Retrieval of Precipitation from Satellites
267
• The satellite data of the comparison data base are never free from errors ranging from
spurious complete radiometer failure (which must be detected) over calibration problems
(these especially affect algorithms based on radiative transfer simulations) to navigation
errors when earth location of the data is crucial (e.g. when data sets are blended with
each other or results from different data sources are compared).
• There are by now no accepted large scale validation data sets. Ground-truth data obtained
from rain-gauges are desirable but the network densities are much too low, especially over
the oceans. Monthly means can be compared at best because of the different sampling
characteristics of satellite (instant area averages) and rain-gauge measurements (timeintegrated point measurements). Radar measurements provide areal averages, which can
be better compared with satellite measurements, but the results of radar-derived rainfall
estimates have their own problems, which must be carefully evaluated before accepting
the data for validation purposes. There are strong indications that radar data are biased
as a function of the distance from radar caused by the increasing beam distance rom the
surface (Petty, personal communication). One of the most recent of the larger satellite
retrieval programs (PIP-I, see below) decided not to use radar data at all for their purpose.
• Lacking adequate validation data sets evaluation of algorithm performance is mostly based
on algorithm intercomparison. While it can be agreed on the type of results to compare
(monthly means, standard deviations, range etc.) and the means how to compare (simple
differences, correlation coefficients, etc.) there is no commonly agreed basis for ranking
the different algorithms.
11.7.1 Programs
There have been many attempts to compare and validate satellite rain retrieval algorithms. Two
of the most recently completed comparison projects involving a range of algorithms and research
groups from many laboratories are the First Algorithm Intercomparison Project (AlP /1) within
the Global Precipitation Climatology Project (GPCP) (Arkin and Xie, 1994) and the First
Precipitation lntercomparison Project (PIP-I) of the NASA-led Wet Net Project (Barrett et
al., 1994; 1995). The former concentrates on the evaluation of IR-based algorithms while the
latter is directed more to passive microwave methods. In the following we will shortly describe
these projects and summarize their mayor results.
AlP /1 (GPCP)
GPCP was established by the World Climate Research Programme (WCRP) to produce global
analyses of area- and time-averaged precipitation for use in climate research. To achieve this
goal satellite-based VIS/IR-retrieval algorithms are used primarily to be able to make use of
the high time and space sampling frequency obtained from geostationary satellites. Within
AlP /1 rainfall estimates derived from the visible and IR measurements of the Geostationary
Meteorological Satellite (GMS) and from DMSP passive microwave observations are compared
with rainfall derived from a combination of data from 15 precipitation radars and over 1300
automated raingauges over the Japanese islands and the adjacent ocean regions during the
June and Mid-July to Mid-August periods of 1989. Of the 17 compared rainfall estimates 7 are
based on IR-estimates alone, four combine IR and visible data, three combine SSM/I and IR
data two use model forcasts and one uses climatological information together with IR data.
We have learned in earlier chapters that IR-methods are based on the statistical relation between rainfall intensity at the surface and the areal and time coverage of clouds with cloud-top
267
• The satellite data of the comparison data base are never free from errors ranging from
spurious complete radiometer failure (which must be detected) over calibration problems
(these especially affect algorithms based on radiative transfer simulations) to navigation
errors when earth location of the data is crucial (e.g. when data sets are blended with
each other or results from different data sources are compared).
• There are by now no accepted large scale validation data sets. Ground-truth data obtained
from rain-gauges are desirable but the network densities are much too low, especially over
the oceans. Monthly means can be compared at best because of the different sampling
characteristics of satellite (instant area averages) and rain-gauge measurements (timeintegrated point measurements). Radar measurements provide areal averages, which can
be better compared with satellite measurements, but the results of radar-derived rainfall
estimates have their own problems, which must be carefully evaluated before accepting
the data for validation purposes. There are strong indications that radar data are biased
as a function of the distance from radar caused by the increasing beam distance rom the
surface (Petty, personal communication). One of the most recent of the larger satellite
retrieval programs (PIP-I, see below) decided not to use radar data at all for their purpose.
• Lacking adequate validation data sets evaluation of algorithm performance is mostly based
on algorithm intercomparison. While it can be agreed on the type of results to compare
(monthly means, standard deviations, range etc.) and the means how to compare (simple
differences, correlation coefficients, etc.) there is no commonly agreed basis for ranking
the different algorithms.
11.7.1 Programs
There have been many attempts to compare and validate satellite rain retrieval algorithms. Two
of the most recently completed comparison projects involving a range of algorithms and research
groups from many laboratories are the First Algorithm Intercomparison Project (AlP /1) within
the Global Precipitation Climatology Project (GPCP) (Arkin and Xie, 1994) and the First
Precipitation lntercomparison Project (PIP-I) of the NASA-led Wet Net Project (Barrett et
al., 1994; 1995). The former concentrates on the evaluation of IR-based algorithms while the
latter is directed more to passive microwave methods. In the following we will shortly describe
these projects and summarize their mayor results.
AlP /1 (GPCP)
GPCP was established by the World Climate Research Programme (WCRP) to produce global
analyses of area- and time-averaged precipitation for use in climate research. To achieve this
goal satellite-based VIS/IR-retrieval algorithms are used primarily to be able to make use of
the high time and space sampling frequency obtained from geostationary satellites. Within
AlP /1 rainfall estimates derived from the visible and IR measurements of the Geostationary
Meteorological Satellite (GMS) and from DMSP passive microwave observations are compared
with rainfall derived from a combination of data from 15 precipitation radars and over 1300
automated raingauges over the Japanese islands and the adjacent ocean regions during the
June and Mid-July to Mid-August periods of 1989. Of the 17 compared rainfall estimates 7 are
based on IR-estimates alone, four combine IR and visible data, three combine SSM/I and IR
data two use model forcasts and one uses climatological information together with IR data.
We have learned in earlier chapters that IR-methods are based on the statistical relation between rainfall intensity at the surface and the areal and time coverage of clouds with cloud-top