You are on page 1of 12

Atmospheric Research 145–146 (2014) 57–68

Contents lists available at ScienceDirect

Atmospheric Research
journal homepage:

An exploratory investigation of an adaptive neuro fuzzy

inference system (ANFIS) for estimating hydrometeors
from TRMM/TMI in synergy with TRMM/PR
Tanvir Islam a,b,c,⁎, Prashant K. Srivastava c,d,e, Miguel A. Rico-Ramirez c, Qiang Dai c,
Dawei Han c, Manika Gupta f
NOAA/NESDIS Center for Satellite Applications and Research, College Park, MD, USA
Cooperative Institute for Research in the Atmosphere, Colorado State University, Fort Collins, CO, USA
Department of Civil Engineering, University of Bristol, Bristol, UK
NASA Goddard Space Flight Center, Greenbelt, MD, USA
Earth System Science Interdisciplinary Center, University of Maryland, College Park, MD, USA
Department of Civil Engineering, Indian Institute of Technology Delhi, New Delhi, India

a r t i c l e i n f o a b s t r a c t

Article history: The authors have investigated an adaptive neuro fuzzy inference system (ANFIS) for the
Received 19 December 2013 estimation of hydrometeors from the TRMM microwave imager (TMI). The proposed
Received in revised form 17 March 2014 algorithm, named as Hydro-Rain algorithm, is developed in synergy with the TRMM
Accepted 18 March 2014 precipitation radar (PR) observed hydrometeor information. The method retrieves rain rates
Available online 27 March 2014
by exploiting the synergistic relations between the TMI and PR observations in twofold steps.
First, the fundamental hydrometeor parameters, liquid water path (LWP) and ice water path
Keywords: (IWP), are estimated from the TMI brightness temperatures. Next, the rain rates are estimated
Liquid water contents from the retrieved hydrometeor parameters (LWP and IWP). A comparison of the
Ice water contents
hydrometeor retrievals by the Hydro-Rain algorithm is done with the TRMM PR 2A25 and
Global precipitation measurement (GPM)
GPROF 2A12 algorithms. The results reveal that the Hydro-Rain algorithm has good skills in
Rain rate retrieval
Passive microwave radiometry estimating hydrometeor paths LWP and IWP, as well as surface rain rate. An examination of
Precipitation radar the Hydro-Rain algorithm is also conducted on a super typhoon case, in which the Hydro-Rain
has shown very good performance in reproducing the typhoon field. Nevertheless, the passive
microwave based estimate of hydrometeors appears to suffer in high rain rate regimes, and as
the rain rate increases, the discrepancies with hydrometeor estimates tend to increase as well.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction radiation and retaining the temperature to comfort level. The

atmosphere is mainly composed of nitrogen and oxygen gasses,
The earth is surrounded by the “atmosphere”, which is a covering 99% of the atmospheric constituents. However, the
blanket of air containing various gasses that act as a protective remaining 1%, a mixture of carbon dioxide, water vapor, ozone
shield for the earth's life to survive by absorbing ultraviolet and other gasses, is responsible for producing hydrometeors. In
fact, the hydrometeor related parameters (e.g. cloud, ice, and
rain water contents) are one of the most critical drivers for
⁎ Corresponding author at: NOAA/NESDIS/STAR, NOAA Center for Weather atmospheric radiative effects (Brandau et al., 2010; Du et al.,
and Climate Prediction, 5830 University Research Ct., College Park, MD
20740-3818, USA.
2011; Shen et al., 2011; Zhou and Ran, 2012). The accurate
E-mail addresses:, simulation and prediction of geophysical parameters through
(T. Islam). numerical weather prediction and climate models also require
0169-8095/© 2014 Elsevier B.V. All rights reserved.
58 T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68

accurate information of hydrometeor variables (Hazra et al., scans ±17° of vertical at intervals of 0.35° providing data over a
2013; Michel et al., 2011; Van Weverberg et al., 2012). swath 215 km wide at the earth's surface with a horizontal
The active microwave sensors such as radars are capable of footprint of about 4 km and a vertical resolution of 250 m at
estimating hydrometeors more accurately than passive micro- nadir. Detail description of the TRMM sensor package address-
wave sensors. In particular, the active sensors at microwave ing the technical specifications of the TMI and PR instruments
(and millimeter wave) frequencies measure the radar reflectiv- can be found in Kummerow et al. (1998).
ity factor Z, which is proportional to the sixth moment of In this work, the raw calibrated brightness temperatures
raindrop size distribution (DSD) providing that the raindrops are taken from TRMM 1B11 product. The PR hydrometeor
are small compared to the instrument's wavelength (Islam et al., variables are derived from the TRMM 2A25 product. In
2012b; Islam et al., 2012d). This Z factor can be related to deriving the hydrometeor variables LWP and IWP, the
the liquid/ice water contents in each atmospheric layer, and methodology mentioned in Masunaga et al. (2002) and
ultimately can provide vertically integrated liquid water path Iguchi et al. (2000) has been employed. The Z at each PR
(LWP) and ice water path (IWP) information. In contrary, the cell has been converted to liquid water content and ice water
passive microwave (PMW) sensors measure the integrated content based on the Z–M relationship:
effects of liquid and ice in the atmospheric column through
brightness temperature observations from emission and scat- M ¼ aZ ð1Þ
tering channels (Islam et al., 2012a; Spencer et al., 1989; You
and Liu, 2012). Therefore, accurate hydrometeor profiling in a where, M is the water/ice content for a given range in gm−3
vertical column is fairly limited. Further, this could lead to and Z is the attenuation corrected reflectivity factor in mm6m−3.
inaccurate estimate of LWP/IWP, and integrally, total water The coefficients a and b are dependent upon particle temper-
path (TWP) information. Commonly, PMW measurements can ature and the type of precipitation (e.g. stratiform/convective).
penetrate the thin clouds, but not the thick clouds. This may lead In 2A25, the coefficients are stored for different nodes based on
to the underestimation of column water contents (Crewell et al., the scenarios for a particular profile (e.g. with/without bright
2009; Saavedra et al., 2012). band). Fig. 1 represents a typical schematic illustration of the
In this study, we stress out the aforementioned limita- nodes for which the coefficients are given. In order to compute
tions and propose an empirical approach, as an alternative to the water/ice content for each range gate, the coefficients
traditional hydrometeor estimation from PMW observations. between the nodes are obtained through linear interpolation.
The algorithm, hereinafter named as Hydro-Rain algorithm, is Once, the liquid and ice water contents are computed, the LWP
designed to produce active sensor like retrieval of hydrome- and IWP are obtained by simply vertically integrating the liquid
teor variables LWP and IWP, and in turn, the hydrometeor water and ice water contents, respectively. For LWP, the water
variables are used to produce instantaneous rain rates. The content from surface to freezing level height, and for IWP, the
Hydro-Rain algorithm is based on an adaptive neuro fuzzy ice content from freezing level height to the top of the storm are
inference system (ANFIS). The active microwave sensor used considered. Besides, the surface rain rate has been taken from
in this work is the TRMM precipitation radar (PR) and the the TRMM 2A25 product. Note that the surface rain rate in 2A25
passive microwave sensor is the TRMM microwave imager is defined as the rain rate at the lowest point in the clutter-free
(TMI). The key advantage is that both active and passive region (Islam et al., 2012c).
sensors are in the same platform, measuring coincident In addition, for the sake of comparison, hydrometeor
observations. Therefore, development of an ANFIS model for estimates produced by the TMI GPROF algorithm are used,
the TMI observations in synergy with the PR measurements taken from TRMM 2A12 product. In GPROF, the hydrometeor
has been possible. related variables are retrieved by relating the radiometric
The remainder of this manuscript is organized as follows. brightness temperatures to the cloud resolving model produc-
Section 2 describes the datasets used in this study. In ing an offline database of hydrometeor parameters by using a
Section 3, the limitations of PMW sensors for hydrometeors Bayesian inversion scheme. Similar strategy is conducted for
estimate are discussed. The Hydro-Rain algorithm is intro- retrieving near surface rain rates. Note that, the 2A12 product
duced in Section 4 which exploits the retrieval of hydrome- contains cloud water path (CWP), rain water path (RWP) and
teors and rain rate information. The validation results are ice water path (IWP). Therefore, the LWP is computed by simply
outlined in Section 5. Conclusions are drawn in Section 6. adding the CWP and RWP for a particular bin.

2. Datasets 3. Limitations of PMW estimate

The major datasets used in this study are the TRMM/TMI The accuracy of the PMW instrument based hydrometeor
calibrated brightness temperatures (TBs) and TRMM/PR hydro- estimates is somewhat limited, mainly due to the fact that PMW
meteor data, obtained from NASA Distributed Active Archive instruments cannot sense the full vertical structure of a storm,
Center (DAAC). The TRMM microwave imager (TMI) is a passive thus providing macro-physical information of hydrometeors.
microwave system that measures upwelling microwave radi- Furthermore, they have coarser measurement resolution in
ances at five frequencies between 19 and 85 GHz. Each comparison to active sensors. Typically, the integrated water
frequency is associated with one vertically polarized and one amounts measured by PMW instruments are often less than
horizontally polarized channel, with the exception of 21.3 GHz physical amounts, or those measured by active sensors. This can
that has vertically polarized channel only. The TRMM PR is a be confirmed from Fig. 2, in which the scatterplots between
nadir looking active sensor operated at 13.8 GHz frequency LWP and IWP are shown for PR 2A25 and TMI 2A12 based
(Ku-band). The scanning is done with cross-track strategy. It retrievals. This particular figure is constructed from one orbit
T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68 59

Fig. 1. The definition of nodes A to D in the TRMM PR 2A25 algorithm (with and without bright band scenarios).

sample data. As such, both sensors sample the same storm, so, illustrate the bivariate kernel density estimate of LWP and IWP
theoretically, the LWP/IWP ranges should be similar, if not the in logarithmic form. The kernel is assumed to be Gaussian.
same. Nevertheless, as the figure shows, the dynamic ranges Mostly, the PR LWP distribution lies in high dynamic ranges
significantly vary between two sensors. In Fig. 2, the left panels while for TMI, estimates stay in low dynamic ranges. The right

Fig. 2. (Left) The bivariate kernel density estimate of LWP and IWP (in log scales) for the PR 2A25 and TMI 2A12 out-comes. The kernel is assumed to be Gaussian.
(Right) The scatterplots between LWP and IWP (in natural scales) for the PR 2A25 and TMI 2A12 are shown.
60 T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68

panels show the scatterplots in physical space, and as can be Overall, one can argue that the monotonic correlations
seen that up to 25 mm (LWP) magnitude of integrated between the integrated hydrometeor amounts and individual
quantities are estimated from the PR. In contrary, the TMI channels are somehow moderate. Nevertheless, the combi-
estimated LWP is less than 5 mm. One should remember that nation of all nine channels should enhance the information
the TRMM PR has a sensitivity of 17 dBZ. This implies that a contents needed to estimate PR like hydrometeors. In the
significant portion of ice particles cannot be detected by the PR. next section, we describe the development of the Hydro-Rain
Nevertheless, the LWP estimate from the PR can be considered algorithm using TMI only information.
as fairly accurate. Therefore, the advantage from an empirical
algorithm in synergy with the PR will mostly be gained in LWP, 4. Hydro-Rain algorithm
thus in TWP, and ultimately in rain rate retrieval.
Fig. 3 represents the Spearman correlation coefficients as The Hydro-Rain algorithm is based on an adaptive neuro
a function of nine TMI channels for the LWP and IWP fuzzy inference system (ANFIS), by which the TMI calibrated
quantities over the ocean and land surfaces. The Spearman TBs are converted to PR like atmospheric column integrated
correlation coefficient is defined as: quantities (LWP and IWP). In turn, the water paths are
converted to near surface rain rate from the retrieved inte-
1 X n grated water content information.
ðx −xÞðyi −yÞ The ANFIS is a multi-layer adaptive network based fuzzy
n−1 i¼1 i
ρxy ¼ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð2Þ Sugeno model that can facilitate learning and adaption of
u u
u 1 X u 1 X
n n complex dynamical system introduced by Jang (1993). It
ðxi −xÞ  t
2 2
t ðy −yÞ integrates the principles of neural networks and fuzzy logic
n−1 i¼1 n−1 i¼1 i
consisting of a number of nodes connected through directional
links. The ANFIS is composed of five layers, namely, fuzzification
where, x and y represent the ranks of data pairs, x and y layer, rule layer, normalization layer, defuzzification layer and
denote the means of x and y ranks, respectively, and n is the an output layer. Fig. 4 represents a typical architecture of an
number of samples. The key advantage of the Spearman ANFIS model. In this system, x and y are the inputs, Ai and Bi are
correlation coefficient over traditional Pearson correlation the fuzzy sets and fi is the output produced by the fuzzy rule.
coefficient is that possible non-linear relations can be In the first layer, every node is an adaptive node. The
reflected in this measure. It ranges between − 1 and 1. The outputs of this layer are the fuzzy membership grade of the
sign of ρxy expresses the direction of association between x inputs, given by:
and y. Fig. 3 is prepared from the collocated TMI brightness
temperatures and PR hydrometeor outputs. Note that the 1
Oi ¼ μ Ai ðxÞ ð3Þ
collocations are made to the high resolution PR grid. Clearly,
the ocean water path is positively associated with the TBs,
with the exception of the 85 GHz channels (8 and 9), while Oi ¼ μ Bi ðyÞ ð4Þ
the land water path is negatively associated with the TBs.
Remarkably, the low frequency emission channels are poorly
where, μ is the weight obtained through fuzzy membership
correlated to hydrometeor contents over land. Despite, the
function. The μAi and μBi can adopt any given membership
emission channels are rather correlated to ocean water
functions. In this study, a Gaussian membership function is
x−bi 2
0.6 μ Ai ðxÞ ¼ exp − ð5Þ
Spearman Correlation Coeifficient

where, ai and bi are the premise parameters of the member-
0.2 ship functions.
In the second layer, every node is a fixed node labeled Π,
0 indicating the firing strength of each rule calculated with a
simple multiplier:
Oi ¼ wi ¼ μ Ai ðxÞμ Bi ðyÞ: ð6Þ
In the third layer, every node is a fixed node labeled N,
−0.6 Ocean LWP indicating the normalized firing strength of each rule from
Land LWP
Ocean IWP
Land IWP
the previous layer:
1 2 3 4 5 6 7 8 9 3 wi
TMI Channel # Oi ¼ wi ¼ : ð7Þ
w1 þ w2
Fig. 3. The Spearman correlation coefficients for the LWP and IWP at nine
TMI channels. The nine TMI channels represent 10V, 10H, 19V, 19H, 21V, In the fourth layer, every node is an adaptive node,
37V, 37H, 85V, and 85H GHz channels, respectively. indicating the contribution of the ith rule through the
T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68 61

Fig. 4. Typical architecture of ANFIS.

product of the normalized firing strength and a first order the Hydro-Rain algorithm, the TMI TBs and the PR hydrome-
Sugeno model: teor quantities (LWP, IWP and surface rain rate) data are
collocated to a single footprint domain. Subsequently, for the
4 retrieval of LWP and IWP (left side in the figure), ocean and
Oi ¼ wi f i ¼ wi ðpi x þ qi y þ r t Þ ð8Þ
land surface databases are prepared separately, by which the
ANFIS is trained. During training course, the subtractive
where, pi, qi, and ri are the consequent parameters.
clustering method is employed. This method is the extension
In the fifth layer, only one node is present, which is a fixed
of the mountain clustering method proposed by Chiu (1994).
node labeled Σ. This node indicates the overall outputs by
The foremost advantage of this method is that it can reduce
performing the summation of all incoming signals:
the computational complexity of the model during training.
X The data points are clustered in an unsupervised way by
wi f i
X measuring the potential of data points in the feature space.
Oi ¼ wi f i ¼ X
: ð9Þ
wi The assumption is that each data point is a potential cluster
i center. The data point with the highest potential, based on
density of surrounding points, is selected as the first cluster
Fig. 5 shows the schematic view of the development of the center, and remaining potential of data points are destroyed.
Hydro-Rain algorithm with ANFIS. It is also worth mention- Consequent cluster centers are found by revising the
ing that the TMI and PR have different scanning strategy potential of data points. This cluster method determines the
resulting in different resolutions. Furthermore, the TMI's low number of fuzzy rules and premise fuzzy membership
frequency channels also differ in resolution to the high functions for the ANFIS model. For the sake of clarity, the
frequency channels. Therefore, during the development of membership functions on nine input channels representing

Fig. 5. Flowchart illustrating the development of the Hydro-Rain algorithm using adaptive neuro fuzzy inference system (ANFIS).
62 T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68

the clusters identified by subtractive clustering method for (RMSE), correlation coefficient (r), and standard deviation
LWP and IWP are plotted in Figs. 6 and 7, respectively. The (STD). Apparently, the Hydro-Rain algorithm is well trained, as
top panels are given for ocean database and bottom panels the table indicates. The calculated correlation coefficients are
are given for the land database. As can be seen, three clusters in the range 0.61–0.68 for hydrometeor paths (FSE, 0.69
are generated for ocean case and two clusters for land case. to 1.75 mm). For rain rate, the correlation is fairly higher
For the development of rain rate retrieval model (right (r = 0.92, FSE = 0.69 mm), as the TRMM PR surface rain rate is
side in Fig. 5), the grid partition method is employed. The highly correlated with the integrated hydrometeor contents,
reason for using grid partition is that, the number of features especially LWP.
in this case is only two (LWP and IWP). As such we can afford
to train the ANFIS with reasonable computational powers 5. Validation results
using grid partition method. In other words, the curse of
dimensions should not be introduced in this case. The grid Two types of validations are presented hereafter in order
partition method divides the input space into rectangular to evaluate the performance of the Hydro-Rain algorithm.
subspaces using axis-paralleled partition with given number First, the validation is performed from a quantitative point of
of membership functions and their types in each dimension. view by applying the Hydro-Rain algorithm to TMI TBs, and
For reference, the membership functions of LWP and IWP comparing the Hydro-Rain retrieval to PR outputs. Second
variables for rain rate estimate using the grid partition validation is based on a case study from a qualitative point of
method are displayed in Fig. 8. view, in which, the super typhoon Haiyan is examined with
Following the determination of fuzzy rules and membership the Hydro-Rain algorithm.
functions, the fuzzy sets and parameters are calculated in least
square manner. In this study, a total of ~250,000 vectors from 5.1. Quantitative validation
several orbital samples from a wide range of meteorological
events are used for training the ANFIS models. Table 1 tabulates For the sake of quantitative validation, the Hydro-Rain
the statistical measures obtained during the training procedure. algorithm is applied to the TMI brightness temperature
The quantification of the algorithm performance is done measurements collocated to the PR footprint resolution. The
through various quantitative metrics, including, mean bias PR retrieval has been considered as the “truth”. Since the TMI
(MB), normalized mean bias (NMB), fractional standard error datasets are collocated to the PR footprint, point-to-point
(FSE), mean absolute error (MAE), root mean squared error comparison between two different fields has been possible.

1 1 1 1 1

0.5 0.5 0.5 0.5 0.5

0 0 0 0 0
180 220 260 100 200 200 240 280 150 200 250 200 250
in1 in2 in3 in4 in5
1 1 1 1

Cluster 1
0.5 0.5 0.5 0.5 Cluster 2
Cluster 3
0 0 0 0
220 240 260 280 150 200 250 150 200 250 150 200 250
in6 in7 in8 in9
1 1 1 1 1

0.5 0.5 0.5 0.5 0.5

0 0 0 0 0
200 250 300 150 200 250 250 300 180 220 260 250 300
in1 in2 in3 in4 in5
1 1 1 1

0.5 0.5 0.5 0.5 Cluster 1

Cluster 2

0 0 0 0
220 240 260 280 180 220 260 150 200 250 150 200 250
in6 in7 in8 in9

Fig. 6. The membership functions of the nine input TMI channels for the LWP, (top nine panels) over the ocean surface, (bottom nine panels) over the land
surface. The membership functions are generated through sub clustering method.
T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68 63

1 1 1 1 1

0.5 0.5 0.5 0.5 0.5

0 0 0 0 0
170 210 250 290 100 200 200 240 280 150 200 250 200 250
in1 in2 in3 in4 in5
1 1 1 1

Cluster 1
0.5 0.5 0.5 0.5 Cluster 2
Cluster 3
0 0 0 0
220 240 260 280 150 200 250 150 200 250 150 200 250
in6 in7 in8 in9
1 1 1 1 1

0.5 0.5 0.5 0.5 0.5

0 0 0 0 0
220 260 300 150 200 250 220 260 300 180 220 260 250 300
in1 in2 in3 in4 in5
1 1 1 1

Cluster 1
0.5 0.5 0.5 0.5
Cluster 2

0 0 0 0
200 220 240 260 280 180 220 280 150 200 250 150 200 250
in6 in7 in8 in9

Fig. 7. Same as Fig. 6, but for IWP.

in1mf1 in1mf2 in1mf3

Degree of membership





5 10 15 20 25
input LWP

in2mf1 in2mf2 in2mf3

Degree of membership





5 10 15 20 25
input IWP

Fig. 8. The membership functions of the LWP and IWP for the rain rate retrievals. The membership functions are generated through grid partition method.
64 T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68

Table 1 can be seen that the Hydro-Rain retrieval agrees reasonably well
The statistical measures of the hydrometeor estimates by the Hydro-Rain with the PR retrieval. Of course, the performance of the
algorithm obtained from the training datasets.
Hydro-Rain has been somewhat degraded from the training
MB NMB FSE MAE RMSE r STD statistics, as expected. The Hydro-Rain retrieval is better
correlated to the PR over the ocean than over land. Nevertheless,
LWP Ocean 0.01 0.02 1.25 0.53 1.10 0.65 1.10
Land 0.01 0.01 1.45 0.58 1.26 0.61 1.26 it is apparent that the PMW based Hydro-Rain algorithm suffers
IWP Ocean 0.00 0.02 1.16 0.16 0.30 0.68 0.30 significantly to produce good estimate of hydrometeor contents
Land 0.00 0.01 1.75 0.23 0.61 0.64 0.61 in high rain rate regimes (higher LWP and IWP values).
Rain rate Ocean & land 0.00 0.00 0.69 1.08 2.63 0.92 2.63
Nonetheless, overall, the error metrics are reasonably small.
Over the ocean, the FSE is calculated as 1.35 mm (correlation
0.57) for LWP and 1.31 mm (correlation 0.63) for IWP. More
Note that, the validation database has been created by com- importantly, the Hydro-Rain retrieval agrees much better with
prising a large number of randomly selected TMI measurements the observations from TRMM PR, compared to the results from
obtained from several orbital samples in different meteorolog- TRMM 2A12 algorithm. For instance, the reported FSE over the
ical conditions during the year of 2013 (~50 orbits). Of course, ocean is much higher than the Hydro-Rain algorithm, revealing
the validation database is independent from the training as 1.49 and 2.86 mm for LWP and IWP, respectively.
database, which used for the development of the ANFIS models. The validation results can be also visualized in terms of
Fig. 9 provides the scattergrams of the Hydro-Rain LWP and IWP empirical cumulative distribution histograms. In Fig. 10, we
outcomes against the PR retrieval over ocean and land surfaces. show the empirical cumulative distribution histogram of
The corresponding statistical metrics are tabulated in Table 2. It hydrometeor estimates by the Hydro-Rain algorithm in

Fig. 9. Scatterplots of Hydro-Rain vs PR 2A25 derived hydrometeor estimates over the ocean (left) and the land (right) surfaces.

Table 2
The statistical measures of the LWP and IWP estimates by the Hydro-Rain and 2A12 algorithms obtained from the validation datasets.


Hydro-Rain LWP Ocean −0.15 −0.14 1.35 0.70 1.47 0.57 1.47
Land −0.02 −0.03 1.55 0.61 1.32 0.49 1.32
IWP Ocean 0.01 0.02 1.31 0.16 0.32 0.63 0.32
Land 0.03 0.09 1.80 0.22 0.54 0.54 0.54
2A12 LWP Ocean −0.33 −0.30 1.49 0.73 1.64 0.42 1.61
IWP Ocean 0.61 2.55 2.86 0.65 0.92 0.53 0.69
T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68 65

Ocean Land
1 1

0.8 0.8

0.6 0.6

0.4 0.4

Ocean Hydro−Rain LWP

0.2 Ocean PR 2A25 LWP 0.2 Land Hydro−Rain LWP
Ocean TMI 2A12 LWP Land PR 2A25 LWP
0 0
0 2 4 6 8 10 0 2 4 6 8 10

1 1

0.8 0.8

0.6 0.6

0.4 0.4

Ocean Hydro−Rain IWP

0.2 0.2 Land Hydro−Rain IWP
Ocean PR 2A25 IWP
Ocean TMI 2A12 IWP Land PR 2A25 IWP
0 0
0 2 4 6 8 10 0 2 4 6 8 10

Fig. 10. The empirical cumulative distribution of the LWP and IWP estimates by different algorithms.

comparison with the TRMM PR. The histograms demonstrate clearly, the Hydro-Rain agrees better with the PR than the
that the Hydro-Rain estimate is in good agreement with the 2A12 algorithm.
TRMM PR, although, it overestimates the LWP retrieval in low In order to understand the performance of Hydro-Rain in
LWP regime (b1.5 mm), but underestimates in high LWP different rain rate regimes, we construct the mean ocean
regime (N 1.5 mm). Same holds for the IWP. Nevertheless, surface hydrometeor estimate as a function of TRMM PR rain

20 PR 2A25
TMI 2A12
LWP (mm)



0 10 20 30 40 50 60 70 80 90 100
Rain Rate (mm/h)
4 PR 2A25
TMI 2A12
IWP (mm)

0 10 20 30 40 50 60 70 80 90 100
Rain Rate (mm/h)

Fig. 11. Mean hydrometeor estimates (LWP and IWP) as a function of PR averaged rain rate using three different types of algorithms (Hydro-Rain, PR, and 2A12).
The error bars represent the ±1 standard deviations.
66 T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68

Fig. 12. Scatterplots of Hydro-Rain vs PR 2A25 derived rain rate estimates over the ocean (top) and the land (bottom) surfaces.

rate for three different types of algorithms (Hydro-Rain, PR, 5.2. Qualitative validation (typhoon case)
and 2A12). The error bars represent the ± 1 standard
deviations (See Fig. 11.) The figure demonstrates that the The algorithm is now tested qualitatively on super typhoon
integrated hydrometeor contents (LWP and IWP) are in Haiyan that occurred on the 7th of November over the
increasing trend as the rain rate increases. Especially, to a Philippines during 2013 Pacific typhoon season. It is the
certain degree, there is a cubic growth in PR LWP against the second-deadliest Philippine typhoon on record, originated
rain rate observations. Nonetheless, there are clear difficul- from an area of low pressure, and then developed into a tropical
ties in estimating hydrometeor contents by the passive depression. The system has subsequently transformed into
microwave based algorithms in high rain rate regimes. As category 5 hurricane pouring rain over the region. Exceptional
the rain rate increases, the inconsistencies between PMW high wind speed has been reported during the typhoon event.
and PR estimates also rise. Taking PR as a reference, both The qualitative evaluation is displayed in Fig. 13. This
Hydro-Rain and 2A12 underestimate the LWP amounts in particular snapshot is taken from orbit no 91024.7 on the 7th
high rain rate regimes, to a significant extent. For IWP of November 2013. For the sake of comparison, the typhoon
estimate, overestimation is seen by the 2A12 algorithm, but fields as seen by the TMI 2A12 and PR are also incorporated in
underestimation is observed by the Hydro-Rain algorithm in the figure. It is quite encouraging to see that the Hydro-Rain
higher rain rate regimes. Overall, the results clearly highlight algorithm has been able to detect and produce the body of
the limitation of PMW based retrieval in higher rain rate the Typhoon reasonably well. The agreement of Hydro-Rain
regimes in comparison with the PR. Nevertheless, one should derived rain rate with the TRMM PR rain product confirms
agree that Hydro-Rain algorithm agrees relatively better with the robustness of the algorithm, even in hurricane/typhoon
the TRMM PR than the 2A12 algorithm. case. Indeed, as the figure shows, the Hydro-Rain induces a
In Fig. 12, we demonstrate the rain rate validation for the
Hydro-Rain algorithm. The Hydro-Rain rain rate retrieval is
solely based on the information of LWP and IWP, and does not
use any other information (e.g. TBs or auxiliary datasets). Hence, Table 3
The statistical measures of the rain rate estimates by the Hydro-Rain and
the validation of rain rate should also be considered as an
2A12 algorithms obtained from the validation datasets.
indirect validation of the hydrometeor contents (LWP and IWP).
Remarkably, the Hydro-Rain retrieved rain rate is in good Mean NMB FSE MAE RMSE r STD
agreement with the PR retrieval. The statistical measures are bias
reported in Table 3. The calculated FSE over ocean and land Hydro-Rain Ocean −0.42 −0.11 1.37 2.47 5.09 0.56 5.08
surface types are 1.37 mm and 1.67 mm respectively for the Land −0.21 −0.07 1.67 2.23 5.10 0.43 5.10
Hydro-Rain algorithm. The statistical scores are somewhat 2A12 Ocean −0.93 −0.30 1.59 2.54 5.39 0.51 5.70
Land −0.30 −0.12 1.81 2.88 5.86 0.40 5.89
lower for the TMI 2A12.
T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68 67

well-organized and varied distribution of rain rates within retrieved hydrometeor estimates with the in situ measure-
the typhoon field. ments, in order to understand the agreement and/or differences
between them.
6. Concluding remarks Nevertheless, there are certainly several advantages in
using the Hydro-Rain algorithm towards the next generation
An empirically based hydrometeor estimation (LWP, IWP, PMW era. For instance, the GPM will have a constellation of
and rain rate) algorithm, named as Hydro-Rain, is proposed PMW instruments, and the Hydro-Rain can easily be adapted
for the TRMM microwave imager (TMI). The Hydro-Rain to those instruments. On the other hand, passive microwave
algorithm is based on the adaptive neuro fuzzy inference instruments cover large swath width, as such; a better spatial
system (ANFIS). The TRMM PR hydrometeor information has coverage is achieved than active instruments. Consequently,
been taken as the “reference” in order to develop as well as better information of hydrometeors and rain rate can be
validate the Hydro-Rain algorithm. gained on outer swaths, where precipitation radar (PR) and
It should be noted that the TRMM PR hydrometeor infor- dual frequency precipitation radar (DPR) do not have any
mation is based on power law relationship between Z and coverage. More importantly, during the GPM era, the DPR
the water content assuming a drop size distribution model. will provide more accurate hydrometeor information with
However, radar reflectivity factor is related to the sixth moment enhanced sensitivity. Therefore, there is a scope to improve
of DSD, while the water content is related to the third moment the hydrometeor estimates, by re-training the Hydro-Rain
of DSD. Therefore, there will always be some uncertainties algorithm, when DPR will be on-board.
associated with hydrometeor estimate from the PR itself. Hence,
the purpose of this study was not to provide “actual physical” Acknowledgments
hydrometeor estimate from the TMI, but to put forward an
approach that would make the retrieved variables consistent The TRMM data used in this effort were acquired as part of
with the PR like estimate. Arguably, in-situ measurements from the NASA's Earth–Sun System Division and archived and
balloon and aircraft platforms could provide more detailed distributed by the Goddard Earth Sciences (GES) Data and
information of hydrometeor profiles. Therefore, in the future, Information Services Center (DISC) Distributed Active Ar-
it might be interesting to compare the PR and Hydro-Rain chive Center (DAAC). The views expressed here are those of

Fig. 13. The rain rate imagery of the super typhoon Haiyan on the 7th of November 2013 as derived by the TMI Hydro-Rain, the TMI 2A12, and the PR 2A25
68 T. Islam et al. / Atmospheric Research 145–146 (2014) 57–68

the authors solely and do not constitute a statement of policy, Islam, T., Rico-Ramirez, M.A., Thurai, M., Han, D.W., 2012d. Characteristics
of raindrop spectra as normalized gamma distribution from a Joss–
decision, or position on behalf of NOAA/NASA or the authors' Waldvogel disdrometer. Atmos. Res. 108, 57–73.
affiliated institutions. Jang, J.S.R., 1993. ANFIS — adaptive-network-based fuzzy inference system.
IEEE Trans. Syst. Man Cybern. 23, 665–685.
Kummerow, C., Barnes, W., Kozu, T., Shiue, J., Simpson, J., 1998. The Tropical
References Rainfall Measuring Mission (TRMM) sensor package. J. Atmos. Ocean.
Technol. 15, 809–817.
Brandau, C.L., Russchenberg, H.W.J., Knap, W.H., 2010. Evaluation of ground- Masunaga, H., Iguchi, T., Oki, R., Kachi, M., 2002. Comparison of rainfall
based remotely sensed liquid water cloud properties using shortwave products derived from TRMM microwave imager and precipitation
radiation measurements. Atmos. Res. 96, 366–377. radar. J. Appl. Meteorol. 41, 849–862.
Chiu, S.L., 1994. Fuzzy model identification based on cluster estimation. J. Michel, Y., Auligne, T., Montmerle, T., 2011. Heterogeneous convective-scale
Intell.Fuzzy Syst. 2, 267–278. background error covariances with the inclusion of hydrometeor
Crewell, S., Ebell, K., Lohnert, U., Turner, D.D., 2009. Can liquid water profiles variables. Mon. Weather Rev. 139, 2994–3015.
be retrieved from passive microwave zenith observations? Geophys. Saavedra, P., Battaglia, A., Simmer, C., 2012. Partitioning of cloud water and
Res. Lett. 36. rainwater content by ground-based observations with the Advanced
Du, P., Girard, E., Bertram, A.K., Shupe, M.D., 2011. Modeling of the cloud and Microwave Radiometer for Rain Identification (ADMIRARI) in synergy
radiation processes observed during SHEBA. Atmos. Res. 101, 911–927. with a micro rain radar. J. Geophys. Res.-Atmos. 117.
Hazra, A., Mandal, V., Chen, J.P., 2013. Study of cloud microphysical Shen, X.Y., Wang, Y., Li, X.F., 2011. Radiative effects of water clouds on
properties over India during CAIPEEX using a mesoscale model with rainfall responses to the large-scale forcing during pre-summer heavy
new cloud microphysical scheme—part I. J. Atmos. Sol. Terr. Phys. 93, rainfall over southern China. Atmos. Res. 99, 120–128.
29–44. Spencer, R., Hood, R., Goodman, H., 1989. Precipitation retrieval over land
Iguchi, T., Kozu, T., Meneghini, R., Awaka, J., Okamoto, K., 2000. Rain-profiling and ocean with the SSM/I — identification and characteristics of the
algorithm for the TRMM precipitation radar. J. Appl. Meteorol. 39, scattering signal. J. Atmos. Ocean. Technol. 6, 254–273.
2038–2052. Van Weverberg, K., van Lipzig, N.P.M., Delobbe, L., Vogelmann, A.M., 2012.
Islam, T., Rico-Ramirez, M.A., Han, D., Srivastava, P.K., 2012a. Using S-band The role of precipitation size distributions in km-scale NWP simulations
dual polarized radar for convective/stratiform rain indexing and the of intense precipitation: evaluation of cloud properties and surface
correspondence with AMSR-E GSFC profiling algorithm. Adv. Space Res. precipitation. Q. J. R. Meteorol. Soc. 138, 2163–2181.
50, 1383–1390. You, Y.L., Liu, G.S., 2012. The relationship between surface rainrate and water
Islam, T., Rico-Ramirez, M.A., Han, D.W., 2012b. Tree-based genetic program- paths and its implications to satellite rainrate retrieval. J. Geophys. Res.-
ming approach to infer microphysical parameters of the DSDs from the Atmos. 117.
polarization diversity measurements. Comput. Geosci. 48, 20–30. Zhou, Y.S., Ran, L.K., 2012. Torrential rainfall responses to vertical wind
Islam, T., Rico-Ramirez, M.A., Han, D.W., Srivastava, P.K., Ishak, A.M., 2012c. shear, radiation and ice clouds: a rainfall partitioning analysis based on
Performance evaluation of the TRMM precipitation estimation using surface rainfall budget. Atmos. Res. 108, 1–8.
ground-based radars from the GPM validation network. J. Atmos. Sol.
Terr. Phys. 77, 194–208.