REANALYSIS OF MARS ORBITER LASER ALTIMETER ATMOSPHERIC
FEATURES WITH MACHINE LEARNING ALGORITHMS.
V. Caillé, LATMOS, Sorbonne Université, UVSQ Paris-Saclay, CNRS, Paris, France (vincent.caille@latmos.ipsl.fr),
A. Määttänen, LATMOS, Sorbonne Université, UVSQ Paris-Saclay, CNRS, Paris, France, A. Spiga, LMD/IPSL, Sorbonne Université, Paris, France - Institut Universitaire de France, France, L. Falletti, LATMOS, Sorbonne Université,
UVSQ Paris-Saclay, CNRS, Paris, France, G. A. Neumann, NASA Goddard Space Flight Center, Greenbelt, Maryland,
USA.
Introduction
The Martian atmosphere is a mix of diverse kinds of
aerosols structures and clouds with different compositions such as dust, water or CO2 ice. Processes involved
in their formations are complex and their understanding have been enhanced by observations from the last
decades mission. Between 1996 and 2006, Mars Global
Surveyor (MGS) carried three instruments, the Mars
Orbiter Camera (MOC), the Thermal Emission Spectrometer (TES) and the Mars Orbiter Laser Altimeter
(MOLA) that have all been able to observe clouds during the same period with different methods. Gathering
and comparing results from these three datasets could
give an appreciation of what has happened in the Martian atmosphere during 1,5 martian years. However,
previous studies of MOLA observations of clouds [Neumann et al., 2003, Ivanov and Muhleman, 2001] were
numerically restrained and we suggest that reanalysing
the dataset with recent methods could give more clouds
and dust observations. Since then, both MOC and TES
datasets have also been analysed, allowing us to compare all three instruments observations. We could also
compare our results with observations from missions
launched after MGS and modeling results.
MOLA [Zuber et al., 1992] was an altimeter aboard
MGS which first goals were to draw precise Mars topography, roughness and albedo at 1064 µm-wavelength
maps using a pulsing laser directed towards Mars’ surface [Smith et al., 2001]. Nevertheless, martian clouds
were dense enough to trigger MOLA receptor and previous studies showed that some laser returns were signatures of atmospheric features. Their distinction of
the different kinds of returns (surface, noise or atmospheric features) have been made using stringent detection criteria to deal with the consequent amount of data.
Thought it ensured that there were as few false positives
as possible in their atmospheric features observations, it
eventually led to missing an important part of them.

Method & Validation
Machine learning algorithms, and especially clustering
methods are an interesting way to extract specific features from massive dataset. K-means method is usually

a good first approach for analysing dataset with such
methods because it is certain to converge [Selim and Ismail, 1984] and quickly enough [Har-Peled and Sadri,
2004]. In order to apply it to MOLA dataset, we first
have to optimize both the distinguishing variables and
the desired number of clusters k. Following [Neumann
et al., 2003] work, the product of surface reflectivity
with the two-way atmosphere transmissivity, rT 2 , can
be used as a returned laser energy, normalised by MOLA
performances over the mission duration and has a clear
signature for atmospheric returns when plotted against
time : while surface returns form a continuum and noise
returns are sporadic sudden variations, atmospheric features returns cause a dip lasting for several consecutive
returns. Three independent methods, elbow, silhouette
score and gap statistic, allow to get the optimised number of clusters by evaluating the clustering performances
for different k. It allowed us to see that both k = 5 and
k = 6 had similar performances but we decided to go
for k = 6 because it eased interpretation of each cluster
and lowered the chance of getting false positives in our
atmospheric features cluster. We subsequently applied
K-means method to all 12 mission orbital phases. Clustering results (example given for one orbital phase in
figure 1) present the expected clusters, and especially
one made of atmospheric features returns.
Validation of our distinction method results is made
by comparing our atmospheric features seasonal and latitudinal distribution (fig 2) with the ones obtained in
reference study [Neumann et al., 2003]. Areas or seasons with the highest density of atmospheric returns,
like north pole during northern autumn, cloudy episode
between 30◦ S and 60◦ at the end of southern winter or
the almost global coverage around Ls 230◦ are present
in both distributions and show that our results are in total
agreement with previous results. However, we end up
having almost 33 times more atmospheric returns. We
regrouped consecutive atmospheric returns into atmospheric structures, that can be dust, water or CO2 ice
clouds and saved them into a catalog in order to ease
their visualisations.

Reanalysis of Mars Orbiter Laser Altimeter Atmospheric Features with Machine Learning Algorithms

Figure 1: Clustering structure for one orbital phase (top) and
example of an atmospheric feature in rT 2 against time plot
(bottom)

what highlights some areas of interest such as Hellas
Basin, Valles Marineris, Syrtis Major...
These maps also allow to see some well-known phenomenon that occur in the Martian atmosphere that were
not observed that well by MOLA previously. By combining results from both years, every stages of the aphelion cloud belt can be seen, from its formation around
Ls 15-30◦ till its density peak around Ls 120◦ . Even
if MOLA analysis does not provide us any information
about atmospheric structures compositions, comparison
with TES [Hale et al., 2011] and MOC [Wang and Ingersoll, 2002] observations allow us to determine that
most of these clouds are water clouds. The development of the south polar hood and its correlation with
the decline of the south polar cap toward the south pole
during southern spring can also be observed while being
absent in MOC observations, what is coherent with assumptions of these structures being CO2 ice clouds. We
are also coherent with global climate model dust profiles that established an almost global coverage between
Ls 220◦ and Ls 260◦ during MY24 [Montabone et al.,
2015], what is observed in our results. Formation and
evolution of clouds above the poles are also analysed
and linked to other missions observations.
Acknowledgements
We thank the Agence National de la Recherche for
funding (projet MECCOM, ANR-18-CE31-0013). This
work was performed using HPC computing resources
from GENCI-CINES (Grant 2021-A0100110391), and
resources at the ESPRI mesocentre of the IPSL institute

Figure 2: Seasonal and latitudinal distribution of atmospheric
features, colors represent number of returns in 1◦ x1◦ bin
around each point

Atmospheric Features Variability & Diversity
We divided the dataset into windows of 15◦ of solar
longitude to draw maps of longitudinal and latitudinal
distribution of atmospheric features during the whole
mission. This gives a final results of what could be observed in the low altitude atmosphere by MOLA during
MY24 from Ls 103◦ (fig 3) and the first half of MY25
till Ls 187◦ (fig 4) and allow us to analyse interannual
variability as well as seasonal variability. By comparing
the only period we have in common for both years, ie
southern winter, we can see that there were way more
atmospheric structures during MY25. However, their
distributions are very similar for both years, apart from
Tharsis Montes where big structures (at least 150 km
long) only formed during MY25. These big structures
distribution do not follow the global distribution (fig 2),

REFERENCES

Figures

References
[Hale et al., 2011] Hale, A., Tamppari, L., Bass, D., and
Smith, M. (2011). Martian water ice clouds: A view
from mars global surveyor thermal emission spectrometer. Journal of Geophysical Research, 116.
[Har-Peled and Sadri, 2004] Har-Peled, S. and Sadri,
B. (2004). How fast is the k-means method? Algorithmica, 41(3):185â202.
[Ivanov and Muhleman, 2001] Ivanov, A. B. and Muhleman, D. O. (2001). Cloud reflection observations:
Results from the mars orbiter laser altimeter. Icarus,
154(1):190–206.
[Montabone et al., 2015] Montabone, L., Forget, F.,
Millour, E., Wilson, R., Lewis, S., Cantor, B., Kass,
D., Kleinböhl, A., Lemmon, M., Smith, M., and
et al. (2015). Eight-year climatology of dust optical
depth on mars. Icarus, 251:65â95.

Figure 3: 15◦ longitudinal and latitudinal maps of atmospheric
features during MY24. Colorbar represents the size of atmospheric structures in log-scale

[Neumann et al., 2003] Neumann, G. A., Smith, D. E.,
and Zuber, M. T. (2003). Two Mars years of clouds
detected by the Mars Orbiter Laser Altimeter. Journal
of Geophysical Research (Planets), 108(E4):5023.
[Selim and Ismail, 1984] Selim, S. Z. and Ismail, M. A.
(1984). K-means-type algorithms: A generalized
convergence theorem and characterization of local
optimality. IEEE Transactions on Pattern Analysis and
Machine Intelligence, PAMI-6(1):81–87.
[Smith et al., 2001] Smith, D. E., Zuber, M. T., Frey,
H. V., Garvin, J. B., Head, J. W., Muhleman, D. O.,
Pettengill, G. H., Phillips, R. J., Solomon, S. C.,
Zwally, H. J., Banerdt, W. B., Duxbury, T. C.,
Golombek, M. P., Lemoine, F. G., Neumann, G. A.,
Rowlands, D. D., Aharonson, O., Ford, P. G., Ivanov,
A. B., Johnson, C. L., McGovern, P. J., Abshire, J. B.,
Afzal, R. S., and Sun, X. (2001). Mars orbiter laser
altimeter: Experiment summary after the first year
of global mapping of mars. Journal of Geophysical
Research: Planets, 106(E10):23689–23722.
[Wang and Ingersoll, 2002] Wang, H. and Ingersoll,
A. P. (2002). Martian clouds observed by mars global
surveyor mars orbiter camera. Journal of Geophysical
Research: Planets, 107(E10):8–1–8–16.

Figure 4: 15◦ longitudinal and latitudinal maps of atmospheric
features during MY25. Colorbar represents the size of atmospheric structures in log-scale

[Zuber et al., 1992] Zuber, M. T., Smith, D. E.,
Solomon, S. C., Muhleman, D. O., Head, J. W.,
Garvin, J. B., Abshire, J. B., and Bufton, J. L. (1992).
The mars observer laser altimeter investigation. Journal of Geophysical Research: Planets, 97(E5):7781–
7797.