nasa-ibm-ai4science/Sombench-NAC-Crater-Detection

SOMBench Benchmark: Hand-Labeled Crater Detection in NAC data (nac_craters_dataset)

Science theme: Impact cratering

Task: Single-class object detection (COCO bounding boxes)

Dataset Summary

A crater-detection benchmark built from expert hand-labeled crater outlines on LROC NAC orthophotos across six lunar NAC PHO sites (Apollo 15 SIVB impact, Apollo 17, Highland Photom, King Ejecta, March 17 Impact, Reiner Gamma). Subject matter expert (SME) annotators exhaustively labeled craters inside fixed 1024 × 1024 m study boxes; each box is tiled into 256 × 256 px chips at the native image resolution and paired with COCO-format bounding boxes. Two annotation passes per box were merged (IoU > 0.5 de-duplication). The set is split into train / validation / test at the box level, so overlapping tiles never cross splits.

Two label completeness regimes exist (label_set field per image): standard boxes are labeled at all crater diameters; large boxes contain only craters ≥ 10 m in diameter.

Dataset Structure

Directory layout

nac_craters_dataset/
├── README.md
├── images/                        # 408 .npy tiles (uint8, 256×256)
├── annotations_train.json         # per-split COCO files (all craters)
├── annotations_val.json
├── annotations_test.json
├── annotations_min5px_train.json  # per-split, craters ≥ 5 px only
├── annotations_min5px_val.json
├── annotations_min5px_test.json
├── annotations.json               # combined (split field per image)
├── annotations_min5px.json        # combined, craters ≥ 5 px
├── splits.json                    # box → split mapping
├── crater_catalog.parquet         # one row per crater (diam m/px, source tiff, split)
└── previews/                      # rendered sample grids with boxes

Tiles are named by provenance:

{SITE}__{label_set}_box{n}__r{row}_c{col}.npy

Images are single-band uint8 NumPy arrays holding the exact NAC orthophoto DN values (no stretching, no format conversion). Each COCO image entry carries site, label_set, box_no, source_image, native_resolution_m, effective_resolution_m, nodata, split, and a geotransform mapping tile pixels back to the lunar projected coordinates of the source image. Each annotation carries bbox (xywh, px), diam_m, diam_px, truncated (crater cut by a tile edge), and src_file (originating shapefile).

Contents

Per-split annotation counts

Known Limitations

• —Labels exist only inside study boxes. Tiles never extend past a box boundary, but detections should not be evaluated outside the tiled area.
• —Sub-pixel labels at coarse resolution: the median crater is ~6.5 px across; 27 % of annotations are below 5 px (44 % below 6 px). The *_min5px_* files remove annotations under 5 px; the full files keep everything.
• —`large` boxes are complete only for craters ≥ 10 m — mixing label sets in evaluation without filtering will inflate apparent false positives on large tiles.
• —Two boxes were labeled on NAC frames absent from the source delivery: one large-crater box (Highland Photom) is excluded, and one (Apollo 15 SIVB) uses a co-registered substitute frame.
• —One source shapefile (King Ejecta box 2) contains corrupt geometry; 1,763 of its 1,764 craters were recovered.
• —Craters larger than a tile are clipped to tile bounds and flagged truncated.

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.