ZhaoAnran/mb-surface_multi_label_cls

---
annotations_creators:
- expert-generated
language_creators:
- found
language:
- en
license:
- cc-by-4.0
multilinguality:
- monolingual
size_categories:
- 1K<n<10K
source_datasets:
- original
task_categories:
- image-classification
task_ids:
- multi-label-image-classification
pretty_name: MER - Mars Exploration Rover Dataset
---

# MER - Mars Exploration Rover Dataset

A multi-label classification dataset containing Mars images from the Mars Exploration Rover (MER) mission for planetary science research.

## Dataset Metadata

* **License:** CC-BY-4.0 (Creative Commons Attribution 4.0 International)
* **Version:** 1.0
* **Date Published:** 2025-10-23
* **Cite As:** TBD

## Classes

This dataset uses multi-label classification, meaning each image can have multiple class labels.

The dataset contains the following classes:

- **rah** (0): Rock Abrasion Tool (RAT) Hole
- **cla** (1): Clasts
- **dur** (2): Dunes/Ripples
- **soi** (3): Soil
- **roc** (4): Rock Outcrops
- **clr** (5): Close-up Rock
- **rab** (6): Rock Abrasion Tool (RAT) Brushed Target
- **div** (7): Distant Vista
- **rod** (8): Rover Deck
- **bso** (9): Bright Soil
- **flr** (10): Float Rocks
- **art** (11): Artifacts
- **pct** (12): Pancam Calibration Target
- **arh** (13): Arm Hardware
- **rrf** (14): Rock (Round Features)
- **sph** (15): Spherules
- **ohw** (16): Other Hardware
- **ast** (17): Astronomy
- **nbs** (18): Nearby Surface
- **rmi** (19): Rocks (Misc)
- **rtr** (20): Rover Tracks
- **sky** (21): Sky
- **rpa** (22): Rover Parts
- **rlf** (23): Rock (Linear Features)
- **sot** (24): Soil Trench
## Statistics

- **train**: 1762 images
- **val**: 443 images
- **test**: 739 images
- **few_shot_train_10_shot**: 128 images
- **few_shot_train_15_shot**: 175 images
- **few_shot_train_1_shot**: 16 images
- **few_shot_train_20_shot**: 220 images
- **few_shot_train_2_shot**: 30 images
- **few_shot_train_5_shot**: 67 images
- **partition_0.01x_partition**: 19 images
- **partition_0.02x_partition**: 33 images
- **partition_0.05x_partition**: 81 images
- **partition_0.10x_partition**: 184 images
- **partition_0.20x_partition**: 361 images
- **partition_0.25x_partition**: 447 images
- **partition_0.50x_partition**: 878 images

## Few-shot Splits

This dataset includes the following few-shot training splits:

- **few_shot_train_10_shot**: 128 images
- **few_shot_train_15_shot**: 175 images
- **few_shot_train_1_shot**: 16 images
- **few_shot_train_20_shot**: 220 images
- **few_shot_train_2_shot**: 30 images
- **few_shot_train_5_shot**: 67 images

Few-shot configurations:

- **10_shot.csv**
- **15_shot.csv**
- **1_shot.csv**
- **20_shot.csv**
- **2_shot.csv**
- **5_shot.csv**
## Partition Splits

This dataset includes the following partition splits:

- **partition_0.01x_partition**: 19 images
- **partition_0.02x_partition**: 33 images
- **partition_0.05x_partition**: 81 images
- **partition_0.10x_partition**: 184 images
- **partition_0.20x_partition**: 361 images
- **partition_0.25x_partition**: 447 images
- **partition_0.50x_partition**: 878 images

Partition configurations:

- **0.01x_partition.csv**
- **0.02x_partition.csv**
- **0.05x_partition.csv**
- **0.10x_partition.csv**
- **0.20x_partition.csv**
- **0.25x_partition.csv**
- **0.50x_partition.csv**
## Format

Each example in the dataset has the following format:

```
{
'image': Image(...), # PIL image
'label': List[int], # Multi-hot encoded binary vector (1 if class is present, 0 otherwise)
'feature_name': List[str], # List of feature names (class short codes)
}
```

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("Mirali33/mb-surface_multi_label_cls")

# Access an example
example = dataset['train'][0]
image = example['image'] # PIL image
label = example['label'] # Multi-hot encoded binary vector

# Example of how to find which classes are present in an image
present_classes = [i for i, is_present in enumerate(label) if is_present == 1]
print(f"Classes present in this image: {present_classes}")
```

## Multi-label Classification

In multi-label classification, each image can belong to multiple classes simultaneously. The labels are represented as a binary vector where a 1 indicates the presence of a class and a 0 indicates its absence.

Unlike single-label classification where each image has exactly one class, multi-label classification allows modeling scenarios where multiple features can be present in the same image, which is often the case with Mars imagery.