Vladimir M. Krasnopolsky

General Sciences Corporation/SAIC,

Laurel, MD 20705

William H. Gemmill and Laurence C. Breaker

National Centers for Environmental Prediction

Washington, D.C. 20233

Submitted to Remote Sensing of Environment

April 1999

ABSTRACT

A new empirical multi-parameter SSM/I retrieval
algorithm based on the neural network approach which retrieves wind speed,
columnar water vapor, columnar liquid water, and sea surface temperature
simultaneously using only SSM/I brightness temperatures is compared with
existing global SSM/I retrieval algorithms. In terms of wind speed retrieval
accuracy, the new algorithm systematically outperforms all algorithms considered
under all weather conditions where retrievals are possible with an algorithm
rms error of 1.0 m/s under clear, and 1.3 m/s under clear plus cloudy conditions.
It also generates high wind speeds with acceptable accuracy. This improvement
in accuracy is coupled with increased areal coverage with obvious benefits
for operational applications. In terms of columnar water vapor and columnar
liquid water, the new algorithm reproduces the results of existing algorithms
closely. SST retrievals are less accurate and have low spatial resolution;
however, by including SST as an additional output, the accuracy in retrieving
wind speed is improved especially at high wind speeds. The new algorithm
has been tested and accepted for operational use at the NWS's National
Centers for Environmental Prediction. This algorithm produces a positive
impact on forecast winds through data assimilation into NCEP's numerical
weather prediction models.

1.0 INTRODUCTION

We evaluate an empirical neural network
(NN) algorithm which simultaneously retrieves several geophysical parameters
over the ocean from brightness temperature (BT) measurements acquired by
the Special Sensor Microwave Imager (SSM/I) (Hollinger *et al.*, 1987)
with primary emphasis on retrieving wind speed. This NN-based algorithm
(OMBNN3) which retrieves wind speed, columnar water vapor, and columnar
liquid water (see *http://polar.ncep.noaa.gov/marine.meteorology/marine.winds/*) has been recently
developed and is described elsewhere (Krasnopolsky et al., 1996, 1997,
1999). Here we present a comparison of this algorithm with other global
SSM/I wind speed, *W *(*m/s*), columnar water vapor, *V *(*mm*),
and cloud liquid water, *L *(*mm*), retrieval algorithms. We
also examine different BT retrieval flags and their relation to different
forms of atmospheric moisture, an important topic which is usually bypassed
in studies of this type. We also present various cross sections of wind
speed retrieval error analysis.

In Section 2, we describe existing global
SSM/I retrieval algorithms which were used in this study. In Section
3, our new multi-parameter empirical approach and the NN-based SSM/I retrieval
algorithm, OMBNN3 (the corresponding
FORTRAN code is available upon request at:
*Vladimir.Krasnopolsky@noaa.gov*),
are briefly described. In Section 4, we compare OMBNN3 with existing algorithms
and perform error analyses using different retrieval flags. Finally,
we discuss some operational applications of the geophysical fields which
are retrieved by the new algorithm^{(2)} and
present a recent example.

2.0 ALGORITHMS SELECTED FOR COMPARISON

For comparison we have selected: (i) the original global operational (cal/val) wind speed retrieval algorithm developed by Goodberlet et al. (1989) (GSW); (ii) the current operational algorithm (GSWP) which is the same as the GSW algorithm but corrected for water vapor (Petty, 1993); (iii) a physically-based (PB) retrieval algorithm (F. Wentz, 1997).

The GSW empirical wind speed retrieval algorithm, which is based on
multiple linear regression, uses a simple linear combination of four SSM/I
brightness temperatures to approximate the SSM/I wind speed transfer function
(TF):

*W _{GSW}* =

Because the actual SSM/I TF is nonlinear, several attempts have been made to improve the performance of the GSW algorithm using nonlinear approaches. The nonlinear water vapor correction, V, introduced by Petty (1993) is another type of linear regression which provides a nonlinear approximation for the transfer function with respect to SSM/I brightness temperatures. Nonlinear functions of BT, introduced in the linear regression in this case, represent the nonlinear behavior of the TF more closely:

(2)

Here *W _{GSW}* is given by (1) and the terms (-2.13 + 0.2198

*w = f ( T)* (3)

where *w* is wind speed, ** T** is a vector of SSM/I BTs,
and

The PB retrieval algorithm (Wentz, 1997) numerically inverts a physically-based
forward model. Therefore, it requires a physically-based forward model
as a necessary prerequisite which is not required for the empirical algorithms
described above, and, as a consequence, the PB algorithm requires a considerable
amount of empirical data for its development (a significantly larger amount
than the number of matchups normally used to develop competing empirical
algorithms). Since PB algorithms do not produce an explicit transfer function,
they require that a forward model to be inverted for each satellite measurement.
Symbolically, they can be represented as,

** g** =

where ** g** = {

3.0 A NEW MULTI-PARAMETER EMPIRICAL NN RETRIEVAL ALGORITHM

The NN algorithm presented here is a multi-parameter NN algorithm called
OMBNN3 (Krasnopolsky et al., 1996, 1997). It contains three new elements.
First, it is a multi-parameter retrieval algorithm. The co-variability
of related atmospheric and surface parameters which can be extracted from
the same set of brightness temperatures is taken into account: wind speed,
columnar water vapor, columnar liquid water, and SST are retrieved simultaneously,

** g** =

where ** g** = {

(6)

The horizontal bar above the symbols on
the left-hand side implies averaging over *V, L, *and* T _{S}*
which are not known for single-parameter algorithms,

Second, a method of NN training which enhances learning at high wind
speeds was used (Krasnopolsky et al., 1996). In preparation for training
the NN, the weights, *,* in the error function were generated using
the following formula,

where* p(W)* is the observed wind speed probability distribution
and *C* is a normalization constant. This choice of weights allows
us to assign higher values to the tails of the distribution and effectively
produces a distribution which is approximately uniform. By introducing
the square root in the denominator, we have restricted the rate of increase
in the weights which reduces noise-like influences at the highest wind
speeds.

Third, an extensive new buoy/SSM/I matchup database containing raw matchup
data sets for the F8, F10, and F11 SSM/I sensors (provided
by the Naval Research Laboratory, Washington, D.C.) and augmented
with additional matchup data for high latitudes containing a significant
number of high wind speed events (up to 26 m/s) from European Ocean Weather
Stations MIKE and LIMA (provided by D. Kilham,
Bristol University) , was used in developing this algorithm ( >
15,000 matchups overall). Both buoy and Ocean
Weather Station data include wind speed and SST; however, they do not include
columnar water vapor and columnar liquid water. These two parameters were
also required for training and validation of the new algorithm. Because
collocated buoy/radiosonde/SSMI data were not available for developing
this algorithm, we obtained columnar water vapor and columnar liquid water
using cal/val (Alishouse et al., 1990), and Weng and Grody (1994) algorithms.
Such single-parameter empirical algorithms, of course, only generate proxies
for the "true" data. These single-parameter retrievals may be biased and
do contain significant random errors which affect NN training and allow
only an approximate separation of the *V*- and *L*-signals from
the other components. However, as shown below (Table 1 and Fig. 2), even
such an approximate separation of signals allows us to reduce the random
and systematic errors in wind speed and reduce the dependence of the wind
speed bias on *V* and *L. *Because the
Weng and Grody (1994) algorithm for cloud liquid water utilizes
the 85 GHz channel, and this channel was not available from F8 during the
time of creating the matchup database, all comparisons which involve cloud
liquid water have been performed on a subset which includes only F10 and
F11 BTs (> 12,000 matchups). The OMBNN3 algorithm uses five SSM/I channels:
19 GHz and 37 GHz (horizontal and vertical polarizations) and 22 GHz (vertical
polarization).

From a mathematical point of view, the multi-parameter retrieval algorithm
(5) corresponds to a continuous mapping. It maps a vector of SSM/I BTs,
** T,**
onto a vector of retrieved geophysical parameters,

(7)

where the matrix *W _{ji }*and
the vector

4.0 COMPARISONS AND VALIDATIONS

4.1 **Wind speed**

SSM/I wind speed retrieval algorithms are
usually used together with so-called retrieval flags which are based on
various BT criteria. These retrieval flags serve as delimiters for BT over
the ranges of applicability for a given algorithm. They also are used to
indicate reliability and accuracy of retrieved wind speed. At least two
different sets of retrieval flags have been developed for SSM/I wind speed
retrievals, and they are presented in Table1 below:

**Table 1**. Two sets of retrieval flags
used for SSM/I wind speed retrievals: RF (Goodberlet et al., 1989) and
SF (Stogryn et al., 1994)

RF |
Criteria |
SF |
Criteria |

0 | T37V - T37H > 50 K
T19H < 165 K |
0 | T37V - T37H > 50 K |

1 | T37V - T37H < 50 K
T19H > 165 K |
1 | T37V - T37H 50 K
T19V < T37V T19H 185 T37H 210 K |

2 | T37V - T37H < 37 K | 2 | Everything else |

3 | T37V - T37H < 30 K |

The first set of criteria was developed by Goodberlet et al. (1989) and is closely related to the linear regression algorithm they developed. This set consists of four retrieval flags (RF0, ..., RF3). The second set was developed by Stogryn et al. (1994) for use with NN algorithms and consists of three flags (SF0, ..., SF2). These two sets of flags are similar but not identical. For obvious reasons we have adopted Stogryn's retrieval flags in our algorithm development. We excluded BTs which corresponded to flag SF2 (the "very cloudy" case) from consideration because, under these conditions, according to Stogryn, "wind speed effects on brightness temperatures are so overshadowed by atmospheric attenuation that retrievals should not be attempted". As a result, we have performed our algorithm development work, validation and comparisons only for "clear" (SF0) and for "cloudy" (SF1) conditions.

For wind speed, high quality ground truth
observations, including buoy and ocean weather station wind speed measurements,
are available to create matchups with the SSM/I brightness temperatures.
Evaluating the different algorithms was performed by comparing them with
ground truth observations. Table 2 shows the summary wind speed statistics
for clear (SF0) and clear + cloudy (SF0 + SF1) conditions. This table compares
four algorithms, GSW, GSWP, PB, and OMBNN3 using all available matchups
(> 15,000) for three SSM/I instruments F8, F10, and F11. The table presents
wind speed bias, RMS algorithm error, total RMSE, and RMSE for high wind
speeds ( *W* > 15 m/s). Additional analyses have been performed to
estimate the algorithm error itself (Wentz, 1997; Krasnopolsky, 1997).
The sensor and observation errors were estimated separately and removed
from the total RMSE. The residual, after removing the sensor and observation
errors, corresponds to the algorithm error per se. In terms of algorithm
error, the improved performance of the OMBNN3 algorithm (RMS errors of
1 m/s for clear-, and 1.3 m/s for clear+cloudy conditions) becomes even
more apparent. Table 2 also shows that the NN algorithm provides significant
improvement in retrieval accuracy at high wind speeds.

**Table 2**. Error budget (m/s) for
different SSM/I wind speed algorithms for clear, and clear+cloudy (in parentheses)
cases, and separately for higher wind speeds.

Algorithm |
Bias |
Algorithm RMSE |
Total RMSE |
W > 15 m/s RMSE |

GSW^{1} |
-0.2 (-0.5) | 1.4 (1.8) | 1.8 (2.1) | (2.7) |

GSWP^{2} |
-0.1 (-0.3) | 1.3 (1.6) | 1.7 (1.9) | (2.6) |

PB^{3} |
0.1 (-0.1) | 1.3 (1.8) | 1.7 (2.1) | (2.6) |

OMBNN3^{4} |
-0.1 (-0.2) | 1.0 (1.3) | 1.5 (1.7) | (2.3) |

^{1}Goodberlet et al., 1989; ^{2}Petty,
1993; ^{3}Wentz, 1997; ^{4}Krasnopolsky et al., 1996

Figure 2. Systematic (wind speed bias -
upper row) and random (standard deviation (SD) of wind speed error - lower
row) errors in wind speed retrieved by four different algorithms as functions
of columnar water vapor, *V*, (a and b); columnar liquid water, *L*,
(c and d); and SST (e and f). Solid line - OMBNN3, dotted line - GSWP,
dashed line - GSW, and dash-dotted line the Wentz (1997) algorithm.

As mentioned earlier, one of the advantages
of the OMBNN3 algorithm is its ability to retrieve not only wind speed
but
also three other parameters: columnar water vapor *V*, columnar liquid
water *L*, and *SST *simultaneously. The accuracies of these
additional retrievals are discussed below. Here we show how simultaneous
retrievals
of the entire vector of related geophysical parameters
(5) improves the accuracy of the wind speed retrievals by taking into account
the co-variability of these parameters. Fig. 2 shows the systematic and
random errors (bias and SD) in wind speed retrieval as functions of *V,L,
*and
*SST* for the GSW, GSWP, PB, and OMBNN3 algorithms. These statistics
were calculated using more than 12,000 matchups. Including the nonlinear
water vapor correction in GSWP reduces the bias and its dependence on water
vapor concentration (and partly on SST which is closely related to water
vapor); however, it does not reduce the dependence on liquid water concentration.
The OMBNN3 and PB algorithms, which both employ the simultaneous multi-parameter
retrieval approach, reduce the bias, and the dependence of the bias, on
both water vapor and cloud liquid water concentrations. The PB algorithm
does not retrieve SST but requires SST as an input parameter. The OMBNN3
algorithm also demonstrates the best performance with respect to SST co-variability.
The random errors for the OMBNN3 algorithm are significantly smaller and
demonstrate weaker dependencies on the related atmospheric and surface
(*SST) *parameters than do the errors for the other algorithms.

Figure 3. (a) Number of matchups corresponding
to different Figure 4. Same as for
Fig. 3 but only for

retrieval flags. Wind speed bias (b),
RMSE (c), and
standard
wind
speeds > 15 m/s.

deviation (SD) of wind speed differences
(differences

between buoy and SSM/I wind speeds) for
different retrieval

flags, and for two different retrieval
algorithms, GSWP

(dashed line) and OMBNN3 (solid line).

We have selected two algorithms, GSWP (currently
used operationally by Fleet Numerical Meteorological and Oceanographic
Center) and OMBNN3 (currently used operationally at National Centers for
Environmental Prediction) which demonstrate the best overall performance
in accordance with Table 2. Next, we perform a more detailed comparison
and error analysis for these two algorithms. Fig 3 shows such a comparison
for different retrieval flags (for both SF and RF, see Table 1). Panel
(a) shows the number of matchups for different flags (matchups corresponding
to SF2 were eliminated from the database at the start). Panels (b), (c),
and (d) show wind speed bias, RMSE, and standard deviation (SD) of the
error for different flags. Fig. 4 shows a similar comparison for high wind
speeds (*W* > 15 m/s). Flags RF0 and SF0 are similar (SF0 includes
a small portion of RF1 in addition to RF0) both in terms of the number
of matchups included, and the corresponding statistics. The SF1 flag includes
the rest of RF1 and portions of RF2 and RF3 which do not overlap SF2 (the
overlapping portions are not included).

Figure 5. Mean values of columnar liquid water, L, (a), and columnar water vapor, V, (b) corresponding to different retrieval flags.

Using a much more limited data set, Krasnopolsky
et al. (1995) demonstrated that BT retrieval flags are closely related
to the amount of cloud liquid water in the atmosphere. Fig. 5 further supports
this conclusion for a much larger data set. This figure shows mean amounts
of water vapor (*V*) and cloud liquid water (*L*) for each retrieval
flag. There is a strong correlation between the retrieval flags and *L*.
Retrieval flags RF2 and RF3 clearly correspond to higher levels of *L*
than does RF1. RF retrieval flags are more discriminative from this point
of view than the SF flags. However, Figs. 3 and 4 show that OMBNN3, which
outperforms GSWP under all conditions, demonstrates much better performance
than GSWP especially at SF1, RF2, and RF3 when the amount of cloud liquid
water increases significantly. Thus, we conclude that the SF flags are
well suited for use with the NN algorithms. Of considerable significance
operationally is the fact that the OMBNN3 algorithm has extended the retrieval
domain from clear (SF0 or RF0), to clear plus cloudy (SF0 + SF1)
conditions, yielding an increase in areal coverage of 15% globally. This
result is particularly important for obtaining more complete coverage of
synoptic weather systems such as extratropical cyclones which are typically
characterized by higher levels of moisture and wind speed. In such areas
(Fig. 6 below) the increase in areal coverage can exceed the global average
increase significantly.

4.2 **Columnar water vapor, liquid water,
and SST**

Because radiosonde observations which are
collocated with buoy and SSM/I measurements are not available, we have
used simulated data for *V* and *L* for training the OMBNN3 algorithm.
Values of *V* generated by the cal/val algorithm (Alishouse et al.,
1990) and values of *L* from the algorithm developed by Weng and Grody
(1994) (WG) were used. Thus, we do not consider OMBNN3 as a new algorithm
for retrieving *V* and *L* but rather consider them as ancillary
parameters which improve the accuracy of the wind speed retrievals. We
have performed comparisons for *V* obtained from the cal/val, PB,
Petty [1993] with the OMBNN3 algorithm. OMBNN3 agrees well with the other
algorithms for *V*, with an RMS difference of less than 2 mm. We also
compared values of *L* obtained from the WG and OMBNN3 algorithms.
These algorithms also compare favorably, yielding an RMS difference of
less than 0.02 mm. When new collocated buoy/radiosonde/SSM/I data become
available, the NN algorithm will be retrained to produce improved retrievals
for *V* and *L*.

SST is also an ancillary parameter which
improves the retrieval accuracy of wind speed. Because the dependence of
SSM/I BTs on SST is weak in the 19- to 37 GHz band, the accuracy and resolution
of retrieved SSTs is low compared with SSTs retrieved from other instruments
such as the Advanced Very High Resolution Radiometer.

4.3 **Operational data validation**

OMBNN3 has been carefully evaluated using operational data from the F10, F11, F13 and F14 SSM/I instruments since mid-1997. Positive impact has been demonstrated from the assimilation of wind speed and columnar water vapor retrieved by the NN algorithm into an operational numerical weather prediction (NWP) model (Yu et al., 1997). The wind speed, columnar water vapor and columnar liquid water fields retrieved by OMBNN3, taken together, depict significant information concerning weather patterns over the ocean (Gemmill and Krasnopolsky, 1999). Comparisons of these fields with other satellite and surface observations, and with various subjective (manual, human generated) and objective (computer generated using NWP models) analyses, show that the new algorithm generates high wind speeds (> 20 m/s) in areas where they occur. It is clear that this algorithm effectively separates the wind speed, columnar water vapor, and columnar liquid water signals contained in the SSM/I brightness temperatures. The multi-parameter retrievals, in addition to characterizing the instantaneous state of the atmosphere, preserve the correct physical relationships between the retrieved parameters. For example, there is a relation between areas with higher amounts of cloud liquid water and higher wind speeds , but they are not necessarily identical. The areas with the highest wind speeds do not usually correspond to the areas with the highest amounts of liquid water and vice versa. Higher amounts of columnar liquid water are related to areas of water vapor convergence which are closely associated with active fronts. Fronts and cyclonic patterns can be seen in both the SSM/I wind speed and columnar liquid water fields, and their locations are well supported by subjective and objective analyses. Conversely, the structure of the water vapor field differs from both wind speed and liquid water, and its higher values in this case are related to a moist air mass which most likely originates in the subtropics (see Gemmill and Krasnopolsky,1999) .

Figure 6. (a) SSM/I-derived ocean surface wind speed data, (b) SSM/I-derived columnar liquid water data, (c) SSM/I columnar water vapor data, (d) buoy and ship wind data, (e) ERS2 scatterometer wind vector data, and (f) sea level pressure analysis over Eastern North Pacific Ocean from 21 March 1998 at 20UTC. Each panel covers an area from 15 to 65 N and from 115 to 165 W.

Fig 6. shows an example of cross validation performed during a synoptically active period in the Eastern North Pacific on March 21, 1999. A series of six panels is shown: (a) SSM/I ocean surface winds speed, (b) SSM/I columnar water vapor, (c) SSM/I columnar liquid water, (d) ocean surface wind data from buoys and ship, (e) ERS2 scatterometer wind vector data, and (f) a sea level pressure analysis available from the NCEP Global Data Analysis System. The satellite data were acquired within a ± 3 hour window about the analysis time. The SSM/I data are a composite from three DMSP satellites (F11, F13 and F14), which together provide almost complete coverage of the region. There is an intense, low pressure system in the NE Pacific centered at 37 N and 131 W, with a central pressure of 983 mb. The SSM/I wind speed data show a broad area of strong winds that exceed gale force ( > 17 m/s) especially on the southern and western sides of the system. A swath of ERS-2 wind vector data shows a similar wind speed pattern south of the storm, with wind speeds again well above gale force. This area of high wind speed corresponds closely to the area of high sea level pressure gradients seen in the objective analysis (Fig. 6f). Both the SSM/I and ERS-2 data show the center of the storm itself. The SSM/I data suggest that the center is located near the area of wind speed minimum (7.5 m/s) just north of the sea level pressure minimum (Fig. 6f). The ERS-2 wind vectors locate the circulation center in about the same area. The buoy and ship wind data depict the general wind circulation pattern around the storm, but are much too sparse to identify details of the system. Two ships reporting winds above gale force, one near 38N, 126 W in front of the storm, agrees with the SSM/I wind speed pattern, and the other near 37 N, 132 W behind the storm, agrees with the SSM/I and ERS-2 wind patterns. The water vapor and liquid water patterns show that the storm is well mixed, i.e., with no strongly contrasting air masses. Higher values of water vapor occur at lower latitudes, with the highest values of liquid water being associated with the storm front approaching from the west.

After evaluating many cases in different oceanic regions, we find the overall agreement between the SSM/I derived parameter fields and those obtained from other sources to be excellent.

5.0 CONCLUSIONS

A new empirical, multi-parameter SSM/I retrieval algorithm based on the neural network approach (OMBNN3) is compared with other global SSM/I retrieval algorithms. This algorithm simultaneously retrieves wind speed, columnar water vapor, columnar liquid water, and SST using only SSM/I brightness temperatures. The accuracy of the wind speed retrievals from the new OMBNN3 algorithm (algorithm RMS error 1.0 m/s under clear, and 1.3 m/s under clear plus cloudy conditions) is systematically higher than the accuracy of wind speed retrievals for the other algorithms tested, for all SSM/I instruments. This improvement is valid under all weather conditions and for all wind speeds where retrievals are possible. The OMBNN3 algorithm has extended the areal coverage by 15% globally and even higher in areas of increased meteorological activity (storms and fronts) which have higher levels of moisture and wind speed .

The new algorithm successfully separates
the wind speed, columnar water vapor, columnar liquid water, and SST signals
contained in the SSM/I brightness temperatures. Multi-parameter retrievals
preserve the correct physical relationships between the retrieved parameters
by partitioning the variance among the variables in an appropriate manner.
The simultaneous retrieval of related atmospheric and surface parameters
is also beneficial when cloudy and very cloudy weather conditions are present.
Because the brightness temperature retrieval flags which we use are essentially
statistical (based on global statistics), they are not highly sensitive
to local conditions. In some cases this may lead to degraded retrievals;
therefore, any additional information about related local atmospheric and
surface conditions which can be derived from the same brightness temperatures
may improve these brightness temperature retrieval flags. Local atmospheric
and surface parameters (*V, L, *and *SST*), which the OMBNN3
algorithm now produces simultaneously with wind speed, help to describe
the instantaneous state of the atmosphere more completely, and, therefore
may help to improve the retrieval flags and to further improve the accuracy
of retrievals under cloudy conditions.

OMBNN3 represents a new generic approach
for developing multi-parameter empirical retrieval algorithms based on
the NN technique. This approach may find use in many other remote sensing
applications as well.

ACKNOWLEDGMENTS

We thank D.B. Rao for his thorough and critical review of this paper. Also, we want to thank those who provided us with expanded collocated SSM/I - buoy data sets; Gene Poe of the Naval Research Laboratory for providing a preliminary raw version of the new NRL matchup database, David Kilham of Bristol University for providing us with additional matchup data for high latitudes, and Michael McPhaden and Linda Magnum for providing information concerning TOGA-TAO buoys. Without this comprehensive data set, our results could not have been extended through high wind speeds.

REFERENCES

Alishouse, J.C., et al. (1990), Determination
of oceanic total precipitable water from the SSM/I. *IEEE Trans. Geosci.
Remote Sens., GE *23:811-816

Gemmil, W.H. and Krasnopolsky, V. (1998),
Weather patterns over the ocean retrieved by neural network multi-parameter
algorithm from SSM/I, In *Proceedings of the Fifth Internatio nal Conference
on Remote Sensing for Marine and Coastal Environment*, San Diego, California,
pp. I-395 - I-402, 5-7 October

Gemmil, W.H. and Krasnopolsky, V. (1999),
The use of SSM/I data in operational marine analysis, *Weather and Forecasting*,
in press

Chen, T., and Chen, H. (1995), Approximation capability to functions
of several variables, nonlinear functionals and operators by radial basis
function neural networks, *Neural Networks*, 6:904-910

- Universal Approximation to nonlinear operators by neural networks
with arbitrary activation function and its application to dynamical systems*,
Neural Networks*, 6:911-917

Goodberlet, M.A., Swift C.T., and Wilkerson,
J.C. (1989), Remote sensing of ocean surface winds with the special sensor
microwave imager. *JGR*, 94:14574 - 14555.

Hollinger, J., et al. (1987), Special Sensor Microwave/Imager user's guide, Technical Report, 120 pp, Nav. Res. Lab., Washington, D.C.

Krasnopolsky, V., Gemmill, W.H., and Breaker, L.C. (1996), A New transfer function for ssm/I based on an expanded neural network architecture. Technical Note, OMB contribution No. 137, NCEP/NOAA

Krasnopolsky, V., Gemmill, W.H., and Breaker, L.C. (1997),
Ocean surface retrievals from the SSM/I using neural networks, In
*Proceedings
of the Fourth International Conference on Remote Sensing for Marine and
Coastal Environment*, Orlando, Florida, p. II-164, 17-19 March

Krasnopolsky, V. (1997), Neural networks for standard and variational satellite retrievals. Technical Note, OMB contribution No. 148, NCEP/NOAA

Petty, G.W. (1993), A comparison of SSM/I algorithms for the estimation
of surface wind,* Proceedings Shared Processing Network DMSP SSM/I Algorithm
Symposium*, 8-10 June

Weng, F., and Grody N.G. (1994),
Retrieval of cloud liquid water using the special sensor microwave
imager (SSM/I). *J. Geophys. Res., *99:25,535-25,551

Weng, F., et al. (1977), Cloud liquid
water climatology from the special sensor microwave/imager. *J. Clim.*,
10:1086-1098

Wentz F.J. (1997), A well-calibrated ocean algorithm for special sensor
microwave / imager, *JGR*, 102:8703-8718

Yu, T.-W., Iredell, M.D. and Keyser, D.
(1997), Global data assimilation and forecast experiments using SSM/I wind
speed data derived from a neural network algorithm.* Weather and Forecasting*,
12:859-865

1. OMB Contribution No.^{
159}

^{2.
}Since
April 1998, OMBNN3 has been running as the operational SSM/I retrieval
algorithm in the data assimilation system at the National Centers for Environmental
Prediction.