---
license: cc-by-4.0
tags:
- space
- mars
- craters
---

## Hello!
**cMars12k** is a collection of 11,355 200x200 .jpg images of Martian craters, as well as their corresponding likelihoods of having subsurface ice at a depth of 0-1m.  
This is an updated and more balanced version of [cMars8k](https://huggingface.co/datasets/evans44/cMars8k).

**A short explanation of how the dataset was derived:**  
- The [THEMIS IR Day Map](https://www.mars.asu.edu/data/thm_dir/) was used to create this dataset, owing to its lightweight nature and scientific reputability.  
- [Robbins & Hynek's 2020 Mars crater database](https://craters.sjrdesign.net/) was used to source the geographic coordinates of the craters' (approximate) centers.  
- NASA's [SWIM Project 0-1m Ice Consistency map](https://swim.psi.edu/SWIM4MIMProducts.php) was used to source the likelihood of subsurface ice at the location of each crater.
- All of these individual datasets were combined together by overlaying them on top of each other inside of QGIS, and then utilizing a python script to mass-export the images.
- Each category contains exactly 2,000 images, except for `very_likely_confirmed`, which contains 1,355.

## Data Formatting
As stated, each image is exactly 200x200 pixels and is stored as a .jpg file within cMars8k/images.
The `ice_likelihood` value is a float64 ranging from approximately `-0.375` to `1.000`, with `1.000` representing an extremely high probability or confirmed presence of subsurface ice.  

Each image has been sorted into one of six classes based on the `ice_likelihood` value, as follows:  
`-0.375 to -0.20`: `very_unlikely`  
`-0.20 to 0.00`: `unlikely`  
`0.00 to 0.20`: `possible`  
`0.20 to 0.50`: `somewhat_likely`  
`0.50 to 0.80`: `likely`  
`0.80 to 1.00`: `very_likely_confirmed`  
  
**Binary classifiers** are recommended to use >=0.50 as the cutoff for Likely and <0.50 for Unlikely, reflective of the original NASA/SWIM Project data. Note that this split is imbalanced (8000 Unlikely > 3355 Likely), however has shown successful results in training with a from-scratch binary classifier as well as one based on the ResNet50 architecture.

## Accessing Data
This dataset can be accessed via:  
```
git clone https://huggingface.co/datasets/evans44/cMars12k
```
or via the Hugging Face API.