The Role of Combination Techniques in Maximizing the Utility of Precipitation Estimates from Several Multi-Purpose Remote-Sensing Systems

2 Pages·2001·0.19 MB·English

by  Huffman, George J.

7.6 THE ROLE OF COMBINATION TECHNIQUES IN MAXIMIZING
THE UTILITY OF PRECIPITATION ESTIMATES FROM SEVERAL
MULTI-PURPOSE REMOTE-SENSING SYSTEMS
George J. Huffman .1'2, Robert F.Adler 1, David T. Bolvin 1'2, Scott Curtis 1,3
1" NASA/GSFC Laboratory for Atmospheres
2: Science Systems and Applications, Inc.
3: JCET/Univ. of Maryland, Baltimore County
1. INTRODUCTION almost 2 overpasses per day. These return periods
are long compared to the correlation times of
precipitation events, so time-average estimates
Multi-purpose remote-sensing products from
based on a single satellite contain substantial
various satellites have proved crucial in developing
uncertainty. Furthermore, the return periods are long
global estimates of precipitation. Examples of these
compared to the timescale of the diurnal cycle, so a
products include low-earth-orbit and geosyn-
single satellite has the strong likelihood of providing
chronous-orbit infrared (leo- and geo-IR), Outgoing
biased estimates of the daily total precipitation,
Longwave Radiation (OLR), Television Infrared
particularly when the orbit is sun-synchronous. As a
Operational Satellite (TIROS) Operational Vertical
result, data from geosynchronous-orbit satellites are
Sounder (TOVS) data, and passive microwave data
preferred because their observation interval is
such as that from the Special Sensor Microwave/
typically 1-3 hr.
Imager (SSM/I). Each of these datasets has served
In common with many, but not all, researchers who
as the basis for at least one useful quasi-global
compute combination precipitation estimates, the
precipitation estimation algorithm; however, the
authors adopt the philosophy that the apparently least
quality of estimates varies tremendously among the
biased and most accurate precipitation estimates
algorithms for the different climatic regions around the should be used to calibrate other, less-trusted
globe. estimates before combinations are computed. We
emphasize performing the calibration using data
2. GENERAL ISSUES
subsets that are as nearly coincident in space and
time as possible. For example, our Adjusted Geosyn-
It is "obvious" that the various precipitation
chronous Operational Environmental Satellite (GOES)
datasets can be combined in some fashion that draws
Precipitation Index (AGPI) uses approximately time/
on the strengths of each dataset to provide a "best"
_pace-matched SSM/I and geo-IR precipitation esti-
estimate of the global precipitation patterns, but the
barriers to practical implementation are substantial mates over a month to derive spatially varying
corrections which are then applied to the full GPI data
The preeminent difficulty is the lack of quality
validation data over the range of climatic conditions set. Once the AGPI is computed we decline to
for which wewish to make estimates. Dense arrays of combine its monthly average with the monthly average
rain gauges are considered the most reliable of the calibrating SSM/I estimates because the SSM/I
validation, and such sites mostly occur in developed estimates are likely diurnally biased and noisier.
countries at mid-latitudes over land. It is difficult to The combination techniques are designed to
correctly extend these regionally-biased, scattered produce better answers globally than the individual
validation results to the vast expanse of the globe for input datasets, but it is clear from the outset that
which validation data are poor (many land areas) or users interested in a specific regions must consider
totally lacking (most oceanic regions). As a result, the datasets' performance for their own needs.
researchers are left to base the improvement and
choice of algorithms on relative comparisons among 3. MONTHLY ESTIMATES
the various candidate datasets. Within these limita-
tions, the SSM/I-based algorithms are preferred over Combinations at the monthly 2.5°x2.5 °
the other algorithms because the remotely sensed latitude/longitude resolution have almost a decade of
signal (upwelling radiant energy invarious microwave heritage because atthose scales the sampling errors
frequencies) is more closely related to the occurrence are contained for low-orbit satellites and relatively
of hydrometeors than inthe other sensors. good for geosynchronous-orbit satellites. Most
A second substantial problem is that low- recently, the Global Precipitation Climatology Project
earth-orbit satellites only sparsely sample a given (GPCP) Version 2 combination product was developed
location. A single SSM/I averages 1.2 overpasses per and released bythe authors (an extension of Huffman
day at the equator, while the National Oceanic and etal. 1997), providing a globally complete estimate of
Atmospheric Administration (NOAA) series averages monthly precipitation on a 2.5°x 2.5°grid for the 21-
year period 1979 - (delayed) present.
The combination of satellite data into a multi-
satellite (MS) product iscarried outdifferently during 3
* Corresponding author address: George J. periods according to data availability. Strong efforts
Huffman, NASA/GSFC Code 912, Greenbelt, MD, were made tohomogenize the data record:
20771' e-mail: [email protected]
boundaries. Validation against dense raingauges for
•
1979 to 1985(0LR Precipitation Index - OPI) The
individual grid boxes values shows a very high RMS
OPI was calibrated against the 1988-1998 GPCP error, as expected, but statistics improve when
satellite-gauge (SG) product for calibration. That
time/space averaging is performed.
is, at each gridbox month-to-month OLR anomalies
from climatology are regressed against SG
5. THREE-HOURLY ESTIMATES
anomalies, and afallback direct regression of OLR
against SG is computed.
Currently we are working towards global 3-hourly
• 1986 to Mid-1987 (geo-IR, OPI) The MS is
precipitation fields. At present we are experimenting
composed of geo-IR GPI scaled by SG-scaled OPI
with the geo-IR zone (40°N-S). Similar to the 1DD, we
in 40°N-S and the SG-scaled OPI elsewhere.
• Mid-1987 to the Present (geo-IR, TOVS, SSM/I) • scale SSM/I estimates to TRMM TMI estimates;
TOVS is merged in with SSM/I where the SSM/I is • combine the TMI and scaled SSM/I to create a
suspect (outside about 45°N-S) or missing. In the high-quality (HQ) estimate;
band 40°N-S thebias of thegeo-IR GPI is adjusted • make geo-IR pixel rainrates a function of the
to that of the SSM/I, producing the AGPI. The MS depression of the To below the HQ-based Tbo
is composed of AGPI inthe band 40°N-S and the threshold; and
merged TOVS-SSM/I elsewhere. • replace geo-IR estimates with the HQ estimates,
In each of the periods the MS and a gauge analysis where available.
are linearly combined into the SG combination using There are substantial questions about how to perform
weighting by inverse estimated mean-square error for the combination, but working at the instantaneous
each individual estimate.
level avoids the diurnally sampling bias issue.
Asimilar dataset is being computed for the Tropical
Rainfall Measuring Mission (TRMM) in the band 40°N-S 6.FUTURE PROSPECTS
using theTRMM Microwave Imager (TMI)-Precipitation
Radar (PR) multisensor precipitation estimate to The continued acceleration of computing power
calibrate the GPI. It covers the TRMM period [1998 - has now enabled the production of essentially full-
(delayed) present]. resolution, half-hourly (when available) global geo-IR
datasets. At the same time, the National Aeronautics
4. DAILY ESTIMATES
and Space Administration has proposed development
of the Global Precipitation Mission (GPM). GPM is
GPCP began collecting a finer scale global geo-IR planned to provide global 3-hourly passive microwave
dataset in late 1996, enabling development of a
data by adding new, low-cost satellites to fill the gaps
correspondingly finer scale set of precipitation incurrent and planned passive microwave satellites,
estimates. The GPCP One-Degree Daily (1DD) data- and to provide aTRMM-like satellite to calibrate all the
set is the authors' first technique for estimating global
passive microwave estimates on an on-going basis.
daily precipitation at the 1°xl ° scale from observa- We expect a permanent role for the geo-IR in filling the
tionally based data. It is currently available for 1997 - inevitable gaps in microwave coverage, as well as
(delayed) present (Huffman et al. 2001). enabling sub-3-hourly precipitation estimates at fine
Where possible, (40°N-S) the Threshold-Matched spatial scales. As well, we expect a great deal of
Precipitation Index (TMPI) provides GPl-like precipita- development of theory and application for estimating
tion estimates in which: errors inthe various estimates. Realizing the potential
• the geo-IR Tbthreshold (Tbo)is set locally from of these datasets will require a long-term research
month-long accumulations of time/space effort and continued cultivation of long-term, high-
coincident geo-IR Tb and SSM/I-based precipita- quality raingauge networks for validation and direct
tion frequency; and combination with the satellite-based estimates.
• the conditional rain rate is set locally from the
resulting frequency of T_<Tbofor the entire month 7, REFERENCES
and the GPCP SG.
Huffman, G.J., R.F. Adler, S. Curtis, M. Morrissey,
Scaled low-orbit IRGPI is used to fill holes in individual
J.E. Janowiak, B McGavock, J. Susskind, 2001:
3-hrly geo-IR images.
Global Precipitation at One-Degree Daily Resolution
Outside 40°N-S the 1DD is based on TOVS:
from Multi-Satellite Observations. J. Hydrometeor.,
• the number of TOVS rain days are reduced to in revision.
match TMPI rain days; Huffman, G.J., R.F. Adler, P. Arkin, A. Chang, R.
• TOVS precipitation by zeroing the (1-ratio) Ferraro, A. Gruber, J. Janowiak, A. McNab, B.
smallest daily TOVS rain accumulations; and Rudolf, U. Schneider, 1997: The Global
Precipitation Climatology Project (GPCP) Combined
• the remaining (non-zero) TOVS rain days are
Precipitation Data Set. Bull. Amer. Meteor. Soc.,
rescaled to sum to the monthly SG.
78, 5-20.
The time series of the global images shows good
continuity in time and space across the data