CoChat

Image-Based Lunar Hazard Detection in Low Illumination Simulated Conditions via Vision Transformers

Open access
Ghilardi, Luca, Furfaro, Roberto · 2023 · Sensors

The DOI matches its Crossref record; 32 of the 36 references it cites were found in a registry or an index; 2 of the 5 sources its claims cite were checked for support.

Abstract

Hazard detection is fundamental for a safe lunar landing. State-of-the-art autonomous lunar hazard detection relies on 2D image-based and 3D Lidar systems. The lunar south pole is challenging for vision-based methods. The low sun inclination and the terrain rich in topographic features create large areas in shadow, hiding the terrain features. The proposed method utilizes a vision transformer (ViT) model, which is a deep learning architecture based on the transformer blocks used in natural language processing, to solve this problem. Our goal is to train the ViT model to extract terrain features information from low-light RGB images. The results show good performances, especially at high altitudes, beating the UNet, one of the most popular convolutional neural networks, in every scenario.

Overview

CoChat Summary

TL;DR

The paper trains a vision transformer (ViT) for image-based lunar hazard detection under low-illumination simulated conditions and compares it with a UNet. The ViT outperforms the UNet on IoU and precision, especially at higher altitudes, though it is slower and evaluated only on synthetic Blender-rendered data.

Key findings

  • The ViT model outperformed the UNet for both IoU and precision on the test set, while recall values were very similar; the authors conclude the UNet classification is biased toward hazards.
  • Both models performed better at higher altitudes, with ViT performance best between 5000 and 9000 m; on a “cleaned” test set that discarded 4891 of 25,000 images with no information, both models improved mainly at lower altitudes, and the ViT IoU distribution peaked around 80% for 5000–9000 m.
  • Computational comparison: ViT mean inference time was 0.346 s per image (max 0.500 s, min 0.318 s), versus UNet mean 0.012 s (max 0.054 s, min 0.006 s).
  • The ViT had 105 million parameters versus 17 million for the UNet.

Methods

  • The authors used a ViT-Base encoder pre-trained on ImageNet and trained the decoder from scratch for semantic segmentation of safe versus hazardous lunar terrain; training used ADAM with learning rate 1 × 10−5, cross-entropy loss, batch size 5, and 100 epochs.
  • The dataset contains 125,000 Blender-rendered RGB images of the lunar south pole (90 × 90 km area) at 384 × 384 resolution, generated from a NASA digital terrain model with a sun lamp inclination of 1.5 deg and camera altitudes from 500 m to 10 km; labels mark pixels safe if slope is less than 8 deg and roughness less than 10%.
  • The dataset was split 60% training, 20% validation, and 20% testing; the test set contained 25,000 images. Evaluation used precision, recall, and IoU, and the UNet was trained on the same dataset and augmentation with batch size 20 and 100 epochs.

Limitations

  • The paper states that no noise effects were considered, the Blender camera model differs from a real CMOS sensor, and the authors lacked real measurements or accurate lab data to create a realistic photometric model; therefore they cannot claim how the dataset compares with real data.
  • The paper also notes the model does not consider Moon curvature, cannot substitute an active sensor like LiDAR, and is not optimized for real-world deployment.
  • Evident from the text: evaluation is entirely on simulated Blender-rendered images with no real lunar imagery or real-camera validation, and the ViT’s much slower inference time (0.346 s vs. 0.012 s per image) limits practical deployment compared with the UNet.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/s23187844

Cites

32 of 32 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Workflow diagram for the dataset creation and network training and testing. The optical images are simulated starting from a digital terrain model (DTM) of the south pole.
Figure 1 Workflow diagram for the dataset creation and network training and testing. The optical images are simulated starting from a digital terrain model (DTM) of the south pole.CC BY · PubMed Central
Figure 2 The network architecture of ViT [27].
Figure 2 The network architecture of ViT [27].CC BY · PubMed Central
Figure 3 Vision transformer network architecture. This picture is a modified version of the architecture overview presented in [26].
Figure 3 Vision transformer network architecture. This picture is a modified version of the architecture overview presented in [26].CC BY · PubMed Central
Figure 4 Image of the surface 3D model.
Figure 4 Image of the surface 3D model.CC BY · PubMed Central
Figure 5 Samples of the dataset. In the first row, the input images are shown; in the second, the labels.
Figure 5 Samples of the dataset. In the first row, the input images are shown; in the second, the labels.CC BY · PubMed Central
Figure 6 IoU plot on the test set. The transparent area in the IoU plot represents the standard deviation.
Figure 6 IoU plot on the test set. The transparent area in the IoU plot represents the standard deviation.CC BY · PubMed Central
Figure 7 In the sample test dataset, the input images are shown in the first row; in the second, the ground truth; and in the third row, the Vision Transformer model’s predictions. The “white” pixels are labeled as safe, while the “black” ones are labeled as hazardous.
Figure 7 In the sample test dataset, the input images are shown in the first row; in the second, the ground truth; and in the third row, the Vision Transformer model’s predictions. The “white” pixels are labeled as safe, while the “black” ones are labeled as hazardous.CC BY · PubMed Central
Figure 8 IoU plot on the test set “cleaned”. The transparent area in the IoU plot represents the standard deviation.
Figure 8 IoU plot on the test set “cleaned”. The transparent area in the IoU plot represents the standard deviation.CC BY · PubMed Central
Figure 9 On the left, the IoU distribution for the ViT and UNet models. On the right, the IoU distribution for the ViT model on the “cleaned” and “uncleaned” datasets.
Figure 9 On the left, the IoU distribution for the ViT and UNet models. On the right, the IoU distribution for the ViT model on the “cleaned” and “uncleaned” datasets.CC BY · PubMed Central

Verification

Claims the paper makes

2 Valid · 5 Not valid · 2 Not checked · 33 Own results and interpretations

16 more claims

Create a free account to see all 42 claims and to verify the 1 reference still open.

See all claims — free account
  • It may state two separate claims as one. 8
  • Probably not a substantial claim. 15
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 32 references

Research works printed but not linked: 3

  1. Ghilardi, L., D’Ambrosio, A., Scorsoglio, A., Furfaro, R., and Curti, F. (February, January 31). Image-based lunar landing hazard detection via deep learning. Proceedings of the 31st AAS/AIAA Space Flight Mechanics Meeting, Virtual.
  2. Scorsoglio, A., D’Ambrosio, A., Ghilardi, L., Furfaro, R., Gaudet, B., Linares, R., and Curti, F. (2020, January 9–13). Safe Lunar landing via images: A Reinforcement Meta-Learning application to autonomous hazard avoidance and landing. Proceedings of the 2020 AAS/AIAA Astrodynamics Specialist Conference, Virtual.
  3. Boncelet, C. (2009). The Essential Guide to Image Processing, Elsevier.

Cited by

See all 14

Metadata

Authors
Ghilardi, L., & Furfaro, R.
Affiliations
University of Arizona (US)
Corresponding author
Furfaro, Roberto
Venue
Sensors
Publisher
MDPI AG
Published
2023-09-13
Volume
23
Issue
18
Pages
7844
ISSN
1424-8220 (ISSN-L)
OpenAlex
W4386713326
Cited by
14
References
36
Citations by year
2026: 6 · 2025: 7 · 2024: 1
FWCI
1.27
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Robotics and Sensor-Based Localization
  • Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
Keywords
vision transformer, low-light conditions, terrain feature extraction, lunar south pole, autonomous landing

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Image-Based Lunar Hazard Detection in Low Illumination Simulated Conditions via Vision Transformers. https://app.cochat.ai/papers/10.3390/s23187844
The paper
Ghilardi, L., & Furfaro, R. (2023). Image-Based Lunar Hazard Detection in Low Illumination Simulated Conditions via Vision Transformers. Sensors. https://doi.org/10.3390/s23187844

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.