Expanding the Global Martian Landslide Inventory with Multimodal DINOv2 Feature Fusion and SVM Classification

Abstract Landslides are globally distributed on Mars and resemble terrestrial analogues in morphology. Understanding the spatial distribution of Martian landslides provides key constraints on Martian surface processes, tectonic activity, and the environmental conditions that govern landslide mechanics. However, existing global inventories remain incomplete because capturing small, overlapping, or morphologically ambiguous deposits solely relying on manual mapping is challenging, leaving uncertainty in understanding their spatial distribution. In this study, we propose Mars-DiSVM, a landslide identification framework based on multimodal images, which fuses features extracted from optical, thermal, and topographic data using a DINOv2 backbone, and the features are classified by a downstream Support Vector Machine (SVM) classifier. The framework using concatenated feature representation achieved a classification accuracy of 97.5%. Applying Mars-DiSVM to four areas of interest (AOIs), 25 previously unmapped, small-scale landslides have been identified over an area of 224,676 km 2 . These landslides are clustered within impact craters and along valley slopes, where manual mapping commonly underrepresents small deposits. Three of the landslides occur in close proximity to impact craters, implying impact as a possible trigger. The results demonstrate the potential of the framework to refine the global landslide inventory on Mars. The expanded inventory produced by Mars-DiSVM reveals that small landslides constitute a substantial and previously underrepresented component of Mars mass wasting, providing critical constraints on landslides occurrence and triggering mechanism and offering new opportunities to reassess Martian surface activity and crater degradation processes.