Detecting Clouds in Mars Orbiter Laser Altimeter/Mars Global Surveyor Data with K-Means Methods
Résumé
The Mars Orbiter Laser Altimeter (MOLA) was an instrument aboard the Mars Global Surveyor spacecraft in charge of building a precise map of Mars’ topography using laser altimetry. However, sensitivity was better than expected, leading to the detection of features that could not be assigned to the surface. In particular, MOLA was the first instrument to detect polar winter CO2 ice clouds. Previous studies have attributed some laser returns to cloud signatures coming from the atmosphere. Due to the large amount of data to analyse, those studies required to very strict distinction criteria, but today there are efficient means for analysing huge datasets.
K-means clustering methods seem to be good options to computionally analyse MOLA data. We proceed by applying the method on a single PEDR data file (about 10 % of data) as a test case. We first explore the observed parameters to find those that allow us to clearly separate surface and atmosphere (cloud) returns. We then use three independent optimisation methods, elbow, gap statistic and average silhouette, to determine the best number of clusters. We then apply the method to the whole data set, and eventually can plot spatial and temporal cloud distributions once the cloud cluster has been identified.
Following the Neumann and al. paper (2003), we find that the product of surface reflectivity and two-way transmissivity of the atmosphere is the best parameter for distinguishing cloud and surface returns. Our three optimisation methods converge to an unique number of clusters for our test case. Plotting the clusters shows that one of them clearly identifies clouds returns, while other ones allow for identification of noise and surface returns. Our method allows us to find more clouds than previous studies due to less stringent detection criteria. Cloud distributions are in agreement with reference studies and tend to confirm the viability of our method. Next step is to work within the cloud cluster to separate different kinds of clouds (absorptive/reflective, CO2/water/dust...), possibly by using other unsupervised machine learning methods.