gremlin97/mars-multi-label-classification

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: 2025-05-10
• 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

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

Format

Each example in the dataset has the following format:

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

Usage

from datasets import load_dataset

dataset = load_dataset("gremlin97/mars-multi-label-classification")

# Access an example
example = dataset['train'][0]
image = example['image']  # PIL image
labels = example['labels']  # 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(labels) 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.