MMLSv2: A Multimodal Dataset for Martian Landslide Detection in Remote Sensing Imagery

TL;DR

MMLSv2 tackles the scarcity and generalization gap in Martian landslide segmentation by introducing a seven-channel multimodal dataset (RGB, DEM, slope, thermal inertia, grayscale) across 664 images plus an isolated 276-image test set for spatial generalization. The authors implement a careful data fusion and labeling workflow, co-registration, and a patch-based, spatially independent partitioning strategy to provide robust, reproducible benchmarks. Across multiple segmentation architectures, the baseline performance is competitive ($mIoU$ around $0.81$–$0.83$ on the standard split) but drops on the isolated test set, highlighting regional generalization challenges; adding modalities consistently improves performance, with the full seven-band configuration achieving the strongest delineation of landslide boundaries. This dataset and evaluation protocol advance reproducible research in planetary geomorphology by enabling robust, multimodal landslide mapping and fair assessment of generalization under distribution shifts.

Abstract

We present MMLSv2, a dataset for landslide segmentation on Martian surfaces. MMLSv2 consists of multimodal imagery with seven bands: RGB, digital elevation model, slope, thermal inertia, and grayscale channels. MMLSv2 comprises 664 images distributed across training, validation, and test splits. In addition, an isolated test set of 276 images from a geographically disjoint region from the base dataset is released to evaluate spatial generalization. Experiments conducted with multiple segmentation models show that the dataset supports stable training and achieves competitive performance, while still posing challenges in fragmented, elongated, and small-scale landslide regions. Evaluation on the isolated test set leads to a noticeable performance drop, indicating increased difficulty and highlighting its value for assessing model robustness and generalization beyond standard in-distribution settings. Dataset will be available at: https://github.com/MAIN-Lab/MMLS_v2

Figures (8)

• Figure 1: Location of Valles Marineris on Mars used in this study for constructing the MMLSv2 dataset.
• Figure 2: Distribution of foreground ratio intervals across the splits of MMLSv2. The three subsets exhibit highly consistent patterns, indicating that the proposed partitioning strategy preserves comparable levels of sample difficulty, with no split dominated by either nearly empty patches or highly foreground-dense samples.
• Figure 3: Representative scenes from the MMLSv2 dataset. Each row corresponds to a different image tile, while columns show the individual input bands composing the multimodal image, followed by the ground-truth landslide mask. The examples illustrate the wide morphological diversity captured by MMLSv2, including small, isolated events (rows a-b), elongated and curved landslides (rows c-d), extensive and continuous failure areas (rows e-f), and fragmented, irregular landslides occurring in complex geomorphological settings (rows g-h). Band order: (1) Red, (2) Green, (3) Blue, (4) DEM, (5) Slope, (6) Thermal inertia, (7) Grayscale.
• Figure 4: Representative scenes from the MMLSv2 isolated test. Compared to the baseline split, these examples highlight noticeable shifts in spatial context, texture, and landslide morphology, including fragmented multi component failures (row a), large continuous regions (row b), discontinuous and irregular patterns (row c), elongated and curved structures (row d), and small isolated events (row e). This shows the out-of-distribution nature of the isolated test set and its role as a challenging benchmark for evaluating model generalization beyond standard in-distribution settings. Band order: (1) Red, (2) Green, (3) Blue, (4) DEM, (5) Slope, (6) Thermal inertia, (7) Grayscale.
• Figure 5: Qualitative comparison of landslide segmentation results obtained with the evaluated models trained using the full set of input bands in MMLSv2. While all models capture the main landslide structures, the highlighted regions reveal persistent errors in boundary delineation, small-scale failures, and fragmented or ambiguous areas, indicating room for improvement.