HiRISE Martian Dust Patches: Binary Classification

Mars years 24–30 (1999–2011) provide the temporal coverage for this dataset of 40,000 image patches from HiRISE observations. It is designed to train a classifier to detect 'dusty' versus 'not dusty' 100-by-100 pixel patches from both full-resolution (EDR) and browse-resolution (RDR) images. The dataset is pre-split into train, validation, and test sets, with balanced classes and manual vetting of a validation subset to estimate label noise.

Use Cases

• Training binary image classifiers to detect atmospheric dust obscuration in Martian orbital imagery.
• Evaluating model robustness against label noise, as a subset of validation patches has been manually vetted.
• Studying the visual characteristics of global and regional dust storms on Mars during specific time periods.
• Developing preprocessing filters for HiRISE image pipelines to flag low-quality, dust-obscured observations.

Strengths

• Contains 40,000 total image patches (20,000 from EDR and 20,000 from RDR products) with balanced classes.
• Includes a manually vetted subset of the 5,000 validation patches to enable estimation of label noise rates.
• Splits (10,000 train, 5,000 validation, 5,000 test) are constructed to prevent data leakage from the same HiRISE observation.

Limitations

• Label noise is acknowledged, as patches from mostly dusty/clear images may not perfectly match their class label.
• Key metadata such as exact row count per split, file size, and last update date are not provided across sources.
• The dataset's creation relies on manual inspection for initial image selection, which may limit scalability and introduce subjective bias.

Provenance

Source

HiRISE (High Resolution Imaging Science Experiment) instrument on Mars Reconnaissance Orbiter.

Collection Method

Patches were extracted from manually identified dusty and non-dusty HiRISE RDR nomap browse images acquired during known dust storm periods from Mars years 24–30.

Time Range

Mars years 24–30 (corresponding to Earth years 1999–2011).

Geography

Mars surface, locations corresponding to HiRISE observations during selected time ranges.