# MMLSv2: A Multimodal Dataset for Martian Surface Landslide Detection

> MMLSv2 is a multimodal remote sensing dataset for Martian surface landslide detection, containing 7 bands including RGB, elevation model, slope, thermal inertia, etc. It has been accepted by the CVPR 2026 AI4Space Workshop and serves as the official dataset for the MARS-LS Challenge.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-05-31T09:06:00.000Z
- 最近活动: 2026-05-31T09:23:10.119Z
- 热度: 163.7
- 关键词: MMLSv2, 火星滑坡, 多模态数据集, 遥感分割, CVPR 2026, AI4Space, 行星地质, 深度学习, 计算机视觉, 深空探测
- 页面链接: https://www.zingnex.cn/en/forum/thread/mmlsv2
- Canonical: https://www.zingnex.cn/forum/thread/mmlsv2
- Markdown 来源: floors_fallback

---

## Introduction / Main Floor: MMLSv2: A Multimodal Dataset for Martian Surface Landslide Detection

MMLSv2 is a multimodal remote sensing dataset for Martian surface landslide detection, containing 7 bands including RGB, elevation model, slope, thermal inertia, etc. It has been accepted by the CVPR 2026 AI4Space Workshop and serves as the official dataset for the MARS-LS Challenge.

## Original Authors and Source

- Original Authors/Maintainers: MAIN-Lab (Sidike Paheding, Abel Reyes-Angulo, Leo Thomas Ramos, etc.)
- Source Platform: GitHub
- Original Title: MMLS_v2
- Original Link: https://github.com/MAIN-Lab/MMLS_v2
- Source Publication/Update Time: 2026-05-31T09:06:00Z

## Research Background

Martian geological research is an important topic in human deep space exploration. Landslide landforms are widely distributed on the Martian surface, and these landslides record clues about Mars' climate history and hydrological activities. Accurate identification and segmentation of Martian landslide areas have important scientific value for understanding Martian geological evolution and assessing the risk of site selection for future manned missions.

However, Martian landslide detection faces unique challenges:

- **Data Scarcity**: Compared to Earth's remote sensing data, the number of high-resolution Mars image samples is limited
- **Multi-source Heterogeneity**: It is necessary to fuse data from multiple sensors such as visible light, elevation, and thermal characteristics
- **Large Domain Differences**: Geological features vary significantly across different regions, making model generalization difficult

## Introduction to the MMLSv2 Dataset

MMLSv2 (Multimodal Dataset for Martian Landslide Detection) is a multimodal remote sensing dataset specifically designed for the Martian landslide segmentation task. This dataset has been accepted by the CVPR 2026 AI4Space Workshop and serves as the official dataset for the first Martian Landslide Segmentation Challenge (MARS-LS).

## Data Composition

The dataset contains remote sensing data of 7 bands:

| Band | Content | Description |
|------|------|------|
| B1 | Red | Red visible light |
| B2 | Green | Green visible light |
| B3 | Blue | Blue visible light |
| B4 | DEM | Digital Elevation Model |
| B5 | Slope | Slope data |
| B6 | Thermal Inertia | Thermal Inertia |
| B7 | Grayscale | Grayscale channel |

Multi-band fusion allows models to simultaneously utilize spectral information, topographic features, and thermophysical properties, significantly improving the accuracy of landslide identification.

## Data Scale and Division

The dataset contains a total of 940 images, divided as follows:

| Division | Number of Images | Average Foreground Ratio | Foreground Ratio Range |
|------|--------|--------------|--------------|
| Training Set | 465 | 35.41% | 0.02% - 99.52% |
| Validation Set | 66 | 31.53% | 0.08% - 90.32% |
| Test Set | 133 | 33.82% | 0.10% - 90.67% |
| Independent Test Set | 276 | 21.83% | 0.01% - 71.95% |

The independent test set comes from geographically disjoint regions of the main dataset and is specifically used to evaluate the spatial generalization ability of models.

## Data Format

Each sample is a multi-channel image of size `(128, 128, 7)` with data type `float32`, and pixel values are normalized to the range `[0.0, 1.0]`. The corresponding mask is a single-channel `(128, 128)` `uint8` image, with pixel values 0 or 1 representing background or landslide areas respectively.

## Benchmark Test Results

The paper reports the performance of 6 mainstream segmentation models on MMLSv2: