P2.9
SATELLITE DATA ESTIMATION OF OCEAN LATENT HEAT FLUX;
A NEURAL NETWORK APPROACH
Charles Jones, Pete Peterson and Catherine Gautier
Institute for Computational Earth System Science (ICESS)
University of California
Santa Barbara, CA 93106-3060
1. INTRODUCTION
The coupling between the ocean and the atmosphere over climatic time scales
poses an interesting challenge for a thorough understanding of the climate
system. The exchange of energy, momentum and mass between the two sub-systems
has to be continually monitored in order to provide adequate data sets to
detect possible climate changes. In this regard, latent heat flux from the
ocean's surface, as one of the dominant terms in the surface energy balance,
plays a key role in the energy transfer between the atmosphere and ocean
as well as in the hydrological cycle. Nevertheless, the global observation
of surface latent heat flux is a difficult problem, and long records of
in-situ estimation are usually limited to ship tracks of Volunteer Observing
Ships (VOS) and mooring buoys.
The uniform spatial and temporal sampling of satellite measurements provide
a unique way to remotely derive surface latent heat fluxes over open oceans
and complement in-situ observations. The main difficulty with the satellite
method however is the estimation of the near surface specific humidity and
air temperature. The pioneer work of Liu (1986) (hereafter L86) derived
an empirical relationship between the total precipitable water obtained
from satellite microwave measurements and near surface specific humidity
from radiosonde measurements. Using this empirical relation along with satellite
derived surface wind speeds and sea surface temperature (SST), the determination
of surface latent heat flux over remote areas in the open oceans became
feasible (Liu, 1988). In spite of improvements obtained by recent studies
(Jourdan and Gautier, 1995, hereafter JG95; Chou et al., 1995), there still
exist significant systematic and random differences between satellite and
in-situ estimations of surface latent heat fluxes and these arise primarily
from uncertainties in the near surface moisture field (Esbensen et al.,
1993). This paper describes a new methodology to derive surface specific
humidity (Qa) and air temperature (Ta) from satellite data which provides
a much better agreement with ship observations than earlier studies. We
limit the discussion here to the derivation and validation of Qa and Ta.
Work is in progress to improve the preliminary methodology and assess the
impact on the estimation of surface latent heat flux.
2. DATA
The model developed to derive Qa and Ta uses as input parameters total precipitable
water (W) and sea surface temperature (SST). A total of 6 years (1988-1993)
of monthly averages of W from the Special Sensor Microwave Imager (SSM/I)
(Wentz, 1992) and SST from the National Center for Environmental Protection
(NCEP) optimal interpolation reanalysis scheme (Reynolds and Smith, 1994)
are used at 2°x2° latitude/longitude resolution. Monthly averages
of Qa and Ta from the Comprehensive Ocean-Atmosphere Data Set (COADS) (Woodruff
et al., 1987) for the same time period and spatial resolution are used to
develop and later validate our model.
3. METHODOLOGY
The model development involves a series of steps which are briefly summarized.
First, the initial data sets of W (SSM/I), SST (NCEP), Qa and Ta (COADS)
are divided into two sub-samples. A two twelve months period, August 1988-July
1989 and August 1992-July 1993, called testing sample is withdrawn from
the larger sample and is later used for validation purposes. This period
is chosen to include two contrasting years, i.e. La Niña and El Niño
events, respectively, and therefore can further test the validity of our
algorithm under different climatic regimes. The remaining 48 months from
the 6 years form the training sample and are used to develop the model components
for Qa and Ta.
The model component for Qa is done in the following way. We first extract
from the training sample W, SST and Qa from all grid cells in the analysis
domain (all longitudes from 70°S to 70°N) in which the number
of COADS observations to estimate the monthly Qa are greater than 20 per
month. This is necessary to ensure that the COADS monthly averages of Qa
used to develop the model do not suffer from undersampling (see JG95 for
details). The pairs of observations W/SST are then classified into six classes
using a K-Means cluster algorithm (Hartigan, 1975) where each class contains
approximately 11,000 W/SST observations. Figure 1 shows the scatterplot
of W and SST as well as the six K-Means classes. These observations span
a large range of W and SST values and represent different physical regimes.
For example, observations falling in class 1 are in general representative
of mid-latitude conditions, i.e. low SST and W values and consequently low
Qa. In contrast, observations in class 6 are typically characteristic of
tropical conditions with high SST and high W and Qa as in the western Pacific
warm pool. Further investigation indicate that the six classes closely follow
on average the climatology of Qa (not shown). The next step involves the
random extraction of 1,000 W, SST and Qa observations from each class and
the development of six supervised 2-6-2-1x Artificial Neural Networks (ANN),
i.e. one ANN for each class (see Hertz et al., 1991 for details on ANN).
Each ANN takes two inputs, W and SST, contains two hidden layers with 6
and 2 nodes, respectively, and is trained to find the target Qa.
Figure 1
The model component to derive Ta is performed following the methodology
of the Qa model component. The final procedure results in six other classes
of W/SST and six other supervised 2-6-2-1x ANN's that takes W and SST as
inputs and are trained to find the Ta target. Our preliminary results show
that the six classes defined by the K-Means algorithm, as well as the ANN
training with 1,000 observations in each class, offer an efficient computational
procedure to derive Qa and Ta from W and SST. One of the main advantages
of our procedure is that the inputs W and SST have global coverage consistent
with high temporal sampling and captures the physical relationships between
W, SST and Qa.
4. RESULTS
To validate the model components for Qa and Ta, we applied the model to
the testing sample, August 1988-July 1989 and August 1992-July 1993, and
compared against COADS observations. In addition, we compared our estimates
with other empirical parameterizations so that improvements upon previous
methodologies can be assessed. The validation of Qa is compared against
the W-Qa empirical relationship of L86, while the Ta comparison is done
against the W-Ta relation used by JG95.
Figure 2
Figure 3
Figure 2a shows the mean bias between COADS Qa and ANN Qa, while Fig. 2b
displays COADS Qa minus L86 Qa. Evident is the much smaller mean bias in
Qa derived from ANN (on average less than 1 g kg-1), even in regions of
very low COADS sampling such as in the southern oceans. Figures 3a and 3b
show the root-mean-square differences (rms) between COADS and ANN and L86
respectively, as well the average COADS sampling for this time period (Fig.
3c). Very small rms values (less than 0.8 g kg-1) are observed between Qa
ANN and Qa COADS over regions of very high COADS sampling and low Qa values
such as the north Pacific and Atlantic Oceans. In other regions the ANN
Qa rms is slightly higher, but still less than 1.2 g kg-1. A substantial
improvement is observed when compared against L86 relation (Fig. 3b), specially
in the Arabian Sea and the southern hemisphere oceans, a problem discussed
in Esbensen et al. (1994).
The mean bias between COADS Ta and ANN Ta, Fig. 4a, is less than ±1°
C over most regions, while the mean bias between COADS Ta and LJ95 Ta is
much higher than ± 1° C. In some regions the mean bias in LJ95
Ta can be as high as 3-5° C as over the Arabian Sea and the southeast
Pacific and Atlantic Oceans. The rms between ANN Ta and COADS Ta, Fig. 5a,
is fairly homogeneous over most regions and on the order of 0.6-1.2°
C. In contrast, the rms between COADS Ta and JG95 Ta is drastically higher
reaching values as high as 3-4° C, even in regions of high COADS sampling
(Fig. 5c). Table 1 summarizes the global rms for Qa and Ta using the ANN
model considering all COADS grid cells with the number of observations greater
than 20 per month.
5. CONCLUSIONS
This paper presents a new method to derive Qa and Ta from satellite data
and NCEP reanalysis. The major strength of our methodology can be summarized
as follows. We use two inputs, W and SST, that are available globally with
consistently high spatial and temporal resolution. The initial development
of the model is tuned to regions with high number of observations (Qa and
Ta) over the ocean, while other methodologies use radionsonde measurements
(L86, Chou et al., 1995) over island continents. In addition, our classification
of W/SST separates different physical regimes, i.e. mid-latitudes versus
tropical regions, while other methods use a global relationship (e.g. L86).
This new method also improves our previous research to develop global fields
of surface latent heat fluxes (JG95), which blends Qa and Ta from satellite
estimations with COADS observations. The major weakness of such procedure
is that the improvement in the blended surface latent heat flux is limited
to the areas of high ship sampling. This new method has the potential to
develop global surface latent heat fluxes with more spatially homogeneous
accuracy. This can be readily seen in the improvements obtained in the Qa
and Ta derived from ANN over the southern oceans.
We are currently working to improve the method described here to assess
the impact that Qa and Ta derived from our model have on global surface
latent heat flux. Preliminary results show that surface latent heat fluxes
derived with Qa and Ta from the ANN model gives a significantly better agreement
with COADS observations. It should be finally noted that systematic and
random errors are present in any data set including in-situ observations
such COADS. An improved satellite estimation of surface latent heat flux,
which includes high temporal and spatial sampling, can be a valuable resource
to complement other data sets such as in-situ measurements and reanalysis
products.
Figure 4
Figure 5
Table 1
ACKNOWLEDGMENTS
The authors would like to specially thank Dr. Tim Liu for his continuing
support, and Don Tveter for providing the neural network computer code.
The authors also greatly acknowledge data support provided by the National
Center for Atmospheric Research (NCAR), which is sponsored by the National
Science Foundation (NSF). This work is supported by National Science Foundation
research grants ATM9319483 and National Aeronautics and Space Administration
grant NASA-JPL959177.
REFERENCES
Chou, S., R. M. Atlas, C. Shie, and J. Ardizzone, 1995: Estimates of surface
humidity and latent heat fluxes over oceans from SSM/I data. Mon. Wea. Rev.,
123, 2405-2425.
Esbensen, S. K., D. B. Chelton, D. Vockers, and J. Sun, 1994: An analysis
of errors in Special Sensor Microwave Imager evaporation estimates over
the global oceans. J. Geophy. Res., 98, 7081-7101.
Hartigan, J. A., 1975: Clustering algorithms. John Wiley & Sons, New
York, 351 pp.
Hertz, J., A. Krogh, and R. G. Palmer, 1991: Introduction to the theory
of neural computation. Addison Wesley Pub. Co., New York, 327 pp.
Liu, W. T., 1986: Statistical relation between mean precipitable water and
surface level humidity over the global oceans. Mon. Wea. Rev., 114, 1591-1602.
________ 1988: Moisture and latent heat flux variabilities in the tropical
Pacific derived from satellite data. J. Geophys. Res., 93, 6749-6770.
Jourdan, D., and C. Gautier, 1995: Comparison between global latent heat
flux computed from multisensor (SSM/I and AVHRR) and from in situ data.
J. Atmos. Oceanic Technol., 12, 46-72.
Reynolds, R. W., and T. M. Smith, 1994: Improved global sea surface temperature
analyses using optimum interpolation. J. Clim., 7, 929-948.
Wentz, F. J., 1992: User's manual SSM/I monthly ocean tape. Remote Sensing
Tech. Rep., 20 pp.
Woodruff, S. D., R. J. Slutz, R. L. Jenne, and P. M. Steurer, 1987: A Comprehensive
Ocean-Atmosphere Data Set. Bull. Amer. Meteor. Soc., (68), 1239--1250.
_____________________________________________________________________________
Corresponding author address: Dr. Charles Jones, ICESS, University of California,
Santa Barbara, CA 93106-3060; E-mail: cjones@icess.ucsb.edu