AMTAtmospheric Measurement TechniquesAMTAtmos. Meas. Tech.1867-8548Copernicus GmbHGöttingen, Germany10.5194/amt-8-1217-2015A novel algorithm for detection of precipitation in tropical regions
using PMW radiometersCasellaD.daniele.casella@artov.isac.cnr.ithttps://orcid.org/0000-0001-7203-4232PanegrossiG.https://orcid.org/0000-0002-5170-7087SanòP.https://orcid.org/0000-0001-7059-1043MilaniL.https://orcid.org/0000-0003-0498-1021PetraccaM.DietrichS.https://orcid.org/0000-0003-3808-365XConsiglio Nazionale delle Ricerche – Istituto di Scienze dell'Atmosfera e del Clima (CNR-ISAC), Rome, ItalyD. Casella (daniele.casella@artov.isac.cnr.it)12March201583121712323August201412September201416January20153February2015This work is licensed under a Creative Commons Attribution 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by/3.0/This article is available from https://www.atmos-meas-tech.net/8/1217/2015/amt-8-1217-2015.htmlThe full text article is available as a PDF file from https://www.atmos-meas-tech.net/8/1217/2015/amt-8-1217-2015.pdf
A novel algorithm for the detection of precipitation is described and tested.
The algorithm is applicable to any modern passive microwave radiometer on
board polar orbiting satellites independent of the observation geometry and
channel frequency assortment. The algorithm is based on the application of
canonical correlation analysis and on the definition of a threshold to be
applied to the resulting linear combination of the brightness temperatures in
all available channels. The algorithm has been developed using a 2-year data
set of co-located Special Sensor Microwave Imager/Sounder (SSMIS) and
Tropical Rainfall Measuring Mission precipitation radar (TRMM-PR)
measurements and Advanced Microwave Sounding Unit (AMSU) Microwave Humidity
Sounder and TRMM-PR measurements. This data set was partitioned into
four classes depending on the background surface emissivity (vegetated land,
arid land, ocean, and coast) with the same procedure applied for each surface
class. In this paper we describe the procedure and evaluate the results in
comparison with many well-known algorithms for the detection of
precipitation.
The algorithm shows a small rate of false alarms and superior detection
capability; it can efficiently detect (probability of detection between 0.55
and 0.71) minimum rain rate varying from 0.14 mmh-1 (AMSU over
ocean) to 0.41 (SSMIS over coast) with the remarkable result of
0.25 mmh-1 over arid land surfaces.
Introduction
The Global Precipitation Measurement mission (GPM) (Hou et
al., 2014) started its operational phase on 28 February 2014 with the launch
of the NASA/JAXA GPM Core Observatory. The goal of the mission is to provide
instantaneous precipitation measurements with a coverage of less than
1 h over 60 % of the globe and less than 3 h over
80 % of the globe through the exploitation of a constellation of passive
microwave (PMW) radiometers on board research and operational satellites
provided by the United States, Japan, France/India, and the European
Community. The GPM Core Observatory carries the first spaceborne
dual-frequency precipitation radar, the dual-frequency precipitation radar
(DPR), operating at Ku and Ka bands (13 and 35 GHz, respectively) and
a conical scanning multichannel (10–183 GHz) microwave imager, the
GPM Microwave Imager (GMI). The GPM constellation consists of radiometers
with different scanning geometries (both conical and cross-track scanning),
different assortments of multichannel frequencies and polarizations, and
different spatial resolutions. The microwave imagers and sounders with a
conical scanning observation geometry are (1) the Special Sensor Microwave
Imager/Sounder (SSMIS), which measures microwave energy at 24 discrete
frequencies from 19 to 183 GHz (Kunkee et al., 2008), and (2) the
Advanced Microwave Scanning Radiometer 2 (AMSR-2) (Shimoda, 2005) with
channels ranging from 6.9 to 89 GHz. The cross-track scanners are
(1) the Microwave Humidity Sounder (MHS) (89–190 GHz) on board the
NOAA Polar-Orbiting Environmental Satellites (POES) and the EUMETSAT MetOp
satellites (Edwards and Pawlak, 2000), (2) the Advanced Technology Microwave
Sounder (ATMS) instrument on board the NPOESS NPP satellite (Muth et
al., 2005) with 22 channels in the range of frequency 23.8–183 GHz,
and (3) Sondeur Atmosphérique du Profil d'Humidité Intertropicale par
Radiométrie (SAPHIR) with six channels in the 183 GHz water vapor
absorption band, carried by the French/Indian satellite Megha-Tropiques
(Desbois et al., 2003). It is also worth mentioning the Advanced Microwave
Sounding Unit-A (AMSU-A), which is another cross-track scanner with
15 channels ranging from 23.8 to 89 GHz; this last instrument is not
included in the official GPM constellation, but it is carried by the same
satellites equipped with the MHS radiometer.
Obtaining coherent (and accurate) precipitation estimates from this
assortment of instruments requires a robust intercalibration of the
brightness temperatures (TBs). The GPM has established an international team
called Intersatellite Calibration Working Group (X-CAL) in order to address
this need (Wilheit, 2013). Moreover, the estimates themselves need to be
obtained with common procedures and data sets. The 2014 version of the
Bayesian-based Goddard PROFiling algorithm (GPROF2014) (Kummerow et
al., 2001, 2011) will be applied to all passive microwave radiometers in the
GPM constellation and it will be the official NASA PMW precipitation
retrieval algorithm for GPM. However, the inconsistencies in the
precipitation detection deriving from the use of different radiometers also
might significantly affect the rainfall estimates from such a heterogeneous
constellation and precipitation products derived from the combination of
these estimates (i.e., IR/MW merging techniques). The differences in the
available channels, polarization information, spatial resolutions, and
observation geometry have a strong impact on the possibility of separating
the radiance due to the background surface from the signal related to
precipitation. Therefore, the limits of each sensor in detecting
precipitation should be carefully analyzed in order to establish the degree
of consistency of the precipitation patterns (and retrievals) obtained from
different radiometers. A certain degree of coherence between different
sensors can be accomplished by developing common procedures to be applied to
all radiometers to detect efficiently the presence of precipitation in
different environmental conditions. Consistency and accuracy are also the
priorities in the development of our PMW precipitation retrieval algorithms
used for the operational PMW precipitation products within the EUMETSAT
“Satellite Application Facility on Support to Operational Hydrology and
Water Management” (H-SAF) program (see Mugnai et al., 2013a). The H-SAF PMW
algorithms for cross-track and conical-scanning radiometers are built upon
the same physical foundations and use common procedures for the detection of
precipitation (see Casella et al., 2013; Sanò et al., 2013, 2015; Mugnai
et al., 2013b; Panegrossi et al., 2013).
Detection of precipitation from satellite observations using passive
microwave radiometers is a difficult task, since the brightness temperatures
emerging from precipitating clouds can be similar to those emitted by some
surfaces in clear sky conditions. Precipitation detection is simpler over
ocean because at the low-frequency channels (10–30 GHz) the surface
background appears relatively cold (in terms of brightness temperatures) due
to the low and nearly uniform sea surface emissivity. This allows the
detection of the absorption/emission from large rain drops as relatively warm
areas. Over land, the detection of precipitation is more complex as the rain-layer emission is obscured by the high emissivity of the background surface.
The detection of precipitation still has some open issues (see Munchak and
Skofronick-Jackson, 2013), especially over coast, arid regions, and over snow-covered surfaces. It is worth noting that some algorithms for precipitation
estimation do not include specific procedures for the detection of
precipitation and provide just a “probability of precipitation” (e.g., GPROF
from version 7 of the TMI-GPROF and in the GMI-GPROF; GPM, 2010). However, an
accurate procedure for the detection of precipitation could be very useful
within the estimation schemes in order to separate the problem of identifying
the precipitating areas from the problem of estimating the intensity of the
rainfall. Moreover, in the regions of the Earth where precipitation is
infrequent (e.g., in semi-arid regions in the tropics), precipitation detection
is crucial for drought monitoring. Over the last 20 years, different
precipitation detection procedures have been widely used within precipitation
retrieval algorithms.
It is quite common in the framework of precipitation estimates to refer to
precipitation detection as “screening”. Some widely used approaches are
based on the scattering index (SI) over land surface (i.e., Grody, 1991;
Ferraro, 1997). Another classical approach makes use of the polarization-corrected temperatures (Barrett et al., 1988; Spencer et al., 1989; Kidd,
1998; PCT). Unfortunately, these algorithms cannot be applied to the
cross-track scanning radiometers (e.g., AMSU-A, AMSU-B, MHS, ATMS), because
they need both vertical and horizontal polarized channels at the same
frequency. However, the most recent spaceborne radiometers utilize high-frequency channels (150–160 GHz) and sounding channels in the
60 GHz oxygen absorption band and the 183 GHz water vapor absorption
band. As a consequence, relatively new algorithms for the detection of
precipitation have been developed (e.g., Chen and Staelin, 2003, using both
the 52.3 GHz channel and the 183 GHz channels and Laviola and Levizzani,
2011, using the water vapor absorption band channels around 183 GHz).
Recently more efforts have been made in solving the issue of detection of
precipitation: among others, Grecu and Anagnostou (2001) applied a
neural network approach, Seto et al. (2005) developed a lookup-database
method, Laviola and Levizzani (2009) proposed a simple technique based upon a
threshold on an AMSU-B channel combination, and Islam et al. (2014) have
built a random forest classifier.
In this paper we will describe a precipitation detection algorithm based on
canonical correlation analysis (CCA), hereafter referred to as the CCA
algorithm. The novelty of the algorithm is that it is applicable to all
available PMW radiometers in the GPM constellation, it is in principle very
simple, and it is conditioned to the availability of a large data set of
co-located measurements of multichannel TBs and quality-controlled
precipitation measurements or estimates that can be considered as “truth”.
The algorithm was inspired by the approach of Petty (2013), who developed a
methodology based on a two-stage principal component analysis aimed at
reducing the dimensionality of the input in a Bayesian algorithm for the
estimation of precipitation and “distilling” the relevant information from
a multidimensional set of channels. The approach of Petty (2013) has been
implemented and validated in two subsequent papers (Petty and Li, 2013a,
2013b). In this work we have modified and adapted the approach of Petty
(2013) to the problem of precipitation detection using a simpler surface
classification and without considering the linearization of the TBs into
pseudochannels. We will show the CCA algorithm results and compare them to
those obtained from different well-known precipitation detection schemes used
within precipitation retrieval algorithms. The CCA algorithm described in
this paper is currently applied to the PMW precipitation retrieval algorithms
currently used to deliver operational products of instantaneous precipitation
over European and African regions within the EUMETSAT H-SAF program (Mugnai
et al., 2013a, 2013b; Casella et al., 2013; Sanò et al., 2013, 2015).
This paper is divided into five sections. Section 2 describes the data sets
of co-located measurements of multichannel TBs from SSMIS and AMSU/MHS and
precipitation measurements from the Tropical Rainfall Measuring Mission
precipitation radar (TRMM-PR) used in the training data set of the CCA
algorithm. A description of the procedures followed to separate the data set
depending on the type of background surface is also provided. Section 3
illustrates the methodology used for the definition of the algorithm based on
a training data set. Section 4 shows the results of the application of the
algorithm to a separate test data set in terms of probability of detection
(POD), false-alarm ratio (FAR) and Heidke skill score (HSS) with variable
rain/no-rain thresholds. Results of the comparisons with other screening
algorithms are also provided and the CCA performance is described in terms of
minimum detectable rain and of total error for each background surface type.
Finally Sect. 5 includes the summary of the main results and the conclusions.
Instruments and data set description
The study was carried out in the area between 36∘ S and 36∘ N
in latitude and between 60∘ E and 60∘ W in longitude,
covering the African continent, parts of the Arabian Peninsula, South
America, and the Atlantic and Indian oceans. All the analyses were performed
using coincident observations of the TRMM-PR with observations from SSMIS and
AMSU/MHS radiometers for the years 2011–2013. The TRMM-PR was the first
spaceborne precipitation radar; it provided publicly available data until
October 2014 and can be considered the precursor to GPM DPR. It is a
13.8 GHz radar scanning between -17 and +17∘ with a swath width
of 247 km (after the satellite was boosted to higher orbit in 2001).
SSMIS is a conical-scanning radiometer with a scanning angle of 45∘
and a swath width of 1707 km, measuring passive microwave radiation
in 24 channels with frequencies ranging from 19 to 183 GHz. The 19,
37, and 91 GHz channels are in the vertical and horizontal polarization,
while the 22 GHz channel is present only in the vertical polarization and
the 150 GHz is present only in the horizontal polarization as the
channels in the water vapor absorption band (around 183.31 GHz).
Finally, the channels in the oxygen absorption band (50.3–63.2 GHz)
are horizontally polarized (in the radiometer carried by the Defense
Meteorological Satellite Program (DMSP) F16, channels 1–7 were incorrectly
designed as V polarized) or right-circular polarized. The SSMIS radiometer is
carried on board four satellites of the DMSP: the F16, F17, F18, and F19; the
launch of F20 satellite is planned for 2020. AMSU-A and MHS are both
cross-track scanning radiometers on board four satellites: NOAA-18, NOAA-19,
Metop A, and Metop B. AMSU-A has 23 channels between 23.8 and 89 GHz
while MHS has a channel at 89 GHz, at 157 GHz, and
three channels in the 183 GHz water vapor absorption band. The polarization
measured by every channel changes with the scan angle which varies between
-48.95 and 48.95∘, and the swath width of the radiometer is about
1920 km.
We selected all available coincidences (within a 15 min time window) of the
SSMIS radiometer with the TRMM-PR in the area of interest considering the
DMSP-F16 and DMSP-F17 satellites (i.e., the DMSP-F18 has the 150 GHz channel
malfunctioning since February 2012). The same was done for the available
AMSU/MHS radiometers for the Metop-A, NOAA-18, and NOAA-19 satellites.
Therefore, two data sets have been created, one of SSMIS-PR coincidences and
one of AMSU/MHS-PR coincidences, with a total of 3889 and 2581 coincident
overpasses for the AMSU/MHS and for the SSMIS data set, respectively.
To obtain co-located vectors of SSMIS or AMSU/MHS TBs and PR rainfall
estimates (TRMM product 2A25), the horizontal resolution variation with
frequency in conically scanning MW radiometer (SSMIS) and with the
observation angle in the cross-track scanning radiometers (AMSU/MHS) needs to
be taken into account. Moreover, there are sampling differences between
AMSU-A and MHS and between the SSMIS components (Imager, Environmental, Lower
and Upper Atmospheric Sounder). Therefore, the TRMM-PR rainfall rate at the
surface was downscaled to the SSMIS and MHS nominal resolution, defined as
the IFOV size of the 91 GHz/89 GHz channel for the SSMIS/MHS radiometer,
respectively. All the channels with a coarser spatial grid were resampled
using a nearest-neighbor approach; for the AMSU-A/MHS sensors we have used an
IFOV size variable with the scan angle as described by Bennartz (2000).
The resulting data sets were divided into three classes depending on the
background surface – land, ocean, and coast – using a digital land/sea map at
2 s of arc resolution. The land class was subdivided into vegetated
land and arid land (desert). The arid land pixels have been identified by
looking at the mean annual difference between the SSMIS 19 GHz V and H channels.
Grody (1991) has shown how the difference of the V and H polarized channels
is very effective in identifying desert. Although the presence of clouds may
reduce the polarization difference, this effect may be minimized by averaging
the TB difference over a long period. In this study we have looked at 1
year (2011) of SSMIS observations over the area of interest, selecting the
observations over land, and remapping them on a regular grid in latitude and
longitude (with 0.5∘ spacing). The difference of the TBs of the
19 GHz V and H channels (dTB19) was calculated for each grid point and
then averaged over 1 year. The area corresponding to each grid point was
identified as desert (or arid land) if the mean annual difference of
dTB19 was higher than 15 K:
dTB‾19=TB19V-TB19H‾>15K.
Figure 1 shows the results of this procedure over the area of interest. It is
clear how the Sahara and the Arabian desert have been correctly
identified as arid land. Smaller areas of arid land also appear in Iran
(including the Dasht-e Kavir and the Dasht-e lut deserts) and over the
African continent (including the Kalahari desert in southwest Africa and
arid regions in the continental Horn of Africa). However, some small deserts
near to the coast have been not correctly identified, i.e., the Namib
in Namibia and the Danakil Desert on the African coast of the Red Sea. The
coast pixels have been excluded from this test to eliminate the polarization
difference due to the sea surface emissivity.
The SSMIS-PR and AMSU/MHS-PR data sets have been both divided into two
separate data sets: a training set including all data from 2011 and 2012 and
a test data set including data from 2013.
Methodology
This section is dedicated to the procedure used to define the CCA algorithm
for which the training data set relative to the years 2011–2012 has been
used. We have considered all channels of both data sets (AMSU/MHS and SSMIS)
except the channels in the 50–60 GHz band with a peak of the weighting
function too high for the scope of this work, i.e., we have excluded the
channels in the 50–60 GHz band with a frequency higher than 55.5 GHz.
Map of arid land identified using the annual mean of the 19 GHz
channels difference (V–H). Dark grey areas are the regions identified as
desert or arid land.
Description of the seven-candidate discriminant functions.
Discriminant functionShort descriptionComments/goal1PCANEOF analysis on “no rain” pixels – first component chosenTo enhance the signal from the background surface2PCAREOF analysis on “rain” pixels – first component chosenTo enhance the signal from precipitation.3PCAN-PCAREOF analysis on “no rain” observations,To mask the background signal and enhance the signaldiscarding the first three components, and EOF analysisfrom precipitation following Petty (2013) .on “rain” pixels selecting the first component4CCARCCA of TBs for “rain” pixels – first canonical variableTo enhance the signal from rain.5PCAN-CCARAs in row three, with the difference that the secondSame goal as row three, with change fromprocedure is a CCA of TBs for “rain” pixelsPetty (2013) in the second step.6PCAR-CCAR(1) EOF analysis of “rain” pixels –Same goal as row three. As suggested by Wilks (1995), CCAdiscarding the last three components.is supposed to be more stable if applied to(2) CCA application to the selectedselected components (eliminating those morecomponents – first canonical variableaffected by random noise).7PCAN-PCAR-CCARAs in row three with one more step (as row six):To mask the background signal and enhance the signala CCA is applied to the PCAR afterfrom precipitation, row six and three procedures are joined.the last three components are discarded
The first test was performed in order to select the combination of channels
that is more suitable to discriminate the signal deriving from the
precipitating cloud (“rain”) from the signal deriving from the background
surface (“no rain”). We call this monodimensional combination of channels
the discriminant function. In order to obtain and test different discriminant
functions we have applied two well-known multivariate methods: empirical
orthogonal function (EOF) analysis and canonical correlation analysis (Wilks,
1995). Both procedures consist of a projection of a multidimensional space
(TBs) in a new set of coordinates, and, while the principal components
(resulting from an EOF analysis) are ordered by the increasing variance, the
canonical variables (resulting from a CCA) are ordered by the correlation
with a third variable (in our case the logarithm of the surface rainfall
rate). We have calculated seven candidate discriminant functions and tested
them in a simple discriminant analysis test. The test consisted in applying
three multivariate statistical procedures either on the part of the data sets
(SSMIS and AMSU/MHS) with precipitating pixels (“rain”) or on the part of
the data sets with non-precipitating pixels (“no rain”) (“rain” or “no
rain” pixel definitions are based on the TRMM-PR 2A25 product). The first
procedure consisted of applying an EOF analysis on the TBs corresponding to
“no rain” pixels, with the resulting combination of channels hereafter
called PCAN. Once the principal components were calculated we
discarded the first components (linked mostly to the signal deriving from the
surface), in order to obtain a set of independent components (combination of
TBs) with minimum signal coming from the surface. A second procedure
consisted of applying an EOF analysis on the TBs corresponding to “rain”
pixels, obtaining a combination of channels hereafter called
PCAR, and discarding the last components. A third procedure
consisted of calculating the CCA on the “rain” fraction of the data set in
order to select the linear combination of TBs with the maximum correlation to
rainfall (CCAR). We have made comparisons with the results
obtained performing the CCA with respect to log(R) (where R is the
rainfall rate) and to R, finding better performances for log(R).
Therefore we have used the logarithm of rainfall rate as a reference variable
for CCAR. The objective of the last two procedures was to enhance
the signal emitted by precipitating clouds. In the work of Petty (2013), two
of these procedures (PCAN and PCAR) were applied to a
data set of TBs in order to enhance the signal from precipitation and to mask
the signal coming from the surface background. In this study we have tested
seven candidate discriminate functions consisting of different combinations
of the three procedures described above (PCAN, PCAR
and CCAR). Table 1 shows a synthetic description of the
seven candidate discriminant functions.
Results of the choice of the discriminant function over different
background surface for the SSMIS data set: the labels of the x axis identify
the discriminant function used (see text for details) and the value
represented in the y axis is the Heidke skill score.
The ability of each of the seven functions to discriminate the “rain” and “no
rain” pixels has been tested with a simple linear discriminant analysis
(Wilks, 1995). First, the SSMIS data set has been considered with a PR-based
rain/no-rain threshold of 0.1 mmh-1 for each surface class. Then
the same analysis was repeated for the AMSU/MHS data set (for each surface
class). The choice of 0.1 mmh-1 threshold for TRMM-PR was based
on the evidence that the minimum detectable rain rate of this instrument,
estimated by Kirstetter et al. (2014) as 0.53 mmh-1 at PR
resolution, corresponds to the peak of the probability density function (pdf)
of the rainfall rates in our data set at full resolution (not shown). By
downscaling the PR rainfall rates to the SSMIS or MHS nominal resolution, the
peak of the pdf of the rainfall rates shifts to lower values nearly equal to
0.1 mmh-1. The results in terms of Heidke skill score
(which measures the fractional improvements over random chance) for the SSMIS
data set are shown in Fig. 3. The HSS is defined as
HSS=2hc-fmh+mm+c+h+ff+c,
where h, m, f, and c are the fractional hits,
misses, false detections, and correct rejections
in a contingency table, respectively (Wilks, 1995). The HSS falls within a [-∞,+1]
range, where 1 indicates the perfect score.
Looking at Fig. 2 it is clear how PCAN and PCAR have
relatively low skills over all background surfaces except for vegetated land.
However, CCAR, PCAR-CCAR, and
PCAN-PCAR-CCAR have the best skill score
over all surfaces and behave quite similarly. Finally,
PCAN-PCAR and PCAN-CCAR have
a skill variable depending on the surface type. CCAR shows a high
HSS and is a relatively simple procedure, being the result of a canonical
correlation analysis on the TBs associated with rainfall as observed from the
TRMM-PR. Similar results of the HSS were obtained for the AMSU/MHS data set.
From the results of this analysis we have chosen the CCAR as the
discriminant function of the CCA algorithm for detection of precipitation.
The CCA algorithm was trained separately for the two data sets of
coincidences of SSMIS with PR and coincidences of AMSU/MHS with
PR using the full set of 2011–2012 observations. We have also tested how
the detection algorithm would perform with a pseudo-GMI radiometer by
selecting only the channels from the SSMIS data set more similar to those of
the GMI radiometer (i.e., the SSMIS channels in the 50–60 GHz absorption
band and at 183±1 GHz were discarded).
The CCA algorithm has been defined in two steps for each of the three
data sets (SSMIS-PR, AMSU/MHS-PR, and pseudo GMI-PR) and considering the four
different background surface types. First, we have carried out the
CCAR (i.e., on the TBs of “rain” pixels) to find the
coefficients (ai) of the linear combination of TB channels best
correlated to precipitation (canonical variables, CV), defined as
CV=∑i=1naiTBi-TB‾mi,
where the index i spans over the n available channels of the radiometer
(SSMIS, AMSU/MHS, or pseudo-GMI), TBi is the brightness temperature in
each pixel, and TB‾Ri is the mean brightness
temperature over the full “rain” data set. Then, we found the threshold
value of CV to discriminate between “rain” and “no rain” pixels
(CVth). This was computed by analyzing the variability of HSS
over the full data set for different threshold values CVth
(ranging between -2 and 8 K), using the TRMM-PR rainfall product as
“truth” with minimum “rain” threshold of 0.1 mmh-1. The
threshold value CVth, which maximizes the HSS, was selected. This
procedure was repeated for each surface background and radiometer data set.
Therefore, for each surface and for each radiometer a set of coefficients
ai and a threshold value CVth was determined to discriminate
between “rain” and “no rain” pixels.
The CCA algorithm marks those pixels as “rain” where
CV=∑i=1naiTBi-TB‾Ri>CVth.
The resulting coefficient of the CCAR analysis (ai) and the
chosen threshold CVth are therefore defined for each radiometer
(SSMIS, AMSU/MHS, or pseudo-GMI) and for each type of surface background, and
they are summarized in Appendix A.
Results
This section shows the results of the application of the algorithm to the
test data set relative to the year 2013. In the tuning of the algorithms we
classified the precipitating pixels adopting an arbitrary rainfall rate
threshold of 0.1 mmh-1; however, it is possible that in some
conditions the radiometric signal is not suited to detect such light
precipitation. Moreover, an algorithm for the detection of precipitation is
always a compromise between the need for detecting the lower minimum
threshold of rain rate and the requirement of low detection errors in terms
of both false alarms and misses. In this section we analyze the results of
the CCA algorithm for different rainfall rate thresholds (RRth),
using the TRMM-PR 2A25 product as ground truth for “rain” pixels. We
compare the results with those obtained from widely used screening algorithms
(presented in Sect. 1), applying them on the same data sets used for the CCA
algorithm. In particular, we have used four other procedures: (1) the
scattering index (Ferraro, 1997), hereafter F97-SI, i.e., the scattering
index over land and ocean considering also the estimated columnar water vapor
from 19 to 37 GHz over ocean; (2) the polarization-corrected temperature algorithm from Spencer et al. (1989) (hereafter SGH-PCT) in which the PCT is calculated with β=0.45, considering the pixels with PCT <255K as “rain”; (3) the AMSU/MHS screening algorithm from Chen and Staelin (2003),
which uses differences between the 183 GHz channels and the 53.6 GHz and
applicable over each type of surface background considered in this study
(hereafter CS03); and (4) the methodology developed by Grody and Weng (2008)
and employed by many authors (e.g., Laviola and Levizzani, 2009, 2011), which
uses the TB difference between the MHS channels at 89 GHz (or SSMIS and
GMI-like 91.6 GHz) and 150 GHz as a simple mask in order to detect the
scattering signal from precipitation (hereafter GW08). The “rain” threshold
on this TB difference has been set to 5 K. Table 2 summarizes the screening
algorithms used for comparison in this study.
The results are evaluated in terms of HSS, POD,
and FAR, defined as
POD=hh+m;FAR=ff+h,
with the same reference to a contingency table as in Eq. (3).
Algorithms used for comparison in the study. * Used only for
SSMIS and pseudo-GMI data sets.
AcronymBrief descriptionReferenceF97-SI*Scattering index over land and ocean using 19 andFerraro (JGR, 1997)22 GHz and estimated IWV over ocean using 37 GHz.SGH-PCT*Polarization-corrected temperature (PCT) algorithm;Spencer et al. (JAOT, 1989)we have calculated the PCT with β=0.45,considering the pixels with PCT <255 K as “rain”.CS03Considers differences between the 183 GHz channels andChen and Staelin (TGRS, 2003)the 53.6 GHz and is applicable over each type of surfacebackground considered in this study.GW08TB difference between the MHS channel at 89 GHzGrody and Weng (TGRS, 2008)(or 91.6 GHz for SSMIS and pseudo-GMI) and 150 GHzto detect the scattering signal from precipitation.
Comparison of the CCA-SSMIS and CCA-GMI algorithms with other
similar algorithms for the detection of precipitation using the PR rain rate
with a variable threshold (represented in the x axis) as ground truth for
the precipitating pixels. The results are shown in four columns of panels
(one for each surface type) in terms of POD (upper row of panels), FAR
(middle row), and HSS (lower row). The scales are the same for every plot.
Solid lines show the statistical indexes resulting from the application of a
screening algorithm to the test data set (year 2013); red crosses are the
result of the application of the SSMIS-CCA algorithm to the training data set
(years 2011–2012).
Same as Fig. 3 but for the AMSU/MHS data set. The CCA-SSMIS (red
solid curve) (results of the CCA algorithm applied to SSMIS test data set
shown in Fig. 3) are shown here for comparison. Red crosses are the result of
the application of the CCA-AMSU algorithm to the training data set (years
2011–2012).
Discussion of skill scores
Figures 3 and 4 show the results for the SSMIS (and pseudo-GMI) data sets and
for the AMSU/MHS data sets, respectively, for all types of surface background.
All tested algorithms have higher POD for higher RRth because
high rain rates are usually associated with precipitating clouds with
a strong radiometric signal (except in some cases such as warm rain over
land). It is worth noting, however, that the impact on the TBs in the
different microwave channels depends on several factors, such as the surface
background, environmental and meteorological conditions, and the
microphysical structure of the cloud. FAR grows with RRth as well
in all of the algorithms considered, as a consequence of the fact that by
increasing RRth the size of the areas considered as precipitating
by the TRMM-PR is reduced, while the areas identified as “rain” by the
detection schemes are unchanged. The overall performance of the detection
schemes can be evaluated by looking at the HSS (last row in each figure).
In Fig. 3, over every surface the CCA algorithm applied to SSMIS and
pseudo-GMI (CCA-SSMIS and CCA-GMI) shows almost identical scores. Moreover,
the comparison of the scores of CCA-SSMIS on the training data set (red
crosses in Fig. 4) and test data set (continuous red line) shows a good
agreement between the two, a sign of the stability of the algorithm. Over
arid land, vegetated land, and coast the CCA-SSMIS performs better (i.e.,
higher HSS) than the other algorithms especially due to lower FAR. The
SGH-PCT shows a low POD and a low FAR over all surface background types. The
major drawback of the SGH-PCT algorithm is that it needs to use both
polarizations of the SSMIS (or pseudo-GMI) 91 GHz channels, and, therefore,
it is not applicable to AMSU/MHS or other cross-track scanning radiometers.
It is also worth noting that the F97-SI algorithm is not suited for detecting
precipitation over desert (arid land) because the use of the SI leads to the
misclassification of the signal deriving from the surface as precipitation
(in fact it was not applied to desert background in the original work by
Ferraro, 1997). Over ocean, the CCA algorithms have higher POD than the others
(while the FAR is comparable to CS03). Over ocean, CCA-SSMIS and CCA-GMI show
higher HSS than the other algorithms except the F97-SI, which has higher HSS
for RRth larger than 0.3 mmh-1. We can conclude that
over ocean the CCA algorithm is more suitable to detect low precipitation
rates (RRth≤0.2mmh-1) than the F97-SI; however,
F97-SI is preferable to detect higher precipitation rates.
Figure 4 shows that the CCA-AMSU has a behavior similar to the CCA-SSMIS
(CCA-GMI). No results for F97-SI and SGH-PCT are presented in this figure
because these two approaches can not be applied to the AMSU/MHS radiometers,
since the 19 GHz channel and the two polarization of the 89 GHz are not
available. Over arid and vegetated land, the CCA-AMSU performs better than the
other algorithms (higher HSS due to the significantly lower FAR). Over
coast,
the CCA-AMSU has skill scores very similar to the other algorithms. Finally,
over ocean the CS03 and GW08 seem to work better (in terms of HSS) for high
values of RRth.
A complete and synthetic representation of the skill scores for all
algorithms but for one rain rate threshold (0.1 mmh-1) is
provided in Fig. 5 (CCA-GMI results have been omitted because they are identical to
CCA-SSMIS scores). From this figure it is evident how the CCA algorithm, both
for AMSU/MHS and for SSMIS, performs well in terms of HSS with respect to the
other tested algorithms and how this is due to the higher POD and to the
lower FAR.
Minimum detectable rate
This section is dedicated to defining the minimum detectable rainfall rate
for the CCA algorithm for each surface type and sensor and to computing the
statistical scores of the algorithm for these thresholds. Figure 6 shows the
results of a binning technique (following Ferraro and Marks, 1995) applied to
the TRMM-PR rainfall rate corresponding to the CV values (Eq. 4) of each
pixel. For each data set (i.e., for SSMIS, AMSU/MHS and pseudo-GMI and for
four different types of surface) the CV range of values has been divided into
bins 0.2 K wide, and the mean and the standard deviation of the TRMM-PR
rainfall rate corresponding to each CV value has been calculated for each
bin. Each panel shows the trend of the mean rainfall rate (and standard
deviation) with the CV binned values. A thick vertical dashed line represents
the CV threshold chosen for each radiometer and surface background type as a
result of the CCA training (i.e, CVth in Eq. 5, Sect. 3; the
values are also provided in Appendix A).
Skill score comparison of the CCA algorithm with other precipitation
detection algorithms with rain/no-rain threshold (“truth” from PR 2A25)
equal to 0.1 mmh-1.
Binning analysis of the rain rate intensity against CV values: mean
(continuous black line) and standard deviation (error bars) of rain rates
inside an interval of CV bins (as large as 0.2 K) are shown. The vertical
dashed line represents the CV threshold (CVth) chosen for each combination of
radiometer and surface background.
Value of the mean rain rate corresponding to the CCA threshold set
for every radiometer and surface background combination. Values in bold refer
to the test data set while the italics values refer to the training data set.
Looking at the trend of mean rain rate with CV and at the rain rate standard
deviation it is evident that increasing values of CV are on average
associated with increasing values of rainfall rate. All the pixels falling in
the bins below CVth are misclassified as “no rain” and they
correspond to the low rainfall values (below 0.5 mmh-1 for all
the data sets). We have considered the minimum detectable rainfall rate for
each data set (RRb) as the mean value corresponding to the
CVth. In Table 3 the values of RRb are reported and the
scores of the CCA algorithm based on these thresholds are also provided. It
is clear from Fig. 6 and Table 3 how the CCA performs better over vegetated
land and over ocean for both SSMIS and AMSU, and RRb is lower over ocean
(less than 0.2 mmh-1) and higher over land (around
0.3 mmh-1). Over arid land it is possible to discriminate lower
values of rain intensity (around 0.25 mmh-1) but the strong
variability in the surface emissivity leads to a higher ratio of misses
(lower POD). Over coast the results are quite different between the SSMIS and
the AMSU algorithms: the RRb over coast for the SSMIS is the highest
among all the data sets (0.41 mmh-1), while the irregular trend
of the mean rain rate over coast for the AMSU/MHS makes the result for
RRb over coast (0.24 mmh-1) uncertain.
Regarding the pseudo-GMI data set, it is worth noting that it was generated
from the SSMIS data set only by discarding the channels not available in the
GMI radiometer (183±1 and the 50–60 GHz band channels). Therefore, the
results are only to a certain extent representative of the performance of a CCA algorithm for the
real GMI radiometer, since we have considered only
the instrument observation geometry (conical scanning) and some of the
channels frequencies and polarizations. The presence of the 10 and 166 GHz
channels with both polarizations (not available on SSMIS) might have a
significant impact on precipitation detection. Moreover, the differences in
the resolution of the GMI and SSMIS sensors have not been considered, and
this may strongly affect the results on real GMI data. It is worth noting
that the almost identical results obtained for the SSMIS and pseudo-GMI
data sets indicate that the inclusion of the sounding channels 183±1 and
50–60 GHz has no impact on the precipitation detection.
Dependence on precipitation regime
In the following subsection we have analyzed the CCA algorithm scores using
the rainType flag as estimated by TRMM product 2A23 (please refer to TRMM,
2011, for details on the rainType flag definitions in the TRMM-PR 2A23
product). Each PR pixel has been classified into four main rain categories as
stratiform (rainType ≥100 and <200)
convective (rainType ≥200 and <250)
shallow convective (rainType ≥250 and <300)
other (rainType ≥300).
In the test data sets this information was downscaled to the SSMIS and MHS
nominal resolutions by calculating the fraction of PR precipitating pixels
corresponding to each rain category within each SSMIS and AMSU/MHS pixel.
Each pixel in the SSMIS-PR and AMSU/MHS-PR data sets has been classified as
mainly convective if at least 50 % of the precipitating part of the pixel
was convective. The same criterion was adopted for the other rain
categories. If a pixel was classified as precipitating with no predominance
of any of the main rain type class, it was flagged as mixed.
Table 4 shows the POD score for each rain type class and for each surface
background. Moreover, Table 5 shows the POD calculated as a function of the
fraction of precipitation in the pixel (for all surface background classes
together). It is clear from Tables 4 to 5 that the CCA algorithm detects convective and stratiform precipitation very
well, while it is almost insensitive
to shallow precipitation over land and coast and gives medium performances
(between 0.30 and 0.58) for other and mixed precipitation types. However,
from Table 5 it is evident that the detection capability of precipitation
grows with the fraction of precipitation within the SSMIS and MHS pixels.
These results indicate that the non-uniform beam filling effect can have a
significant impact on the detection of the precipitation, and the impact
depends on the predominant type of precipitation within each pixel (i.e, the
most significant impact for shallow precipitation, the least significant for
stratiform precipitation, and quite significant for convective
precipitation).
Probability of detection scores classified by surface background and
rain type.
POD ConvectiveStratiformShallowOtherMixedSSMISArid land0.610.68–0.300.33Veg. land0.760.840.080.550.48Coast0.680.820.090.390.36Ocean0.840.930.420.480.55AMSU/MHSArid land0.73*0.77–0.44*0.33*Veg. land0.760.840.07*0.450.47Coast0.750.870.090.300.38Ocean0.890.960.400.380.58
* sample 100–500
pixels.
Probability of detection scores classified by rain type and
percentage of precipitation in the pixel. No results are provided for samples
with less than 100 pixels.
POD with fraction of precipitation FractionConvectiveStratiformShallowOtherMixedof precip.conv.SSMIS>0.10.750.870.380.470.53>0.20.770.880.440.470.53>0.30.790.880.470.470.57>0.40.810.890.490.490.62>0.50.820.910.510.560.69>0.60.840.920.540.590.71>0.70.840.930.610.570.81>0.80.880.950.680.630.90>0.90.890.960.710.700.92AMSU/MHS>0.10.790.890.280.400.52>0.20.810.900.320.410.57>0.30.840.910.360.450.64>0.40.870.930.440.570.70>0.50.910.950.570.730.81>0.60.930.960.590.810.88>0.70.970.970.70*0.840.94>0.80.990.980.75*0.88*0.98>0.91.000.99––1.00
Table 6 shows the results for the false alarms considering the actual
potential presence of rain in the AMSU/MHS and SSMIS pixel based on the
downscaled PR rain/no-rain flag values (2A23 rainFlag). The first column
shows the FAR scores for each surface type, while the other columns show the
percentage of false alarms counts (for each surface type) for three different
classes of rain/no-rain conditions. The three classes are as follows.
Low rain: SSMIS or AMSU/MHS pixels classified as “rain certain” (rainFlag
equals to 20). For this class, CCA false alarms refer to PR downscaled rain
rates less than 0.1 mmh-1.
Potential rain: SSMIS or AMSU/MHS pixels classified as “rain possible”
or “probable” (rainFlag between 10 and 15). For this class, false alarms
refer to PR downscaled rain rates equal to 0 mmh-1. This class
includes very weak echoes (possibly noise) and non-precipitating clouds.
No rain: SSMIS or AMSU/MHS pixels classified as “no rain” ” (rainFlag
equals to 0).
It is clear from Table 6 that only a relatively small percentage of the false
alarms are associated with “no rain” pixels (especially for AMSU/MHS ranging
between 3 and 6 %), while a quite significant number of the false alarms
are
associated with “low rain” PR observations with low rain rates (for the
SSMIS 27–30 % and for AMSU/MHS 39–60 %) .
False alarms ratio scores classified by surface and rain flag.
Percentages represent the fraction of false alarm counts for a given surface
type and rainFlag w.r.t the total number of false alarms for each surface
type.
The CCA algorithm for the detection of precipitation described
in this paper results from the application of two main procedures:
(1) the application of the canonical correlation analysis to a large data set of
coincident PMW multichannel observations and precipitation measurements or
estimates, which can be considered as “truth” to define canonical variables
for different types of surface background; (2) the estimation of a threshold
value for the canonical variables that maximizes the Heidke skill score for a
given rainfall rate threshold. The algorithm has been applied to three large
data sets built from coincident SSMIS and AMSU/MHS measurements and TRMM-PR
rainfall products (SSMIS-PR, AMSU/MHS-PR, and pseudo-GMI-PR). Four different
types of surface background have been considered, and the results have been
compared to other well-known screening algorithms. The resulting CCA
algorithm is simple, and it can be adapted to any PMW radiometer and to any
geographical region where a large data set of coincident precipitation
measurements and PMW observations from that radiometer is available.
The CCA algorithm almost always shows better performance in comparison to
other well-known algorithms, especially in terms of low false-alarm ratios.
The HSS is always higher than all other algorithms tested, except for AMSU/MHS
over ocean for high rainfall rate thresholds > 0.7 mmh-1 and
SSMIS over ocean for rainfall rate thresholds
> 0.3 mmh-1. The estimate of the minimum rain rate that is
efficiently detected by the algorithm shows values varying from
0.14 mmh-1 (AMSU/MHS over ocean) to 0.41 (SSMIS over coast) with
the remarkable result of 0.25 mmh-1 over arid land surface. It
is worth noting that the pseudo-GMI data set was generated by selecting the
channels available in the GMI radiometer from the SSMIS data set. Therefore,
the results from the pseudo-GMI data set need to be looked at with some
caution. The geometry of observation of the two instruments is very similar
and the polarization and frequency of the selected channels almost coincide.
However, the presence of the 10 and 166 GHz channels with both polarizations
(not available on SSMIS) might have a significant impact on precipitation
detection. In this study the almost identical results obtained for the SSMIS
and pseudo-GMI data sets are more related to the significance of the oxygen
band absorption channels for light precipitation detection. Our results
indicate that the inclusion of the sounding channels 183±1 and
50–60 GHz has almost no impact on precipitation detection.
An analysis of the results based on different precipitation regimes (as
identified by the TRMM-PR) has shown how the CCA algorithm has very high
detection capability for convective and stratiform precipitation. Shallow
convective precipitation detection efficiency strongly depends on the surface
background as CCA is almost insensitive to shallow convective rain over coast
and land and gives moderate performances over ocean. Moreover, the results
show that the non-uniform beam filling effect can have a significant impact on
the detection of precipitation, and the impact depends on the precipitation
regime, i.e, highest for shallow precipitation, lowest for stratiform
precipitation, and quite significant for convective precipitation.
The CCA algorithm for the different radiometers (i.e., CCA-AMSU, and
CCA-SSMIS) seems to perform well, with a good detection capability and low
false alarms ratios. It is worth noting that only a relatively small
percentage of the CCA false alarms are related to totally rain-free pixels,
while a quite significant contribution comes from pixels with very low rain
rates, i.e., downscaled PR rain rates lower than the minimum rain/no-rain
threshold used (0.1 mmh-1).
The advent of the GPM era requires the combined use of different PMW
radiometers with different channels and dissimilar observation geometry for
global precipitation monitoring. This poses a difficult challenge to the
scientific community, i.e., the obtaining of coherent estimates of precipitation
from this constellation of radiometers. Much effort is put into achieving
consistency between precipitation pattern and precipitation estimates from
the different sensors. Some fundamental improvements in this direction will
come from the use of common procedures applicable to all radiometers for the
detection (and the retrieval) of precipitation, applicable to all types of
background surfaces. The CCA algorithm is an important step toward this goal,
considering it is suitable for application to any PMW sensor (conical and
cross-track scanning) for which a long time series of data coincidences with
rainfall rate considered as ground truth is available.
The results show a certain level of consistency between the detection
capability of CCA-SSMIS and CCA-AMSU algorithms. It is worth noting that by
using different thresholds and linear combinations of channels for each
sensor, the algorithm optimally exploits the characteristics of each sensor.
However, the available channels, polarization information, spatial
resolutions, and observation geometry, which differ from radiometer to
radiometer, introduce fundamental variations in the precipitation detection
capabilities of each sensor and pose intrinsic limitations to the level of
achievable consistency. In the EUMETSAT H-SAF project, the consistency of the
precipitation estimates between cross-track and conical scanning radiometers
has been strongly improved by the use a common procedure for the detection of
the precipitating clouds (besides the use of the same physical foundation in
the retrieval algorithms; see Mugnai et al., 2013b; Panegrossi et al., 2013).
In the near future we plan to develop the CCA algorithm for GMI, GCOM-W1
AMSR2, and Suomi NPP ATMS. We are planning to take advantage of the imminent
availability of the GPM DPR products in order to extend and test the CCA
procedure between 65∘ N and 65∘ S, especially over snow- and
ice-covered surfaces. Moreover, the estimates of the CloudSat Profiling Radar
may be used as ground truth (especially for snowfall and light precipitation)
in order to create an even larger data set of coincident observations of
active and passive MW satellite-borne instruments and extend the CCA
algorithm to higher latitudes and to the polar regions.
Tables A1–3 show the CV thresholds (CVth), the coefficients,
and mean TB values used by the CCA algorithms for the SSMIS, pseudo-GMI, and
AMSU/MHS data sets. The mean TBs in the pseudo-GMI data set are identical to
the corresponding channels of the SSMIS data set.
List of CCA coefficients and CVth thresholds used by the CCA
algorithms for the SSMIS radiometer.
SSMIS Arid land Vegetated land Coast Ocean CVth2.4 K 0.6 K 1.1 K 1.1 K Ch. #Ch. nameaTB‾RaTB‾RaTB‾RaTB‾R(GHz)1150-0.07274.86-0.08277.65-0.05277.39-0.01277.492183±6.6-0.11272.51-0.02267.49-0.04270.660.00271.993183±30.01263.310.03259.860.03262.60-0.00263.734183±10.04249.90-0.01247.080.00249.520.00249.70591.6 V-0.04280.53-0.01280.98-0.04274.120.04266.26691.6 H0.07268.410.06278.170.04258.09-0.03240.88719 H0.03260.610.03278.43-0.02210.99-0.05144.39819 V0.01290.690.02284.46-0.04246.960.01205.56922.2 V0.02290.240.03285.600.05263.45-0.01237.661037 H-0.02262.77-0.03276.900.03219.560.08164.451137 V-0.03285.72-0.02281.13-0.02251.850.10220.261250.30.02275.23-0.04275.60-0.00262.59-0.01249.791352.80.02265.73-0.01265.05-0.00264.08-0.02262.261453.60.03249.900.01249.960.01249.91-0.02249.401554.40.00217.640.01217.720.01216.620.00217.051655.50.02211.21-0.01210.19-0.00210.870.00211.17
List of CCA coefficient and CVth threshold
used by the CCA algorithms for the pseudo-GMI data set.
List of CCA coefficient and CVth threshold used by the CCA
algorithms for the AMSU/MHS data set.
AMSU A-MHS Arid land Vegetated land Coast Ocean CVth2.3 K 0.6 K 0.9 K 1.0 K Ch. #Ch. name (GHz)aTB‾RaTB‾RaTB‾RaTB‾R1890.04284.020.06285.490.04263.600.08239.742150-0.06285.12-0.05284.45-0.08282.53-0.04279.563183.3±10.03253.96-0.03250.970.01253.00-0.02254.314183.3±30.05267.250.08263.620.02265.740.05267.375183.3±7-0.17278.89-0.11273.11-0.05275.31-0.01276.41623.80.01285.720.04286.060.06239.15-0.04187.14731.4-0.02284.17-0.00283.86-0.06226.900.10166.38850.30.01284.84-0.06283.490.01258.43-0.02231.73952.80.02276.40-0.04274.62-0.00269.23-0.08262.651053.60.04261.160.03260.27-0.00259.34-0.05257.621154.40.04240.450.05240.38-0.00240.430.02239.961254.90.02230.140.04229.96-0.01230.110.06229.861355.5-0.03216.91-0.03216.33-0.02216.600.13216.54Acknowledgements
The authors would like to thank NOAA (www.nsof.class.noaa.gov) for
providing the AMSU-A/MHS and SSMIS radiometers data and NASA
(trmm.gsfc.nasa.gov) for providing the TRMM data. The authors express
appreciation to Stephen Joseph Munchak and two anonymous reviewers for
helpful comments and suggestions that have contributed to significantly
improve the manuscript. This research was supported by EUMETSAT through the
project “Satellite Application Facility on Support to Operational Hydrology
and Water Management” (H-SAF).
Edited by: S. J. Munchak
References
Barrett, E. C., Kidd, C., and Bailey, J. O.: A new instrument with rainfall
monitoring potential, Int. J. Remote Sens., 9, 1943–1950, 1988.
Bennartz, R.: Optimal convolution of amsu-b to amsu-a, J. Atmos. Ocean.
Tech., 17, 1215–1225, 2000.
Casella, D., Panegrossi, G., Sano, P., Dietrich, S., Mugnai, A., Smith,
E. A., Tripoli, G. J., Formenton, M., Di Paola, F., and Leung, W.-Y.:
Transitioning from CRD to CDRD in Bayesian retrieval of rainfall from
satellite passive microwave measurements: Part 2. Overcoming database profile
selection ambiguity by consideration of meteorological control on
microphysics, IEEE T. Geosci. Remote, 51, 4650–4671, 2013.
Chen, F. W. and Staelin, D. H.: Airs/amsu/hsb precipitation estimates,
IEEE T. Geosci. Remote, 41, 410–417, 2003.Desbois, M., Roca, R., Eymar L., Viltard, N., Viollier, M., Srinivasan, J.,
and Narayanan, S.: The Megha-Tropiques mission, in: Atmospheric and Oceanic
Processes, Dynamics, and Climate Change, edited by: Sun, Z., Jin, F.-F., and
Iwasaki, T., International Society for Optical Engineering, SPIE P., 4899,
172–183, 10.1117/12.466703, 2003
Edward, P. and Pawlak, D.: MetOp: The space segment for EUMETSAT's polar
system, ESA Bulletin, 102, 6–18, 2000.Ferraro, R. R.: Special sensor microwave imager derived global rainfall
estimates for climatological applications, J. Geophys. Res., 102,
16715–16735, 10.1029/97jd01210, 1997.
Ferraro, R. R. and Marks, G. F.: The development of SSM/I rain-rate retrieval
algorithms using ground-based radar measurements, J. Atmos. Ocean. Tech., 12,
755–770, 1995.GPM: Algorithm Theoretical Basis Document, NASA, available at:
http://pmm.nasa.gov/sites/default/files/document_files/GPROF_ATBD_1Dec2010.pdf,
last access: 7 September 2014, 2010.
Grecu, M. and Anagnostou, E. N.: Overland precipitation estimation from TRMM
passive microwave observations, J. Appl. Meteorol., 40, 1367–1380, 2001.Grody, N. C.: Classification of snow cover and precipitation using the
Special Sensor Microwave Imager, J. Geophys. Res., 96, 7423–7435,
10.1029/91JD00045, 1991.
Grody, N. C. and Weng, F.: Microwave emission and scattering from deserts:
theory compared with satellite measurements, IEEE T. Geosci. Remote, 46,
361–375, 2008.Hou, A. Y., Kakar, R. K., Neeck, S., Azarbarzin, A. A., Kummerow, C. D.,
Kojima, M., Oki, R., Nakamura, K., and Iguchi, T.: The global precipitation
measurement mission, B. Am. Meteorol. Soc., 95, 701–722,
10.1175/BAMS-D-13-00164.1, 2014
Islam, T., Rico-Ramirez, M. A., Srivastava, P. K., and Dai, Q.:
Non-parametric rain/no rain screening method for satellite-borne passive
microwave radiometers at 19–85 GHz channels with the Random Forests
algorithm, Int. J. Remote Sens., 35, 3254–3267, 2014.
Kidd, C.: On rainfall retrieval using polarization-corrected temperatures,
Int. J. Remote Sens., 19, 981–996, 1998.Kirstetter, P. E., Hong, Y., Gourley, J., Schwaller, M., Petersen, W., and
Cao, Q.: Impact of sub-pixel rainfall variability on spaceborne precipitation
estimation: evaluating the TRMM 2A25 product, Q. J. Roy. Meteor. Soc.,
10.1002/qj.2416, online first, 2014.
Kummerow, C. D., Hong, Y., Olson, W., Yang, S., Adler, R., McCollum, J.,
Ferraro, R., Petty, G., Shin, D.-B., and Wilheit, T.: The evolution of the
Goddard profiling algorithm (GPROF) for rainfall estimation from passive
microwave sensors, J. Appl. Meteorol., 40, 1801–1820, 2001.
Kummerow, C. D., Ringerud, S., Crook, J., Randel, D., and Berg, W.: An
observationally generated a priori database for microwave rainfall
retrievals, J. Atmos. Ocean. Tech., 28, 113–130, 2011.Kunkee, D. B., Poe, G. A., Boucher, D. J., Swadley, S. D., Hong, Y., Wessel,
J. E., and Uliana, E. A.: Design and evaluation of the first special Sensor
Microwave Imager/Sounder, IEEE T. Geosci. Remote, 46, 863–883,
10.1109/tgrs.2008.917980, 2008.
Laviola, S. and Levizzani, V.: Observing precipitation by means of water
vapor absor tion lines: a first check of the retrieval capabilities of p the
183-WSL rain retrieval method, Riv. Ital. Telerilevam., 41, 39–49, 2009.
Laviola, S. and Levizzani, V.: The 183-WSL fast rain rate retrieval
algorithm: Part I: Retrieval design, Atmos. Res., 99, 443–461, 2011.Mugnai, A., Casella, D., Cattani, E., Dietrich, S., Laviola, S., Levizzani,
V., Panegrossi, G., Petracca, M., Sanò, P., Di Paola, F., Biron, D.,
De Leonibus, L., Melfi, D., Rosci, P., Vocino, A., Zauli, F., Pagliara, P.,
Puca, S., Rinollo, A., Milani, L., Porcù, F., and Gattari, F.:
Precipitation products from the hydrology SAF, Nat. Hazards Earth Syst. Sci.,
13, 1959–1981, 10.5194/nhess-13-1959-2013, 2013a.Mugnai, A., Smith, E. A., Tripoli, G. J., Bizzarri, B., Casella, D.,
Dietrich, S., Di Paola, F., Panegrossi, G., and Sanò, P.: CDRD and PNPR
satellite passive microwave precipitation retrieval algorithms:
EuroTRMM/EURAINSAT origins and H-SAF operations, Nat. Hazards Earth Syst.
Sci., 13, 887–912, 10.5194/nhess-13-887-2013, 2013b.
Munchak, S. J. and Skofronick-Jackson, G.: Evaluation of precipitation
detection over various surfaces from passive microwave imagers and sounders,
Atmos. Res., 131, 81–94, 2013.Muth, C., Webb, W. A., Atwood, W., and Lee, P.: Advanced technology microwave
sounder on the National Polar-Orbiting Operational Environmental Satellite
System, in: Geoscience and Remote Sensing Symposium, 25–29 July 2005. IGARSS
'05. Proceedings. 2005 IEEE International, 1, 4 pp.,
10.1109/IGARSS.2005.1526113, 2005.
Panegrossi, G., Casella, D., Dietrich, S., Sanò, P., Petracca, M., and
Mugnai, A.: A verification study over Europe of AMSU-A/MHS and SSMIS passive
microwave precipitation retrievals, Proc. 2013 Joint EUMETSAT/AMS
Meteorological Satellite Conference, 8 pp., 2013.Petty, G. W.: Dimensionality reduction in Bayesian estimation algorithms,
Atmos. Meas. Tech., 6, 2267–2276, 10.5194/amt-6-2267-2013, 2013.Petty, G. W. and Li, K.: Improved Passive Microwave Retrievals of Rain Rate
over Land and Ocean. Part I: Algorithm Description, J. Atmos. Ocean. Tech.,
30, 2493–2508, 10.1175/jtech-d-12-00144.1, 2013a.Petty, G. W. and Li, K.: Improved Passive Microwave Retrievals of Rain Rate
over Land and Ocean. Part II: Validation and Intercomparison, J. Atmos.
Ocean. Tech., 30, 2509–2526, 10.1175/jtech-d-12-00184.1, 2013b.
Sanò, P., Casella, D., Mugnai, A., Schiavon, G., Smith, E. A., and
Tripoli, G. J.: Transitioning from CRD to CDRD in Bayesian retrieval of
rainfall from satellite passive microwave measurements: Part 1. Algorithm
description and testing, IEEE T. Geosci. Remote, 51, 4119–4143, 2013.Sanò, P., Panegrossi, G., Casella, D., Di Paola, F., Milani, L., Mugnai,
A., Petracca, M., and Dietrich, S.: The Passive microwave Neural network
Precipitation Retrieval (PNPR) algorithm for AMSU/MHS observations:
description and application to European case studies, Atmos. Meas. Tech., 8,
837–857, 10.5194/amt-8-837-2015, 2015.
Seto, S., Takahashi, N., and Iguchi, T.: Rain/no-rain classification methods
for microwave radiometer observations over land using statistical information
for brightness temperatures under no-rain conditions, J. Appl. Meteorol., 44,
1243–1259, 10.1175/JAM2263.1, 2005.Shimoda, H.: GCOM missions, Geoscience and Remote Sensing Symposium, 2005.
IGARSS '05. Proceedings. 2005 IEEE International, 25–29 July, Seoul, South
Korea, 6, 4201–4204, 10.1109/IGARSS.2005.1525844, 2005.
Spencer, R. W., Goodman, H. M., and Hood, R. E.: Precipitation retrieval over
land and ocean with the SSM/I: identification and characteristics of the
scattering signal, J. Atmos. Ocean. Tech., 6, 254–273, 1989.TRMM: Tropical Rainfall Measuring Mission (TRMM) Precipitation Radar
Algorithm – Instruction Manual For Version 7, JAXA-NASA,
www.eorc.jaxa.jp/TRMM/documents/PR_algorithm_product_information/pr_manual/PR_Instruction_Manual_V7_L1.pdf,
last access: March 2015, 2011.
Wilheit, T. T.: Comparing calibrations of similar conically scanning
window-channel microwave radiometers, IEEE T. Geosci. Remote, 51, 1453–1464,
2013.
Wilks, D. S.: Statistical methods in the atmospheric sciences: An
introduction, Academic Press, San Diego, 1995.