Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Indian Journal of Radio & Space Physics
Vol. 37, October 2008, pp. 341-352
Estimation of rain parameters from spectral moments of L-band wind profiler
using Multi-Layer Perceptron Network model
Mahen Konwar1, Diganta Kumar Sarma
1, Sanjay Sharma
*, 1, U. K. De
2, S. Pal
3 & Jyotirmoy Das
4
1Department of Physics, Kohima Science College, Jotsoma, Kohima, Nagaland 797 002, India 2Department of Environmental Science, Jadavpur University, Kolkata, India
3ECSU, Indian Statistical Institute, Kolkata 700 108, India 4Institute of Technology and Marine Engineering, Jhinga 743368, India
Received 19 February 2007; revised 3 September 2007; accepted 11 July 2008
Two Multi-Layer Perceptron (MLP) models are developed to estimate the radar reflectivity factor (dBZ) and rain
intensity (R) from the spectral moments of an L-band wind profiler. Out of the four spectral moment inputs of the MLP
models, the backscattered power (P) and Doppler velocity (VD) were found to have better correlation with the rain
parameters. The model results were validated with the Joss Waldvogel Disdrometer (JWD) observations. For the training
and validation data sets of dBZ, the root mean square error (rmse) of estimated and observed data sets were found to be 5.05
and 5.32 dBZ with correlation coefficients of 0.90 and 0.86, respectively. Similarly, for the training and validation data sets
of R, the rmse for estimated and observed values were found to be 4.27 and 7.74 mmh-1 respectively with correlation
coefficients of 0.95 and 0.78. The developed models were validated with a rain event on 22 June 2000, that consisted of rain
from both convective and stratiform regimes. The error between the estimated and observed rain accumulation was found to
be ∼ 5%. The height profiles of estimated dBZ were able to identify the bright band during stratiform rain by virtue of high
values of reflectivity gradient at around 4.0 km height. Though ambiguity in the estimation of R was observed at bright band
level, overall estimated parameters were in good agreement with general characteristics of convective and stratiform rain.
Keywords: L-band wind profiler, Spectral moments, Joss Waldovogel Disdrometer, Radar reflectivity factor, Rain
intensity, Multi-Layer Perceptron network
PACS No.: 92.40.Ea; 92.40.eg; 92.60.gf
1 Introduction Measurement of rain parameters by remote sensing
technique is a fundamental objective in radar
meteorology. There are two types of sensors, i.e.
passive and active, for rain measurement by the
remote sensing technique. The passive sensors, such
as microwave radiometer, measure radiance that is the
end product of the integrated effects of the
electromagnetic absorption-emission and scattering
through the precipitating cloud along the sensor view
path1. However, height assignment of cloud systems
by passive remote sensing is not very specific. On the
other hand, active sensors provide specific height
information of precipitating systems based on the time
delay of back scattered return power1. Presently many
active sensor systems e.g. scanning reflectivity radar,
Doppler weather radar, polarimetric radar and
profiling micro rain radar2-5
are being utilized for rain
measurements by using back scattered power in terms
of radar reflectivity factor (Z). The common approach
for measuring rainfall intensity from backscattered
power is by utilizing the empirical Z-R relations,
where Z is calculated by using the following radar
equation2
2
2
r
ZKCPr = … (1)
where Pr is backscattered
power from the target, K the
dielectric constant of the target, C the radar constant
and r the range of the target. Here, Z is a function of
rain drop size distribution and is independent of radar
frequency.
The wind profilers were originally developed to
measure the height profiles of the three components
of wind velocity. Subsequently however, many
researchers demonstrated its utility for the
identification/classification of precipitating systems6-9
and in measuring the rain parameters, i.e. radar
reflectivity factor and rainfall intensity10-14
. The
precipitating signatures obtained from the wind
profiler are due to Rayleigh scattering from
INDIAN J RADIO & SPACE PHYS, OCTOBER 2008
342
hydrometeors2. The Rayleigh scattering component
observed by a vertically incident profiler results from
convolution of normalized atmospheric turbulent
probability density function with the hydrometeor
reflectivity spectral density10
and it is expressed as
follows
)(]2/)(exp[]2/1[
)()()(
hyd2air
2air
hydairRayleigh
vSv
vSvSvS
×σω−−πσ=
×ω−= … (2)
where v, is the fall velocity of hydrometeor; ϖ, the air
velocity either in the upward or downward direction;
and σair , the spectral broadening of clear air Doppler
spectrum. The hydrometeor spectrum in stationary air
is represented by
vDDDNvS d/d)()(6
hyd = … (3)
where N(D), D and dD/dv are number concentration
of the hydrometeor distribution, the rain drop
diameter and the coordinate transformation from
terminal fall speed to diameter space, respectively.
Extensive research has been carried out to estimate
rain drop size distribution (RDSD) from the Doppler
spectra of either individual or combined use of Very
High Frequency (VHF) and Ultra High Frequency
(UHF) wind profilers13-15
. For RDSD retrievals and
subsequently to estimate the rain parameters, the wind
profilers suffer from certain limitations, viz.
(i) Non-linear response of the receiver during rain -
During clear air measurements, clear air wind
profiler receiver has a linear response. However,
during rain, when backscattered signal is very
strong, profiler receiver has a non-linear
response and an increase in the incident power
produces only a small increase16
in the measured
signal power (S). At some point the receiver
even becomes completely saturated and further
increase in incident power leads to no
measurable increase in S. In addition, the wind
profilers are not calibrated, which implies that
radar reflectivity factor (Z) can not be
determined directly.
(ii) Another limitation is the assumptions made in
certain models17-22
, both physical and empirical.
Also, due to the natural variation of RDSD with
respect to either Z or R, there is an inherent
limitation of the parametric approach to follow a
particular model for the estimation of rain
parameters11
.
In the past, non-parametric technique based Multi-
Layer Perceptron23
(MLP) networks have been
utilized to solve many diverse problems in microwave
remote sensing24–26
. A similar approach is utilized for
the first time to estimate the rain parameters from
Doppler spectra of L-band wind profiler. The
motivating factor of the present study is to enhance
the capability of the wind profiler as a rain profiling
system during rainy situations, when clear air echoes
are completely masked by precipitation echoes. The
main objective of the present study is to develop MLP
network
models to estimate the radar reflectivity
factor (dBZ) and rainfall intensity (R) from the
spectral moments of L-band wind profiler; and to
study the characteristics of the precipitating systems
in terms of height profiles of the estimated
parameters. The multivariable MLP based non-
parametric techniques are important because of their
ability in dealing with the above mentioned errors
common to wind profiler data. First of all, the
inclusion of the multivariable parameters reduce the
total dependency of dBZ and R measurements on
normalized back scattered power. In the present
study, these parameters are estimated with the help of
combined use of all the spectral moments of the
Doppler spectra of L-band wind profiler and DSD
spectra from Joss Waldvogel Disdrometer (JWD).
Moreover, the training of the spectral moments with
dBZ as measured from the independent system will
overcome the limitation of the non-linear response of
the receiver during rainy situations to a certain extent.
Further in the proposed methodology, it is not
necessary to assume a specific model for the retrieval
of rain parameters, and to take into account the
natural variation of the rain parameters.
2 System descriptions and data preparation National Atmospheric Research Laboratory
(NARL), at Gadanki (13.5°N, 79.2°E), India, has the
facility of collocated instruments. For the present
work, observations of L-band wind profiler and Joss
Waldvogel Disdrometer (JWD) for the years 1998-
2000 were utilized. The L-band wind profilers
operated at 1357 MHz is very sensitive to
precipitation particles7
due to Rayleigh scattering,
where scattering cross-section is inversely
proportional to the fourth power of wavelength. The
system description of the L-band wind profiler is
shown in Table 1.
For the extraction of atmospheric parameters from
the back scattered signals of wind profiler it requires
KONWAR et al.: RAIN PARAMETERS FROM WIND PROFILER USING MLP NETWORK MODEL
343
extensive signal processing. The signal processing of
these signals mainly consist of estimation of 0th, 1st
and 2nd moments of Doppler spectrum27
. These
moments correspond to total power (P), mean
Doppler shift (fD) and variance (σ2) of the Doppler
spectra and they are estimated by the following
expressions27
:
∑= iPP … (4)
iiD fPP
f ∑=1
… (5)
22 )(1
Di ffPP I
−=σ ∑
… (6)
Where Pi and fi are the power and Doppler frequency
shift of the ith spectral line in Doppler spectrum,
respectively.
Further, for the vertical incident beam of L-band
wind profiler, fall velocity of the mean raindrop is
given by
)2/(λ= DD fV … (7)
where λ, is the wavelength of the L-band profiler. By
using this expression the Doppler frequency spectra
can be converted to the corresponding velocity
spectra.
Along with these moments the noise level of the
back scattered echoes were also estimated. From the
returned power (P) and noise level (L), the signal-to-
noise ratio (SNR) was calculated by the following
expression
)NFFT/log(10SNR(dB) ×= LP … (8)
where NFFT, is the number of FFT points. For L-
band profiler observations, though the measured
parameters are available from 0.3 km onward, but due
to the presence of ground clutter, data from 0.75 km
onward only have been considered. The moment
estimation of profiler spectra were carried out with
the help of Atmospheric Data Processor (ADP)
software28
developed at NARL, Gadanki.
The other system, i.e. JWD counts the number of
rain drops of different diameters with an integration
time of one minute. For the present study, the dBZ
and R obtained from JWD have been utilized.
Although the JWD underestimates smaller raindrops
during heavy precipitation period due to dead error
time29
, still it is one of the most widely used tools to
study rain characteristics. For the present data set, the
dead error correction was applied to reduce the error
due to the underestimation of the smaller drops during
heavy rains. The dBZ and R were estimated from
JWD observations, by using the following
equations30
:
)Z(log10dBZ 10= … (9)
where
DDDVNATZi
iii ∆= ∑=
20
1
6))(/()/1( … (10)
DDNATR i
i
i ∆π= ∑=
320
1
3 )/1)(10/6.3)(6/( … (11)
where A, is the collecting area of the sensor; T, the
integration time of one minute; Ni, the number of
drops in the ith channel; and Di, the mean drop
diameter of the ith channel.
To make the near simultaneous data points of L-
band profiler and JWD, a time lag due to falling rain
drops from a height of 0.75 km to the ground is
incorporated. The time lag is calculated with the help
of fall velocity parameter (VD) at 0.75 km. Averaging
of the DSD spectra was carried out from the start-to-
stop time observations of the vertical beam of the
profiler to match the different integration time of the
two systems.
3 Methodology
In this study, dBZ and R are estimated from
spectral moments of L-band wind profiler with the
help of DSD observations from JWD by utilizing the
MLP31
network. No prior assumption of any model is
made to the spectrum of the wind profiler. The MLP
networks consist of several layers of neurons, of
which first one is the input layer and the last one the
output layer and in-between layers are called hidden
Table 1―System description of the L-band wind profiler
Parameters Specification
Frequency 1357.5 MHz
Antenna Aperture 3.8m × 3.8m Phased Array
Beam width 4°
Pulse width 1 µs
Inter Pulse 80 µs
Number of FFT points 128
Number of coherent Integration 50
Duty ratio 1.25%
INDIAN J RADIO & SPACE PHYS, OCTOBER 2008
344
layers. Every node, except the input layer nodes,
computes the weighted sum of its inputs and applies
an activation function, a sigmoid function to compute
its output. This output is then transmitted to the nodes
of the next layer. The objective of the MLP learning is
to set the connection weight in such a manner as to
minimize the error between the network output and
the target. The back propagation gradient decent
method32
of learning is utilized for this purpose. The
architectural diagram of the present MLP is shown in
Fig. 1. The present MLP networks consist of three
layers, viz. input layer with four nodes, hidden layer
with 15 nodes and output layer with one node. It is to
be mentioned here that the optimum number of
hidden nodes were found to be 15 in the present MLP
network. The two MLP networks were trained
separately by using the four input vectors from the L-
band wind profiler, i.e. P, VD, σ, and SNR and output
vectors as dBZ and R from the JWD. For the
present study the whole data set consists of ∼ 1000
points. The data set is divided into two parts, i.e. 90%
is utilized for training and remaining 10% for
validation of the trained MLPs. The above network
architecture is realized through the MATLAB toolbox
based on Levenberg-Marquardt (L-M) back
propagation algorithm. The advantage of using this
algorithm is that it converges quickly as compared to
most of the other algorithms. It can train any network
as long as its weight, net input and transfer function
have derivative function. The L-M algorithm uses the
approximation to the Hessian matrix in the Newton
method, where the weight vectors Xt is updated as
t
TT
tt EJIJJXX −µ+−=+ ][1 … (12)
where J, is the Jacobian matrix that contains first
derivative of the network error with respect to weight
and bias; I, the unit matrix; Et, the error at output
node and µ, the learning parameter. Depending upon
the magnitude of µ the method transit smoothly
between the two extremes: Newton method (µ – 0)
and well known steepest decent method (µ - ∞). For
L-M algorithm the performance index to be optimized
is defined as
∑ ∑ −= ])([)( 2
kpkp odwF … (13)
where dkp and okp are the target and output values of
the kth output and pth features.
After the supervised training to the prepared input
data set with respect to target dBZ and R, the weight
matrices for input- hidden layers and hidden-output
layers in both the cases were estimated. These two
developed MLP network models were designated as
MLP_dBZ and MLP_R. Thereafter, with the help of
developed MLP network models the dBZ and R were
estimated for the given set of input spectral moments
data.
4 Results
Simultaneous observations of the L-band wind
profiler and JWD during convection-precipitation
campaign at NARL, Gadanki were utilized for this
work. A rain event on 22-23 June 2000 was selected
to study the typical characteristics of different
moments of Doppler spectra during rainy situations.
This event consists of rain from both the convective
and stratiform regimes. The height-time intensity
(HTI) plots of P, VD, σ and SNR of Doppler spectrum
are shown in Figs. 2(a), (b), (c) and (d), respectively.
On this day, the convective rain started at around
2142 hrs LT and as evident from the profiler
observations, it was characterized by high values of
backscattered power up to the upper height of ~ 6 km
[Fig. 2(a)]. In general, the convective rain is
characterized by high intensity rain at ground. The
stratiform type of rain is identified through the
presence of bright band, which is indicated by a band
of high value of back scattered powers at around 4.2
km. The bright band started to form at around 0131
hrs LT onwards on 23 June 2000. The stratiform rain
at the ground is characterized by low intensity rain for
longer duration of time. Figure 2(b) shows the mean
Doppler velocity of falling rain drops. It was observed
that Doppler velocity of rain drops vary from 1 to 10
m/s. The downward velocity is indicated by –ve sign.
Fig. 1―Architecture of three layers MLP network to estimate the
rain parameters from the spectral moments of L-band profiler
KONWAR et al.: RAIN PARAMETERS FROM WIND PROFILER USING MLP NETWORK MODEL
345
During convective regime the maximum Doppler
velocity is found to be around 10 m/s at higher
altitudes. Figure 2(c) shows the HTI plot of spectral
width of Doppler spectra. Large values of SW present
during convective rain indicate strong turbulence in
the atmosphere. This parameter has positive
correlation with rain as heavy rainfall in the ground
was preoccupied by strong turbulence in the
atmosphere. Figure 2(d) shows the HTI plot of signal-
to-noise ratio during non-rainy and rainy situations.
This parameter is a good indicator of rainy and non-
rainy situations. Though its variation with rain
intensity is not very significant, its value jumped from
-10 dB to ≥ 10 dB from non-rainy to rainy situation.
During this period, the temporal variation of dBZ and
R at ground are shown in Figs. 2(e) and (f),
respectively. During the convective spell the values of
reflectivity factor were varying in the range ∼ 40-50
dBZ, whereas during the stratiform period its values
were < 40 dBZ and minimum values observed up to
20 dBZ. Similarly, the temporal variation of R shows
three spells of convective rain with intensity ranging
from 25 to 78 mmh-1
. These spells started at
2137 hrs LT (JWD data is available from 2143 hrs LT
onwards) and ended at 0110 hrs LT followed by
stratiform type of rain with low rain intensity
< 5 mmh-1
.
The near simultaneous and collocated data points of
the wind profiler and JWD are considered to study
the characteristics of moments of the profiler with
respect to dBZ and R. For this purpose, the scatter
plots for P-dBZ, VD-dBZ, σ-dBZ and SNR-dBZ are
shown in Figs. 3(a), (b), (c) and (d), respectively. For
the scatter plots, the dBZ values were considered from
10 dBZ onward. The power law is fitted to these data
sets and their corresponding equations are provided in
the respective panels. Maximum value of exponent
(0.53) in the power law was observed for the P-dBZ
relation and its values were found to be in descending
order for VD-dBZ, σ-dBZ and SNR-dBZ. The
correlation coefficient (CC) for each scatter plot is
also provided in the respective panels. Maximum
correlation (0.70) was found for VD-dBZ and the CC
were in decreasing order for P-dBZ, SW-dBZ and
SNR-dBZ. Similarly, Figs 4 (a), (b), (c) and (d) show
the scatter plots for P-R, VD-R, σ-R and SNR-R,
respectively. For these scatter plots the rain intensity
values were considered from 0.1 mmh-1
onward. The
power law is fitted to these data sets and their
corresponding equations are provided in the
respective panels. Maximum value of exponent
(0.107) in the power law relation was observed for P-
R plot and its values were in decreasing order for VD-
R, σ-R and SNR-R. The correlation coefficient for
each scatter plot is also provided in the respective
panels. Maximum correlation (0.47) was found for
the VD-R relation and in decreasing order for P-R,
SW-R and SNR-R. Interestingly, though the SNR-R
relation has least values of exponent and correlation
coefficient (CC) but it is a good indicator of rain. It
Fig. 2―HTI plots of (a) returned power, (b) Doppler velocity, (c)
Spectral width, (d) SNR of a Doppler spectrum of L-band wind
profiler, (e) temporal variation of radar reflectivity factor and (f)
temporal variation of rain intensity from JWD
INDIAN J RADIO & SPACE PHYS, OCTOBER 2008
346
Fig. 3―Scatter plots for (a) returned power versus radar reflectivity factor, (b) Doppler velocity versus radar reflectivity factor,
(c) spectral width versus radar reflectivity factor and (d) SNR versus radar reflectivity factor during the rainy situations
Fig. 4―Scatter plots for (a) returned power versus rain intensity, (b) Doppler velocity versus rain intensity, (c) spectral width versus rain
intensity and (d) SNR versus rain intensity during the rainy situations
KONWAR et al.: RAIN PARAMETERS FROM WIND PROFILER USING MLP NETWORK MODEL
347
was observed that as soon as rain started its SNR
jumped from negative to positive values.
After developing the weight matrices of the MLP
network model with the help of supervised training, as
discussed above, the dBZ and R were estimated by
using the spectral moment as inputs from the training
and validation data sets. The estimated dBZ from the
training and validation data sets were further validated
with the observed dBZ. Pixel-to-pixel comparison of
the estimated and observed dBZ for training and
validation data sets are shown in Figs 5(a) and (b),
respectively. The root mean square error (rmse) value
for training data set with respect to observed values
was found to be 5.05 dBZ, with a bias of -0.02 dBZ
and correlation coefficient of 0.90 [Fig. 5(a)]. The
rmse value for validation data set was found to be
5.32 dBZ with a bias of 0.93 dBZ and CC of 0.86
[Fig. 5(b)]. Figures 6(a) and (b) show pixel-to-pixel
Fig. 5―Pixel-to-pixel comparison of estimated and observed radar reflectivity factor for (a) training data set and (b) validation data set
Fig. 6―Pixel-to-pixel comparison of estimated and observed rain intensity for (a) training data set and (b) validation data set
INDIAN J RADIO & SPACE PHYS, OCTOBER 2008
348
comparison of model estimated and observed R for
the training and validation data set. It was observed
that most of the time the peaks of the estimated rain
intensity were following the observed intensity,
though in some cases there were some differences in
magnitudes. The rmse values for training and
validation data set were found to be 5.32 mmh-1
and
7.74 mmh-1
, respectively. The correlation coefficients
for training and validation data set were found to be
0.95 and 0.78, respectively.
An independent validation of the model estimated
dBZ and R with the ground JWD observations was
carried out for a rain event on 22 June 2000. Further,
the model estimated parameters were also compared
with the results from the multivariable regression
method33
. This event consists of both convective and
stratiform rain. The details of this rain event are
provided in Figs 2(a)-(d). The temporal variation of
dBZ as estimated from MLP_ dBZ and multivariable
regression method, along with the observed values are
shown in Fig. 7(b). The rmse for the MLP_ dBZ
estimated values with respect to the observed values
was found to be 5.10 dBZ with a bias of -1.08 dBZ
and CC of 0.84. The rmse for the regression method
was found to be 5.44 dBZ and bias of 1.34 dBZ with a
CC of 0.69. It showed that the results obtained from
the MLP_dBZ model were reasonably improved
compared to that of regression model. Similarly, Fig.
7(a) shows the temporal variation of R as estimated
from MLP_R model and multivariable regression
model along with the observed values on this day.
With respect to observed values, the rmse for the
MLP _R estimated values was found to be 4.40
mmh-1
, with a bias and CC of 0.17 mmh-1
and 0.91,
respectively. In case of regression model, the rmse
value of 8.77 mmh-1
was found with a bias and CC of
3.21 and 0.33 mmh-1
, respectively. The observed and
MLP_R estimated total rain accumulation were found
to be ∼ 21 mm and ∼19 mm, respectively with an
error of ∼ 5.0 %. Here the rain accumulation from the
profiler measurement, i.e. MLP_R was
underestimated compared to rain accumulation from
JWD. Rajopadhya et al.34
, have reported an error,
between estimated and observed rain rate, of around
4% with a correlation coefficient of 0.76. In this
referred work the rain intensity was calculated from
the RDSD, which was estimated by fitting the Gamma
function to the Doppler spectra of UHF profiler at the
height of 1.9 km, by taking into account the vertical
clear air velocity.
Further, on 22 June 2000, the height profiles
of dBZ and R, as estimated by the proposed
Fig. 7―Temporal variation of estimated and observed (a) rain intensity (mmh-1) and (b) radar reflectivity factor (dBZ) [on 22 June 2000
obtained from regression and MLP method]
KONWAR et al.: RAIN PARAMETERS FROM WIND PROFILER USING MLP NETWORK MODEL
349
methodology are shown in Figs. 8(a)-(d) and 8(e)-(h),
respectively. The presented height profiles are at
different stages of convective and stratiform rain.
Figures 8(a) and (b) show the height profiles of dBZ
during the initial and mature stage of the convective
systems, respectively. During initial convective stage,
the reflectivity factor of around 32 dBZ were
observed up to ∼ 2.0 km, whereas during mature
stage the significant values of reflectivity factor of 40
- 50 dBZ were observed up to ∼ 4.5 km. Figures 8(c)
and (d) show the height profiles of dBZ during
different stages of stratiform rain. At the initial
stage of the stratiform rain, at 0154 hrs LT, the
presence of bright band is quite obvious at a height
of around ∼ 4.0 to 4.5 km by virtue of high value of
reflectivity gradient of ∼ 37 dBZ/km. At around 0216
hrs LT, at the later stage of the stratiform rain, the
bright band was not very prominent as concluded by
virtue of low value of reflectivity gradient of ∼ 12
dBZ/km at around same height [Fig. 8(d)].
Interestingly, the values of the estimated dBZ at the
lower heights are quite comparable with the JWD
observations at ground, which are provided at each
panel. Similarly, for the same day Figs 8(e) and (f)
show the model estimated height profiles of R during
the initial and mature stage of the convective systems,
respectively. During initial convective stage, the low
values of R of the order of 2-10 mm/h were observed,
whereas during the mature stage, high values of R, of
the order of 50-90 mm/h were observed up to ∼ 5 km.
Figures 8(g) and (h) show height profiles of R during
different stages of stratiform rain. During stratiform
situation, the overall values of estimated R were low,
i.e. of the order of 2-10 mmh-1
. At the bright band
level, there is an ambiguity of high values of R, which
require a correction for the realistic measurements of
these parameters at bright band level. During the
absence of strong bright band, this ambiguity is
reduced to certain extent [Fig. 8(h)]. Overall, the
values of the model estimated R at the lower heights
is reasonably comparable with the JWD observations
at ground, which are provided in each panel.
On 18 May 1999, a case study was carried out to
compare the height profiles of model estimated dBZ
and R with the Precipitation Radar (PR) observation
[on board of Tropical Rainfall Measuring Mission
(TRMM) satellite]. On this day, the TRMM satellite
passed nearly over the radar site during precipitating
time (TRMM orbit number 8463). For this
comparison, the 2A25 data product of PR was
utilized. The nearly coincident footprints of TRMM-
PR and L-band profiler were selected for analysis.
Fig. 8―Height profiles of MLP estimated-Z during (a), (b) convective and (c), (d) stratiform; and height profiles of R during
(e), (f) convective and (g), (h) stratiform, type of rain on 22 June 2000
INDIAN J RADIO & SPACE PHYS, OCTOBER 2008
350
The pixel resolution of PR is 16.0 km2 with a height
resolution of 250 km and a beam width of 1°. The
profiler and PR estimated height profiles of dBZ and
R are shown in Figs 9(a) and (b), respectively. As
these two systems have different height resolution,
therefore, the MLP and PR estimated rain parameters
are plotted at 150 and 250 m resolutions, respectively.
The important feature of this comparative study of
dBZ is the detection of a bright band at around
~ 3.7 km, by virtue of reflectivity gradient of 11
dBZ/ km signifying the presence of weak bright
band during stratiform precipitation [Fig. 9(a)].
Though the height profile of R from MLP_R is
comparable with PR measurements, it had slightly
underestimated rain rate compared to PR. For the
height profile of R, at the bright band level, the
ambiguity of rain rate estimation from MLP_R is
relatively less due to the presence of weak bright
band [Fig. 9(b)]. Though this comparison may not be
statistically very significant, overall good agreement
was found between the height profiles of model and
PR estimated dBZ and R.
5 Summary and discussions
Artificial Neural Network (ANN) based two MLP
models have been developed to estimate the dBZ and
R from the spectral moments of the L-band wind
profiler at NARL, Gadanki. The characteristics of the
spectral moments of the L-band wind profiler were
studied with respect to rain parameters. It was
observed that back scattered power P was most
closely related with these parameters by virtue of
higher value of exponent in the power law relations
between these two parameters. The fall velocity of
rain drops is most closely related with these
parameters in terms of high value of correlation
coefficient. The SNR parameter of the profiler,
though poorly correlated with rain parameters, was a
good indicator to represent rain or no-rain situations.
Overall good agreement between the estimated rain
parameters from the proposed methodology and the
JWD measured parameters were observed. It was also
found that the MLP model performed better compared
to parametric regression models. The estimated rain
parameters from the proposed methodology were also
able to show the general characteristics of the
convective and stratiform rain by virtue of its high
values of dBZ and R up to upper heights during
convective rain and the presence of bright band
during stratiform rain. It was noticed that during the
presence of strong bright band there is significant
error in the rain rate estimation at that height.
The ambiguity of the rainfall estimation at the
bright band is expected because different
microphysical processes dominate the precipitating
mechanisms. The MLPs estimated dBZ and R were
compared with the PR estimated height profiles over
Gadanki. Significantly the height profiles of dBZ
from both the systems have shown bright band
signature at around 3.7 km height. There are many
reasons to expect the discrepancy in the model
estimated and the observed rain parameters34
. There is
a limitation of sample volumes in the estimation of
height profiles of the rain parameters. In the present
study, the spectral moments at the lower most reliable
range bin (0.75 km) were trained with the JWD
Fig. 9―Comparison of height profiles of (a) Z; and (b) R (from PR observed and MLP estimated)
KONWAR et al.: RAIN PARAMETERS FROM WIND PROFILER USING MLP NETWORK MODEL
351
observations, which is essentially a point observation.
Now to further estimate the rain parameters at higher
range bins, which have increasingly different sample
volume, the errors in the estimation of height profiles
of these parameters are expected to increase. The
difference between the measurements from these two
systems could also be attributed to the ambient
atmospheric conditions such as humidity and
temperature, which determine the rate of evaporation
of the drops that fall from the retrieved heights to the
ground (JWD).
During thunderstorm or squall line seasons there
may be advection and convergence at the surface that
may have effect on the fall velocity of the
hydrometeor particles35
. It is important to mention
here that the TRMM ground validation field
campaigns36
were carried out by utilizing a standard
915 MHz profiler during Texas and Florida
Underflights Experiment (TEFLUN)-A and
TEFLUN-B in 1998. In that experiment the profiler
was calibrated with collocated RD-69 JWD at the
lowest range gates. In the present study, an L-band
wind profiler operated at 1.3 GHz has been utilized to
estimate the reflectivity and rain profiles with the help
of collocated RD-69 disdrometer. The study shows
that the MLP based techniques can be utilized to
estimate the rain parameters from the L-band wind
profilers with a reasonably good accuracy, though it
was originally developed to measure the clear air
velocity. A practical implication of the proposed
methodology is that it will contribute to the ongoing
effort to enhance the capability of the wind profiler
as a rain profiling system, when the clear air echoes
are masked by the precipitation echoes.
Acknowledgements Financial grant from Department of Space to carry
on this work [No. 10/4/362, under RESPOND
program] is gratefully acknowledged. Authors would
like to thank Director, NARL, Gadanki and
Coordinator SVU-UGC Centre for MST Radar
Applications for providing the necessary support to
conduct the experiment at radar site and subsequent
data analysis. The active support from the engineers at
NMRF during the experiments is thankfully
acknowledged. Authors are thankful to NASA-GSFC
for providing the TRMM data. Authors are thankful to
Kohima Science College authorities for providing
necessary facilities to carry out the research work.
Authors are grateful to the three anonymous reviewers
for their valuable comments and suggestions.
References
1 Kummerow C, Barnes W, Kozu T, Shiue J & Simpson J,
The Tropical Rainfall Measuring Mission (TRMM) sensor
package, J Atmos Ocean Technol (USA), 15 (1998) 809.
2 Battan L J, Radar observations of the atmosphere
(University of Chicago Press, Chicago III, USA), 1973,
p. 324.
3 Doviak R J & Zrnic D S, Doppler radar and weather
observations (Academic Press, USA), p. 458.
4 Seliga T A & Bringi V N, Potential use of radar differential
reflectivity measurements at orthogonal polarizations for
measuring precipitation, J Appl Meterol (USA), 15 (1976) 69.
5 Peters P, Fischer B & Anderson, Rain observations with a
vertically looking Micro Rain Radar (MRR), Borealenv Res
(Germany), 7 (2002) 353.
6 Fukao S, Wakasugi K, Sato T , Morimoto T, Tsuda T,
Hirota I, Kimura I & Kato S, Direct measurements of air and
precipitation particle motion by VHF Doppler radar, Nature
(UK), 316 (1985) 712.
7 Ecklund W L, Gage K S & Williams C R, Tropical
precipitation studies using 915 MHz wind profiler, Radio
Sci (USA), 30 (1995) 1055.
8 Ralph F M, Neiman P J & Ruffieux D, Precipitation
identification from radar wind profiler spectral moments
data: Vertical velocity histogram, velocity variance and
signal power – vertical velocity correlations, J Atmos Ocean
Technol (USA), 13 (1996) 545.
9 Rao T N & Rao D N, Tropical precipitating systems
observed with Indian MST radar, Radio Sci (USA), 34 (1999)
1125.
10 Wakasugi K, Mizutani A, Mastsuo M, Fukao S & Kato S, A
direct method for deriving drop size distribution and vertical
air velocities from VHF Doppler radar spectra, J Atmos
Ocean Technol (USA), 3 (1986) 623.
11 Wakasugi K, Mizutani A, Mastsuo M, Fukao S & Kato S,
Further discussion on deriving drop size distribution and
vertical air velocities directly from VHF Doppler radar
spectra, J Atmos Ocean Technol (USA), 4 (1987) 170.
12 Sato T, Doji H, Iwai H & Kimura I, Computer processing for
deriving drop size distribution and vertical velocities from
VHF Doppler radar spectra, Radio Sci (USA), 25 (1990)
961.
13 Currier P E, Avery S K, Balsley B B, Gage K S & Ecklund
W L, Combine use of 50 MHz and 915 MHz wind profilers
in the estimation of raindrop size distribution, Geophys Res
Lett (USA), 19 (1992) 1017.
14 Maguire W B & Avery S K, Retrieval of raindrop size
distributions using two Doppler wind profilers: Model
sensitivity testing, J Appl Meteorol (USA), 33 (1994) 1623.
15 Sharma S, Sarma D K, Konwar M, Das J & Jain A R,
Retrieval of raindrop size distributions from the L-band and
VHF wind profilers during convective and stratiform rain,
Indian J Radio & Space Phys, 37 (2008) 1850.
16 Ralph F M, Neiman P J, Kamp van de D W & Law D C,
Using spectral moment data from NOAA’s 404 MHz radar
wind profilers to observe precipitation, Bull Am Meteorol
Soc (USA), 76 (1995) 1717.
17 Rajopadhyaya D K, Avery S K, May P T & Ciefelli R C,
Comparison of precipitation estimation using single- and
dual- frequency wind profilers simulations and experimental
results, J Atmos Ocean Technol (USA), 6 (1999) 165.
INDIAN J RADIO & SPACE PHYS, OCTOBER 2008
352
18 Cifelli R, Williams C R, Ecklund W L, Rajopadhyaya D K,
Avery S K, Gage K S & May P T, Drop-size distribution
characteristics in tropical mesoscale convective systems, J
Appl Meteorol (USA), 39 (2000) 760.
19 Huser D & Amayenc P, A new method for deducing
hydrometeor size distribution and vertical air motions from
Doppler radar measurements at vertical incident, J Appl
Meteorol (USA), 20 (1981) 547.
20 Huser D & Amayenc P, Exponential size distributions of
raindrops and vertical air motions deduced from vertically
pointing Doppler radar using a new method, J Clim Appl
Meteorol (USA), 22 (1983) 407.
21 Reddy K K, Kozu T, Ohno Y, Jain A R & Rao D N,
Estimation of vertical profiles of rain drop size distribution
from the VHF wind profiler radar Doppler spectra, Indian J
Radio Space Phys, 34 (2005) 319.
22 Kumar N V P K & Rao D N, Deriving drop size distribution
from VHF and UHF radar spectra, Tenth International
Workshop on Technical and Scientific Aspect of MST radar
(Piura, Peru), 2003, pp. 278-281.
23 Haykin Simon, Neural networks: A comprehensive
foundation, 2nd edn (Pearson Education, USA), 2001, p. 156.
24 Xio R & Chandrasekar V, Development of neural network
based algorithm for rainfall estimation from radar
observations, IEEE Trans Geo Sci Remote Sens (USA), 35
(1997) 160.
25 Sarma D K, Konwar M, Das J, Pal S & Sharma S, A soft
computing approach for rainfall retrieval from the TRMM
microwave imager, IEEE Trans Geo Sci Remote Sens (USA),
43 (2005) 2879.
26 Hsu K, Gao X, Sorooshian S & Gupta H V, Precipitation
estimation from remotely sensed information using artificial
neural network, J Appl Meteorol (USA), 36 (1997) 1176.
27 Woodman R F, Spectral moment estimation in MST radars,
Radio Sci (USA), 20 (1985) 1185.
28 Anandan, V K, Balamurlidhar P, Rao P B & Jain A R, A
method for adaptive moments estimation technique applied
to MST radar echoes, Progress in electromagnetic research
symposium (Telecommunication Research Center, City
University of Hong Kong), 1996.
29 Sauvageot, H & Laucaux J P, The shape of averages drop
size distributions, J Atmos Sci (USA), 52 (1995) 1070.
30 Rao T N, Rao D N, Mohan K & Raghavan S, Classification
of tropical precipitating systems and associated Z–R
relations, J Geophys Res (USA), 106 (2001) 17699.
31 Rosenblatt F, The Perceptron: A probabilistic model for
information storage and organization in the brain, Psychol
Rev (USA), 65 (1958) 386.
32 Rumelhart D, Hinton G & Williams, Learning represent-
tations by back propagating errors, Nature (UK), 323 (1986)
533.
33 Sharma S, Konwar M, Sarma D K, Pal S & Das J, Estimation
of rainfall intensity and radar reflectivity factor from Lower
Atmosphere Wind Profiler utilizing artificial neural network
technique, Eleventh International workshop on Technical &
scientific aspects of MST radar, 2006, p. 186.
34 Rajopadhyaya D K, May P T, Cifelli R, Avery S K, Willams
C R, Ecklund W L & Gage K S, The effect of vertical air
motions on rain rates and median volume diameter
determined from combined UHF and VHF wind profiler
measurements and comparison with rain gauge measure-
ments, J Atmos Ocean Technol (USA), 15 (1998) 1306.
35 Cifelli R & Rutledge S A, Vertical motion structure in
maritime continent mesoscale convective systems: results
from a 50 MHz profiler, J Atmos Sci (USA), 51 (1994) 2631.
36 Gage K S, Williams C R, Johnston P E, Ecklund W L, Cifelli
R, Tokay A & Carter D A, Doppler Radar Profilers as
calibration tools for scanning radars, J Appl Meteorol (USA),
39 (2000) 2209.