F1nnSBK/lunar-pits-dataset

lunar_pits_v3: Lunar Pit & Cave Skylight Benchmark Dataset

lunar_pits_v3 is a benchmark dataset of lunar collapse pits, volcanic skylights, and candidate subsurface void entrances extracted from NASA Lunar Reconnaissance Orbiter Camera (LROC Narrow Angle Camera, NAC) orbital observations. It provides 1.0 m/px pushbroom-projected tiles with native Hugging Face viewer integration, geodesic Haversine regional splits, and a 3-stage pit-free negative sampling protocol.

Technical Specifications

• Spatial Resolution: 1.0 m/px isometric (256 × 256 px tiles, 256 × 256 m ground footprint)
• Geometric Pipeline: NAIF SPICE pushbroom ephemeris ray-tracing with SLDEM2015 topographic relief correction
• Illumination Alignment: Rotated by sub-solar azimuth to vertical top-down illumination (−y → +y)
• Observability Gate: 30° ≤ i ≤ 82°, dynamic range ≥ 0.25, core umbral shadow ≥ 0.05%
• Negative Sampling: 3-stage full-catalog LPA masking (r ≥ max(600 m, 4 × D_pit)) stratified 50/50 across moderate relief and uniform background
• Geodesic Splits: 100 km Haversine clustering (1,311 train, 336 val in King crater, 283 held-out test across Aristarchus, Marius Hills, Ohm, Highland)

Observability Gate Breakdown (Positives)

Filter Stage	Criterion	Excluded Tiles	% of Excluded	Physical Rationale
High-Noon Illumination	i < 30°	289	52.9%	Pit floor fully illuminated, zero shadow
Low Dynamic Range	DR < 0.25	118	21.6%	Washed-out regolith, low contrast
Zero Umbral Shadow	shd < 0.05%	78	14.3%	Absence of morphological shadow dipole
Grazing Terminator	i > 82°	3	0.5%	Severe low-sun clipping, noise floor
Pointing / Frame Error	Flagged / std < 0.005	58	10.6%	Pointing drift or blank calibration frame
Total Excluded Positives	Non-observable	546	69.1%	Excluded from training/evaluation
Total Valid Positives	Observable	244	30.9%	Retained in splits

Geodesic Splits (100 km Haversine Linkage)

Split	Total Tiles	Positives	Negatives	Regions	% of Valid	Host Crater Assignments
train	1,311	173	1,138	29	67.9%	29 regional clusters (including Copernicus)
validation	336	38	298	1	17.4%	King crater (cluster_07)
test	283	33	250	6	14.7%	Aristarchus, Marius Hills, Ohm, Highland
excluded	4,812	546	4,266	41	-	Non-observable strips and negatives

Radiometric Distributions (Mean ± Std | Median [IQR])

Cohort	N	Solar Incidence (i)	Dynamic Range (DR)	Shadow Fraction
Valid Positives	244	47.5° ± 9.9° | 47.5° [14.6°]	0.62 ± 0.21 | 0.64 [0.33]	2.8% ± 6.7% | 0.9% [1.6%]
Negatives (Relief)	886	47.1° ± 10.1° | 46.7° [15.0°]	0.40 ± 0.08 | 0.41 [0.11]	0.0% ± 0.0% | 0.0% [0.0%]
Negatives (Background)	800	46.8° ± 10.2° | 46.7° [15.0°]	0.28 ± 0.12 | 0.27 [0.20]	0.0% ± 0.0% | 0.0% [0.0%]

Parquet Schema

Field	Type	Description
image	Image	Decoded 256 × 256 px 8-bit grayscale image
tile_id	string	Observation and sample identifier
product_id	string	LROC NAC observation ID
pit_id	string	Lunar Pit Atlas identifier
pit_name	string	Feature name
region	string	Regional host name
region_group_id	string	100 km Haversine cluster ID
split	string	Split designation (train, validation, test)
label	int8	Binary target (1 = pit, 0 = negative patch)
neg_tier	string	Sample tier (positive, moderate_relief, uniform_background)
lat, lon	float32	Surface coordinates from Lunar Pit Atlas (degrees)
sample, line	float32	Pushbroom pixel coordinates in raw NAC frame
incidence_deg	float32	Solar incidence angle (degrees)
emission_deg	float32	Spacecraft emission angle (degrees)
phase_deg	float32	Phase angle (degrees)
sub_solar_azimuth_deg	float32	Sub-solar azimuth angle used for top-down alignment
dynamic_range	float32	Radiometric dynamic range (quantile_0.99 − quantile_0.01)
shadow_fraction	float32	Umbral shadow fraction in central core
dist_from_anchor_m	float32	Distance in meters from positive anchor (0.0 m for anchor)

Usage Examples

Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("F1nnSBK/lunar-pits-dataset")
print(dataset)

sample = dataset["test"][0]
print(sample["tile_id"], sample["label"], sample["neg_tier"])
sample["image"].show()

PyTorch LunarPitsV3Dataset

from huggingface_hub import snapshot_download
from torch.utils.data import DataLoader
from hole.data.dataset_v3 import LunarPitsV3Dataset, InStripBatchSampler

dataset_path = snapshot_download("F1nnSBK/lunar-pits-dataset", repo_type="dataset")

test_dataset = LunarPitsV3Dataset(
    dataset_dir=dataset_path,
    meta_filename="metadata_splits.parquet",
    split="test",
    is_valid_only=True,
    augment=False,
)

loader = DataLoader(
    test_dataset,
    batch_sampler=InStripBatchSampler(test_dataset, n_strips_per_batch=4, shuffle=False),
    num_workers=2,
)

batch = next(iter(loader))
print(batch["pixel_values"].shape, batch["label"])

Resources

• Model Checkpoint: F1nnSBK/lunar-dinov3-lora
• Source Code: Project HOLE