Expanding the Global Martian Landslide Inventory with Multi-modal DINOv2 Feature Fusion and SVM Classification

Landslides on Mars are abundant and far more mobile than terrestrial landslides. Their exceptional scale and mobility provide key constraints on Martian surface processes, tectonic activity, and the environmental conditions that govern landslide mechanics. However, existing global inventories remain incomplete as small, overlapping, or morphologically ambiguous deposits are difficult to capture through manual mapping alone, leaving uncertainty in understanding their spatial distribution. Automated Martian landslide detection remains challenging due to the data scarcity with only a few thousand labeled samples and the natural morphological complexity of landslides. Therefore, we propose Mars-DiSVM, a landslide identification framework based on multi-modal imagery, which fuses features extracted from CTX, MOLA-HRSC DEM, and THEMIS night-time imagery using a DINOv2 backbone, followed by a downstream SVM classifier. The classification using fused feature representations achieves the top accuracy, up to 97.5%, with precision, recall, and MCC consistently exceeding 90%. Mars-DiSVM was further assessed within four areas of interest (AOIs), including the area with mapped landslides in the existing global inventory [1] and areas without mapped landslides in Noachian/Hesperian highlands.
Our framework identified 25 previously unmapped landslides across four AOIs, where the features are predominantly located on slopes within impact craters and along valley slopes, and are classified as rock avalanches and slump/flow types. These features are generally small, with approximately
half exhibiting runout lengths shorter than 5 km, sizes which are often underrepresented in manual mapping due to limited visibility or morphological degradation. The newly mapped landslides display diagnostic morphological characteristics, including lateral levees, tongue-shaped deposits, and longitudinal ridges within the deposits. Notably, three of the detected landslides occur adjacent to impact craters, implying impact events as the possible trigger. These findings highlight the importance of improving the completeness of global inventory, providing clues to their potential triggering mechanism.
Mars-DiSVM is implemented at the global scale to generate a preliminary expanded global inventory of Martian landslides. The resulting dataset will provide new constraints on the spatial distribution of landslides, thereby improving our understanding of their relationship with key controlling factors, such as the presence of ice or water and seismic activity [2]. In addition, we plan to monitor recent Martian landslide activity by incorporating newly acquired CTX imagery, thereby gaining insights into Martian recent geological activity and triggering mechanisms.

[1] Crosta, G. B., Frattini, P., Valbuzzi, E., & De Blasio, F. V. 2018, Earth and Space Science, 5, 89, doi: 10.1002/2017EA000324

[2] Roback, K. P., & Ehlmann, B. L. 2021, Journal of Geophysical Research: Planets, 126, e2020JE006675, doi: 10.1029/2020JE006675