Planetary and Space Science 70 (2012) 46–58

Contents lists available at SciVerse ScienceDirect

Planetary and Space Science

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

Image processing algorithm for the identification of Martian dust devil tracks
in MOC and HiRISE images
Thiago Statella a,n, Pedro Pina b, Erivaldo Antônio da Silva c
a
Instituto Federal de Educac- ao,
~ Ciência e Tecnologia de Mato Grosso - IFMT 95 Zulmira Canavarro 780025-200, Cuiabá, Brazil
b
Centro de Recursos Naturais e Ambiente (CERENA), Instituto Superior Técnico – IST, Av. Rovisco Pais 1049-001 Lisboa, Portugal
c
Universidade Estadual Paulista, Faculdade de Ciências e Tecnologia – FCT, 305 Roberto Simonsen 19060-900, Presidente Prudente, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a method for automatic identification of dust devils tracks in MOC NA and HiRISE
Received 3 September 2011 images of Mars. The method is based on Mathematical Morphology and is able to successfully process
Received in revised form those images despite their difference in spatial resolution or size of the scene. A dataset of 200 images
31 May 2012
from the surface of Mars representative of the diversity of those track features was considered for
Accepted 1 June 2012
developing, testing and evaluating our method, confronting the outputs with reference images made
Available online 13 June 2012
manually. Analysis showed a mean accuracy of about 92%. We also give some examples on how to use
Keywords: the results to get information about dust devils, namelly mean width, main direction of movement and
Dust devils coverage per scene.
Feature detection
& 2012 Elsevier Ltd. All rights reserved.
Mathematical Morphology
Mars

1. Introduction are believed to help in maintaining the observed background dust

Dust devils are vortices caused by unstable wind convection
processes near the planetary surfaces, due to solar heat. They
have been studied on Earth for more than a century (Baddeley,
1860; Brooks, 1960) and were first observed on Mars in the 1970s
in orbital images taken by the Viking orbiters (Balme et al., 2003).
In Mars they are more frequent in the South hemisphere where
solar radiation is higher due to the planet eccentric orbit.
Substantial mixing and expansion of the dust devil vortex occurs
near the surface. Warm near-surface air moves horizontally
toward the low pressure center until it reaches the dust column.
Then, it rises rapidly. Within the column the radial velocity nearly
vanishes and the presence of dust particles in the dust devil inner
core is suppressed by both a descending motion and centrifugal
forces. When the surface is composed of loose materials, dust
particles might become airborne making the dust devil visible
(Rennó et al., 1998). The phenomena can achieve miles in width
and height, and knowledge about their activity contributes to the
understanding of Martian climate, geology and surface modifica-
tion which is essential to plan future manned missions (Balme
et al., 2003; Balme and Greeley, 2006). Dust devils peak activity is
around noon (Mars local time) during spring and summer. They
n
Corresponding author. Tel.: þ55 18 32225042.
E-mail addresses: thiago.statella@cba.ifmt.edu.br (T. Statella),
ppina@ist.utl.pt (P. Pina), erivaldo@fct.unesp.br (E.A. da Silva).
load in the Martian atmosphere and their local contribution to the
amount of suspended dust might reach 50% (Drake et al., 2006;
Reiss et al. 2011).
According to Örmo and Komatsu (2003), air circulation is one
of the currently active processes which model the surface of Mars
and some researches show that these vortices are responsible for
most of the linear and curvilinear surface features of the planet.
Moreover, the inference of the wind direction based on dust
devils tracks detection is one of the few procedures for verifying
circulation models of the atmosphere. The General Circulation
Model (GCM) of the atmosphere indicates a wind behavior
according to a pattern N–S and E–W (Greeley et al., 1992),
contrasting with directions inferred from dust devils tracks
observation which follow NE–SW, NW–SE directions (Balme
et al., 2003; Fenton et al., 2005; Örmo and Komatsu, 2003). This
fact suggests that more research on aeolian processes is needed.
The direction of dust devils tracks can be used to get information
on wind circulation, and it can be done by image analysis (Fenton
et al., 2005; Örmo and Komatsu, 2003).
Many researchers have being studying dust devils in an
attempt to better understand the phenomena. Generally, the
research fields comprise mechanic and numerical simulation of
dust devils in laboratories (Babak and Taylor, 2011; Neakrase and
Greeley, 2010; Rennó, 2008), methodologies for recognition of
dust devils plumes from the rovers on Mars surface (Castano
et al., 2008; Metzger et al., 2000; Lorenz et al., 2009), detection of
plumes (Stanzel et al., 2008; Towner, 2009) and tracks (Balme and

0032-0633/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.pss.2012.06.003
 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58 47

Greeley, 2006; Balme et al., 2003; Cantor et al., 2006; Drake et al.,
2006; Fenton et al., 2005; Fisher et al., 2005; Örmo and Komatsu,
2003; Whelley and Greeley, 2008) from orbital images. Despite
the number of papers regarding the subject, none of them
addresses the automatic detection of dust devils tracks from
remotely sensed imagery.
All those works regarding identification, counting and analysis of
dust devils tracks use a manual method and nothing is said about the
tracks counting and marking process. As the number of well
succeeded missions launched to study Mars rises, so does the number
of orbital images. There are hundreds of high resolution images
depicting Martian surface being taken and the amount of information
on them grew at a rate greater than the human capability to analyze
and extract relevant information from these products to characterize
the planet under study. As examples of the difficulty in analyzing

Table 1
Summarized information about the image dataset.

Lat1 Lon1 Ls1 Width (pixels) High (pixels)

Min  14.47 0.70 134.17 138.00 179.00

Mean  53.39 156.06 300.56 2754.85 2457.47
Fig. 1. Distribution of the images considered in our study on the surface of Mars. Max  64.80 358.80 353.90 9058.00 7526.00
Triangles: HiRISE; Circles: MOC. Image credits: NASA/JPL/MSSS.

Fig. 2. Flowchart of the method.

Fig. 3. Filtering step: (a) original image; (b) filtered image; (c) and (d) details of (a) and (b). The rectangle annotated in (a) indicates the region selected for the details
shown in (c) and (d). Image credits: NASA/JPL/University of Arizona.
48 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58

manually so many images, Balme et al. (2003), Fisher et al., 2005,
Cantor et al. (2006), Örmo and Komatsu (2003), Drake et al. (2006),
Whelley and Greeley (2006, 2008), had to search for tracks in (3000),
(8116), (1238), (1700), (1734) and (6002; 167,254) MOC images,
respectively; Stanzel et al. (2008) searched in (23) HRSC images and
Towner (2009) did that in (3079) THEMIS images.
Regarding the amount of images to be analyzed and the
importance of identifying dust devils tracks, we started working
on an automated procedure to identify them in Mars Global
Surveyor (MGS) Mars Orbiter Camera (MOC) and Mars Reconnais-
sance Orbiter (MRO) High Resolution Imaging Science Experiment
(HiRISE) images. Several experiments were conducted showing at
once interesting results, although on small datasets. This method
was based on Mathematical Morphology and targeted the struc-
tures aligned along given directions (radial closings approach) with
small curvilinear behavior (Statella and Silva, 2011a, 2011b, 2011c).
A generalization of the approach was obtained afterwards using a
procedure that detected elongated structures independently of its
curvature (path closing approach), but that was unable to binarize
every image with the same high correctness (Statella et al., 2011).
Finally, we are able to present in this paper a consistent improved
method, with a skeleton that incorporates some of those contribu-
tions presented at conferences and whose robustness is demon-
strated by the high mean global accuracy of  92% on a dataset of
200 high spatial resolution images (MOC NA and HiRISE) covering
the diversity of dust devil tracks on Mars.

Fig. 4. Path closing (a) and top-hat by closing (b).

Fig. 5. Binarization: (a) Otsu; (b) Otsu constrained to the interval ½kmean ,kmax ; (c) Histogram of the top-hat in Fig. 4b.
 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58 49

2. Image datasets
Based on some evidences (Balme et al., 2003; Fenton et al.,
2005; Örmo and Komatsu, 2003; Whelley and Greeley, 2008)
verified that dust devils are more likely to occur in the southern
hemisphere and on the fact that they form during the spring and
summer, a search for MOC narrow angle (Malin Space Science
Systems database) and HiRISE (The University of Arizona) images
with solar longitudes ranging between 1801 and 3601 (a few
images out of that range we knew beforehand containing tracks
were also considered as part of the dataset) containing tracks of
dust devils was performed in the regions Aeolis, Noachis, Argyre,
Eridania and Hellas.
More than a thousand images were searched and initially, a
total of 124 images (75 MOC narrow angle panchromatic band
and 49 HiRISE red band) showing dark dust devils tracks were
considered. The albedo of the tracks varies significantly from
scene to scene as does the morphology and landform. Based on
that, the choice for these 124 images was driven by the attempt to
represent as much as possible the high albedo variability of the
tracks. In order to decrease the processing time (mainly because
of the size of HiRISE images) and discard irrelevant information
(like areas with no tracks) the images were trimmed to its regions
of interest (HiRISE images were also resampled by a factor of 1/3),
making a set of 200 images (90 MOC and 110 HiRISE), distributed
as follows: 4 MOC and 1 HiRISE images depicting regions in

Fig. 6. HiRISE and MOC NA images related in Table 2. Image credits: NASA/JPL/University of Arizona (HiRISE) and NASA/JPL/MSSS (MOC).
50 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58

Aeolis, 20 MOC and 19 HiRISE images depicting regions in
Noachis, 16 MOC and 55 HiRISE images depicting regions
in Argyre, 25 MOC and 20 HiRISE images depicting regions in
Eridania and 24 MOC and 16 HiRISE images from Hellas. Table 1
summarizes some characteristics of the images. The spatial
resolution of HiRISE images before resampling is either 0.25 m
(  90% of the images) or 0.50 m while the MOC spatial resolution
varies from 1.43 m to 8.75 m (mean  5 m).
In Table 1, Ls1 stands for solar longitude. The solar longitude is
the Mars–Sun angle, measured from the Northern Hemisphere
spring equinox where Ls ¼01. Fig. 1 shows the distribution of the
initial set of images according to their coordinates in the
planetocentric system. Circles and triangles locate, respectively,
MOC and HiRISE images on Mars surface.
3. Method
The method we are introducing to identify dust devil tracks in
Martian images benefits from preliminary developments and
tests, where several algorithmic options were evaluated in paral-
lel. The current version is a more complete and robust approach
whose main steps are summarized in Fig. 2. It is constituted by a

Fig. 7. Dust devil tracks detection for images shown in Fig. 6.

 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58 51

sequence of the following four main tasks: filtering, track candi- interval ½kmean ,kmax of the histogram. This was defined in a
dates selection, tracks recovery and tracks detection. preliminary analysis of the binarization performance. A set of 60
The initial filtering consists in applying the morphological MOC and HiRISE images was binarized manually (hence, the
surface area opening and surface area closing operators (Soille, thresholds are considered as ideals) and the result was compared
2004) to the images to, respectively, attenuate the high reflec- to Otsu’s method: with no constrains (O1); constrained to the
tance of, for instance, boulders, ripples and other small features, interval ½kmean ,kmax (O2); applied to the Laplacian of the image
and dark spots caused by boulder shadows. Also, some possible (O3); applied to the Laplacian of the image and then constrained
noise introduced by the image acquisition process is expected to to the interval ½kmean ,kmax (O4). The correlation coefficients
be filtered out. Surface area opening and closing operators remove between the thresholds obtained manually and those obtained
from the image any connected component whose size (specified by methods Oi ði ¼ 1,:::,4Þ were calculated and resulted in
in number of pixels) is smaller than a reference number l given as CorðM,O1 Þ  81:46%, CorðM,O2 Þ  88:71%, CorðM,O3 Þ  82:10%
a parameter. The openings affect image objects (bright patterns) and CorðM,O4 Þ  88:88%, where M is the set of thresholds defined
and the closings affect the background (dark patterns). The value manually for the 60 images. Once CorðM,O2 Þ  CorðM,O4 Þ, O2 was
l of the area considered for the opening/closing varies with the chosen for being easier and faster to calculate. Again, using the
spatial resolution R of the images. For the surface area opening we image PSP_006163_1345A as example, Fig. 5 shows the binariza-
fixed l ¼ 500=R and for the area closing this value is equal to l=2 tion by method O1 (Fig. 5a) and by method O2 (Fig. 5b) and the
(to avoid an excessive attenuation of the dark patterns). The grey level histogram of the image in Fig.4c. The threshold defined
factor 500 was chosen empirically from our experiments with by O1 was at value 27 (light gray line in Fig. 5c) whereas the one
images from both sensors. Fig. 3b shows the result of the filtering defined by O2 was 72 (dark gray line in Fig. 5c), which clearly
step on a portion of the HiRISE image PSP_006163_1345A (Fig. 3a) demonstrates that this improved version of Otsu’s method per-
where not only the illuminated faces of the ripples are suppressed mits to detect most of the dust devil tracks with less false positive
but also the dark regions of the track of the dust devil are structures.
strengthened or compacted. All the mathematical formalism of morphological transforma-
The same portion of this resampled image is used to illustrate tions, their properties and many application examples can be
the next steps of our method: it has a dimension (after resam- found in Serra (1982) and Soille (2004).
pling) of 1844  2954 pixels, 0.75 m of spatial resolution, is
located at 45.31S and 316.31E in Argyre region and the date of
acquisition is 11/19/2007 at 14:28 h Mars local time with a solar 4. Results and discussion
longitude of 349.81.
Next, in the second step of the processing sequence, a path closing The method presented in the previous section was applied to
(Hendriks, 2010) is applied to select all possible dark paths. Anyhow, 90 MOC NA and 110 HiRISE images.
this is performed on an indirect way, since we are first suppressing We selected some examples from this dataset to representa-
the desired long dark structures (Fig. 4a). This transformation does tively illustrate the variety of detections obtained. These results
not filter the image by the size of the components only (as does the are obtained for HiRISE PSP_006163_1345B, PSP_005780_1215,
surface area closing). Instead, it filters linear paths with specified size PSP_002548_1255 and PSP_004249_1255 images (Fig. 6(a)–(d)
and orientation. The search for the paths is carried out in the and for MOC NA M10-01206, E11-00582, E10-04027 and R13-
directions 01, 451, 901 and 1351 so paths in these directions are 01467 images (Fig. 6(e)–(h) and are shown in Fig. 7. The relevant
suppressed. The considered paths are not strictly straight line as they information about these images is summarized in Table 2.
are allowed to deflect inside a 901 aperture cone centered in each of The performance of the results, and consequent analysis, is
the four adopted directions of search. The maximum lengths of the made according to a procedure proposed in Bandeira et al. (2011)
paths to search for are defined by dimensions
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi of the images and which is based on the following global measure:
obtained by the quantity given by m2 þn2  2, where m and n are,
Accuracy ¼ ðTP þTNÞ=ðm  nÞ ð1Þ
respectively, the number of columns and rows of the image (although
the path closings being used here are the constrained ones defined by where TP stands for the number of true positives and TN for
Hendriks (2010), they still may zig-zag a little so the biggest possible number of true negatives (both amounts refer to the number of
path in the worse case would be the image diagonal times two). pixels). TP and TN are defined relative to a ground truth or
The dark paths are then recovered and enhanced, in the third reference image.
step of the method, by a closing top-hat transformation (Fig. 4b). For a given processed image PI and a ground truth image GT, TP
The closing top-hat is the arithmetic difference between the and TN are calculated as:
closed image (in which the tracks were removed by the path
TP ¼ AreaðGT \ PIÞ
closing) and the original image. All dark paths in the original
image are enhanced by the transformation. TN ¼ Areað  GT \  PIÞ: ð2Þ
This grey level output must be binarized to obtain an image
containing only the dust devil tracks. Since the intensity of the Table 2
tracks changes from object to object an automated approach Summarized information about images on Fig. 6. In the Table, ‘‘Res.’’ stands for
‘‘Spatial resolution’’ and letters (a) through (h) in the field ‘‘Name’’ correspond to
should be applied to detect for each image the threshold value the images shown in Fig. 6.
that best separate our structures of interest from the rest of the
image. One of most successful methods to perform an automated Name Sensor Date Res. (m) Size (pixels) Region
thresholding based on the histogram of the image is Otsu’s
PSP_006163_1345B(a) HiRISE 11/19/07 0.75 4415  7184 Argyre
method (1979). It is simple method to apply, but has some
PSP_005780_1215(b) HiRISE 10/20/07 0.75 2168  1423 Argyre
drawbacks related to the contents of the image, namely, if the PSP_002548_1255(c) HiRISE 02/10/07 0.75 4069  2779 Noachis
grey level distribution is not bimodal or if the weights of the two PSP_004249_1255(d) HiRISE 06/23/07 0.75 3616  3448 Noachis
classes are very unequal. Thus, we decided to evaluate Otsu M10-01206(e) MOC NA 12/11/99 5.55 363  829 Argyre
method and other alternatives in our image dataset to obtain E11-00582(f) MOC NA 12/04/01 5.81 512  590 Noachis
E10-04027(g) MOC NA 11/25/01 4.53 672  1118 Eridania
the higher segmentation performance. In the binarization step, R13-01467(h) MOC NA 01/10/04 1.43 2077  2855 Aeolis
Otsu’s algorithm is constrained to look for a threshold in the
52 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58

where \ and  are the operators intersection and negation, marked when making the ground truth was based on the contrast
respectively. and thickness of the tracks: tracks with a thickness less than
For each of the 200 images processed, a ground truth image 3 pixels were not considered for ground truth. Fig. 8 shows some
was made manually. The decision on which tracks should be examples of the type of tracks that were considered (a)–(f) and

Fig. 8. Examples of tracks marked (g)–(l) and not marked (a)–(f) in ground truth.
T. Statella et al. / Planetary and Space Science 70 (2012) 46–58 53

Fig. 9. Ground truth of the images shown in Fig. 6.

54 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58

some examples of tracks that were not considered (g)–(l) when a length bigger than the length considered for the opening and, at
making the ground truth. Fig. 9 shows the ground truth made for the same time, its orientation is the same one specified for the
the images of Fig. 6 and Table 3 gives the accuracies of detection. opening, then it is not removed. After applying openings by
In Table 3, pFP and pFN are, respectivelly, the probabilities of straight lines oriented in four directions and counting the amount
false positivies and false negatives and are calculated as follows: of pixels removed we expect the main direction of the tracks to be
the one with fewer pixels filtered out. The amount of pixels
pFN ¼ FN=ðFN þ TPÞ
removed in each direction can be visualized in a rose of direc-
pFP ¼ FP=ðFP þ TNÞ: ð3Þ tions. Fig. 14 shows the rose of directions for image
PSP_006163_1345B. Each concentric circle gives the number of
The mean accuracy of the 200 images processed was about
pixels removed from the image.
92%. The worst accuracy was 69% (for image R13-01467 shown in
According to Fig. 14, the main directions followed by the dust
Fig. 6h) and the best was above 99% (for image PSP_006163_
devils in this scene were NE–SW and N–S, which agrees with the
1345A shown in Fig. 3).
visual analysis of image PSP_006163_1345B. The length of lines in
The lower accuracy of image R13-01467 is due to the poor
the opening transformation varied from image to image accord-
contrast between the tracks and their neighbourhood. The edges
ingly to the maximum width of tracks calculated previously by
of the tracks are too smooth and the threshold calculated by the
the granulometric analysis.
method O2 was 36 whereas the ideal one (defined manually) was
Table 4 shows some features calculated for the images in
27. We illustrate the good performance of the method in difficult
Fig. 7. The features are: percentage of area covered by tracks
situations: Fig. 10 shows examples in which the algorithm
‘‘Area’’, main direction of the tracks ‘‘Dir’’ and diameters of the
succeeds in detecting the tracks even in the presence of boulder
tracks ‘‘Diameter’’.
shadows (does not take them for tracks) (Fig. 10(a)–(c)) or where
Besides assessing the mean global accuracy of the method, it is
the tracks have their edges smooth and poorly defined
important to know how reliable it is when it comes to calculating
(Fig. 10(d)–(i)).
features like those in Table 4 from the results. In order to assess
It is also important to notice that the accuracy was not affected
the accuracy of the data in Table 4, the exact same features were
by the variation in spatial resolution of the images. The method
also calculated for the ground truth images. The results obtained
works for MOC NA as much as for HiRISE. Besides, the method is
are shown in Table 5.
not sensitive to variations in latitude and solar longitude. This can
Table 6 presents the absolute discrepancies between the
be verified by observing that no clusters were formed when
features calculated from the processed and the ground truth
plotting accuracy versus latitude (Fig. 11) and versus solar long-
images. As expected, the biggest difference in the measured
itude (Fig. 12) for MOC NA and HiRISE images.
direction, which is 451 (difference between E–W and NW–SE
Once tracks have been detected, the respective binary images
directions), occurs for the image with lower accuracy (R13-
can be used to compute dust devils characteristics. For instance,
01467) in the detection process.
the area covered by the tracks for each scene is easily obtained by
The same features were calculated for the complete dataset of
counting object pixels in the binary output. Besides, a morpholo-
200 ground truth images and the results were compared with the
gical granulometrical analysis is carried out to infer the mean
features calculated from the outputs of the method using the
diameter of the tracks and directional openings infer the main
correlation coefficient for the characteristics ’’Area’’, ’’Mean Dia-
direction of the vortices.
meter’’ and ’’Maximum Diameter’’, which were approximately
Granulometrical analysis gives information on size distribu-
85%, 86% and 71%, respectively. For the directions, an estimative
tion of the connected components in the image (Serra, 1982).
of the accuracy can be calculated by counting how many direc-
A series of openings is used to filter those components whose
tions inferred from processed images agree with those ones
shape does not fit disks of increasing size. The first opening
inferred from the ground truth images, which was about 88% of
operation filters out all image components that do not contain a
the whole dataset of 200 images. One can consider from these
disk of 1 pixel radius. Then a second opening operation by a disk
results that the method we are proposing is robust and reliable
of 2 pixels radius is applied and so forth. The amount of pixels
and has the same performance for both sensors.
removed by each opening can be plotted versus the radius of the
The rose of directions for regions Aeolis, Noachis, Argyre,
disk in the successive transformations in a graphic called pattern
Eridania and Hellas are shown in Fig. 15. Fenton et al. (2005)
spectrum. The pattern spectrum for the dust devil tracks of image
observed dunes and ripples in Proctor crater, Noachis region,
PSP_006163_1345B (Fig. 6a) is shown in Fig. 13.
formed by winds in the direction ENE–WSW, dust devil tracks in
The main direction of the vortices can be inferred by applying
the directions ENE–WSW, NE–SW and NW–SE as well as bright
openings by straight lines aligned in given directions. We selected
bedforms eroded by winds in the direction ENE–WSW. We found
now 4 main directions: 01 (E–W), 451 (NE–SW), 901 (N–S) and
a wind direction according to a strong pattern N–S for Noachis
1351 (NW–SE). The openings will filter out straight linear image
(Fig. 15b). Balme et al. (2003) observed wind streaks in the
components whose length and orientation do not fit the orienta-
direction ENE–WSW in Argyre and dust devil tracks in the
tion and length specified in the transformation. When a track has
direction NW–SE in Hellas whereas we found directions E–W
for Argyre (Fig. 15c) and E–W, N–S for Hellas (Fig. 15e). Drake
Table 3
Accuracies calculated for the images shown in Fig. 7.
et al. (2006) observed dust devil tracks and wind streaks in a
latitude band of 651N–721N in the North Polar region and related
Name pFP pFN Accuracy winds in the directions NW–SE (2501E–2701E), N–S, E–W, NE–SW
(1201E–1401E) and WNW–ESE (651E–851E). For Aeolis (Fig. 15a)
PSP_006163_1345B(a) 0.0047 0.1488 0.9896
we inferred direction E–W and for Eridania (Fig. 15d) the wind
PSP_005780_1215(b) 0.0159 0.0793 0.9816
PSP_002548_1255(c) 0.0100 0.4300 0.9058 direction was mainly N–S with some tracks showing E–W direc-
PSP_004249_1255(d) 0.0138 0.5893 0.7117 tion. The directions inferred by our method agree to the general
M10-01206(e) 0.0087 0.2090 0.9798 pattern N–S, E–W predicted by the GCM (Greeley et al., 1992).
E11-00582(f) 0.0167 0.3467 0.9415 Our method can help understanding local variations in wind
E10-04027(g) 0.0336 0.3794 0.9166
R13-01467(h) 0.0503 0.7113 0.6915
direction pattern which contrast with the predicted ones by the
GCM. The track coverage may be used to estimate mass flux of
 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58 55

Fig. 10. Examples of the good performance of the method: (a)–(c) shadows of boulders were not detected due to the initial filtering; (d)–(f) and (g)–(i) detection of tracks
with low contrast. The annotated rectangles in (d) and (g) indicate the area of the details shown in (e) and (f) and (h) and (i).
56 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58

Fig. 11. Accuracy versus latitude: there is no clustering so the method is latitude
independent. Fig. 14. Rose of directions for image PSP_006163_1345B.

Table 4
Features calculated from processed images shown in Fig. 7.

Name Area (%) Dir Diameter (m)

Mean Max

PSP_006163_1345B(a) 3.76 NE–SW 10.40 33.75

PSP_005780_1215(b) 5.11 N–S 12.70 26.25
PSP_002548_1255(c) 14.27 N–S 19.21 83.25
PSP_004249_1255(d) 20.31 N–S 22.84 189.75
M10-01206(e) 5.55 N–S 63.84 300.85
E11-00582(f) 9.72 NE–SW 89.45 191.73
E10-04027(g) 11.80 N–S 16.48 176.67
R13-01467(h) 14.34 E–W 23.83 170.17

Table 5
Features calculated from ground truth.

Name Area (%) Dir Diameter (m)

Mean Max
Fig. 12. Accuracy versus solar longitude: the results do not show any dependence
to the variation in solar longitude. PSP_006163_1345B(a) 19.59 NE–SW 10.30 32.25
PSP_005780_1215(b) 3.89 N–S 16.83 27.75
PSP_002548_1255(c) 21.40 N–S 13.96 81.75
PSP_004249_1255(d) 47.69 N–S 18.84 213.75
M10-01206(e) 5.75 N–S 64.31 205.35
E11-00582(f) 12.64 NE–SW 80.48 203.35
E10-04027(g) 14.40 N–S 50.14 167.61
R13-01467(h) 39.06 NW–SE 19.55 218.79

Table 6
Discrepancies between features calculated from ground truth and from detection.

Name Accuracy DArea (%) DDir (1) Diameter (m)

Dmean Dmax

PSP_006163_1345B(a) 0.9896 0.12 0 0.09 1.50

PSP_005780_1215(b) 0.9816 1.22 0 4.12 1.50
PSP_002548_1255(c) 0.9058 7.13 0 5.26 1.50
PSP_004249_1255(d) 0.7117 27.38 0 4.00 24.00
M10-01206(e) 0.9798 0.20 0 0.47 55.50
E11-00582(f) 0.9415 2.92 0 8.97 11.61
E10-04027(g) 0.9166 2.60 0 33.66 9.06
R13-01467(h) 0.6915 24.72 45 4.28 48.62
Fig. 13. Pattern spectrum of image PSP_006163_1345B.
 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58 57

Fig. 15. Rose of directions: (a) Aeolis, (b) Noachis, (c) Argyre, (d) Eridania and (e) Hellas.

material entrained by dust devils and the track width calculated
by our method can help estimating the area over which a dust
devil acts.
5. Conclusion
Studying Martian dust devils is important to provide a better
understanding of aeolian processes, low atmospheric behavior
and regolith, among others. Such importance may be asserted by
the huge amount of papers being published about the subject. But
none of them proposes an automatic method for the detection of
dust devils tracks. This has been done manually until now and is a
time-consuming task. This paper presented a method for detect-
ing Martian dust devils tracks automatically which achieved 92%
of accuracy for 200 MOC NA and HiRISE images in 5 quadrangles
of the southern hemispehere of Mars where those features are
very common. It succeeds in detecting tracks despite the variation
in size and spatial resolution of the images, and works also
successfully for a great range of albedo variation. Also, the paper
shows examples of features that can be calculated from the
detections. The accuracy of these features can be assessed by
58 T. Statella et al. / Planetary and Space Science 70 (2012) 46–58
comparing them to the same features derived from the ground
truth. For the whole set of images the correlation coefficients
between the features estimated from ground truth and the ones
estimated from the outputs of the method were about 85% for
’’Area’’, 86% for ’’Mean Diameter’’ and 71% for ’’Maximum Dia-
meter’’. The accuracy for the inference of the directions was
near 88%.
The results show that the mean global accuracy of 92% for the
detection process allows the features to be estimated accurately.
We can conclude that our method shows high and robust
performances and can be a very useful tool for intensive mapping
of dust devil tracks and thus largely increasing our understanding
of these aeolian features.
In a near future, the method is to be applied to images
covering every region of Mars and the directions in the low
atmosphere to be inferred from tracks will be compared to the
directions predicted by the Global Circulation Model. Besides,
from the detections provided by the method we are developing an
approach to calculate a measure of relative contrast between the
tracks and their surroundings. Such information would be impor-
tant to understand the relative age of the tracks and also some
surface characteristics once the brightness of a track is intrinsi-
cally related to the composition of the regolith and hence the
capability of the vortex to rise sand and dust.
Appendix A. Supporting information
Supplementary data associated with this article can be found
in the online version at http://dx.doi.org/10.1016/j.pss.2012.06.
003.
References
Babak, T.G., Taylor, P.A., 2011. Large eddy simulation of typical dust devil-like
vortices in highly convective Martian boundary layers at the Phoenix lander
site. Planetary and Space Science 59, 43–50.
Baddeley, P.F.H., 1860. Whirlwinds and Dust Storms of India, ed. Bell and Daldy,
London.
Balme, M.R., Whelley, P.L., Greeley, R., 2003. Mars: dust devil track survey in
Argyre Planitia and Hellas basin. Journal of Geophysical Research 108 (E8).
Balme, M.R., Greeley, R., 2006. Dust devils on Earth and Mars. Reviews of
Geophysics 44, RG3003.
Bandeira, L., Marques, J.S., Saraiva, J., Pina, P., 2011. Automated detection of
Martian Dune fields. IEEE Geoscience and Remote Letters 8, 626–630.
Brooks, H.B., 1960. Rotation of dust devils. Journal of Meteorology 17, 84–86.
Cantor, B.A., Kanak, K.M., Edgett, K.S., 2006. Mars Orbiter Camera observations of
Martian dust devils and their tracks (September 1997 to January 2006) and
evaluation of theoretical vortex models. Journal of Geophysical Research 111,
1–49.
Castano, A., Fukunaga, A., Biesiadecki, J., Neakrase, L., Whelley, P., Greeley, R.,
Lemmon, M., Castano, R., Chien, S., 2008. Automatic detection of dust devils
and clouds on Mars. Machine Vision and Applications 19, 467–482.
Drake, N.B., Tamppari, L.K., Baker, R., Cantor, B.A., Hale, A.S., 2006. Dust devil tracks
and wind streaks in the north polar region of Mars: a study of the 2007
Phoenix Mars Lander sites. Geophysical Research Letters 33, 1–4.
Fenton, L.K., Toigo, A.D., Richardson, M.I., 2005. Aeolian processes in Proctor Crater
on Mars: mesoscaling modeling of dune-forming winds. Journal of Geophysi-
cal Research 110, E06005.
Fisher, J.A., Richardson, M.I., Newman, C.E., Szwast, M.A., Graf, C., Basu, S., Ewald,
S.P., Toigo, A.D., Wilson, R.J., 2005. A survey of Martian dust devil activity using
mars global surveyor mars Orbiter Camera images. Journal of Geophysical
Research 110, E03004.
Greeley, R., Lancaster, N., Lee, S., Thomas, P., 1992. In: Kieffer, H.H. (Ed.), Martian
Aeolian Processes, Sediments and Features. Mars. University of Arizona Press,
TucsonIn: Kieffer, H.H. (Ed.), Martian Aeolian Processes, Sediments and
Features. Mars. University of Arizona Press, Tucson.
Hendriks, C.L.L., 2010. Constrained and dimensionality-independent path open-
ings. IEEE Transactions on Image Processing 19, 1587–1595.
Lorenz, R.D., Jackson, B., Barnes, J.W., 2009. Inexpensive time-lapse digital cameras
for studying transient meteorological phenomena: dust devils and playa
flooding. Journal of Atmospheric and Oceanic Technology 27, 246–256.
Metzger, S.M., Carr, J.R., Johnson, J.R., Parker, T.J., Lemmon, M.T., 2000. Techniques
for identifying dust devils in Mars Pathfinder images. IEEE Transactions on
Geoscience and Remote 38, 870–876.
Neakrase, L.D.V., Greeley, R., 2010. Dust devil sediment flux on Earth and Mars:
laboratory simulations. Icarus 206, 306–318.
Örmo, J., Komatsu, G., 2003. Mars Orbiter Camera observation of linear and
curvilinear features in the Hellas Basin: indications for multiple processes of
formation. Journal of Geophysical Research 108, 1–13.
Otsu, N., 1979. A threshold selection method from gray-level histograms. IEEE
Transactions on Systems, Man, and Cybernetics 9, 62–66.
Reiss, D., Zanetti, M., Neukum, G., 2011. Multitemporal observations of identical
active dust devils on Mars with the High Resolution Stereo Camera (HRSC) and
Mars Orbiter Camera (MOC). Icarus 215, 358–369.
Rennó, N.O., 2008. A thermodynamically general theory for convective vortices.
Tellus 60, 688–699.
Rennó, N.O., Burkett, M.L., Larkin, M.P., 1998. A simple thermodynamical theory
for dust devils. Journal of the Atmospheric Sciences 55, 3244–3252.
Serra, J., 1982. Image Analysis and Mathematical Morphology, ed. Academic Press,
London.
Soille, P., 2004. Morphological Image Analysis. Springer-Verlag, Berlin.
Stanzel, C., Patzold, M., Williams, D.A., Whelley, P.L., Greeley, R., Neukum, G., 2008.
Dust devils speeds, directions of motion and general characteristics observed
by the Mars Express High Resolution Stereo Camera. Icarus 197, 39–51.
Statella, T., Silva, E.A., 2011a. Mathematical morphology for detecting Martian dust
devils tracks and inferring their directions. Revista Brasileira de Cartografia 63,
211–219.
Statella, T., Silva, E.A., 2011b. Dust devils detection and inference of their move-
ment directions. In: Proceedings of 42nd Lunar and Planetary Science
Conference, Houston, TX, n. 1001, March 7–11.
Statella, T., Silva, E.A., 2011c. Directions of Martian dust devils movement inferred
from automatic detection of their tracks using mathematical morphology. In:
Proceedings of 34th International Symposium for Remote Sensing of the
Environment, Sydney, Australia, April 10–15.
Statella T., Pina P., Silva E.A., 2011. A study on automatic methods based on
mathematical morphology for Martian dust devils tracks detection. In: LNCS
7042, 533-540, 16th Iberoamerican Congress, Pucón, Chile, November 15–18.
Towner, M.C., 2009. Characteristics of large Martian dust devils using mars
odyssey thermal emission imaging system visual and infrared images. Journal
of Geophysical Research 114, 1–11.
Whelley, P.L., Greeley, R., 2006. Latitudinal dependency in dust devil activity on
Mars. Journal of Geophysical Research 111, E10003.
Whelley, P.L., Greeley, R., 2008. The distribution of dust devil activity on Mars.
Journal of Geophysical Research 113 (E7).