nasa-ibm-ai4science/Sombench-Ice-Prospectivity-Regression

SomBench Benchmark: Polar Ice Prospectivity Regression

Science theme: Polar volatiles

Task: Regression

Dataset Summary

A polar, multi-layer benchmark for predicting near-surface water-ice prospectivity within ~10° latitude of each pole at 240 m/pixel. Following the ice-prospectivity workflow of Coyan et al. (2025), the dataset includes a group of physically motivated evidential layers (thermophysical, illumination, and terrain) alongside a continuous prospectivity target. Each polar patch is stored as a set of individual GeoTIFF layers, for both the north (80N) and south (80S) poles, with train/val/test split lists.

Task framing: regression of the prospectivity value (0–1) from the evidential layers.

Dataset Structure

Directory layout

prospectivity_dataset/
├── README.md                                         # data source overview
├── band_statistics.json                              # per-layer normalization statistics
├── train_filtered.txt                                # train split patch lists
├── val_filtered.txt                                  # val split patch lists
├── test_filtered.txt                                 # test split patch lists
└── patch_{RRRR}_{CCCC}[_S]_{80N|80S}_{LAYER}.tif     # patch files

• Patch naming: patch_{row}_{col}_{pole}_{LAYER}.tif, where the pole tag is 80N for the north pole and S_…_80S for the south pole. Each patch index therefore has one GeoTIFF per layer.
• band_statistics.json: per-layer statistics for normalization.
• train_filtered.txt / val_filtered.txt / test_filtered.txt: patch-id lists defining the splits.

Layers (one GeoTIFF per layer, per patch)

Suffix	Layer	Group
DICE	Diviner ice stability depth	Thermophysical
TMAX	Maximum surface temperature	Thermophysical
LPSR	Permanently shadowed regions (PSR) mask	Illumination
LPSR_DIS	Distance to nearest PSR	Context
LPSR_DEN	PSR density (local areal density)	Context
SLOPE	Slope	Terrain
CUR	Curvature	Terrain
ASP_SIN_COS	Aspect (sine + cosine)	Terrain
ONLYCOS	Aspect (cosine-only variant)	Terrain
PRO	Ice prospectivity, regression target/label	Target

These correspond to the nine evidential layers described in the SOMBench paper (ice stability depth, maximum temperature, PSRs, slope, curvature, aspect, distance to PSRs, and PSR density) plus the prospectivity target. Aspect is provided in two versions with different input channels (ASP_SIN_COS, ONLYCOS); choose the input channels appropriate to your model.

Contents

Item	Value
Resolution	240 m/pixel
Extent	within ~10° latitude of each pole (north 80N, south 80S)
Layer format	single-band GeoTIFF, one file per layer per patch
Files	1,782 GeoTIFFs ≈ 162 patches × 11 layers (both poles)
Splits	Train 108 · Val 24 · Test 25 patches (162 total), listed in train_filtered.txt / val_filtered.txt / test_filtered.txt
Regression target	PRO (continuous prospectivity, 0–1)

Known Limitations

• Prospectivity is a model-derived, tunable product: weights encode assumed relationships between ice concentration and the evidential layers and may evolve as in-situ constraints improve.
• Extreme polar illumination and sparse ground truth make evaluation sensitive to metric choice and label assumptions.

Citation

@article{fraccaro2026lfm,
  title  = {Multimodal-Multiresolution Foundation Model for Lunar Remote Sensing},
  author = {Fraccaro, Paolo and Nyirjesy, Gabby and Szwarcman, Daniela and Patil, Himanshu
            and Gaur, Vishal and Lal, Rohit and Slank, Rachel A. and Dawson, Geoffrey
            and Debary, Hiyam and Dionelis, Nikolaos and Barker, Michael K. and Annex, Andrew
            and Viswanathan, Vishnu and Morse, Zachary and Schaefer, Ethan I. and Kumar, Ankur
            and Watson, Campbell D. and Dawson-Rigas, Rebekah I. and Maskey, Manil
            and Roy, Sujit and Ramachandran, Rahul and Bernab\'e-Moreno, Juan},
  year   = {2026}
  howpublished = {\url{https://huggingface.co/collections/nasa-ibm-ai4science/nasa-ibm-lunar-fm-and-downstream-models}}
}

@misc{sombench2026collection,
  author = {Patil, Himanshu and Nyirjesy, Gabby and Slank, Rachel A. and Gaur, Vishal
          and Szwarcman, Daniela and Fraccaro, Paolo and Dionelis, Nikolaos and Barker, Michael K.
          and Annex, Andrew and Viswanathan, Vishnu and Morse, Zachary and Schaefer, Ethan I.
          and Debary, Hiyam and Kumar, Ankur and Lal, Rohit and Dawson, Geoffrey
          and Watson, Campbell and Dawson-Rigas, Rebekah I. and Maskey, Manil
          and Bernab\'e-Moreno, Juan and Ramachandran, Rahul and Roy, Sujit},
  title        = {{SomBench}: Benchmark Dataset for Advancing Machine Learning in Lunar Science},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/collections/nasa-ibm-ai4science/lunar-fm-ml-ready-benchmark-dataset-sombench}}
}

License

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.