Presentation Information
[PPS01-P14]Machine learning based terrain classification of Saturn’s moon Titan using SAR and VIMS data
*Kumpei Goromaru1, Hitoshi Hasegawa1, An-Sheng Lee1 (1.Kochi University)
Keywords:
Titan,Machine Learning,Autoencoder,Gaussian Mixture Model
Titan, Saturn's largest moon, is currently recognized as the only solar system body besides Earth to possess a stable liquid cycle on its surface. The Cassini spacecraft's Synthetic Aperture Radar (SAR) and Visual and Infrared Mapping Spectrometer (VIMS) played a key role in global mapping Titan's terrain distribution over 13 years of observations (Lopes et al., 2019 Nature Astronomy). However, existing research faces challenges in quantitatively classifying the complex terrains observed on Titan's surface. This study attempts to classify and map Titan's surface terrain using machine learning. Here, an unsupervised learning approach was adopted to achieve an objective and quantitative classification (Gao et al., 2021 Proceedings of the IEEE/CVF conference).
The SAR mosaic was divided into 128x128 pixel meshes (approximately 1°x1°), with each mesh treated as an independent analysis unit. For each mesh, feature learning was performed using a Convolutional Autoencoder, and the resulting latent feature representations were clustered using a Gaussian Mixture Model (Gao et al., 2021 Proceedings of the IEEE/CVF conference). Each cluster was subsequently compared to the six terrain units defined by previous research and the best-matching unit for each cluster was identified through visual inspection. The meshes were then colored according to their corresponding terrain units. This workflow was applied to two regions: Xanadu and Selk crater.
The terrain classification maps generated in this study were compared with the results of Lopes et al. (2019). In the Xanadu and Selk crater regions, the clustering results generally agreed with previous studies, primarily comprising three main terrain units: dunes, hummocky, and plains. Conversely, labyrinths, craters, and lakes were integrated into other units rather than being extracted as independent clusters. While discrepancies with previous research were observed in some areas, these are likely due to the use of different datasets. Lopes et al. (2019) determined terrain units by integrating correlations among multiple datasets, utilizing not only SAR data but also emissivity, altimeter, and VIMS data. Therefore, efforts are currently underway to attempt a classification that integrates both SAR and VIMS data by employing a four-dimensional dataset consisting of SAR images and the RGB components of VIMS as input.
The SAR mosaic was divided into 128x128 pixel meshes (approximately 1°x1°), with each mesh treated as an independent analysis unit. For each mesh, feature learning was performed using a Convolutional Autoencoder, and the resulting latent feature representations were clustered using a Gaussian Mixture Model (Gao et al., 2021 Proceedings of the IEEE/CVF conference). Each cluster was subsequently compared to the six terrain units defined by previous research and the best-matching unit for each cluster was identified through visual inspection. The meshes were then colored according to their corresponding terrain units. This workflow was applied to two regions: Xanadu and Selk crater.
The terrain classification maps generated in this study were compared with the results of Lopes et al. (2019). In the Xanadu and Selk crater regions, the clustering results generally agreed with previous studies, primarily comprising three main terrain units: dunes, hummocky, and plains. Conversely, labyrinths, craters, and lakes were integrated into other units rather than being extracted as independent clusters. While discrepancies with previous research were observed in some areas, these are likely due to the use of different datasets. Lopes et al. (2019) determined terrain units by integrating correlations among multiple datasets, utilizing not only SAR data but also emissivity, altimeter, and VIMS data. Therefore, efforts are currently underway to attempt a classification that integrates both SAR and VIMS data by employing a four-dimensional dataset consisting of SAR images and the RGB components of VIMS as input.