CLOUD DETECTION FOR MERIS MULTISPECTRAL IMAGES

Luis Gomez-Chova1, Gustavo Camps-Valls1, Julia Amoros1, Jose D. Martın1, Javier Calpe1,Luis Alonso2, Luis Guanter2, Juan C. Fortea2, and Jose Moreno2

1GPDS, Electronic Engineering Department, University of Valencia∗2LEO, Earth Sciences Department, University of Valencia∗

*Doctor Moliner 50, 46100, Burjasot (Valencia), Spain.

ABSTRACT

Two of the key features of the MEdium Resolu-tion Imaging Spectrometer (MERIS) instrument onboard the ESA Envisat environmental satellite areits temporal resolution of three days and its spatialcoverage. The amount of images acquired over theglobe every day makes inevitable that many of theseimages present cloud covers. From an operationalpoint of view, an automatic and accurate methodfor cloud detection in MERIS scenes is a key issuesince, with no accurate cloud masking, undetectedclouds are the most significant source of error for trueground reflectance, affecting a wide range of remotesensing applications. By masking only the image ar-eas affected by cloud covers, the whole image is notnecessarily discarded.

The main objective of this work is to develop andvalidate a method for cloud detection using self-contained information provided by MERIS Full Res-olutiuon Level 1b products and a Digital Eleva-tion Model (DEM). The method must be capableof: detecting clouds accurately providing probabilityand cloud abundance rather than flags; and betterdescribing detected clouds (cloud abundance, type,height, subpixel coverage) in order to include thisinformation in radiation models.

Key words: cloud detection, cloud masking, multi-spectral images, MERIS, unsupervised classification,spectral unmixing.

1. INTRODUCTION

Accurate and automatic detection of clouds in satel-lite scenes is a key issue for a wide range of remotesensing applications. Clouds significantly affect theheat fluxes of the atmosphere and constitute one of

the most important transient phenomena to be in-corporated into climate models. Moreover, withoutan accurate cloud masking, undetected clouds in thescene are a significant source of error in both sea andland cover biophysical parameter retrieval [1].

The simplest approach to cloud detection in a sceneis the use of a set of static thresholds (albedo, tem-perature) applied to every pixel in the image. Thesemethods can fail for several reasons, such as sub-pixel clouds, high reflectance surfaces, illuminationand observation geometry, sensor calibration, vari-ation of the spectral response of clouds with cloudtype and height, etc [1]. Spatial coherence meth-ods have an advantage over static threshold meth-ods because they use the local spatial structure todetermine cloud free and cloud covered pixels [1].However, spatial coherence methods can fail whenthe cloud system is multilayered (which often is thecase), the clouds over the scene are smaller than theinstrument spatial resolution, or the scene presentscirrus clouds (which are not opaque). As a conse-quence, researchers have turned to developing adap-tive threshold cloud-masking algorithms [2].

The approaches for cloud detection and maskingheavily depend on the characteristics of each sensor.Obviously, its spectral and spatial resolution, or thespectral range are critical. For example, the presenceof channels in the thermal infrared enables detectionbased on thermal contrasts [3, 4]. Optical sensorswith spectral channels beyond 1 µm have demon-strated good capabilities to perform cloud masking.Some other recent hyperspectral sensors can not ex-ploit this spectral range, such as the MEdium Res-olution Imaging Spectrometer (MERIS) instrumenton board ESA Envisat environmental satellite. How-ever, one can take advantage of the high spectral andradiometric resolution, and the specific band loca-tions to increase the cloud detection accuracy, andto properly describe detected clouds [5].

In this context, the main objective of this paper

is to develop and validate a method for cloud de-tection using self-contained information provided byMERIS Level 1b products and a Digital ElevationModel (DEM). The method must be capable of: de-tecting clouds accurately providing probability andcloud abundance instead of flags; and better describ-ing detected clouds (cloud abundance, type, height,subpixel coverage) in order to include this informa-tion in radiation models [6].

2. DATA MATERIAL

A dataset consisting of four acquisitions over threesites has been selected since both Level 1b andLevel 2 products were available for all MERISFull Resolution (FR) images (300 m). In partic-ular, the site of Barrax (BR, Spain) was selectedas the main test site since it has been the coresite of previous Earth observation campaigns andthe analyzed cloudy images are part of the dataacquired in the framework of the SPARC 2003and 2004 ESA campaigns (ESA-SPARC Project,contract ESTEC-18307/04/NL/FF). These two im-ages were acquired the same day of two consec-utive years (July 14th, 2003 and 2004). Addi-tionally, a set of MERIS/AATSR sample productstaken over France (FR) and Finland (FI) have beenincluded in the study in order to take into ac-count their different characteristics: geographic lo-cation (latitude/longitude); date and season; type ofcloud (cumulus, cirrus, stratocumulus); and surfacetypes (soil, vegetation, sand, ice, snow, rivers, sea,etc). The selected images represent different scenar-ios very useful to validate the performance of themethod, including different landscapes; soils coveredby vegetation or bare; and two critical cases giventhe especial characteristics of the induced problems:ice and snow.

Although the proposed method can be applied onthe top of aerosols reflectance provided in Level 2products, it is only applied to the MERIS Level 1bproducts (top of atmosphere radiance) because Level2 products are processed from Level 1b using a cloudpixel classification that could be inaccurate. There-fore, Level 2 products are only used for validationpurposes by comparing its cloud flag product withthe cloud mask produced by the presented method.

3. IMAGE PRE-PROCESSING

Since the method must be general and must work un-der many situations, image information format mustbe homogeneous and with accurate data for all im-ages. Therefore, in order to remove the dependenceon particular illumination conditions (day of the yearand angular configuration) and illumination effects

due to rough terrain (cosine correction), Top Of At-mosphere (TOA) apparent reflectance is estimatedaccording to [7]:

ρ(λ) =π ·R(λ)

cos(θS) · S(λ), (1)

where R(λ) is the provided at sensor upward TOAradiance, S(λ) is the extraterrestrial instantaneoussolar irradiance, and θS is the angle between theillumination vector and a vector perpendicular tothe surface. In this work, θS is computed for eachpixel using the Sun Azimuth and Sun Zenith angles(provided in the Tie Point Location and AuxiliaryData of the MERIS product) and the vector per-pendicular to the surface, which is computed fromthe GETASSE30 DEM (included in the BEAM ESAsoftware). The Sun irradiance, S(λ), is taken fromThuillier et al. [8], corrected for the day of year whenthe acquisition takes place, and convolved with theMERIS spectral channels. (Fig.1).

Finally, one of the key features extracted fromMERIS in this work is obtained from the oxygen ab-sorption band that is extremely narrow. Therefore,before obtaining the TOA reflectances, in order tocorrect small variations of the spectral wavelength ofeach pixel along the image (smile effect), the BEAMSmile Correction Processor was used to calculate cor-rected radiances from the MERIS L1b products.

4. CLOUD DETECTION ALGORITHM

In this work, we present a cloud detection procedurewhich is constituted by the following steps:

1. Feature extraction: physically-inspired featuresare extracted to increase separability of cloudsand any-other surface type.

2. Region of interest : growing maps are built fromcloud-like pixels in order to select regions whichpotentially could contain clouds.

3. Image clustering : an unsupervised clustering al-gorithm is applied using all extracted featuresin order to obtain all the existing clusters overthe previously identified areas providing a prob-abilistic membership of pixels to each cluster.

4. Cluster labeling : the obtained clusters are la-beled into geo-physical classes taking into ac-count the spectral signature of the cluster cen-ters, which allows to merge all the cloud-clustersproviding a probabilistic cloud index.

5. Spectral unmixing : an spectral unmixing algo-rithm is applied to the segmented image in or-der to obtain an abundance map of the cloudcontent in the cloud pixels.

300 400 500 600 700 800 900 1000 11000

1000

2000

3000Solar Irradiance (mW·m−2·nm−1)

wavelength (nm)

300 400 500 600 700 800 900 1000 11000.5

0.6

0.7

0.8

0.9

1Spectral efficiency of MERIS channels

wavelength (nm)

Cosine correction: Illumination effects in rough terrain.

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

400 450 500 550 600 650 700 750 800 850 90020

40

60

80

100Radiance (mW⋅m−2⋅nm−1⋅sr−1)

wavelength (nm)

400 450 500 550 600 650 700 750 800 850 9000.1

0.2

0.3

0.4

0.5TOA Apparent Reflectance

wavelength (nm)

Figure 1. Left: Sun irradiance corrected for the day of the year and the spectral channels. Center: Cosineof the angle between the illumination vector and a vector perpendicular to the surface. Right: TOA apparentreflectance estimated from the at sensor radiance.

4.1. Feature Extraction

The measured spectral signature depends on the illu-mination, the atmosphere, and the surface. Spectralbands free from atmospheric absorptions contain in-formation about the surface reflectance, while othersare mainly affected by the atmosphere. Therefore,a series of considerations have been taken into ac-count, as it is explained in the following paragraphs.For illustration purposes, the figures of this sectionwill show maps of the extracted features from one ofthe images (BR-2003-07-14).

Regarding the reflectance of the surface, one of themain characteristics of clouds is that they presentbright and white spectra.

• A bright spectrum means that the intensity ofthe spectral curve (related to the albedo) mustpresent relatively high values. Therefore, cloudbrightness is calculated as the squared sum ofspectral bands normalized by the number ofbands. Bands that correspond to severe atmo-spheric absorptions are not included in orderto consider only the target, thus avoiding at-mospheric absorption effects. Considering theintensity independently in the VIS and NIRranges discriminates clouds better since the restof the surfaces have less reflectance in the VISrange (Fig.2).

• A white spectrum means that the first deriva-tive of spectral curve must present low values.As previously discussed, bands on atmosphericabsorptions are not included in order to consideronly the surface contribution. Noise or calibra-tion errors may reduce the accuracy in the esti-mation of the spectrum flatness. Smoothing oraverage ratios between the VIS and NIR can bemore robust. For the present method, the meanspectral derivative has been chosen as one of thefeatures (Fig.2).

Spectral intensity

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

Mean spectral derivative

−15

−10

−5

0

5

x 10−4

Figure 2. Cloud brightness ( left) and Mean SpectralDerivative of the TOA reflectance of the BR-2003-07-14image ( right) of the BR-2003-07-14image.

Regarding the atmospheric absorptions present inthe spectrum of a pixel, another meaningful featureis the fact that clouds are at a higher altitude thanthe surface. It is worth noting that atmosphericabsorption depends on the atmospheric componentsand the optical path. Since light reflected on highclouds crosses a shorter section of the atmosphere,the consequence would be an abnormally short op-tical paths, thus weaker atmospheric absorption fea-tures. Atmospheric oxygen or water vapour absorp-tions may be used to estimate this optical path.

The light transmitted through a non-dispersivemedium can be expressed by (Bouguer-Lambert-Beer law):

ρ(λ) = ρ0(λ) · exp(−k(λ) · d) , (2)

where the term exp(−k(λ) · d) is the transmittancefactor; k(λ) is the atmospheric optical depth for avertical path; and d is a factor accounting for thepath crossed by the radiation (product of the com-ponent concentration and the distance crossed by theradiation that will be approximately 1 when the lightcrosses one atmosphere with a vertical path).

Assuming that there are no horizontal variations inthe atmospheric component concentrations, one canconsider that the optical path only depends on the

750 755 760 765 770 775 780 7850

0.5

1

Normalized Atmospheric Transmittance in the 760.625nm absorption band

750 755 760 765 770 775 780 785

0.4

0.6

0.8

1

1.2

Spectral efficiency of sensor channels

wavelength (nm)400 450 500 550 600 650 700 750 800 850 900

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5Estimated absorption at 760.625nm

wavelength (nm)

TO

A r

efle

ctan

ce

Oxygen absorption

0

0.1

0.2

0.3

0.4

0.5

0.6

Figure 3. Intermediate products used to estimate the optical path, d, from the O2 absorption band. Left:effective atmospheric vertical transmittance, k(λ), estimated for the MERIS channels. Center: the interpolated,ρ0(λ), and measured, ρ(λ), reflectances inside the oxygen band, and the estimated reflectance at the maximumabsorption (λ=760.625 nm). Right: Estimation of the optical path from the oxygen absorption band for theBR-2003-07-14image.

Digital Elevation Model (m)

0

200

400

600

800

1000

1200

1400

1600

1800

2000

⇒

Difference betwen the DEM and the O2 height estimation (m)

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

5500

Figure 4. Left: DEM altitude of the BR-2003-07-14image. Center: Estimation of the topographic contributionto the optical path from the oxygen absorption band by successive linear regressions. Rigth: Difference betweenthe DEM altitude and the O2 height estimation for the BR-2003-07-14image.

altitude. Therefore, a robust estimation of the atmo-spheric absorption in the so-called Oxygen-A bandprovides a measure of the optical path. Even thoughMERIS has two specific spectral bands in this region(bands 10 and 11), estimating the optical path, d, isnot straightforward. The approach followed in thispaper can be devised from Fig.3. Since the O2 ab-sorption band is extremely narrow, the effective at-mospheric vertical transmittance, k(λ), is estimatedfor the MERIS channels from a high resolution curve.The spectral signature without absorption, ρ0(λ), isestimated by interpolating the nearby samples thatare unaffected by this process. Finally, the ratios be-tween the interpolated and the measured reflectanceat the affected bands provided by the instrument (λi)are used to estimate d.

d = average(− ln(ρ(λi)/ρ0(λi))

k(λi)

)(3)

However, a short optical path can be obtained overhigh altitude locations (light crosses less atmospherewhen is reflected at a high mountain and low valuesof d are found). Therefore, when a good DEM isavailable, the extracted features can be improved by

removing topographic contributions with an empiri-cal model. Considering only the cloud free pixels, theoptical path, (1 − d), can be related to the DEM inorder to assess the real altitude of targets for a pos-terior removal of topographic effects (see details inFig.4), thus obtaining relative measurements to thesurface (approximately zero in cloud free pixels). Inthis process, we neglect dependence on the temper-ature and the pressure since the objective is not anaccurate or unbiased height estimation but a goodrelative measure for cloud detection.

An additional estimation of the optical path can beobtained from the water vapour absorption in theNIR close to the end of the valid range of the sen-sor (900 nm). In this case, the maximum watervapour absorption (940 nm) is located outside theMERIS range and the water vapour distribution isextremely variable, thus it is not straightforward torelate this feature to the real altitude. However,it is still valid for relative measurements inside thesame image since almost all the atmospheric watervapour is distributed in the first 2-3 km of the at-mosphere below most of the cloud types. Moreover,

Water and Cloud Mask

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2Region of Interest

Figure 5. Result of the threshold-basedcloud/land/water classification ( left) and the grow-ing algorithm ( right) for the BR-2003-07-14image(significant pixels in white).

snow presents higher absorption than clouds at 900nm and this behaviour can be appreciated and is usedin the extracted feature. The same approach than inthe O2 case has been followed to obtain this featurebut using bands 13, 14 and 15.

4.2. Region of interest

As previously discussed, static thresholds appliedto every pixel in the image can fail due to sub-pixel clouds, sensor calibration, variation of the spec-tral response of clouds with cloud type and height,etc. On the other hand, unsupervised segmentationmethods can find clusters of similar pixels in the im-age but direct relation of clusters with desired classes(e.g. clouds) is not ensured. In order to mitigatethis drawback, in section 4.1, features have been ex-tracted to increase separability of clouds from any-other surface type. Moreover, if clouds were not sta-tistically representative in a given image, clusteringmethods could not find small clouds or could mixthem with other similar classes. Therefore, in addi-tion to using representative features, clustering im-proves if it is only applied over the regions of theimage where clouds are statistically representative.

In order to find these regions that potentially couldcontain clouds, hard non-restrictive thresholds areused to provide a first map of cloud-like pixels. Theseabsolute thresholds were obtained empirically andwere applied to well-defined features: the brightnessin the VIS and the NIR region, the estimated watervapour absorption, and the NDVI (in order to ex-clude areas with vegetation). Then, a region growingalgorithm is carried out, along with a morphologicalprocess that dilates cloudy areas, in order to ensurethat all possible clouds and their contiguous areaswill be considered in the clustering. The result isnot intended to be a classification, it just providesa mask which delimits the region of interest (ROI),where clouds will be significant for the later cluster-ing (Fig.5).

4.3. Image Clustering and Labeling

The clustering algorithm is applied only to the re-gions selected as ROI using the extracted features:the brightness in the VIS and the NIR region, themean spectral derivative of the TOA reflectance,and the estimated oxygen and water vapour absorp-tions. We use the Expectation-Maximization (EM)algorithm [9] since it considers the full relationshipamong variables and provides probability maps foreach cluster when subsequently applying a GaussianMaximum Likelihood (GML) classifier. If the cloudROI has been correctly found, even using a low num-ber of clusters N , some of them should correspondto different cloud types.

Once clusters are determined, the spectral signatureof each cluster, si(λ), is estimated as the average ofthe spectra of the cluster pixels but excluding thosewith abnormally low membership values (posteriorprobability Pik)1. It is important to emphasize thatthese spectral signatures of each cluster could dif-fer a lot from the spectra obtained when applyingthe EM algorithm over the image using the spec-tral bands rather than the extracted features. Theextracted features used to find the clusters are op-timized to increase separability between clouds andany-other surface type while in the spectral domainthese clusters could present a high degree of overlap.At this point of the process, the obtained clusterscan be labeled into geo-physical classes taking intoaccount three complementary sources of information(Fig.6): the thematic map with the distribution ofthe clusters in the scene, the spectral signatures ofthe cluster, si, and the location in the image of thepixels with the spectral signature closer to si. Thisinformation can be either analyzed directly by theuser or compared to a spectral library with represen-tative spectra of all the classes of interest.

Once all clusters have been related to a class witha geo-physical meaning (Fig.7), it is straightfor-ward to merge all the clusters belonging to a cloudtype. Since the EM algorithm provides posteriorprobabilities1 (Pik ∈ [0, 1] and

∑Ni=1 Pik = 1), a

probabilistic cloud index, based on the clustering ofthe extracted features, can be computed as the sumof the posteriors of the cloud-clusters: Cloud Prob-abilityk =

∑i Pik ∀i classified as cloud. However, if

the clusters are well separated in the feature space,the posteriors decrease drastically from one to zero inthe boundaries between clusters (Fig.7). Therefore,this Cloud Probability index indicates the probabilitythat one pixel belongs more to a cloud-cluster thanto one of the other clusters found in the image, butit does not give information about the cloud contentat subpixel level (very important when dealing withthin clouds or partially covered pixels).

1The posterior probability of cluster ci given the sample

xk is given by the Bayes theorem by P (i|xk) =p(xk|i)·P (i)

p(xk),

where P (i) is the a priori probability of cluster ci.

Clustering of selected areas

0

1

2

3

4

5

6

7

8

400 450 500 550 600 650 700 750 800 850 9000

0.2

0.4

0.6

0.8

1

1.2

1.4Cluster centers

wavelength (nm)

TO

A A

ppar

ent R

efle

ctan

ce

12345678

1

23

45

6

7

8

Figure 6. Thematic map with the distribution of the clusters in the scene ( left), the spectral signatures of theclusters ( center), and the location in the image of the pixels with the most similar spectra ( right).

Classification of selected areas

0:background

1:soil

2:cirrus

3:soil

4:soil

5:clouds

6:bright clouds

7:soil

8:soil

Cloud Probability

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1Unmixing Cloud Abundances

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 7. Left: Thematic map with the distribution in the scene of the classes of the clusters. Center: CloudProbability index computed from the posteriors of the cloud-clusters. Right: Cloud Abundance computed fromthe unmixing coefficients of the cloud clusters.

4.4. Spectral Unmixing

In order to obtain a cloud abundance map for everypixel in the image, rather than flags or a binary clas-sification, a spectral unmixing algorithm 2 is appliedto the MERIS image using the full spectral informa-tion. In our case, we are going to consider the spec-tral signatures of the clusters, si, as the representa-tive pixels of the covers present in the scene, and theyare used to build matrix M. The Fully ConstrainedLinear Spectral Unmixing (FCLSU) [10] algorithmis used to perform the spectral unmixing. This algo-rithm solves a constrained linear least-squares prob-lem minimizing the norm of (M · ak − ρk) where thevector, ak, of independent variables is restricted tobeing nonnegative (since it represent the abundancesor contributions of reflectance spectral signatures)and its sum being one (since it is supposed that Mcontains at least one spectrum for all the compoundsin the image). If the ROI where the clusters werefound did not cover the whole image, but the abun-

2The spectral unmixing algorithm expresses each pixel ofthe image, ρk, as a linear combination of a set of basis vectors,M, being the coefficients, ak, of this combination, ρk = M·ak,the unmixing coefficients, which can be interpreted as theabundances of the spectral components expressed in M (usu-ally called endmembers).

dance map for all the image has to be obtained, it isnecessary to include in M some representative pixelsof the non analyzed areas by performing a cluster-ing in these areas outside the ROI. Otherwise, theassumption that motivates the second constraint ofthe FCLSU algorithm can be false if there are newnon-considered spectral classes outside the ROI, pro-viding misleading abundances in these areas due tothe high residuals (M · ak − ρk).

The vector ak contains the abundances of the spec-tral signatures of the clusters, which are related toa class with a geo-physical meaning, for the sam-ple pixel k. As it happens with the probabilities ofthe clusters, the abundance of cluster ci for the sam-ple k is aik ∈ [0, 1] and

∑Ni=1 aik = 1. Therefore,

the abundance of cloud is computed as the sum ofthe abundances of the cloud-clusters: Cloud Abun-dancek =

∑i aik ∀i classified as cloud (Fig.7). As

in the case of the probabilities a threshold of 0.5would give a good cloud mask, but some false de-tections could appear since the unmixing has beenperformed on the basis of spectral signatures thatcould be non-pure pixels or non completely indepen-dent. An improved cloud abundance map can beobtained when combining the Cloud Abundance andthe Cloud Probability by means of a pixel-by-pixel

multiplication. In this way, we combine two com-plementary sources of information processed by in-dependent methods: the degree of cloud abundanceor mixing (obtained from the spectral information)and the cloud probability that is close to one in thecloud-like pixels and close to zero in remaining areas(obtained from the extracted features).

5. RESULTS

The presented method was tested in a previous work[11] on CHRIS/Proba hyperspectral images in orderto propose and validate cloud detection methodolo-gies. The use of this data allowed us to assess al-gorithm performance in favorable spatial resolution(34 m) and number of bands (62 channels). How-ever, MERIS products allow us to take advantageof: the illumination and observation geometry, anoverlapped DEM, and an accurate oxygen absorp-tion estimation. In this section, results of the pro-posed scheme for the aforementioned images will beshown. The by-pass images have been shown in theprevious section (Fig.2 to Fig.7), and now we will an-alyze its performance by comparing the final cluster-ing classification and cloud abundance product withthe RGB composite of the MERIS images and theMERIS Level 2 Cloud Flag respectively (Fig.8).

The two images over the Barrax (Spain) site are agood example of an easy cloud detection problem,when dense clouds are well contrasted with soil andvegetation. The ROI selection can be easily appre-ciated in the classification images, being more im-portant in the 2004-07-14 image where small cloudscloud be mixed in a cluster with other classes if thewhole image is considered. At the 2003-07-14 image,dry soil pixels belong to a cluster labeled as clouddue to their high reflectance and whiteness, but theypresent low probabilities and abundances. The 2004-07-14 image presents thin and small clouds over landand over sea, which are well detected since a specificcluster describe them. One of the weak points of thealgorithm is the use of thresholds to select the ROI,because some thin or small clouds can be excludedfrom the ROI. A possible solution for this is to relaxthresholds at the risk of considering the whole imageas ROI. But, even in this case, results are good ifclouds cover a sufficient percentage of the image orthe number of clusters is high enough (as in the imageof Finland). The presence of bright pixels is one ofthe critical issues in cloud detection (e.g. ice/snowin the surface), since these pixels and clouds havea similar reflectance behavior. However, the atmo-spheric absorption suffered by cloud pixels is lowerthan for the surface pixels due to their height, anddifferent clusters are found for these two classes inthe image. Thanks to the extracted atmospheric fea-tures, ice/snow pixels present low Cloud Probabilityvalues although the Cloud Abundance provided bythe spectral unmixing could be relatively high due

to the spectral similarities. In consequence, bothinformation types are combined improving the finalclassification accuracy.

Finally, figures in last column of Fig.8 show a com-parison of MERIS Level 2 Cloud Flag with the resultsfrom our algorithm. Pixels where both algorithmsagree are in white for the cloudy pixels and blue forthe cloud free pixels. From the results, two maindiscrepancies can be found. On the one hand, whenour algorithm detects cloudy pixels they are plotedin red, showing a good agreement with cloud bor-ders. Therefore, one can assume that the proposedmethod provides better recognition in cloud bordersand in small and thin clouds. On the other hand,discrepancies when our algorithm classify as cloudfree are shown in yellow, and one can see that theseareas correspond only to ice covers (Finland image)and snow over high mountains (Pyrenees and Alpsin the France image) indicating the goodness of ourapproach.

6. CONCLUSIONS

This work presents a new technique that faces theproblem of accurately identifying the location andabundance of clouds in multispectral images. Theproposed algorithm has demonstrated a clear im-provement in the classification of hard-to-detectcloud pixels in ENVISAT MERIS FR images, espe-cially thin cirrus clouds, and clouds over ice/snow.One critical feature for the improved results was theuse of the O2 and H2O absorption bands togetherwith an overlaped DEM, therefore it would be advan-tageous to see those bands included in future sensors.In any case, the procedure can serve as a test to de-velop a cloud masking algorithm for other spectralsensors of the VNIR spectral range.

Further refinements can be introduced in order toenhance the robustness of the procedure:

• using dynamic thresholds could be useful to findthe regions to analyze;

• final maps could be further analyzed throughtexture algorithms;

• sun position could be taken into account in orderto relate cloud and shadow positions, by detect-ing edges of the clouds and shadow regions, andthen apply morphology operators that extractinformation on the structure of the image;

• DEM could be used together with the obtainedcloud mask to correct the image and estimatecloud heights;

• coastline map could be used in order to improveresults by considering clouds over land or water,separately.

Figure 8. MERIS images over the test sites of BR-2003-07-14, BR-2004-07-14, FI-2005-02-26, and FR-2005-03-19 displayed in rows from left to right. First row: RGB composite with an histogram stretching such that 10%of data is saturated at both low and high reflectance (10%-90%) in order to increase the contrast of the cloudyimages. Second row: Classification of the relevant regions. Third row: Cloud abundance product. Fourth row:Comparison of MERIS Level 2 Cloud Flag with the obtained cloud mask (discrepancies are shown in red whenour algorithm detects cloud and in yellow when pixels are classified as cloud free).

ACKNOWLEDGMENTS

The authors wish to thank ESA for the availabilityof the images aquired in the framework of the ESA-SPARC Project (ESTEC-18307/04/NL/FF).

This paper has been partially supported by the Span-ish Ministry for Education and Science under projectESP2004-06255-C05-02 and by the “Grups Emer-gents” programme of Generalitat Valenciana underproject HYPERCLASS/GV05/011.

REFERENCES

1. Simpson, J. Improved cloud detection andcross-calibration of ATSR, MODIS and MERISdata. In ESA-SP-479, editor, ATSR Interna-tional Workshop on the Applications of the ERSalong track scanning radiometer, ESRIN, Fras-cati, Italy, June 1999.

2. Vittorio, A. D. and Emery, W. An automated,dynamic threshold cloud-masking algorithm fordaytime AVHRR images over land. IEEE Trans.Geosci. Remote Sensing, 40(8), 2002.

3. Papin, C., Bouthemy, P., and Rochard, G. Un-supervised segmentation of low clouds from in-frared METEOSAT images based on a contex-tual spatio-temporal labeling approach. IEEETrans. Geosci. Remote Sensing, 40(1):104–114,Jan. 2002.

4. McIntire, T. and Simpson, J. Arctic sea ice,cloud, water, and lead classification using neuralnetworks and 1.6 µm data. IEEE Trans. Geosci.Remote Sensing, 40(9):1956–1972, Sept. 2002.

5. Diner, D., Clothiaux, E., and Girolamo, L. D.MISR Multi-angle Imaging Spectro-RadiometerAlgorithm Theoretical Basis. Level 1 Cloud De-tection. Jet Propulsion Laboratory, JPL D-13397,California Institute of Technology, 1999.

6. Tian, B., Shaikh, M., Azimi, M., Haar, T.,and Reinke, D. A study of cloud classificationwith neural networks using spectral and textu-ral features. IEEE Trans. on Neural Networks,10(1):138–151, Jan. 1999.

7. Liou, K. N. An Introduction to Atmospheric Ra-diation, 2nd Ed. Academic Press, 2002.

8. Thuillier, G., Hers, M., Simon, P., Labs, D., Man-del, H., Gillotay, D., and Foujols, T. Observationof the visible solar spectral irradiance between350 and 850 nm during the ATLAS I missionby the SOLSPEC spectrometer. Solar Physics,177:41–61, 1998.

9. Dempster, A. P., Laird, N. M., and Rubin, D. B.Maximum likelihood from incomplete data viathe EM algorithm. Journal of the Royal Statisti-cal Society, Series B, 39:1–38, 1977.

10. Chang, C. Hyperspectral Imaging: Techniquesfor Spectral Detection and Classification. KluwerAcademic/Plenum Publishers, 2003.

11. Gomez-Chova, L., Amoro, J., Camps-Valls, G.,Martın, J., Calpe, J., Alonso, L., Guanter, L.,Fortea, J., and Moreno, J. Cloud detection forCHRIS/Proba hyperspectral images. In Schafer,K. and Comeron, A., editors, Remote Sensing ofClouds and the Atmosphere X, volume 5979 ofProceedings of SPIE, page in press, 2005.