SystemOnline
Craters
CRATER CANDIDATE
22.054°N 85.875°W · NEARSIDE · NORTHERN HEMISPHERE
- DiameterEstimated diameter
- 3.04 km
- ConfidenceModel confidence
- 0.99
- Diameter rankDiameter rank
- Top 19.7%
MOONATLAS derived
The NASA-IBM LFM crater detection model predicts a crater candidate about 3.0 km across with confidence 0.99. It is a model detection, not a reference catalog entry. No Robbins (2019) reference crater matched this detection under the IoU ≥ 0.5 rule; that does not make it a new or previously unknown crater. This is a training sample: the model saw this area during training, so the prediction is not independent evidence of model performance.
Why MOONATLAS selected this
Unmatched with Robbins
No Robbins (2019) reference crater matched under the IoU ≥ 0.5 rule. It does not mean a new or previously unknown crater. Product curation over model outputs, not a scientific ranking.
Dataset split
The model saw this area during training, so this prediction is not independent evidence of performance. The observation and the model output are equally real.

M1481381795CE · 56 predicted boxes · 57 Robbins (2019) reference boxes
A model detection on one analyzed tile, not a catalogued crater and not a unique physical crater: SomBench tiles overlap, so the same crater can be detected on more than one tile.
Model outputs
- CONFIDENCEAI prediction · NASA-IBM LFM
- 0.990
- DIAMETERAI prediction · NASA-IBM LFM
- 3.044 km
- REFERENCE MATCHMOONATLAS derived metric
- No
Provenance
- Model
- NASA-IBM Lunar Foundation Model
- Task
- Crater detection on LROC WAC imagery
- Checkpoint
- WAC_ni_lfm_ps8_lora_s46.ckpt
- Dataset
- sombench-wac-crater-detection
- Split
- Training sample
- Source
- M1481381795CE_r11504_c480.tif
- Processing
- science-0.3.1
- Data origin
- Real model output
Scientific details
- Model id
- lfm-crater-detection
- Model revision
- f09ebeae2099
- Reference
- robbins-2019
- Object id
- wac-m1481381795ce-r11504-c480#15
- Catalogued
- 2026-09-17
Values are model output as produced, or MOONATLAS derived metrics computed from it; reference annotations are shown separately and never mixed into predictions. How MOONATLAS works.