Automatic Classification of Aeolian Dunes on Mars Using an Autoencoder

The morphological analysis of aeolian dunes from satellite and spacecraft imagery has traditionally relied on human expert interpreters. This approach has been often impeded by the need for manual input, which limits the breadth of scope, and could be sensitive to human bias [1-3]. In recent years, machine learning techniques have revolutionized the automatic analysis of images, particularly for the purpose of object detection [4, 5] – but require optimization (“training”) on large manually labeled datasets. In this study, we use an autoencoder, a convolutional neural network which learns from context – and does not require manual labels – to analyze the morphology of dunes in the north polar erg of Mars.

Like the better-known principal component analysis technique (PCA), an autoencoder synthesizes information through dimensionality reduction. By compressing input data and reconstructing it back from the compressed version, the autoencoder effectively extracts important features from non-linear data like images (similar to principal components of a linear PCA) without human guidance. In this study, we employ an autoencoder to analyze images of martian dunes obtained from the global mosaic processed from calibrated images obtained by the Mars Reconnaissance Orbiter (MRO) Context Camera (CTX) (5 m/px) [6]. After optimizing the model weights on 8000 images, we use the autoencoder to automatically classify dune morphology.

The compact representation of images of martian dunes in this principal component space shows clustering by dune morphology (Figure 1); not only by dune type, but also by varying morphology of the same type of dunes. Our synthesized data, which does not require discrete categorical classification (unlike, e.g., [7]) demonstrates the continuous transition between isolated barchan dunes and connected barchanoidal ridges. For example, our model identifies the higher density of barchanoidal ridges in Olympia Undae and near Escorial crater, which is in the convergence between Chasma Borelae and the circumpolar erg (red points), previously manually mapped [8].

In the meeting, we will present refined autonomous mapping and morphological analysis of dunes on Mars using a state-of-the-art Mask Autoencoder [9] and apply our model to satellite images of terrestrial dunes.

 

References:

[1] Bond et al., GSA (2007).

[2] Robbins et al., Icarus (2014).

[3] Bergen et al., Science (2019).

[4] Rubanenko et al., IEEE JSTARS (2021).

[5] Ali-Dib et al., Icarus (2020).

[6] Dickson et al., ESS (2024).

[7] Du Pont et al., ESR (2024).

[8] Hayward et al., Journal of Geophysical Research: Planets (2007).

[9] Kaiming et al., Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022).

 

Plain-language Summary

We propose using machine learning, specifically autoencoders, to analyze the morphology of aeolian dunes on Mars in satellite imagery without manual labeling. By training the model on thousands of images, we automatically identified and grouped dune morphologies, revealing transitions like those between isolated barchan dunes and connected ridges. Key findings include clustering of specific dune types in regions like Olympia Undae and Escorial Crater. This method provides an efficient, unbiased approach to studying Martian and terrestrial dunes.