Mirali33/MOMO-pretraining-data

MOMO Pre-training Dataset

This dataset contains the pre-training data used to train MOMO (Mars Orbital Model), the first multi-sensor foundation model for Mars remote sensing, presented at CVPR 2026.

📄 Paper: MOMO: Mars Orbital Model Foundation Model for Mars Orbital Applications

💻 Code: github.com/kerner-lab/MOMO

🤗 Model Checkpoints: huggingface.co/Mirali33/MOMO

🏆 Mars-Bench (Downstream tasks): mars-bench.github.io

MOMO trains separate models on HiRISE, CTX, and THEMIS data and merges them into a single foundation model capable of diverse Mars orbital tasks.

---

Dataset Overview

The dataset consists of grayscale image crops (.tif, non-georeferenced) collected from three Mars orbital instruments, organized by their corresponding USGS Geologic Map of Mars (GMoM) unit. The data spans a wide range of Martian surface types, resolutions, and acquisition conditions.

Instrument	Resolution	Coverage	# Samples
HiRISE	0.25, 0.5, 0.696, 1 m/pixel	~4.5% of Mars surface	~16M (4M used)
CTX	5 m/pixel	>99.5% of Mars surface	~10M (4M used)
THEMIS	100 m/pixel	Global	~4M

---

Data Collection

• HiRISE: Grayscale crops from the RED band of map-projected Reduced Data Record (RDR) products, acquired from November 2006 through May 2025 (Primary and Extended Science Phases). Source: HiRISE PDS RDR Archive
• CTX: Crops from the Murray Lab open-source CTX global mosaic (updated March 2023), rendered at 5.0 m/pixel. Source: Murray Lab CTX Mosaic
• THEMIS: Crops from Projected Brightness Temperature (PBT) products (ODTGEO_v2) from the THEMIS archive, collected from October 2002 to April 2025. Source: THEMIS PBT Archive

For details on data curation, geographic distribution, quality filtering, and stratified sampling strategy, please refer to Section 4 and Appendix A.1 of the paper.

Illustrative samples from the three Mars orbital instruments — HiRISE, CTX, and THEMIS.

---

Repository Structure

MOMO-pretraining-data/
├── HiRISE/
│   ├── Aa.tar
│   ├── AHi.tar
│   └── ... (one .tar per GMoM unit)
├── CTX/
│   ├── Aa.tar
│   └── ...
├── THEMIS/
│   ├── Aa.tar
│   └── ...
└── metadata/
    ├── HiRISE_metadata.csv
    ├── CTX_metadata.csv
    └── THEMIS_metadata.csv

Each .tar file corresponds to one GMoM (Geologic Map of Mars) unit and contains .tif image crops for that unit. The metadata/ folder contains per-image statistics and acquisition information for all samples across all three instruments.

---

Metadata

Each instrument has an associated metadata CSV with per-image statistics and acquisition information. The columns are described below.

Column Descriptions

Spatial & Identification

Column	HiRISE	CTX	THEMIS	Description
Filename	✅	✅	✅	Unique filename for each image crop
Center Latitude	✅	✅	✅	Latitude of the image center in decimal degrees
Center Longitude	✅	✅	✅	Longitude of the image center in decimal degrees
Top Left Latitude	✅	✅	✅	Latitude of the top-left corner of the image crop
Top Left Longitude	✅	✅	✅	Longitude of the top-left corner of the image crop
Top Right Latitude	✅	✅	✅	Latitude of the top-right corner of the image crop
Top Right Longitude	✅	✅	✅	Longitude of the top-right corner of the image crop
Bottom Left Latitude	✅	✅	✅	Latitude of the bottom-left corner of the image crop
Bottom Left Longitude	✅	✅	✅	Longitude of the bottom-left corner of the image crop
Bottom Right Latitude	✅	✅	✅	Latitude of the bottom-right corner of the image crop
Bottom Right Longitude	✅	✅	✅	Longitude of the bottom-right corner of the image crop
GMoM_Unit_acronym	✅	✅	✅	Acronym of the USGS Geologic Map of Mars unit the crop falls within (e.g., lHl, mNh)
which_instrument	✅	✅	✅	Source instrument: HiRISE, CTX, or THEMIS

Basic Image Statistics

Column	HiRISE	CTX	THEMIS	Description
mean	✅	✅	✅	Per-image mean pixel intensity over the normalized image (values in [0, 1])
std	✅	✅	✅	Per-image standard deviation of pixel intensities over the normalized image
black_percentage	✅	✅	✅	Fraction of pixels with zero intensity; used to detect no-data or border regions

Image Quality Metrics

All quality metrics below are computed on the grayscale, min-max normalized version of each image (scaled to [0, 255] then normalized to [0, 1] for metric computation).

Column	HiRISE	CTX	THEMIS	Description
SSIM	✅	✅	✅	Structural Similarity Index between the original image and a Gaussian-blurred version (kernel 5×5, σ=1.5). Measures how much structural content is preserved relative to a smoothed reference. Lower values indicate noisy or artifact-heavy images.
PSNR	✅	✅	✅	Peak Signal-to-Noise Ratio between the original and Gaussian-blurred image (in dB). Higher values indicate cleaner images.
SNR	✅	✅	✅	Signal-to-Noise Ratio, computed as the ratio of mean pixel intensity to standard deviation. Measures the relative strength of the signal versus background variation.
Noise Estimate	✅	✅	✅	Noise level estimated by applying a Laplacian-like high-pass filter and computing σ = (sum of absolute filtered values) × √(0.5π) / (6 × (W−2) × (H−2)). Lower values indicate a cleaner image.
Blur	✅	✅	✅	Variance of the Laplacian of the image. Higher values indicate sharper images; lower values indicate blurry images.
Edge Density	✅	✅	✅	Fraction of pixels detected as edges by the Canny edge detector (thresholds 100 and 200). Higher values indicate images with richer spatial structure.
Contrast	✅	✅	✅	Standard deviation of normalized pixel intensities. Higher values indicate greater tonal variation within the image.
High Frequency Ratio	✅	✅	✅	Fraction of FFT magnitude energy outside a central low-frequency disk (radius = min(H,W)/8). Higher values indicate greater high-frequency content, which can signal noise or fine texture detail.
FFT Ratio	✅	✅	✅	Ratio of horizontal to vertical frequency band energy in the FFT magnitude spectrum (±5 pixel bands through the center). Values significantly above 1.0 indicate horizontal striping or stretching artifacts.
Gradient Ratio	✅	✅	✅	Ratio of the variance of horizontal Sobel gradients to vertical Sobel gradients. Values significantly above 1.0 suggest dominant horizontal structure, which may indicate scan-line artifacts.
Flagged	✅	✅	✅	Binary flag (1/0) indicating whether the image was detected as a known satellite processing artifact — specifically, images with float32 dtype and a median pixel value of exactly 91.43115.

Acquisition Metadata

Column	HiRISE	CTX	THEMIS	Description
GSD	✅	✅	✅	Ground Sampling Distance in meters per pixel, derived from MAP_SCALE in the PDS label
Acquisition Date	✅	❌	✅	UTC datetime of the observation start, derived from START_TIME in the PDS label
Start Time	✅	❌	✅	UTC datetime of the observation stop, derived from STOP_TIME in the PDS label
Incidence Angle	✅	❌	❌	Angle between the incoming solar ray and the surface normal at acquisition time (degrees), from VIEWING_PARAMETERS in the HiRISE PDS label
Emission Angle	✅	❌	❌	Angle between the surface normal and the direction toward the sensor at acquisition time (degrees), from VIEWING_PARAMETERS in the HiRISE PDS label
Phase Angle	✅	❌	❌	Angle between the incoming solar ray and the direction toward the sensor (degrees); affects image brightness and shadowing. From VIEWING_PARAMETERS in the HiRISE PDS label
Local Time	✅	❌	❌	Local solar time at the image center at acquisition (in units of local day / 24), from VIEWING_PARAMETERS in the HiRISE PDS label
Solar Longitude	✅	❌	❌	Solar longitude (Ls) in degrees, representing the Martian season at acquisition (0° = northern spring equinox). From VIEWING_PARAMETERS in the HiRISE PDS label

THEMIS-specific

Column	Description
Minimum Brightness Temperature	Minimum projected brightness temperature (Kelvin) recorded in the image, from MINIMUM_BRIGHTNESS_TEMPERATURE in the PDS label
Maximum Brightness Temperature	Maximum projected brightness temperature (Kelvin) recorded in the image, from MAXIMUM_BRIGHTNESS_TEMPERATURE in the PDS label
Band Number	THEMIS IR spectral band number used for this image crop (e.g., band 9 = 12.57 µm), from BAND_NUMBER in the PDS label

Note on CTX metadata: CTX data in this dataset is sourced from the Murray Lab global mosaic rather than individual PDS RDR products. As a result, per-image acquisition metadata (viewing angles, local time, solar longitude, acquisition date) is not available for CTX. Only image quality metrics, spatial coordinates, and GSD are provided.

---

Data Filtering

To filter out low-quality images, two primary metrics are recommended, as used in the MOMO paper:

Examples of low-quality (top) and high-quality (bottom) samples from HiRISE, CTX, and THEMIS.

• SSIM (Structural Similarity Index): measures structural content preservation relative to a Gaussian-smoothed version of the image.
• Noise Estimate: measures the absolute noise level via a Laplacian-based high-pass filter.

Both metrics range from 0 to 1, where lower values indicate poorer image quality. In MOMO, samples with either metric below 0.4 were discarded, a threshold determined through human verification. This step effectively removes images with satellite artifacts, excessive noise, or blur.

You may also consider additional metadata columns for custom filtering — for example, black_percentage to remove crops with significant no-data regions, FFT Ratio and Gradient Ratio to filter horizontal scan-line artifacts, Flagged to remove known processing artifacts, or acquisition metadata such as Incidence Angle to restrict to specific illumination conditions.

---

Citation

If you use this dataset, please cite:

@InProceedings{Purohit_2026_CVPR,
    author    = {Purohit, Mirali and Gajera, Bimal and Mehta, Irish and Tokas, Bhanu and Adler, Jacob and Lu, Steven and Dickenshied, Scott and Diniega, Serina and Bue, Brian and Rebbapragada, Umaa and Kerner, Hannah},
    title     = {MOMO: Mars Orbital MOdel Foundation Model for Mars Orbital Applications},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {27772-27782}
}