Directory UMM :Data Elmu:jurnal:A:Atmospheric Research:Vol55.Issue1.Sept2000:
Atmospheric Research 55 Ž2000. 65–83
www.elsevier.comrlocateratmos
Determination of cloud microphysical properties
from AVHRR images: comparisons of
three approaches
Anne Fouilloux a,) , Jean-François Gayet a , Karl-Theodor Kriebel b
a
Laboratoire de Meteorologie
Physique, UPRESA 6016 CNRS, UniÕersite´ Blaise Pascal, Aubiere,
´´
` France
b
Institut fur
¨ Physik der Atmosphare,
¨ Deutsche Forschungsanstalt fur
¨ Luft-und Raumfahrt,
Oberpfaffenhofen, Germany
Received 8 January 1998; accepted 1 March 2000
Abstract
This contribution to the EUCREX mission 206 series aims at evaluating the performances of
three methods considered for the retrieval of cloud microphysical properties Žoptical depth, droplet
effective radius and liquid water path ŽLWP.. from Advanced Very High Resolution Radiometer
ŽAVHRR. satellite images. These procedures, namely a set of empirical parameterizations, a
physically based inversion code, and a neuronal approach, are tested with the mission 206 case
study. The influence of horizontal averaging on the retrieved cloud properties is also investigated
to assess the limitations of the comparison between satellite observations and aircraft measurements performed at a smaller scale. q 2000 Elsevier Science B.V. All rights reserved.
Keywords: Cloud radiative properties; Remote sensing retrieval techniques
1. Introduction
For usual satellite remote sensing applications, the classical inversion problem
consists of extracting pertinent parameters that characterize the medium from a set of
)
Corresponding author. LIVIC, Bat 140, 13 route de la Miniere,
` 78000 Versailles, France. Tel.: q33-1-4043-29-07; fax: q33-1-40-43-29-30.
E-mail address: [email protected] ŽA. Fouilloux..
0169-8095r00r$ - see front matter q 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 9 - 8 0 9 5 Ž 0 0 . 0 0 0 5 7 - 0
66
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
radiative measurements, namely the radiances sampled in several illumination and
observation conditions andror in different spectral bands. As part of meteorological and
climate studies, the correct interpretation of such observations in terms of cloud
microphysical properties requires the development of tools that are in the meantime
powerful, robust and fast ŽArking and Childs, 1985; Rossow, 1989..
Among the various retrieval procedures that were developed for deriving cloud
optical thickness, liquid water path ŽLWP. or droplet effective radius ŽKriebel et al.,
1989; Nakajima and King, 1990; Parol et al., 1991; Iaquinta and Pinty, 1997., three
kinds of approaches emerged: empirical parameterizations derived from in situ measurements, physical methods using optimization techniques or look-up tables, and a new
generation of algorithms built with neuronal networks. These techniques have often been
validated for homogeneous clouds, but their performances on heterogeneous situations
are less documented.
The aim of the present paper is to retrieve values of cloud optical thickness, droplet
effective radius and LWP within the stratocumulus observed on 18 April 1994, during
the European Cloud and Radiation Experiment ŽEUCREX. mission 206. This case is
particularly interesting since the cloud system was rather homogeneous over land
surfaces, but progressively broken towards the North, over the Atlantic ocean. This case
thus provides an opportunity to test the three retrieval procedures over a large range of
heterogeneity of the cloud structure. The methods tested here are: Ži. the AVHRR
Processing Scheme Over Clouds, Land and Ocean ŽAPOLLO. parameterizations of
Kriebel et al. Ž1989.; Žii. the Cloud Retrieval ŽCRTVL. inversion code developed by
Nakajima and Nakajima Ž1995.; and Žiii. the Adjustable Combination of Neuronal
Networks ŽACNN. algorithm recently elaborated by Fouilloux and Iaquinta Ž1998..
Parameterizations specifically developed for the determination of cloud microphysical
properties is an attractive solution since the method is straightforward and does not
necessitate tedious computations. However, care must be taken to avoid using the
parameterizations on situations that may differ significantly from those for which the
empirical formulas have been developed.
Inversion techniques, such as CRTVL, require a preliminary modeling step and their
accuracy depends upon the efficiency of the method at recovering the optimal fit
between the actual radiances and those predicted by the radiative transfer model
ŽNakajima and Tanaka, 1986.. In order to avoid laborious and repeated calculations at
each iteration, it is advantageous to keep only relevant information into look-up tables
ŽNakajima and King, 1990.. Therefore, the intrinsic quality of the retrieval depends on
the model accuracy, but it is also strongly conditioned by the resolution of the look-up
table in the parameter space.
The third method is based on a combination of self-organised neuronal networks
ŽFouilloux and Iaquinta, 1998. and performs a classification where the accuracy to be
achieved is one of the parameters driving the procedure. That allows to explicitly design
the network by increasing the number of classes, according to the desired accuracy.
In the following sections, the satellite data available on 18 April 1994 are described,
the three methods are briefly presented, and their performances at retrieving the
stratocumulus cloud microphysical characteristics are evaluated.
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
67
2. Description of the retrieval schemes
2.1. Satellite data
The radiance data considered hereafter are taken from the Advanced Very High
Resolution Radiometer ŽAVHRR. instrument mounted on board the National Oceanic
and Atmospheric Administration ŽNOAA.-11 satellite.
The NOAA-AVHRR satellite image of 18 April 1994 was recorded at around 08:25
UTC, while in situ measurements with the instrumented aircraft were performed
between 09:30 and 12:00 UTC ŽPawlowska et al., 2000b.. The following analysis is
conducted in five spectral bands ŽTable 1. with a spatial resolution of about 1.1 km. Fig.
1 shows the satellite image in the first visible channel. The bordered area of 300 = 300
pixels is the region of interest for the present study and the white straight line represents
the aircraft sampling track ŽPawlowska et al., 2000b.. Fig. 1 reveals that the cloud
system is more heterogeneous towards the northwest of the sampling leg, above the
Atlantic Ocean than towards the southeast, over the coast of Brittany.
The calibration procedures applied to this data follow standard techniques. In
particular, since the visible channel of the sensor has been degrading from the pre-launch
condition, the corrected calibration constants of Kaufman and Holben Ž1993. have been
used to obtain the radiances from the satellite-recorded digital counts.
2.2. APOLLO parameterizations for cloud microphysical properties
The first step in this approach consists in making a cloud classification ŽSaunders and
Kriebel, 1988. to identify pixels, which are fully cloudy. These are then categorized into
thick and thin clouds. The thick clouds are then divided into low, medium and high
clouds according to their temperature. Thick clouds are treated as water clouds. For the
second step, the optical thickness and LWP are computed on a pixel-by-pixel basis using
parameterizations according to Stephens Ž1978. and Stephens et al. Ž1984. Žsee also
Kriebel et al., 1989.. They are based on the observation that the radiation fields are the
most sensitive to cloud optical thickness in the visible window Ž0.4 to 0.8 mm. ŽCurran
and Wu, 1982; Arking and Childs, 1985; Nakajima and King, 1990.. The formulas are
thus valid in the solar spectral range, at wavelengths below 0.7 mm, and they are used
Table 1
Characteristics of the AVHRR sensor
AVHRR channel
Wavelength
Domain
1
2
3
4
5
0.58–0.68 mm
0.725–1.0 mm
3.55–3.93 mm
10.3–11.3 mm
11.5–12.5 mm
Visible
Visible
Near-infrared
Infrared
Infrared
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
68
Fig. 1. Visible channel of the NOAA-AVHRR image at 08:24 UTC on 18 April 1994. The image encompasses
the stratocumulus cloud area covered by the AVHRR measurements. The area of interest is indicated by a
white frame and the leg flown by the instrumented aircraft is marked by a white segment.
with the AVHRR channel 1 data. The directional–hemispherical cloud reflectance in
Channel 1, R c Ž m 0 ., is related to cloud optical thickness t by the following relationship:
Rc s
b Ž m 0 . t Ž m 0 . rm 0
1 q b Ž m 0 . t Ž m 0 . rm 0
Ž 1.
where b Ž m 0 . is the modified backscattering coefficient given for water clouds with
optical thickness ranging from 1 to 500 and m 0 is the cosine of the solar zenith angle.
The values of b Ž m 0 . that are considered hereafter are listed in the paper of Stephens et
al. Ž1984..
The cloud LWP Žin g my2 . is derived from the optical thickness by the following
formula ŽKriebel et al., 1989.:
log Ž LWP . s Ž 0.5454 = t .
0.234
Ž 2.
Fig. 2 shows the dependence between directional–hemispherical cloud reflectance
R c Ž m 0 . and optical thickness obtained with the APOLLO parameterization for the
clouds observed on 18 April 1994, within the area indicated in Fig. 1. Note that there are
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
69
Fig. 2. Optical thickness values retrieved with the APOLLO code as a function of the actual directional
hemispherical cloud reflectance, R c Ž m 0 ..
no values of R c Ž m 0 . below 0.3 because such pixels are not classified as water clouds
Žthick clouds. and are therefore eliminated.
2.3. CRTVL inÕersion code
The CRTVL inversion code ŽNakajima and Nakajima, 1995. has been implemented
to simultaneously retrieve optical thickness and droplet effective radius of cloudy pixels
from the NOAA-AVHRR satellite images. This algorithm computes the microphysical
quantities from the radiances at visible Ž0.58 to 0.65 mm. and near-infrared Ž3.55 to 3.93
mm. wavelengths by making use of look-up tables that saves computational time, yet
conserving the accuracy of the analysis. For the simulation of the satellite measured
radiances, we applied the radiative transfer model developed by Nakajima and Tanaka
Ž1986. with an atmosphere typical of winter oceanic mid latitudes. The number of
vertical levels are set to four Žwith boundaries at 1, 2, 12 and 20 km.. That does not
particularly alter the results and drastically reduces the computational time. According to
aircraft observations, the cloud layer was set at 1 km, and we assumed a Lambertian
surface Žalbedo of 0.06 . for the underlying oceanic surface.
Fig. 3 shows the simulation of reflected solar radiances in the AVHRR spectral bands
1 and 3 as a function of cloud optical thickness t and droplet effective radius reff , for a
70
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
Fig. 3. Simulation of reflected solar radiances in AVHRR channels 1 and 3 as functions of cloud optical
thickness Žt s1, 2, 4, 8, 16, 32, 64. and effective radius Ž reff s 2, 4, 8, 16, 32., with Ž u 0 s608, u s 408 and
w s 508.. The isolines of optical thickness and effective radius are represented by solid and dashed lines,
respectively.
solar zenith angle of 608, a satellite zenith angle of 408 and a relative azimuth of 508.
Near-vertical and near-horizontal lines illustrate iso-optical thickness and iso-effective
radius radiances, respectively. Ground-reflected and thermal radiations are not taken into
account, and the LOWTRAN-7 model associated with a mid-latitude model is considered to properly take into account the gaseous aerosol absorption ŽNakajima and King,
1990.. The principle of the inversion procedure is deduced from this figure as follows:
The t value is sought so as to minimize the difference between the actual and theoretical
Žstored in the look-up tables. cloud-reflected radiances in the first channel Žthe threshold
being 0.1%.. Afterwards, t is set to this value and reff is modified until the difference
between measured and tabulated radiances in the third channel becomes less than 0.1%.
This process is repeated until both differences Žin the visible and near infrared channels.
become lower than 0.1%, which usually requires three to four iterations. LWP Žg my2 .
is then calculated from the droplet effective radius and cloud optical thickness values by
using the following equation:
2
LWP s rt reff
Ž 3.
3
where r is the liquid water density Žg my3 ..
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
71
2.4. ACNN neuronal network
The main problem with the previous approaches is that, with APOLLO, the determination of optical thickness and LWP is fast, but rather inaccurate, while with CRTVL,
the retrieval requires large computer storage because of the Žhuge. look-up tables. A
compromise between accuracy and fastness seems to be possible through a combination
of neural networks as described in Fouilloux and Iaquinta Ž1998..
The principle of the method is as follows: The determination of optical thickness and
droplet effective radius is split into several steps, each being dedicated to a particular
task. The first step consists in classifying the pixels of an AVHRR image into cloud
categories. Actually, the same algorithm as in APOLLO has been used for cloud
classification, for a better comparison between the two methods. The second step allows
differentiation of the pixels as a function of the geometry of illumination and observation. One or two further steps Ždepending on the desired accuracy. are then required for
the determination of the retrieved parameters. They are performed with self-organized
neuronal networks with two layers Žfor the entry and for the output cells. so as to let the
system converge near a stable configuration corresponding to the required degree of
accuracy. The accuracy depends on the quantity of information that the entry vectors
convey and also on a ‘‘ vigilance’’ factor Žalso called the accuracy factor. j Ža value
between 0 and 1. fixed by the operator. If this factor is close to 1, the system is forced to
solve with a great accuracy and inversely, when j tends to 0, the system is relaxed to a
lower accuracy. Nevertheless, the accuracy of the restitution depends also on both the
radiative transfer model used and the degree of noise included in the initial data.
The inversion process is divided into two stages: a learning phase during which the
connection weights of the neuronal networks are fixed and a generalization phase that
consists in using the neuronal network with actual remotely sensed data.
Regarding the learning phase, several synthetic databases, containing radiances in the
two visible and near infrared AVHRR channels for various values of the illuminationr
observation geometry and cloud parameters are generated by running the radiative
transfer model ŽNakajima and Tanaka, 1986.. In order to limit the size of these learning
data sets Ži.e., the range of possible values for the retrieved variables., a sorting is
performed by means of a ‘‘rough’’ cloud classification of AVHRR imagery. Such a
discrimination considers the algorithm proposed by Saunders and Kriebel Ž1988., which
relies on five tests conducted on the two visible and two infrared AVHRR channels.
Four categories have been identified, for which specific distributions of cloud optical
thickness and droplet effective radius were defined. On 18 April 1994, most of the
cloudy pixels are recognized as low clouds, which is in agreement with the aircraft
observations ŽPawlowska et al., 2000a.. For this cloud category, the following characteristics for the droplet size distribution have been selected: a number concentration of 350
cmy3 , a mean radius of 4 mm Žthe size distribution ranges from 0 to 20 mm. and a LWC
value of 0.09 g my3 ŽCotton and Anthes, 1989.. These values are derived from a large
set of in situ data made during several experiments ŽCotton and Anthes, 1989. and they
are considered as typical of mid-latitude stratocumulus clouds. The selected droplet size
distribution is similar to the distributions measured in situ during mission 206 ŽPawlowska
et al., 2000a..
72
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
The first neural network is specially designed to discriminate the pixels as a function
of the geometry of illumination and observation. These parameters have been added as
Fig. 4. Optical thickness values issued from the ACNN classification vs. values fixed in the radiative transfer
model, with j s 0.5 in Ža. and j s 0.8 in Žb..
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
73
neural network inputs to visible and near-infrared radiances and the accuracy parameter
j has been set to a value of 0.7 in order to distinguish 25 classes. A second network
allows to determine cloud optical thickness and droplet effective radius. Fig. 4a and b
shows the sensitivity of the neuronal network to values of optical thickness and of the
parameter j Žtwo values — 0.5 and 0.8 — are used.. It appears that in both cases, the
steps clearly depend on optical thickness, and that the accuracy is improved when j
increases Žthe various steps are less pronounced.. Similar results are found with the
droplet effective radius when using a learning database sensitive to this parameter. LWP
is finally computed using Eq. Ž3..
2.5. Accuracy assessment
The application of APOLLO parameterizations is straightforward and very fast, but
the uncertainty on the estimated parameters is of the order of 35%. Nevertheless, these
parameterizations have been extensively validated against in situ measurements ŽKriebel,
1989; Kriebel et al., 1989; Kaestner et al., 1993.. The CRTVL inversion code, which
provides an accuracy of about 10% Žwith synthetic data., has been tested against actual
measurements and no systematic bias was found. The uncertainties were however not
systematically computed. The ACNN and CRTVL inversion procedures have been
tested on a large set of synthetic clouds of various types, and a very good agreement has
been found Žthe discrepancies are lower than 10%. for values of optical thickness lower
than 25 and values of droplet effective radius lower than 20 mm. The discrepancies
between the two methods increase to 30% and 45%, for optical thickness and droplet
effective radius, respectively, in the case of cumulonimbus clouds, that are characterized
by much higher values of optical thickness and effective radius. In fact, the model that is
used for generating either the look-up tables ŽCRTVL. or the learning database ŽACNN.,
is not efficient at simulating radiative transfer in very thick clouds.
3. Comparisons
3.1. APOLLO Õs. CRTVL
The comparison is conducted over the area of 300 = 300 pixels Žwhite frame in Fig.
1. covered by the stratocumulus cloud of interest. Fig. 5a shows the values of optical
thickness derived from the CRTVL code vs. values calculated with the APOLLO
parameterization. Values of the same order of magnitude are derived with both techniques, with minimum values of 2 and 3.6, and maximum values of 12 and 15, for
CRTVL and APOLLO, respectively. Nevertheless, the agreement between the two
methods is poor as indicated by the correlation factor of 0.6. Fig. 5b shows the
comparison of the LWP values. The differences are even larger than for optical
thickness, and the correlation factor is only 0.52. With the CRTVL inversion code, the
mean effective radius is about 7.5 mm, which is in agreement with in situ measurements
ŽPawlowska et al., 2000a..
74
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
75
This demonstrates that the parameterization of LWP in the APOLLO method is not
well adapted to the type of clouds examined here.
3.2. ACNN Õs. CRTVL
Fig. 6a shows values of optical thickness retrieved with the ACNN approach vs.
values derived from the CRTVL code, on the same 300 = 300 pixels area as in the
previous section. The agreement is quite good, with a mean relative difference of 10%.
The largest discrepancies are observed for small values of t Žlower than 3.. At such
values, the inversion with CRTVL is ambiguous, as the same values of reflectances are
produced by the model with two different sets of values of t and reff ŽFig. 3.. In
constrast, the distribution of values of optical thickness and effective radius used for the
generation of the learning ACNN database has been restricted to values representative of
a stratocumulus cloud ŽFouilloux and Iaquinta, 1998.. Therefore, the ACNN solutions
are likely to be more realistic than some of the ambiguous CRTVL solutions.
It must also be noted that, for both the CRTVL and ACNN methods, cloud
heterogeneities are not explicitly taken into account in the model since the radiative
transfer scheme is plane-parallel Žcloud parameters uniform horizontally.. Therefore,
additional errors are produced in the most heterogeneous parts of the mission 206
stratocumulus cloud, namely in the northern part of the system, over the ocean, where
the lowest values of optical thickness are also observed.
Fig. 6b illustrates the comparison of the LWP values retrieved with ACNN vs.
CRTVL. The mean relative difference is 7%. This is greater than with synthetic data, but
much lower than with the APOLLO parameterization. At small values of LWP Žbetween
2 and 25 g my2 ., the discrepancies are slightly larger than for the optical thickness. This
corresponds to small values of both optical thickness and effective radius, for which
multiple solutions are possible with CRTVL, while ACNN is restricted to realistic
values. In order to eliminate this inconsistency in CRTVL, a cloud classification
ŽSaunders and Kriebel, 1988. can be conducted as a pre-processing for the CRTVL
approach, for an a priori selection of the most probable solutions Žt , reff ..
Direct comparison of the APOLLO calculations with the ACNN retrieved values has
not been attempted here, since the results would be very similar to those obtained with
the CRTVL code ŽSection 3.1.. Actually, the remaining question is to assess which of
these three retrieval approaches is the most accurate compared to in situ measurements
of cloud characteristics. In the 18 April 1994 case study, the time difference between
aircraft and satellite observations is approximately 2 h. This prevents any point-to-point
comparison. The results of the ACNN method will however be tested vs. in situ
measurements and other remote sensing techniques in the summary paper ŽPawlowska et
al., 2000b.. The area selected for such a comparison is restricted to the region, which is
likely to be advected across the aircraft sampling leg between 10:00 and 12:00 UTC,
based on constant wind speed and direction in the boundary layer, of 10 msy1 and 508,
Fig. 5. Comparison of the values derived with CRTVL vs. APOLLO: optical thickness in Ža. and effective
radius in Žb..
76
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
77
respectively ŽPawlowska et al., 2000b.. This procedure is necessary for the comparison
between AVHRR and in situ measurements, but it is based on the crude hypothesis that
the cloud system and the wind are stationary over a 2-h period. Because of the spatial
uncertainty of this extrapolation procedure, it is no longer justified to keep the best
spatial resolution and AVHRR data can be averaged over larger samples to reduce the
variability of the retrieved parameters. The comparison can then be performed on a
statistical basis. The various quantities Žradiances, t and reff . have accordingly been
averaged over boxes of 5 = 5 and 10 = 10 pixels. The effects of averaging are discussed
in the next section.
3.3. Horizontal aÕeraging
The effects of averaging depend upon cloud homogeneity, that must be quantified. It
is measured by H, the ratio of the standard deviation of the parameter of interest, either
optical thickness, or effective radius, to its mean value, over areas of either 5 = 5 or
10 = 10 pixels.
In a first step, the satellite radiances in channels 1, 3 and 4 are averaged over the
5 = 5 Ž10 = 10. pixels areas. Cloud optical thickness tAI and droplet effective radius rAI
are then retrieved with the ACNN method Žthe index ‘‘AI’’ indicates that the data are
averaged before inversion.. In a second step, the parameters t IA and r IA are estimated
from the satellite image Žat the spatial resolution of 1.1 km. and then averaged over the
5 = 5 Ž10 = 10. pixels areas Ž‘‘IA’’ means inversion before averaging.. Similar results
were obtained with 5 = 5 and 10 = 10 pixels averaging, so that the 10 = 10 case will not
be further discussed. Fig. 7 shows the ratio t IA rtAI as a function of H calculated over
5 = 5 pixels. Fig. 8 is similar for the droplet effective radius. These graphs reveal that,
for large values of H, t IA is lower than tAI , while r IA is greater than rAI . Intuitively,
this result is not very surprising because the albedo is smaller over heterogeneous clouds
than over the equivalent homogeneous area Žwith the same averaged optical properties t
and r ., whereas the transmission is larger.
Nevertheless, no tendencies can be observed for small values of H since the ranges
of values of the ratios t IA rtAI and r IA rrAI are very large. This can be explained by the
fact that the parameter H does not take into account particular clouds structures Ži.e.,
specific orientation with respect to the geometry of illumination, etc... A better
characterization of the cloud heterogeneities could be made by using the textural
properties of clouds that can be computed using Gray Level Cooccurrence Matrices
ŽHaralick, 1979., Gray Level Difference Vector ŽWeszka et al., 1976. or Sum and
Difference Histogram ŽUnser, 1986.. Among these textural properties Žall of them could
not be computed because of the amount of computational time required., those that are
really pertinent for the characterization of cloud fluctuations still have to be settled.
Therefore, further investigations must be undertaken in order to improve the analysis of
Fig. 6. Comparison of the values derived with ACNN vs. CRTVL: optical thickness in Ža. and effective radius
in Žb..
78
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
Fig. 7. Ratio of the values of optical thickness derived before averaging to those derived from averaged
radiances, vs. cloud homogeneity.
Fig. 8. Same as Fig. 7 for the droplet effective radius.
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
79
heterogeneous clouds as in Garand Ž1988. or Bankert Ž1993., where the cloud textural
properties are taken into account.
The problems we faced are the same as those to be encountered when comparisons
between aircraft data and satellite observations are performed. Indeed, the optical
thickness and the droplet effective radius are derived from in situ measurements with a
horizontal resolution of about 100 m Žaircraft speed of 100 msy1 and sampling
frequency of 1 Hz. and then averaged at the scale of the satellite data. The optical
thickness issued from the aircraft data should thus be lower than that computed from the
AVHRR satellite image, while the values of droplet effective radius measured by the
aircraft should be greater than those derived from the satellite observations. These are
very significant results that we have to keep in mind for the comparisons between data
of different spatial scales.
3.4. ACNN learning database of in situ measurements
Different sources of information can be combined to increase the accuracy of a neural
network. This is done here with the values of cloud optical thickness and droplet
effective radius that have been derived from in situ microphysical measurements
ŽPawlowska et al., 2000b.. It is, however, not possible to relate those values directly to
Fig. 9. Ratio of the values of optical thickness derived with the ACNNq network to those derived with the
original method.
80
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
the values of radiances measured with AVHRR, because of the time lag between the
AVHRR image and the in situ measurements. A first attempt is made here by assuming
that the cloud system is stationary and that it is advected with the mean wind in the
boundary layer ŽPawlowska et al., 2000b.. Each pixel of the AVHRR image is advected
assuming a 2-h difference between the two sets of measurements. Only values measured
at the top of the convective cells are selected since the adiabatic estimation of optical
thickness and effective radius from in situ data is restricted to such regions ŽPawlowska
et al., 2000a..
Hence, the neural network is enriched with this new information without performing
another complete learning phase. Clearly, any additional material can be further introduced if necessary as soon as it is available. The computational time required for the
learning phase is therefore reduced, while the accuracy of the inversion can be
improved.
Retrievals have then been conducted over boxes of 5 = 5 pixels with this new ACNN
q
version Žcalled ACNNq. . Fig. 9 represents the ratio tAC
NNrtACNN as a function of H
q
and Fig. 10 represents the ratio rAC NNrrACNN . Fig. 9 shows that the values derived in
homogeneous cloud regions, with the synthetic learning database, only and those derived
after implementation of the in situ data learning base agree reasonably well, whereas, for
more heterogeneous cases, the new values of optical thickness are smaller than the
original ones. This is quite surprising because theoretical considerations show that in the
Fig. 10. Same as Fig. 9 for droplet effective radius.
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
81
case of heterogeneous clouds, implementation of in situ data in the learning base should
lead to greater values than the ones derived with the synthetic learning database only.
The same tendency as in Fig. 7 is observed, namely an under-estimation of optical
thickness. In contrast, the over-estimation of the droplet effective radius apparent in Fig.
q
8 is not observed with the ACNNq method in Fig. 10 Žthe ratio rAC
NNrrACNN is always
close to 1..
This study shows that comparison between data sampled at different spatial scales is
not that simple: It requires new campaigns of measurements with simultaneous in situ
and remote sensing measurements of different cloud types. Indeed, the experimental
approach is necessary to improve and validate retrieval procedures, which aim at taking
into account cloud heterogeneities. Since there is no systematic bias, it is not conceivable to make simplistic corrections to adjust the parameters derived from satellite
observations with in situ measurements.
4. Summary and conclusion
In this paper, three retrieval procedures have been compared, which are designed for
the retrieval of cloud optical thickness, droplet effective radius, and LWP, in the case of
a stratocumulus observed on 18 April 1994, during the EUCREX mission 206. The
geographical region selected in this study is homogeneous over the southern part of the
region and heterogeneous over the northern part. This case is thus well suited for the
comparison between the various methods.
The main drawback of the APOLLO parameterizations set is that the uncertainty on
the retrieved parameters is large. For the stratocumulus cloud situation analysed here, it
is clear that better relationships should be derived from in situ measurements. The
results obtained with the CRTVL algorithm based on a radiative transfer model are more
general, but they rely upon the representativeness of the model.
Indeed, the ACNN and CRTVL methods give similar results for both cloud optical
thickness and droplet effective radius Žor LWP.. This is partly due to the fact that the
same radiative transfer model is used to build the CRTVL look-up table and the ACNN
learning database. In particular, there is no systematic bias in terms of cloud optical
thickness and droplet effective radius. The discrepancy between the two approaches
becomes significant at the smallest values of optical thickness and effective radius. This
is due to an ambiguity in the inversion with the CRTVL technique, while the ACNN
learning database has been built with non-ambiguous values, that are typical of the type
of cloud analysed here.
The influence of horizontal averaging as a function of cloud heterogeneity has also
been tested. Cloud heterogeneity is characterized by the ratio of the standard deviation
of the parameter of interest, either optical thickness or effective radius, to its mean value
over the averaging area. Values retrieved after averaging cloud radiances have been
compared to the average of the values retrieved at the finest resolution. The results
suggest that the values of cloud optical thickness derived before averaging are smaller
than those derived from the averaged radiances. In contrast, the values of droplet
82
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
effective radius derived before averaging are greater than those derived from the
averaged radiances.
Therefore, comparisons between aircraft measurements and satellite observations will
not be valid if the cloud is heterogeneous Žwhich is the case for most of the clouds.. This
preliminary study suggests that the parameter H, which was used to characterize cloud
homogeneity seems to be inadequate. An attempt has been made to introduce in situ data
in the learning phase of the ACNN method. The result shows no systematic tendency for
the droplet effective radius, while an under-estimation of the cloud optical thickness is
found as a function of H.
In conclusion, it is suggested that the efficiency of the radiative transfer inversion
technique ŽCRTVL. can be significantly improved by using neural networks, without
loosing its capability to retrieve cloud parameters with a high accuracy. A neural
network can also be progressively improved by introducing in situ measurements in its
learning database, especially for solving the problem of heterogeneous clouds.
Acknowledgements
The authors are particularly thankful to T.Y. Nakajima and T. Nakajima for providing
them with their CRTVL inversion scheme. This work was supported by a grant from the
Direction des Recherches Etude et Technique no. 9334075 and was part of the EUCREX
supported by the European Union, under Grant ENVV5V-CT94-0455.
References
Arking, A., Childs, J.D., 1985. Retrieval of cloud cover parameters from multispectral satellite images. J.
Clim. Appl. Meteor. Soc. 69, 618–626.
Bankert, R.L., 1993. Cloud classification of AVHRR imagery in maritime regions using a probabilistic neural
network. IEEE Trans. Geosci. Remote Sens. 33, 909–918.
Cotton, W.R., Anthes, R.A., 1989. Storm and Cloud Dynamics. Academic Press, San Diego, 883 pp.
Curran, R.J., Wu, L.M.C., 1982. Skylab near-infrared observations of clouds indicating supercooled liquid
water droplets. J. Atmos. Sci. 39, 635–647.
Fouilloux, A., Iaquinta, J., 1998. Assessment of clouds characteristics from satellite observations by means of
self-organized neuronal networks. Remote Sens. Environ. 66, 101–109.
Garand, L., 1988. Automated recognition of oceanic cloud patterns: Part I. Methodology and application to
cloud climatology. J. Clim. 1, 20–39.
Haralick, R.M., 1979. Stastical and structural approaches to texture. Proc. IEEE 67, 786–804.
Iaquinta, J., Pinty, B., 1997. Radiation field in a multi-layered geophysical medium: the Ice–Water–Aerosol–
VEgetation–Soil ŽIWAVES. model. J. Geophys. Res. 102, 13627–13642.
Kaestner, M., Kriebel, K.T., Meerkoetter, R., Renger, W., Ruppersberg, G.H., Wendling, P., 1993. Comparison of cirrus height and optical depth derived from satellite and aircraft measurements. Mon. Weather.
Rev. 121 Ž10., 2708–2717.
Kaufman, Y.J., Holben, B.N., 1993. Calibration of the AVHRR visible and near-infrared bands by atmospheric
scattering, ocean glint and desert reflexion. Int. J. Remote Sens. 14, 21–52.
Kriebel, K.T., 1989. Cloud liquid water path derived from AVHRR data using APOLLO. Int. J. Remote Sens.
10, 723–729.
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
83
Kriebel, K.T., Saunders, R., Gesell, G., 1989. Optical properties of clouds derived from fully cloudy AVHRR
pixels. Beitr. Phys. Atmos. 62, 165–171.
Nakajima, T., King, M.D., 1990. Determination of the optical thickness and effective particle radius of clouds
from reflected solar radiation measurements: Part I. Theory. J. Atmos. Sci. 47, 1878–1893.
Nakajima, T.Y., Nakajima, T., 1995. Wide-area determination of cloud microphysical properties from
NOAA-AVHRR measurements for FIRE and ASTEX regions. J. Atmos. Sci. 52, 4043–4059.
Nakajima, T.Y., Tanaka, M., 1986. Matrix formulation for the transfer of solar radiation in plane-parallel
scattering atmosphere. J. Quant. Spectrosc. Radiat. Transfer 35, 13–21.
Parol, F., Buriez, J.C., Brogniez, G., Fouquart, Y., 1991. Information content of AVHRR channels 4 and 5
with respect to the effective radius of cirrus cloud particles. J. Appl. Meteorol. 30, 973–984.
Pawlowska, H., Brenguier, J.L., Burnet, F., 2000a. Microphysical properties of stratocumulus clouds. Atmos.
Res., in press.
Pawlowska, H., Brenguier, J.L., Fouquart, Y., Armbruster, W., Descloitres, J., Fischer, J., Flamant, C.,
Fouilloux, A., Gayet, J.F., Ghosh, S., Jonas, P., Parol, F., Pelon, J., Schuller,
L., 2000b. Microphysical and
¨
radiative properties of stratocumulus cloud: the EUCREX mission 206 case study. Atmos. Res., in press.
Rossow, W.B., 1989. Measuring cloud properties from space: a review. J. Clima. 2, 201–213.
Saunders, R.W., Kriebel, K.T., 1988. An improved method for detecting clear sky and cloudy radiances from
AVHRR data. Int. J. Remote Sens. 9, 123–150.
Stephens, G.L., 1978. Radiation profiles in extended water clouds: II. Parameterization schemes. J. Atmos.
Sci. 35, 2123–2132.
Stephens, G.L., Ackerman, S., Smith, E.A., 1984. A shortwave parameterization revised to improve cloud
absorption. J. Atmos. Sci. 41, 687–690.
Unser, M., 1986. Sum and difference histograms for texture classification. IEEE Trans. Pattern Anal. Mach.
Intell. PAMI-8, 118–125.
Weszka, J.S., Dyer, C.R., Rosenfeld, A., 1976. A comparative study of texture measures for terrain
classification. IEEE Trans. Syst. Man. Cybern. SMC-6, 2269–2285.
www.elsevier.comrlocateratmos
Determination of cloud microphysical properties
from AVHRR images: comparisons of
three approaches
Anne Fouilloux a,) , Jean-François Gayet a , Karl-Theodor Kriebel b
a
Laboratoire de Meteorologie
Physique, UPRESA 6016 CNRS, UniÕersite´ Blaise Pascal, Aubiere,
´´
` France
b
Institut fur
¨ Physik der Atmosphare,
¨ Deutsche Forschungsanstalt fur
¨ Luft-und Raumfahrt,
Oberpfaffenhofen, Germany
Received 8 January 1998; accepted 1 March 2000
Abstract
This contribution to the EUCREX mission 206 series aims at evaluating the performances of
three methods considered for the retrieval of cloud microphysical properties Žoptical depth, droplet
effective radius and liquid water path ŽLWP.. from Advanced Very High Resolution Radiometer
ŽAVHRR. satellite images. These procedures, namely a set of empirical parameterizations, a
physically based inversion code, and a neuronal approach, are tested with the mission 206 case
study. The influence of horizontal averaging on the retrieved cloud properties is also investigated
to assess the limitations of the comparison between satellite observations and aircraft measurements performed at a smaller scale. q 2000 Elsevier Science B.V. All rights reserved.
Keywords: Cloud radiative properties; Remote sensing retrieval techniques
1. Introduction
For usual satellite remote sensing applications, the classical inversion problem
consists of extracting pertinent parameters that characterize the medium from a set of
)
Corresponding author. LIVIC, Bat 140, 13 route de la Miniere,
` 78000 Versailles, France. Tel.: q33-1-4043-29-07; fax: q33-1-40-43-29-30.
E-mail address: [email protected] ŽA. Fouilloux..
0169-8095r00r$ - see front matter q 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 9 - 8 0 9 5 Ž 0 0 . 0 0 0 5 7 - 0
66
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
radiative measurements, namely the radiances sampled in several illumination and
observation conditions andror in different spectral bands. As part of meteorological and
climate studies, the correct interpretation of such observations in terms of cloud
microphysical properties requires the development of tools that are in the meantime
powerful, robust and fast ŽArking and Childs, 1985; Rossow, 1989..
Among the various retrieval procedures that were developed for deriving cloud
optical thickness, liquid water path ŽLWP. or droplet effective radius ŽKriebel et al.,
1989; Nakajima and King, 1990; Parol et al., 1991; Iaquinta and Pinty, 1997., three
kinds of approaches emerged: empirical parameterizations derived from in situ measurements, physical methods using optimization techniques or look-up tables, and a new
generation of algorithms built with neuronal networks. These techniques have often been
validated for homogeneous clouds, but their performances on heterogeneous situations
are less documented.
The aim of the present paper is to retrieve values of cloud optical thickness, droplet
effective radius and LWP within the stratocumulus observed on 18 April 1994, during
the European Cloud and Radiation Experiment ŽEUCREX. mission 206. This case is
particularly interesting since the cloud system was rather homogeneous over land
surfaces, but progressively broken towards the North, over the Atlantic ocean. This case
thus provides an opportunity to test the three retrieval procedures over a large range of
heterogeneity of the cloud structure. The methods tested here are: Ži. the AVHRR
Processing Scheme Over Clouds, Land and Ocean ŽAPOLLO. parameterizations of
Kriebel et al. Ž1989.; Žii. the Cloud Retrieval ŽCRTVL. inversion code developed by
Nakajima and Nakajima Ž1995.; and Žiii. the Adjustable Combination of Neuronal
Networks ŽACNN. algorithm recently elaborated by Fouilloux and Iaquinta Ž1998..
Parameterizations specifically developed for the determination of cloud microphysical
properties is an attractive solution since the method is straightforward and does not
necessitate tedious computations. However, care must be taken to avoid using the
parameterizations on situations that may differ significantly from those for which the
empirical formulas have been developed.
Inversion techniques, such as CRTVL, require a preliminary modeling step and their
accuracy depends upon the efficiency of the method at recovering the optimal fit
between the actual radiances and those predicted by the radiative transfer model
ŽNakajima and Tanaka, 1986.. In order to avoid laborious and repeated calculations at
each iteration, it is advantageous to keep only relevant information into look-up tables
ŽNakajima and King, 1990.. Therefore, the intrinsic quality of the retrieval depends on
the model accuracy, but it is also strongly conditioned by the resolution of the look-up
table in the parameter space.
The third method is based on a combination of self-organised neuronal networks
ŽFouilloux and Iaquinta, 1998. and performs a classification where the accuracy to be
achieved is one of the parameters driving the procedure. That allows to explicitly design
the network by increasing the number of classes, according to the desired accuracy.
In the following sections, the satellite data available on 18 April 1994 are described,
the three methods are briefly presented, and their performances at retrieving the
stratocumulus cloud microphysical characteristics are evaluated.
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
67
2. Description of the retrieval schemes
2.1. Satellite data
The radiance data considered hereafter are taken from the Advanced Very High
Resolution Radiometer ŽAVHRR. instrument mounted on board the National Oceanic
and Atmospheric Administration ŽNOAA.-11 satellite.
The NOAA-AVHRR satellite image of 18 April 1994 was recorded at around 08:25
UTC, while in situ measurements with the instrumented aircraft were performed
between 09:30 and 12:00 UTC ŽPawlowska et al., 2000b.. The following analysis is
conducted in five spectral bands ŽTable 1. with a spatial resolution of about 1.1 km. Fig.
1 shows the satellite image in the first visible channel. The bordered area of 300 = 300
pixels is the region of interest for the present study and the white straight line represents
the aircraft sampling track ŽPawlowska et al., 2000b.. Fig. 1 reveals that the cloud
system is more heterogeneous towards the northwest of the sampling leg, above the
Atlantic Ocean than towards the southeast, over the coast of Brittany.
The calibration procedures applied to this data follow standard techniques. In
particular, since the visible channel of the sensor has been degrading from the pre-launch
condition, the corrected calibration constants of Kaufman and Holben Ž1993. have been
used to obtain the radiances from the satellite-recorded digital counts.
2.2. APOLLO parameterizations for cloud microphysical properties
The first step in this approach consists in making a cloud classification ŽSaunders and
Kriebel, 1988. to identify pixels, which are fully cloudy. These are then categorized into
thick and thin clouds. The thick clouds are then divided into low, medium and high
clouds according to their temperature. Thick clouds are treated as water clouds. For the
second step, the optical thickness and LWP are computed on a pixel-by-pixel basis using
parameterizations according to Stephens Ž1978. and Stephens et al. Ž1984. Žsee also
Kriebel et al., 1989.. They are based on the observation that the radiation fields are the
most sensitive to cloud optical thickness in the visible window Ž0.4 to 0.8 mm. ŽCurran
and Wu, 1982; Arking and Childs, 1985; Nakajima and King, 1990.. The formulas are
thus valid in the solar spectral range, at wavelengths below 0.7 mm, and they are used
Table 1
Characteristics of the AVHRR sensor
AVHRR channel
Wavelength
Domain
1
2
3
4
5
0.58–0.68 mm
0.725–1.0 mm
3.55–3.93 mm
10.3–11.3 mm
11.5–12.5 mm
Visible
Visible
Near-infrared
Infrared
Infrared
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
68
Fig. 1. Visible channel of the NOAA-AVHRR image at 08:24 UTC on 18 April 1994. The image encompasses
the stratocumulus cloud area covered by the AVHRR measurements. The area of interest is indicated by a
white frame and the leg flown by the instrumented aircraft is marked by a white segment.
with the AVHRR channel 1 data. The directional–hemispherical cloud reflectance in
Channel 1, R c Ž m 0 ., is related to cloud optical thickness t by the following relationship:
Rc s
b Ž m 0 . t Ž m 0 . rm 0
1 q b Ž m 0 . t Ž m 0 . rm 0
Ž 1.
where b Ž m 0 . is the modified backscattering coefficient given for water clouds with
optical thickness ranging from 1 to 500 and m 0 is the cosine of the solar zenith angle.
The values of b Ž m 0 . that are considered hereafter are listed in the paper of Stephens et
al. Ž1984..
The cloud LWP Žin g my2 . is derived from the optical thickness by the following
formula ŽKriebel et al., 1989.:
log Ž LWP . s Ž 0.5454 = t .
0.234
Ž 2.
Fig. 2 shows the dependence between directional–hemispherical cloud reflectance
R c Ž m 0 . and optical thickness obtained with the APOLLO parameterization for the
clouds observed on 18 April 1994, within the area indicated in Fig. 1. Note that there are
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
69
Fig. 2. Optical thickness values retrieved with the APOLLO code as a function of the actual directional
hemispherical cloud reflectance, R c Ž m 0 ..
no values of R c Ž m 0 . below 0.3 because such pixels are not classified as water clouds
Žthick clouds. and are therefore eliminated.
2.3. CRTVL inÕersion code
The CRTVL inversion code ŽNakajima and Nakajima, 1995. has been implemented
to simultaneously retrieve optical thickness and droplet effective radius of cloudy pixels
from the NOAA-AVHRR satellite images. This algorithm computes the microphysical
quantities from the radiances at visible Ž0.58 to 0.65 mm. and near-infrared Ž3.55 to 3.93
mm. wavelengths by making use of look-up tables that saves computational time, yet
conserving the accuracy of the analysis. For the simulation of the satellite measured
radiances, we applied the radiative transfer model developed by Nakajima and Tanaka
Ž1986. with an atmosphere typical of winter oceanic mid latitudes. The number of
vertical levels are set to four Žwith boundaries at 1, 2, 12 and 20 km.. That does not
particularly alter the results and drastically reduces the computational time. According to
aircraft observations, the cloud layer was set at 1 km, and we assumed a Lambertian
surface Žalbedo of 0.06 . for the underlying oceanic surface.
Fig. 3 shows the simulation of reflected solar radiances in the AVHRR spectral bands
1 and 3 as a function of cloud optical thickness t and droplet effective radius reff , for a
70
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
Fig. 3. Simulation of reflected solar radiances in AVHRR channels 1 and 3 as functions of cloud optical
thickness Žt s1, 2, 4, 8, 16, 32, 64. and effective radius Ž reff s 2, 4, 8, 16, 32., with Ž u 0 s608, u s 408 and
w s 508.. The isolines of optical thickness and effective radius are represented by solid and dashed lines,
respectively.
solar zenith angle of 608, a satellite zenith angle of 408 and a relative azimuth of 508.
Near-vertical and near-horizontal lines illustrate iso-optical thickness and iso-effective
radius radiances, respectively. Ground-reflected and thermal radiations are not taken into
account, and the LOWTRAN-7 model associated with a mid-latitude model is considered to properly take into account the gaseous aerosol absorption ŽNakajima and King,
1990.. The principle of the inversion procedure is deduced from this figure as follows:
The t value is sought so as to minimize the difference between the actual and theoretical
Žstored in the look-up tables. cloud-reflected radiances in the first channel Žthe threshold
being 0.1%.. Afterwards, t is set to this value and reff is modified until the difference
between measured and tabulated radiances in the third channel becomes less than 0.1%.
This process is repeated until both differences Žin the visible and near infrared channels.
become lower than 0.1%, which usually requires three to four iterations. LWP Žg my2 .
is then calculated from the droplet effective radius and cloud optical thickness values by
using the following equation:
2
LWP s rt reff
Ž 3.
3
where r is the liquid water density Žg my3 ..
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
71
2.4. ACNN neuronal network
The main problem with the previous approaches is that, with APOLLO, the determination of optical thickness and LWP is fast, but rather inaccurate, while with CRTVL,
the retrieval requires large computer storage because of the Žhuge. look-up tables. A
compromise between accuracy and fastness seems to be possible through a combination
of neural networks as described in Fouilloux and Iaquinta Ž1998..
The principle of the method is as follows: The determination of optical thickness and
droplet effective radius is split into several steps, each being dedicated to a particular
task. The first step consists in classifying the pixels of an AVHRR image into cloud
categories. Actually, the same algorithm as in APOLLO has been used for cloud
classification, for a better comparison between the two methods. The second step allows
differentiation of the pixels as a function of the geometry of illumination and observation. One or two further steps Ždepending on the desired accuracy. are then required for
the determination of the retrieved parameters. They are performed with self-organized
neuronal networks with two layers Žfor the entry and for the output cells. so as to let the
system converge near a stable configuration corresponding to the required degree of
accuracy. The accuracy depends on the quantity of information that the entry vectors
convey and also on a ‘‘ vigilance’’ factor Žalso called the accuracy factor. j Ža value
between 0 and 1. fixed by the operator. If this factor is close to 1, the system is forced to
solve with a great accuracy and inversely, when j tends to 0, the system is relaxed to a
lower accuracy. Nevertheless, the accuracy of the restitution depends also on both the
radiative transfer model used and the degree of noise included in the initial data.
The inversion process is divided into two stages: a learning phase during which the
connection weights of the neuronal networks are fixed and a generalization phase that
consists in using the neuronal network with actual remotely sensed data.
Regarding the learning phase, several synthetic databases, containing radiances in the
two visible and near infrared AVHRR channels for various values of the illuminationr
observation geometry and cloud parameters are generated by running the radiative
transfer model ŽNakajima and Tanaka, 1986.. In order to limit the size of these learning
data sets Ži.e., the range of possible values for the retrieved variables., a sorting is
performed by means of a ‘‘rough’’ cloud classification of AVHRR imagery. Such a
discrimination considers the algorithm proposed by Saunders and Kriebel Ž1988., which
relies on five tests conducted on the two visible and two infrared AVHRR channels.
Four categories have been identified, for which specific distributions of cloud optical
thickness and droplet effective radius were defined. On 18 April 1994, most of the
cloudy pixels are recognized as low clouds, which is in agreement with the aircraft
observations ŽPawlowska et al., 2000a.. For this cloud category, the following characteristics for the droplet size distribution have been selected: a number concentration of 350
cmy3 , a mean radius of 4 mm Žthe size distribution ranges from 0 to 20 mm. and a LWC
value of 0.09 g my3 ŽCotton and Anthes, 1989.. These values are derived from a large
set of in situ data made during several experiments ŽCotton and Anthes, 1989. and they
are considered as typical of mid-latitude stratocumulus clouds. The selected droplet size
distribution is similar to the distributions measured in situ during mission 206 ŽPawlowska
et al., 2000a..
72
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
The first neural network is specially designed to discriminate the pixels as a function
of the geometry of illumination and observation. These parameters have been added as
Fig. 4. Optical thickness values issued from the ACNN classification vs. values fixed in the radiative transfer
model, with j s 0.5 in Ža. and j s 0.8 in Žb..
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
73
neural network inputs to visible and near-infrared radiances and the accuracy parameter
j has been set to a value of 0.7 in order to distinguish 25 classes. A second network
allows to determine cloud optical thickness and droplet effective radius. Fig. 4a and b
shows the sensitivity of the neuronal network to values of optical thickness and of the
parameter j Žtwo values — 0.5 and 0.8 — are used.. It appears that in both cases, the
steps clearly depend on optical thickness, and that the accuracy is improved when j
increases Žthe various steps are less pronounced.. Similar results are found with the
droplet effective radius when using a learning database sensitive to this parameter. LWP
is finally computed using Eq. Ž3..
2.5. Accuracy assessment
The application of APOLLO parameterizations is straightforward and very fast, but
the uncertainty on the estimated parameters is of the order of 35%. Nevertheless, these
parameterizations have been extensively validated against in situ measurements ŽKriebel,
1989; Kriebel et al., 1989; Kaestner et al., 1993.. The CRTVL inversion code, which
provides an accuracy of about 10% Žwith synthetic data., has been tested against actual
measurements and no systematic bias was found. The uncertainties were however not
systematically computed. The ACNN and CRTVL inversion procedures have been
tested on a large set of synthetic clouds of various types, and a very good agreement has
been found Žthe discrepancies are lower than 10%. for values of optical thickness lower
than 25 and values of droplet effective radius lower than 20 mm. The discrepancies
between the two methods increase to 30% and 45%, for optical thickness and droplet
effective radius, respectively, in the case of cumulonimbus clouds, that are characterized
by much higher values of optical thickness and effective radius. In fact, the model that is
used for generating either the look-up tables ŽCRTVL. or the learning database ŽACNN.,
is not efficient at simulating radiative transfer in very thick clouds.
3. Comparisons
3.1. APOLLO Õs. CRTVL
The comparison is conducted over the area of 300 = 300 pixels Žwhite frame in Fig.
1. covered by the stratocumulus cloud of interest. Fig. 5a shows the values of optical
thickness derived from the CRTVL code vs. values calculated with the APOLLO
parameterization. Values of the same order of magnitude are derived with both techniques, with minimum values of 2 and 3.6, and maximum values of 12 and 15, for
CRTVL and APOLLO, respectively. Nevertheless, the agreement between the two
methods is poor as indicated by the correlation factor of 0.6. Fig. 5b shows the
comparison of the LWP values. The differences are even larger than for optical
thickness, and the correlation factor is only 0.52. With the CRTVL inversion code, the
mean effective radius is about 7.5 mm, which is in agreement with in situ measurements
ŽPawlowska et al., 2000a..
74
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
75
This demonstrates that the parameterization of LWP in the APOLLO method is not
well adapted to the type of clouds examined here.
3.2. ACNN Õs. CRTVL
Fig. 6a shows values of optical thickness retrieved with the ACNN approach vs.
values derived from the CRTVL code, on the same 300 = 300 pixels area as in the
previous section. The agreement is quite good, with a mean relative difference of 10%.
The largest discrepancies are observed for small values of t Žlower than 3.. At such
values, the inversion with CRTVL is ambiguous, as the same values of reflectances are
produced by the model with two different sets of values of t and reff ŽFig. 3.. In
constrast, the distribution of values of optical thickness and effective radius used for the
generation of the learning ACNN database has been restricted to values representative of
a stratocumulus cloud ŽFouilloux and Iaquinta, 1998.. Therefore, the ACNN solutions
are likely to be more realistic than some of the ambiguous CRTVL solutions.
It must also be noted that, for both the CRTVL and ACNN methods, cloud
heterogeneities are not explicitly taken into account in the model since the radiative
transfer scheme is plane-parallel Žcloud parameters uniform horizontally.. Therefore,
additional errors are produced in the most heterogeneous parts of the mission 206
stratocumulus cloud, namely in the northern part of the system, over the ocean, where
the lowest values of optical thickness are also observed.
Fig. 6b illustrates the comparison of the LWP values retrieved with ACNN vs.
CRTVL. The mean relative difference is 7%. This is greater than with synthetic data, but
much lower than with the APOLLO parameterization. At small values of LWP Žbetween
2 and 25 g my2 ., the discrepancies are slightly larger than for the optical thickness. This
corresponds to small values of both optical thickness and effective radius, for which
multiple solutions are possible with CRTVL, while ACNN is restricted to realistic
values. In order to eliminate this inconsistency in CRTVL, a cloud classification
ŽSaunders and Kriebel, 1988. can be conducted as a pre-processing for the CRTVL
approach, for an a priori selection of the most probable solutions Žt , reff ..
Direct comparison of the APOLLO calculations with the ACNN retrieved values has
not been attempted here, since the results would be very similar to those obtained with
the CRTVL code ŽSection 3.1.. Actually, the remaining question is to assess which of
these three retrieval approaches is the most accurate compared to in situ measurements
of cloud characteristics. In the 18 April 1994 case study, the time difference between
aircraft and satellite observations is approximately 2 h. This prevents any point-to-point
comparison. The results of the ACNN method will however be tested vs. in situ
measurements and other remote sensing techniques in the summary paper ŽPawlowska et
al., 2000b.. The area selected for such a comparison is restricted to the region, which is
likely to be advected across the aircraft sampling leg between 10:00 and 12:00 UTC,
based on constant wind speed and direction in the boundary layer, of 10 msy1 and 508,
Fig. 5. Comparison of the values derived with CRTVL vs. APOLLO: optical thickness in Ža. and effective
radius in Žb..
76
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
77
respectively ŽPawlowska et al., 2000b.. This procedure is necessary for the comparison
between AVHRR and in situ measurements, but it is based on the crude hypothesis that
the cloud system and the wind are stationary over a 2-h period. Because of the spatial
uncertainty of this extrapolation procedure, it is no longer justified to keep the best
spatial resolution and AVHRR data can be averaged over larger samples to reduce the
variability of the retrieved parameters. The comparison can then be performed on a
statistical basis. The various quantities Žradiances, t and reff . have accordingly been
averaged over boxes of 5 = 5 and 10 = 10 pixels. The effects of averaging are discussed
in the next section.
3.3. Horizontal aÕeraging
The effects of averaging depend upon cloud homogeneity, that must be quantified. It
is measured by H, the ratio of the standard deviation of the parameter of interest, either
optical thickness, or effective radius, to its mean value, over areas of either 5 = 5 or
10 = 10 pixels.
In a first step, the satellite radiances in channels 1, 3 and 4 are averaged over the
5 = 5 Ž10 = 10. pixels areas. Cloud optical thickness tAI and droplet effective radius rAI
are then retrieved with the ACNN method Žthe index ‘‘AI’’ indicates that the data are
averaged before inversion.. In a second step, the parameters t IA and r IA are estimated
from the satellite image Žat the spatial resolution of 1.1 km. and then averaged over the
5 = 5 Ž10 = 10. pixels areas Ž‘‘IA’’ means inversion before averaging.. Similar results
were obtained with 5 = 5 and 10 = 10 pixels averaging, so that the 10 = 10 case will not
be further discussed. Fig. 7 shows the ratio t IA rtAI as a function of H calculated over
5 = 5 pixels. Fig. 8 is similar for the droplet effective radius. These graphs reveal that,
for large values of H, t IA is lower than tAI , while r IA is greater than rAI . Intuitively,
this result is not very surprising because the albedo is smaller over heterogeneous clouds
than over the equivalent homogeneous area Žwith the same averaged optical properties t
and r ., whereas the transmission is larger.
Nevertheless, no tendencies can be observed for small values of H since the ranges
of values of the ratios t IA rtAI and r IA rrAI are very large. This can be explained by the
fact that the parameter H does not take into account particular clouds structures Ži.e.,
specific orientation with respect to the geometry of illumination, etc... A better
characterization of the cloud heterogeneities could be made by using the textural
properties of clouds that can be computed using Gray Level Cooccurrence Matrices
ŽHaralick, 1979., Gray Level Difference Vector ŽWeszka et al., 1976. or Sum and
Difference Histogram ŽUnser, 1986.. Among these textural properties Žall of them could
not be computed because of the amount of computational time required., those that are
really pertinent for the characterization of cloud fluctuations still have to be settled.
Therefore, further investigations must be undertaken in order to improve the analysis of
Fig. 6. Comparison of the values derived with ACNN vs. CRTVL: optical thickness in Ža. and effective radius
in Žb..
78
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
Fig. 7. Ratio of the values of optical thickness derived before averaging to those derived from averaged
radiances, vs. cloud homogeneity.
Fig. 8. Same as Fig. 7 for the droplet effective radius.
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
79
heterogeneous clouds as in Garand Ž1988. or Bankert Ž1993., where the cloud textural
properties are taken into account.
The problems we faced are the same as those to be encountered when comparisons
between aircraft data and satellite observations are performed. Indeed, the optical
thickness and the droplet effective radius are derived from in situ measurements with a
horizontal resolution of about 100 m Žaircraft speed of 100 msy1 and sampling
frequency of 1 Hz. and then averaged at the scale of the satellite data. The optical
thickness issued from the aircraft data should thus be lower than that computed from the
AVHRR satellite image, while the values of droplet effective radius measured by the
aircraft should be greater than those derived from the satellite observations. These are
very significant results that we have to keep in mind for the comparisons between data
of different spatial scales.
3.4. ACNN learning database of in situ measurements
Different sources of information can be combined to increase the accuracy of a neural
network. This is done here with the values of cloud optical thickness and droplet
effective radius that have been derived from in situ microphysical measurements
ŽPawlowska et al., 2000b.. It is, however, not possible to relate those values directly to
Fig. 9. Ratio of the values of optical thickness derived with the ACNNq network to those derived with the
original method.
80
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
the values of radiances measured with AVHRR, because of the time lag between the
AVHRR image and the in situ measurements. A first attempt is made here by assuming
that the cloud system is stationary and that it is advected with the mean wind in the
boundary layer ŽPawlowska et al., 2000b.. Each pixel of the AVHRR image is advected
assuming a 2-h difference between the two sets of measurements. Only values measured
at the top of the convective cells are selected since the adiabatic estimation of optical
thickness and effective radius from in situ data is restricted to such regions ŽPawlowska
et al., 2000a..
Hence, the neural network is enriched with this new information without performing
another complete learning phase. Clearly, any additional material can be further introduced if necessary as soon as it is available. The computational time required for the
learning phase is therefore reduced, while the accuracy of the inversion can be
improved.
Retrievals have then been conducted over boxes of 5 = 5 pixels with this new ACNN
q
version Žcalled ACNNq. . Fig. 9 represents the ratio tAC
NNrtACNN as a function of H
q
and Fig. 10 represents the ratio rAC NNrrACNN . Fig. 9 shows that the values derived in
homogeneous cloud regions, with the synthetic learning database, only and those derived
after implementation of the in situ data learning base agree reasonably well, whereas, for
more heterogeneous cases, the new values of optical thickness are smaller than the
original ones. This is quite surprising because theoretical considerations show that in the
Fig. 10. Same as Fig. 9 for droplet effective radius.
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
81
case of heterogeneous clouds, implementation of in situ data in the learning base should
lead to greater values than the ones derived with the synthetic learning database only.
The same tendency as in Fig. 7 is observed, namely an under-estimation of optical
thickness. In contrast, the over-estimation of the droplet effective radius apparent in Fig.
q
8 is not observed with the ACNNq method in Fig. 10 Žthe ratio rAC
NNrrACNN is always
close to 1..
This study shows that comparison between data sampled at different spatial scales is
not that simple: It requires new campaigns of measurements with simultaneous in situ
and remote sensing measurements of different cloud types. Indeed, the experimental
approach is necessary to improve and validate retrieval procedures, which aim at taking
into account cloud heterogeneities. Since there is no systematic bias, it is not conceivable to make simplistic corrections to adjust the parameters derived from satellite
observations with in situ measurements.
4. Summary and conclusion
In this paper, three retrieval procedures have been compared, which are designed for
the retrieval of cloud optical thickness, droplet effective radius, and LWP, in the case of
a stratocumulus observed on 18 April 1994, during the EUCREX mission 206. The
geographical region selected in this study is homogeneous over the southern part of the
region and heterogeneous over the northern part. This case is thus well suited for the
comparison between the various methods.
The main drawback of the APOLLO parameterizations set is that the uncertainty on
the retrieved parameters is large. For the stratocumulus cloud situation analysed here, it
is clear that better relationships should be derived from in situ measurements. The
results obtained with the CRTVL algorithm based on a radiative transfer model are more
general, but they rely upon the representativeness of the model.
Indeed, the ACNN and CRTVL methods give similar results for both cloud optical
thickness and droplet effective radius Žor LWP.. This is partly due to the fact that the
same radiative transfer model is used to build the CRTVL look-up table and the ACNN
learning database. In particular, there is no systematic bias in terms of cloud optical
thickness and droplet effective radius. The discrepancy between the two approaches
becomes significant at the smallest values of optical thickness and effective radius. This
is due to an ambiguity in the inversion with the CRTVL technique, while the ACNN
learning database has been built with non-ambiguous values, that are typical of the type
of cloud analysed here.
The influence of horizontal averaging as a function of cloud heterogeneity has also
been tested. Cloud heterogeneity is characterized by the ratio of the standard deviation
of the parameter of interest, either optical thickness or effective radius, to its mean value
over the averaging area. Values retrieved after averaging cloud radiances have been
compared to the average of the values retrieved at the finest resolution. The results
suggest that the values of cloud optical thickness derived before averaging are smaller
than those derived from the averaged radiances. In contrast, the values of droplet
82
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
effective radius derived before averaging are greater than those derived from the
averaged radiances.
Therefore, comparisons between aircraft measurements and satellite observations will
not be valid if the cloud is heterogeneous Žwhich is the case for most of the clouds.. This
preliminary study suggests that the parameter H, which was used to characterize cloud
homogeneity seems to be inadequate. An attempt has been made to introduce in situ data
in the learning phase of the ACNN method. The result shows no systematic tendency for
the droplet effective radius, while an under-estimation of the cloud optical thickness is
found as a function of H.
In conclusion, it is suggested that the efficiency of the radiative transfer inversion
technique ŽCRTVL. can be significantly improved by using neural networks, without
loosing its capability to retrieve cloud parameters with a high accuracy. A neural
network can also be progressively improved by introducing in situ measurements in its
learning database, especially for solving the problem of heterogeneous clouds.
Acknowledgements
The authors are particularly thankful to T.Y. Nakajima and T. Nakajima for providing
them with their CRTVL inversion scheme. This work was supported by a grant from the
Direction des Recherches Etude et Technique no. 9334075 and was part of the EUCREX
supported by the European Union, under Grant ENVV5V-CT94-0455.
References
Arking, A., Childs, J.D., 1985. Retrieval of cloud cover parameters from multispectral satellite images. J.
Clim. Appl. Meteor. Soc. 69, 618–626.
Bankert, R.L., 1993. Cloud classification of AVHRR imagery in maritime regions using a probabilistic neural
network. IEEE Trans. Geosci. Remote Sens. 33, 909–918.
Cotton, W.R., Anthes, R.A., 1989. Storm and Cloud Dynamics. Academic Press, San Diego, 883 pp.
Curran, R.J., Wu, L.M.C., 1982. Skylab near-infrared observations of clouds indicating supercooled liquid
water droplets. J. Atmos. Sci. 39, 635–647.
Fouilloux, A., Iaquinta, J., 1998. Assessment of clouds characteristics from satellite observations by means of
self-organized neuronal networks. Remote Sens. Environ. 66, 101–109.
Garand, L., 1988. Automated recognition of oceanic cloud patterns: Part I. Methodology and application to
cloud climatology. J. Clim. 1, 20–39.
Haralick, R.M., 1979. Stastical and structural approaches to texture. Proc. IEEE 67, 786–804.
Iaquinta, J., Pinty, B., 1997. Radiation field in a multi-layered geophysical medium: the Ice–Water–Aerosol–
VEgetation–Soil ŽIWAVES. model. J. Geophys. Res. 102, 13627–13642.
Kaestner, M., Kriebel, K.T., Meerkoetter, R., Renger, W., Ruppersberg, G.H., Wendling, P., 1993. Comparison of cirrus height and optical depth derived from satellite and aircraft measurements. Mon. Weather.
Rev. 121 Ž10., 2708–2717.
Kaufman, Y.J., Holben, B.N., 1993. Calibration of the AVHRR visible and near-infrared bands by atmospheric
scattering, ocean glint and desert reflexion. Int. J. Remote Sens. 14, 21–52.
Kriebel, K.T., 1989. Cloud liquid water path derived from AVHRR data using APOLLO. Int. J. Remote Sens.
10, 723–729.
A. Fouilloux et al.r Atmospheric Research 55 (2000) 65–83
83
Kriebel, K.T., Saunders, R., Gesell, G., 1989. Optical properties of clouds derived from fully cloudy AVHRR
pixels. Beitr. Phys. Atmos. 62, 165–171.
Nakajima, T., King, M.D., 1990. Determination of the optical thickness and effective particle radius of clouds
from reflected solar radiation measurements: Part I. Theory. J. Atmos. Sci. 47, 1878–1893.
Nakajima, T.Y., Nakajima, T., 1995. Wide-area determination of cloud microphysical properties from
NOAA-AVHRR measurements for FIRE and ASTEX regions. J. Atmos. Sci. 52, 4043–4059.
Nakajima, T.Y., Tanaka, M., 1986. Matrix formulation for the transfer of solar radiation in plane-parallel
scattering atmosphere. J. Quant. Spectrosc. Radiat. Transfer 35, 13–21.
Parol, F., Buriez, J.C., Brogniez, G., Fouquart, Y., 1991. Information content of AVHRR channels 4 and 5
with respect to the effective radius of cirrus cloud particles. J. Appl. Meteorol. 30, 973–984.
Pawlowska, H., Brenguier, J.L., Burnet, F., 2000a. Microphysical properties of stratocumulus clouds. Atmos.
Res., in press.
Pawlowska, H., Brenguier, J.L., Fouquart, Y., Armbruster, W., Descloitres, J., Fischer, J., Flamant, C.,
Fouilloux, A., Gayet, J.F., Ghosh, S., Jonas, P., Parol, F., Pelon, J., Schuller,
L., 2000b. Microphysical and
¨
radiative properties of stratocumulus cloud: the EUCREX mission 206 case study. Atmos. Res., in press.
Rossow, W.B., 1989. Measuring cloud properties from space: a review. J. Clima. 2, 201–213.
Saunders, R.W., Kriebel, K.T., 1988. An improved method for detecting clear sky and cloudy radiances from
AVHRR data. Int. J. Remote Sens. 9, 123–150.
Stephens, G.L., 1978. Radiation profiles in extended water clouds: II. Parameterization schemes. J. Atmos.
Sci. 35, 2123–2132.
Stephens, G.L., Ackerman, S., Smith, E.A., 1984. A shortwave parameterization revised to improve cloud
absorption. J. Atmos. Sci. 41, 687–690.
Unser, M., 1986. Sum and difference histograms for texture classification. IEEE Trans. Pattern Anal. Mach.
Intell. PAMI-8, 118–125.
Weszka, J.S., Dyer, C.R., Rosenfeld, A., 1976. A comparative study of texture measures for terrain
classification. IEEE Trans. Syst. Man. Cybern. SMC-6, 2269–2285.