Martian Frost in HiRISE Observations: Labeled Image Tiles for Frost Detection

HiRISE satellite imagery of Mars's northern mid-latitude craters is annotated for frost presence. The dataset contains labeled 299x299 pixel image tiles at 0.5 m/pixel resolution, with labels determined by a majority vote over polygonal annotations. It was created by S. Diniega of the Jet Propulsion Laboratory using the Labelbox platform.

Use Cases

• Training image classification models for frost detection based on labeled HiRISE image tiles.
• Developing semantic segmentation models for planetary surface features based on polygonal annotations.
• Analyzing seasonal frost patterns on Mars based on comparisons between 'frost' and 'background' (summer) subframes.

Strengths

• Labels are derived from a majority vote across polygonal annotations, which may improve consistency.
• Includes 'background' subframes taken during summer months, providing a control set for model training.
• Data is provided in TensorFlow TFRecord format, which is convenient for machine learning workflows.

Limitations

• Row count and dataset size are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The 'ambiguous' label category indicates potential uncertainty in the ground truth for some tiles.

Provenance

Source

Jet Propulsion Laboratory

Collection Method

Created using Labelbox to draw polygonal annotations over subframes extracted from HiRISE observations.

Geography

Mars, specifically northern mid-latitude craters.