LROCNet: Lunar Crater Classification Imagery from LRO

A collection of labeled lunar surface images used to train the LROCNet convolutional neural network classifier. The dataset contains three classes—'fresh crater', 'old crater', and 'none'—with images derived from calibrated Lunar Reconnaissance Orbiter Narrow Angle Camera swaths. It was created by Emily Dunkel at the Jet Propulsion Laboratory with volunteer annotations.

Use Cases

• Training crater detection models based on labeled 'fresh crater' and 'old crater' image chips.
• Benchmarking geospatial computer vision algorithms on planetary science tasks using the provided validation and test splits.
• Studying surface feature saliency on the Moon using the saliency detection method described.
• Developing data augmentation pipelines for imbalanced satellite image datasets, referencing the described PyTorch augmentation techniques.

Strengths

• Images are pre-split into train, validation, and test sets with no overlapping swaths, ensuring clean evaluation.
• Class distribution is quantified, with 'fresh crater' making up 11% and 'old crater' 18% of an initial labeled set.
• Includes a detailed labeling guide describing the classification criteria for human annotators.

Limitations

• The dataset was reduced from an initial 7 classes to 3 due to limited examples, narrowing its scope.
• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.

Provenance

Source

Lunar Reconnaissance Orbiter Narrow Angle Cameras (NACs).

Collection Method

Calibrated Data Record browse imagery was processed with a saliency detector to create 227x277 pixel image chipouts, which were then labeled via the Zooniverse platform.

Geography

Global lunar surface.