SUSTech/Mars-Landforms

---
language:
- en
license: cc
size_categories:
- 1K<n<10K
task_categories:
- text-to-image
dataset_info:
features:
- name: image
dtype: image
- name: label
dtype:
class_label:
names:
'0': Aeolian_Bedforms
'1': Aeolian_Dunes
'2': Aeolian_Ripples
'3': Barchan_Dunes
'4': Boulder_Track
'5': Brain_Terrain
'6': Bright_Rays_Craters
'7': Central_Peak_Crater
'8': Chaos
'9': Cliff
'10': Concentric_Crater_Fill
'11': Crater_Chain
'12': Crater_Cluster
'13': Dark_Ray_Craters
'14': Double_Ring_Basin
'15': Doublet_Crater
'16': Dune_Field
'17': Dust_Devil_Tracks
'18': Fan_Shape_Deposit
'19': Fractured_Mounds
'20': Fresh_Crater
'21': Gully
'22': Landslide
'23': Lava_Flow_Front
'24': Lava_Tubes
'25': Layers
'26': Linear_Dunes
'27': Lobate_Debris_Apron
'28': Outflow_Channel
'29': Pancake_Crater
'30': Pedestal_Crater
'31': Pitted_Cone
'32': Pitted_Terrain
'33': Polar_Layered_Deposits
'34': Polygons
'35': Rampart_Crater
'36': Rocky_Ejecta_Crater
'37': Scalloped_Depression
'38': Slope_Streaks
'39': Spider
'40': Swiss_Cheese
'41': Transverse_Aeolian_Ridges
'42': Troughs
'43': Valley_Networks
'44': Volcano
'45': Wind_Streaks
'46': Wrinkle_Ridges
'47': Yardangs
splits:
- name: train
num_bytes: 763505091
num_examples: 1185
download_size: 758103040
dataset_size: 763505091
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
tags:
- planet
- multimodal
- retrieval
---

# Landform Retrieval

[**Paper**](https://huggingface.co/papers/2602.13961) | [**Code**](https://github.com/ml-stat-Sustech/MarsRetrieval)

## Dataset Summary

This dataset is Task 2 of [**MarsRetrieval**](https://github.com/ml-stat-Sustech/MarsRetrieval), a retrieval-centric benchmark for evaluating vision-language models (VLMs) on Mars geospatial discovery. Task 2 evaluates **concept-to-instance generalization** for Martian geomorphology. Given a textual geomorphic concept, the model must retrieve its corresponding visual instances from a curated Martian image gallery. The dataset comprises **1,185** carefully curated image patches collected from CTX and HiRISE imagery. The landforms follow a two-level geomorphology taxonomy:

- **7 major genetic classes** (e.g., Aeolian, Volcanic and Fluvial processes)
- **48 geomorphic subclasses** (e.g., Aeolian Dunes, Dust Devil Tracks, Yardangs)

## Task Formulation

We formulate this task as a **text-to-image multi-positive retrieval problem**:

- A text query describes a geomorphic subclass.
- Multiple image instances in the gallery are considered valid positives.
- The goal is to rank all gallery images by cosine similarity in the embedding space.

### Metrics

We report metrics suitable for long-tailed multi-positive retrieval:

- Macro mean Average Precision (mAP)
- nDCG@10
- Hits@10

## How to Use
```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("SUSTech/Mars-Landforms")

# Access a sample image and its geomorphic label
print(dataset["train"][0]["image"])
print(dataset["train"][0]["label"])
```
For detailed instructions on the retrieval-centric protocol and official evaluation scripts, please refer to our [Official Dataset Documentation](https://github.com/ml-stat-Sustech/MarsRetrieval/blob/main/docs/DATASET.md).

## Citation

If you find this useful in your research, please consider citing:

```bibtex
@article{wang2026marsretrieval,
title={MarsRetrieval: Benchmarking Vision-Language Models for Planetary-Scale Geospatial Retrieval on Mars},
author={Wang, Shuoyuan and Wang, Yiran and Wei, Hongxin},
journal={arXiv preprint arXiv:2602.13961},
year={2026}
}
```