miss light rain events compared to CloudSat (Behrangi et al 2012, 2014b).  Trenberth and Fasullo (2013) evaluate moisture flow from land to ocean in order to validate precipitation data (in combination with evaporation estimates), which captures orographic rain that isn’t captured by the gauge network.  These “missing” rainfall sources may help explain why GPCP has a lower global-mean precipitation than climate models.

 A list of comparable satellite-gauge precipitation datasets was compiled by the International Precipitation Working Group, and can be found here: http://www.isac.cnr.it/~ipwg/data/datasets1.html. CMAP is another monthly dataset based on a different set of inputs (including satellites and also incorporating some NCEP reanalysis data, http://www.esrl.noaa.gov/psd/data/gridded/data.cmap.html#detail).  TRMM 3B43 is another monthly dataset with coverage from 50 N to 50 S with higher spatial resolution.

The uniform spatial grid of this dataset lends itself to comparison with climate models (though, see comments about regridding in GPCP 1dd description).  To undertake this endeavor, one could compare climatologies over the same time period, or compare variations in time by examining model integrations forced with prescribed SSTs (AMIP experiments) and historical forcing.  Trends can be compared between model experiments with historical forcing over the same period of record.

 The most common mistake dealing with all precipitation data arises from its highly variable nature in both space and time.  Precipitation datasets with different spatial or temporal coverage are fundamentally different, so care must be taken, for example, comparing the data taken from one station (a point in space) with a GPCP grid-box (which represents an areal average). In order to compare two precipitation datasets on grids with different resolutions, one should typically regrid one or both of the datasets to a common grid, no finer than the coarsest of the datasets being compared, using a regridding method that conserves the total amount of rain falling in an area. Often, the default interpolation in an analysis software package is bilinear interpolation (eg in Matlab), which is not conservative.  One conservative regridding method is described in Jones (1999).  The NCL language offers two methods to perform conservative regridding: area_conserve_remap  or, the more robust, Earth System Modeling Framework (ESMF) regridding. See also the Climate Data Guide's regridding page (https://climatedataguide.ucar.edu/climate-data-tools-and-analysis/regridding-overview).

Some corrections are made to GPCP v 2.2for changes in satellite and gauge dataset inputs over time.  Different datasets and algorithms were used, with some periods of inter-calibration to minimize inconsistencies.  No corrections were made for OPI’s drift of equator-crossing time (GPCP V2.2 documentation, Huffman and Bolvin 2013, ftp://precip.gsfc.nasa.gov/pub/gpcp-v2.2/doc/V2.2_doc.pdf).

There are many temporal features in the record that users should be aware of. Before and after 1987, the dataset has different inputs and algorithms (Gruber and Levizzani 2008).   Microwave precipitation estimates became available in July of 1987. Different gauge products with different sampling are used before and after 1987.  Different IR datasets are also used before and after 1987.  New satellite data were also incorporated after 1997. See Table 2.2 of Gruber and Levizzani (2008) for a summary of input datasets.

Some known anomalies are also listed in the v2.2 documentation (section 9).  These are:

- June 1990-Dec 1991: loss of SSM/I data (Gruber and Levizzani 2008) resulted in the use of a different algorithm

 - August 1993-Jan 1994: bias due to a decrease in Meteosat IR images

 - Jan 2000, extreme southwestern Greenland: unusually high rainfall estimates due to unusually high gauge rainfall at Nuuk, Greenland.

There are also spatial features in the dataset to be aware of. Satellite data are treated differently in different latitude bands.  Equatorward of 40 degrees, SSM/I data are used.  When missing, they are replaced with adjusted TOVS data.  Between 40 and 70 degrees, a combination of SSM/I and TOVS data with bias adjustment that varies with latitude is used. Poleward of 70 N and S, adjusted TOVS data are used exclusively."##