new dust aerosol identification method for spaceborne lidar measurements

8
New dust aerosol identification method for spaceborne lidar measurements Yingying Ma a , Wei Gong a,n , Pucai Wang b , Xiuqing Hu c a State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, Hubei 430079, China b Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing 100029, China c National Satellite Meteorological Center, China Meteorological Administration, Beijing 100081, China article info Keywords: CALIPSO SVM (the support vector machine) Aerosol Cloud abstract Classification is a critical step in the backscatter lidar data processing to accurately retrieve extinction and backscatter profiles of atmospheric aerosols and clouds. Different schemes, such as the probability distribution functions (PDFs) method, have been used in the cloud and aerosol classification. In this paper, we attempt to use the support vector machine (SVM) to discriminate aerosols from clouds, with a focus on dust aerosol classification in China. To demonstrate the feasibility of the SVM classifier, we chose dust storms that occurred in the Gobi and Taklimakan deserts and observed by the CALIPSO lidar in spring time 2007. The results show that the SVM can correctly identify the dust storms. & 2010 Elsevier Ltd. All rights reserved. 1. Introduction Over East Asia, soil and mineral dust aerosols dominate aerosol loading during dust active seasons [1]. Many researches have been carried out to observe Asian dust. These researches include ground-based observations such as the NIES lidar network using active instrumentations, and AERONET (AErosol RObotic NETwork) using passive instrumentation (sun-photometer). These ground-based networks can obtain more accurate and continuous ob- servation data. However, the coverage of these networks is limited mainly to land. It is difficult to obtain dynamic information of the dust transport on a large scale [2]. On the other hand, spaceborne observations provide a global coverage that enables the track of the dust transport on a global scale. The MODerate resolution Imaging Spectro- radiometer (MODIS) aboard the NASA’s TERRA and AQUA satellites has been widely used in aerosol and climate studies [3]. However, these spaceborne remote sensors are passive facilities and hard to provide vertical distribution of clouds and aerosols. In addition, these passive measure- ments are strongly impacted by the surface conditions and, for the measurements in the visible regime, limited to day time only. The Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) onboard the Cloud Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) is an active instrument [4]. The CALIOP’s high-resolution vertical profiling ability and accurate depolarization measure- ments make it a superb platform for the study of dust aerosols [5]. Therefore, the CALIOP measurements have been used to characterize the dust occurrence over the Tibet Plateau and the surrounding area (e.g. [5]) and validate the dust transport model (e.g. [1]). In the CALIOP production data processing, the classi- fication between aerosols and clouds is the first step of the scene classification algorithm (a detailed description will be given in Section 3). The parameters used in the classification do not include the depolarization ratio in the version 2 algorithm. This parameter is used as an additional dimension of the probability density functions Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/jqsrt Journal of Quantitative Spectroscopy & Radiative Transfer 0022-4073/$ - see front matter & 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.jqsrt.2010.08.004 n Corresponding author. E-mail addresses: [email protected], [email protected] (W. Gong). Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345

Upload: yingying-ma

Post on 26-Jun-2016

212 views

Category:

Documents


0 download

TRANSCRIPT

Contents lists available at ScienceDirect

Journal of Quantitative Spectroscopy &Radiative Transfer

Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345

0022-40

doi:10.1

n Corr

E-m

(W. Gon

journal homepage: www.elsevier.com/locate/jqsrt

New dust aerosol identification method for spacebornelidar measurements

Yingying Ma a, Wei Gong a,n, Pucai Wang b, Xiuqing Hu c

a State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, Hubei 430079, Chinab Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing 100029, Chinac National Satellite Meteorological Center, China Meteorological Administration, Beijing 100081, China

a r t i c l e i n f o

Keywords:

CALIPSO

SVM (the support vector machine)

Aerosol

Cloud

73/$ - see front matter & 2010 Elsevier Ltd. A

016/j.jqsrt.2010.08.004

esponding author.

ail addresses: [email protected], weigong

g).

a b s t r a c t

Classification is a critical step in the backscatter lidar data processing to accurately

retrieve extinction and backscatter profiles of atmospheric aerosols and clouds.

Different schemes, such as the probability distribution functions (PDFs) method, have

been used in the cloud and aerosol classification. In this paper, we attempt to use the

support vector machine (SVM) to discriminate aerosols from clouds, with a focus on dust

aerosol classification in China. To demonstrate the feasibility of the SVM classifier, we

chose dust storms that occurred in the Gobi and Taklimakan deserts and observed by

the CALIPSO lidar in spring time 2007. The results show that the SVM can correctly

identify the dust storms.

& 2010 Elsevier Ltd. All rights reserved.

1. Introduction

Over East Asia, soil and mineral dust aerosols dominateaerosol loading during dust active seasons [1]. Manyresearches have been carried out to observe Asian dust.These researches include ground-based observations such asthe NIES lidar network using active instrumentations,and AERONET (AErosol RObotic NETwork) using passiveinstrumentation (sun-photometer). These ground-basednetworks can obtain more accurate and continuous ob-servation data. However, the coverage of these networks islimited mainly to land. It is difficult to obtain dynamicinformation of the dust transport on a large scale [2]. On theother hand, spaceborne observations provide a globalcoverage that enables the track of the dust transport ona global scale. The MODerate resolution Imaging Spectro-radiometer (MODIS) aboard the NASA’s TERRA and AQUAsatellites has been widely used in aerosol and climate

ll rights reserved.

@lmars.whu.edu.cn

studies [3]. However, these spaceborne remote sensors arepassive facilities and hard to provide vertical distribution ofclouds and aerosols. In addition, these passive measure-ments are strongly impacted by the surface conditions and,for the measurements in the visible regime, limited to daytime only.

The Cloud-Aerosol Lidar with Orthogonal Polarization(CALIOP) onboard the Cloud Aerosol Lidar and InfraredPathfinder Satellite Observations (CALIPSO) is an activeinstrument [4]. The CALIOP’s high-resolution verticalprofiling ability and accurate depolarization measure-ments make it a superb platform for the study of dustaerosols [5]. Therefore, the CALIOP measurements havebeen used to characterize the dust occurrence over theTibet Plateau and the surrounding area (e.g. [5]) andvalidate the dust transport model (e.g. [1]).

In the CALIOP production data processing, the classi-fication between aerosols and clouds is the first step of thescene classification algorithm (a detailed description willbe given in Section 3). The parameters used in theclassification do not include the depolarization ratio inthe version 2 algorithm. This parameter is used as anadditional dimension of the probability density functions

Table 1The horizontal and vertical resolution of data after averaging.

Altitude range

(km)

Horizontal resolution

(km)

Vertical resolution

(km)

30.1–40.0 60 0.3

20.2–30.1 25 0.18

8.2–20.2 15 0.06

0.01–8.2 5 0.03

Y. Ma et al. / Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345 339

(PDFs) of clouds and aerosols in the version 3 algorithm.The depolarization ratio is calculated from the ratio of theperpendicular channel and the parallel channel signals at532 nm [6]. It is a useful indicator that can be used toidentify irregular particles (such as dust particles and icecrystals) from spherical particles. In the current version ofthe algorithm described in this paper, however, thedepolarization ration is not used. This parameter will beused in future. In this study we use the SVM classifier toperform the scene classification. The observation objectchosen is the dust storms that occurred in spring (March30 and May 10) 2007 in the Gobi and Taklimakan desertsin China (40–451N, 109.16–107.471E, 37–421N, and86.91–85.351N), respectively.

Section 2 below describes the data we use. Section 3introduces the new dust aerosol classification method.Section 4 presents the test results of the classification andthe corresponding discussions. Conclusions are summar-ized in Section 5.

2. Data

The CALIPSO satellite is one of the Afternoon satelliteconstellation (A-Train) which consists of several Earthobservation satellites maintained in sun-synchronous orbitsat an altitude of �700 km [4]. They are spaced a few minutesapart from each other so they are considered as providingsynchronous observations. The CALIPSO lidar (CALIOP) is anactive instrument; the signal it receives is the profile of theatmosphere. The SVM algorithm studied in this paper isapplied to the atmospheric features detected by the CALIPSOlidar to classify their type (cloud vs. aerosol). The CALIOPversion 2 products are used. To assist select the requiredtraining samples and test the SVM algorithm, we use theCloudSat data. CloudSat is another satellite in the A-Trainconstellation that is ahead of CALIPSO by �15 s. It carries acloud profiling radar (CPR) that can penetrate optically thickclouds and profile the clouds consisting of large particles. Inthis study, the MODIS imagery data acquired on the Aquasatellite is also used to identify the horizontal distribution ofclouds and dust. The Aqua satellite leads the A-Trainconstellation and is ahead of CALIPSO by � 1 min.

2.1. CALIPSO

The CALIPSO payload consists of three nadir-viewinginstruments: the two-wavelength polarization-sensitivebackscatter lidar (CALIOP), wide field camera (WFC), andimaging infrared radiometer (IIR) [4]. The CALIOP level 1major data products is a set of calibrated profiles of totalattenuated backscatter at 532 nm and its perpendicularcomponent and attenuated backscatter at 1064 nm. TheCALIOP level 2 data products include aerosol and cloudlayer and profile products. The parameters reported inthese products include, for example, vertically resolvedaerosol and cloud layers, extinction, optical depth, aerosoland cloud type, cloud water phase, cirrus emissivity,and particle size, and shape [7]. In this paper, we use theLevel 1B data. We first search the total attenuatedbackscatter profiles to find the location (top and base

heights) of each layer. Then, the mean attenuated back-scatter, total color ratio, and volume depolarization ratioof each layer are calculated. To improve the signal-to-noise ratio (SNR), the L1B CALIOP data are averaged for 15profiles, corresponding to a horizontal resolution of 5 km.The vertical resolution still remains the same as theoriginal ones (i.e., 30, 60, 180, and 300 m in the differentaltitude ranges) as described in Table 1.

2.2. CloudSat

Onboard the CloudSat payload is a 94-GHz, nadir-pointing Cloud Profiling Radar (CPR). The purpose of theCloudSat mission is to measure the vertical structure ofclouds from space and simultaneously observe cloud andprecipitation. It provides vertically resolved informationon cloud location, cloud ice and liquid water content(IWC/LWC), precipitation, cloud classification, radiativefluxes, and heating rates. The vertical resolution is 480 mwith 240 m sampling, and the horizontal resolution isapproximately 1.4 km (cross-track)�2.5 km (along-track)with sampling roughly every 1 km [8].

A unique feature that CloudSat brings to the constella-tion is the ability to fly a precise orbit enabling the fieldsof view of the CloudSat radar to be overlapped with theCALIPSO lidar footprint and the other measurements ofthe constellation. The precision of this overlap creates aunique multi-satellite observing system for studyingthe atmospheric characteristics. In this paper, we usethe 2B-GEOPROF data files. A cloudy range bin in2B-GEOPROF is associated with a confidence mask valuethat ranges from 0 to 40. Values Z30 are confidentlyassociated with clouds and a value of 6 suggests cloudsapproximately 50% of the time. CloudSat can measure thelocation of thick clouds and find the total attenuated area,providing additional information that can be used toidentify the cloud signals measured by CALIOP and selecttraining samples.

2.3. Aqua

Aqua-MODIS views the entire Earth’s surface every1–2 days and acquires data in 36 spectral bands (see theMODIS Technical Specifications). These data will improveour understanding of global dynamics and processesoccurring on the land, in the oceans, and in the loweratmosphere. MODIS is playing a vital role in the develop-ment of validated global interactive Earth system modelsthat is able to predict global changes and accurately assist

w

Y. Ma et al. / Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345340

the policy makers in making sound decisions concerningthe protection of our environment [9].

This MODIS image data acquired during the day time isemployed to help select data samples and test the SVMclassifier.

Margin = 2/ w

H = -1H = 0

H = 1

Fig. 2. Maximum-margin hyperplane and margins for a SVM trained

with samples from two classes. Samples on the margin are called the

support vectors.

3. Methods

To retrieve the extinction and backscatter coefficientprofiles of aerosols and clouds using the existing lidarinversion technique, the extinction-to-backscatter ratio(or lidar ratio) has to be selected in the most cases basedon the classification results of feature types [10].Generally, different types of aerosols have different lidarratios. Consequently, we need to distinguish the aerosoltype. Before subtyping aerosols, we need to distinguishthe aerosol from cloud first. The flow of the cloud andaerosol classification and the following data processing isshown in Fig. 1.

In the CALIOP product data processing, the PDFs basedalgorithm is used [11]. In this study, we use the support

First phase

Find atmospheric layers in 5-kmbackscatter profile

Yes

Subdivide the aerosol by

known model

The subdividedAerosol type

Execute retrievalLayer’s optical parameter

Calculate physical andoptical character of ealayer

SVM

Aerosol?

Fig. 1. Flow chart of the

vector machine (SVM). Unlike the PDF based methods,SVMs do not have to develop PDFs. However, a SVM needsto define a hyperplane in the feature space based on

Second phase

attenuated

No

Subdivide the cloud by other

parameters

Selection of the corresponding lidar ratio

ch

The subdividedCloud type

data processing.

Fig. 3. (a) Aerosol samples, (b) thick cloud samples, (c) thin cloud samples, these samples were chosen from CALIPSO observation, and (d) use the

CloudSat data to validate the thin layer suspended was the cloud top, and the thick cloud induced the laser cannot penetrated.

Y. Ma et al. / Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345 341

Y. Ma et al. / Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345342

carefully selected training samples. As shown in Fig. 2, thereare two classes. After finding the support vector, thehyperplane is also defined. The samples on the brokenlines in Fig. 2 are the support vectors, and the solid line isthe hyperplane. Vector w is a normal vector perpendicularto the hyperplane. x is an n-dimensional real vector,x¼ ðx1,. . .,xnÞ, where xi is a real number. Each xj representsa feature layer and belongs to a class. In this study, x1,:::,xn

are the layer top and base, attenuated backscatter coeffi-cient, and color ratio. Consider a dot product space Rn, inwhich the samples x are embedded, x1,. . .,xm 2 Rn. Anyhyperplane in space Rn can be written as

fx 2 Rn9wUxþb¼ 0g, w 2 Rn,b 2 R ð1Þ

The dot product, wUx, is defined by

wUx¼Xn

i ¼ 1

wixi ð2Þ

A training set of samples is linearly separable, if thereexists at least one linear classifier defined by the pair ðw,bÞwhich correctly classifies all training patterns. Thislinear classifier is represented by the hyperplaneH(wUxþb¼ 0) and defines a region for class +1samples and another region for class �1 samples, when(wUxþb � 0, wUxþb!0). After being trained, the classi-fier is ready to predict the class membership for newsamples different from those used in training. Theclassification of samples depends only on the sign of theexpression wUxþb [12,13].

Samples selection is very important and indispensablestep in the supervised classification. If the classification ofsamples is incorrect or the selected samples are notstatistically significant and representative, the accuracy ofclassification is influenced directly. The CALIPSO lidar hasbeen performing observations in space for more than fouryears. A plenty of observation data can be used. We canselect layers with unambiguous characteristics as thetraining data set. Three different cases are selected, as

Fig. 4. Classification between dust aerosol and cloud b

shown in Fig. 3. These cases include features of aerosol,thick cloud and thin cloud, measured on March 2, 2008,(28–341N), March 19, 2008, (24–301N), and on March 25,2008, (42.5–44.81N), respectively. The correspondingCloudSat Level 2 data observed on March 19 is shown inFig. 3c, which confirms that the layers observed by CALIOPon March 19 (Fig. 3b) are cloud. All the profileshorizontally are averaged to 5 km and 1061 samples intotal (aerosol: 302, cloud: 759) are obtained.

The overall accuracy and Kappa coefficient, twoimportant parameters in the error analysis, are defined as

OA¼

Pki ¼ 1 xii

Nð3Þ

K ¼NPk

i ¼ 1 xii�Pk

i ¼ 1ðxiþ xþ iÞ

N2�Pk

i ¼ 1ðxiþ xþ iÞð4Þ

OA is the overall accuracy and K is the Kappa coefficient,N is the total number of samples, xii is the correctclassification, k represents the number of different classes(in the cloud and aerosol classification, there are only twoclasses, i.e., aerosol and cloud). xiþ and xþ i are theincorrect classification, that is, the cloud mislabeled asaerosol or the aerosol mislabeled as cloud. The inputparameters include the top and base of a layer, color ratio,and attenuated backscatter. A test shows that the overallaccuracy and Kappa coefficient are 0.961357 and0.907237, respectively.

4. Experiments

In this section, we present the implementation of theSVM algorithm and test the SVM classifier using theCALIOP measurements acquired over North China. Majordust sources in China are located in Xinjiang Province,Gansu Province, and Inner Mongolia. A large area in thenorthern China suffers from dust storms in springeach year. Therefore, we chose North China (40–451N,

y SVM (the radial basis function kernel is used).

Y. Ma et al. / Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345 343

109.17–107.471E) and Northwest China (371421N, 86.91–85.351E) for our experiment.

Fig. 4 presents the training samples selected from theCALIOP measurements shown in Fig. 3 and its hyperplane in atwo dimensional space (color ratio and attenuated back-scatter). ‘1’ and ‘2’ represent the cloud and aerosol layers. The‘red plus’ and ‘green asterisk’ symbols are the selectedtraining samples of aerosol and cloud, respectively. The

Fig. 5. (a) A dust storm observed by Aqua-MODIS on March 30, 2007. (b) The att

gray line in (a). (c) The CloudSat measurement along the same track indicate

confidence. (d) The CALIOP vertical feature mask (version 2). In this case the du

SVM algorithm developed in this study.

maximum margin hyperplane are defined using the supportvectors (the circled symbols). As seen in this figure, thetraining samples have the best classification. Finally, weapply the SVM to other samples shown in Fig. 3 and theresults are shown by the ‘magenta plus and cyan asterisk’ inFig. 4. Despite there are some error classifications in trainingsamples (linear nonseparable), this method can correctlyclassify most cases selected.

enuated backscatter observed by CALIOP along the track indicated by the

d by the gray line in (a). The orange-colored areas are cloud with high

st aerosol was misclassified as cloud. (e) The classification result by the

Y. Ma et al. / Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345344

To assess the performance of the SVM classifier, weexamine several cases. Fig. 5 presents a dust storm thatoccurred in the Gobi Desert and in the following days thedust plume blew eastward over the Beijing region andinto the Yellow Sea. The MODIS on the NASA’s Aquasatellite acquired an image of a dust storm on March 30,2007 (Fig. 5a). At the same time, the CALIPSO lidar, whichwas flown behind the Aqua satellite by �1 min, alsoobserved this dust layer (Fig. 5b). However, this dust layerwas misclassified as cloud (Fig. 5d) in the version 2release. As mentioned earlier, the CALIOP cloud andaerosol discrimination (CAD) is a PDF-based scheme. Inthe version 2 release, three-dimensional (3D) PDFsdeveloped before the CALIPSO launching was used[11,14]. The parameters used in the 3D PDFs include thelayer-averaged attenuated backscatter, total color ratio,and altitude. Because very dense dust layers over/near thesource regions fall in the overlap region of the cloud andaerosol PDFs in the 3D space, these dust layers can be

Fig. 6. (a) A dust storm observed by Aqua-MODIS on May 10, 2007, The gray line

of the version 2 release and (c) the classification result by SVM.

misclassified in the version 2 release [14]. Fig. 6 presentsanother dust storm case observed on May 10, 2007. Thedust plume covered most of the Taklimakan Desert asobserved by the MODIS sensor on the NASA’s Aquasatellite. SVM can also correctly classify the dust layers,but misclassify the aerosol as cloud layers over Mt.Kunlun north of �381N.

We also attempted to study the optical properties ofdust aerosols. Fig. 7 presents the calculated depolarizationand color ratio of cloud and aerosol observed by CALIOPon March 30. Clouds are seen at latitude ranges of40–411N and 44–44.51N at �10 and 5 km, respectively.The layer-average value of the volume depolarizationratio is 0.436 and 0.384, respectively, for cloud andaerosol. As shown in Fig. 7b, both the clouds and thedust aerosol have a large color ratio value, indicating largeparticle sizes. The averaged value is 1.03999 and 1.16094,respectively, for the cloud at �10 km altitude and �5 kmaltitude.

is the corresponding CALIPSO orbit. (b) The CALIOP vertical feature mask

Fig. 7. (a) Depolarization ratio and (b) color ratio of cloud and aerosol

observed on March 30, 2007. The resolution is 5 km.

Y. Ma et al. / Journal of Quantitative Spectroscopy & Radiative Transfer 112 (2011) 338–345 345

5. Conclusions

In this paper, we have developed a classifier based on thesupport vector machine (SVM) to discriminate dust fromclouds in North China where two major dust sources (Gobiand Taklimakan) are located. A set of training samples wasselected from the CALIOP measurements based on additionalinformation of layer types provided by the CloudSat radarmeasurements and the MODIS imagery data. The usedCALIOP measurements in the SVM classification includelayer’s top and base heights and layer-averaged attenuatedbackscatter and color ratio. The conducted test showed thatthe SVM classifier can correctly classify most dense dustlayers, but sometimes misclassify some clouds as aerosol. Inaddition, a feature finding algorithm has also been developedand used in this study.

More works are planned to improve the performanceof the SVM classifier in future. These works include thefollowing:

(1)

Depolarization ratio is a major product of the CALIOPmeasurements. It can provide additional informationabout the particle shapes. In general, nonsphericalparticles such as dust particles and ice crystals have alarge depolarization ratio, whereas spherical particleshave a polarization ratio of zero in the absence ofmultiple scattering. The depolarization ratio is cur-rently not used in the SVM classifier and will be usedin the future. With the addition of this parameter, thedust and clouds will have a better separation in the

feature space and therefore the SVM performance isexpected to be improved.

(2)

The capability of the SVM classifier will be extendedto classify other types of aerosols. This requires alarger set of training data with a global coverage. Thismay be done by selecting different geographic loca-tions where a specific type of aerosol typically exists.

(3)

The feature finding algorithm developed in this papercan also be improved with a large set of data.

With the improved detection and classification infuture, the optical and radiative properties of the aerosolsand clouds can be well studied. Our understanding aboutthe interaction between aerosols and clouds and theirimpacts on the climate changes can be improved.

Acknowledgement

Paper supported by 973 project (2009CB723905), 863Project (2009AA12Z107), and NSFC (10978003, 40871171).

References

[1] Yumimoto K, Uno I, Sugimoto N, Shimizu A, Liu ZY, Winker DM.Adjoint inversion modeling of Asian dust emission using lidarobservation. Atmos Chem Phys 2008;8:2869–84.

[2] Myhre G, Stordal F, Johnsrud M, Diner DJ, Geogdzhayev IV,Haywood JM, et al. Intercomparison of satellite retrieved aerosoloptical depth over ocean during the period September 1997 toDecember 2000. Atmos Chem Phys 2008; 4: 8201–8244.

[3] Chu DA, Kaufman YJ, Ichoku C, Remer LA, Tanre D, Holben BN.Validation of MODIS aerosol optical depth retrieval over land.Geophys Res Lett 2002;29(12)10.1029/2001GL013205 MOD2.

[4] Winker DM, Pelon J, McCormick MP. The CALIPSO mission:spaceborne lidar for observation of aerosols and clouds. Proc SPIE2003;4893:1–11.

[5] Liu ZY, Liu D, Huang JP, Vaughan M, Uno I, Sugimoto N, et al.Airborne dust distributions over the Tibetan Plateau and surround-ing areas derived from the first year of CALIPSO lidar observations.Atmos Chem Phys 2008;8:5957–77.

[6] Liu ZY, Voelger P, Sugimoto N. Simulations of the observation ofclouds and aerosols with the Experimental Lidar in Space Equip-ment system. Appl Opt 2000;39:3120–37.

[7] Vaughan MA, Young SA, Winker DM. Fully automated analysis ofspace-based lidar data: an overview of the CALIPSO retrievalalgorithms and data products. Laser radar techniques for atmo-spheric sensing, Gran Canaria, Spain, 2004.

[8] Stephens GL, Vane DG, Boain RJ, Mace GG, Sassen K, Wang Z, et al.The CloudSat mission and the A-Train. Bull Am Meteorol Soc2002;83:1771–90.

[9] Barnes W, Xiong X, Salomonson V. Status of Terra MODIS and AquaMODIS. Adv Space Res 2003;32(11):2099–106.

[10] Bosenberg J, Timm R, Wulfmeyer V. Study on retrieval algorithmsfor a backscatter lidar. Final report, no. 226, Max-Planck-Institut furMeteorologie, Hamburg, 1997.

[11] Liu ZY, Vaughan MA, Winker DM, Hostetler CA, Poole LR, Hlavka D,et al. Use of probability distribution functions for discriminatingbetween cloud and aerosol in lidar backscatter data. J Geophys ResAtmos 2004;109:D15202, doi:10.1029/2004JD004732.

[12] Pal M, Mather P. Support vector machines for classification inremote sensing. Int J Remote Sensing 2005;26(5):1007–11.

[13] Gunn SR. Support vectors machines for classification and regres-sion. Technical report, Image Speech and Intelligent SystemsResearch Group, University of Southampton, 1997. Available from:/http://www.isis.ecs.soton.ac.uk/resource/svminfo/S.

[14] Liu Z, Vaughan MA, Winker DM, Kittaka C, Kuehn RE, Getzewich BJ,et al. The CALIPSO Lidar Cloud and aerosol discrimination: version2 algorithm and initial assessment of performance. J Atmos OceanicTechnol 2009;26:1198–213, doi:10.1175/2009JTECHA1229.1.