CoChat

SegMarsViT: Lightweight Mars Terrain Segmentation Network for Autonomous Driving in Planetary Exploration

Open access
Dai, Yuqi, Zheng, Tie, Xue, Changbin, Zhou, Li · 2022 · Remote Sensing

The DOI matches its Crossref record; 67 of the 67 references it cites were found in a registry or an index; 1 of the 4 sources its claims cite were checked for support.

Abstract

Planetary rover systems need to perform terrain segmentation to identify feasible driving areas and surround obstacles, which falls into the research area of semantic segmentation. Recently, deep learning (DL)-based methods were proposed and achieved great performance for semantic segmentation. However, due to the on-board processor platform’s strict comstraints on computational complexity and power consumption, existing DL approaches are almost impossible to be deployed on satellites under the burden of extensive computation and large model size. To fill this gap, this paper targeted studying effective and efficient Martian terrain segmentation solutions that are suitable for on-board satellites. In this article, we propose a lightweight ViT-based terrain segmentation method, namely, SegMarsViT. In the encoder part, the mobile vision transformer (MViT) block in the backbone extracts local–global spatial and captures multiscale contextual information concurrently. In the decoder part, the cross-scale feature fusion modules (CFF) further integrate hierarchical context information and the compact feature aggregation module (CFA) combines multi-level feature representation. Moreover, we evaluate the proposed method on three public datasets: AI4Mars, MSL-Seg, and S5Mars. Extensive experiments demonstrate that the proposed SegMarsViT was able to achieve 68.4%, 78.22%, and 67.28% mIoU on the AI4Mars-MSL, MSL-Seg, and S5Mars, respectively, under the speed of 69.52 FPS.

Overview

CoChat Summary

TL;DR

SegMarsViT is a lightweight ViT-based encoder-decoder for Martian terrain segmentation, combining a MobileViT backbone with an efficient layer aggregation decoder using cross-scale feature fusion and compact feature aggregation. It reports 68.4% mIoU on AI4Mars-MSL, 78.22% on S5 Mars, and 67.28% on MSL-Seg, at 69.52 FPS.

Key findings

  • On AI4Mars-MSL, SegMarsViT (MobileViT-s) reaches 68.4% mIoU and 92.46% pixelACC with 5.61 M parameters and 8.54 G FLOPs; compared with SegFormer (MIT-B0), it keeps 0.66% better mIoU.
  • On S5 Mars, it reaches 78.22% mIoU, 92.15% pixelACC, 84.74 mFscore, and 85.86 mPrecision at 69.52 FPS with 5.61 M parameters and 8.54 G FLOPs.
  • On MSL-Seg, it reaches 67.28% mIoU, 86.05% pixelACC, 78.69 mFscore, and 78.75 mPrecision at 69.52 FPS; the paper reports improvements of 2.96% pixelACC and 5.17% mIoU over other latest methods.
  • Ablation on MSL-Seg shows adding CFF and CFA improves mIoU and pixelACC by 4.06% and 2.62% over the baseline; ELAD gives 67.28 mIoU at 8.54 G FLOPs versus AMD at 66.69 mIoU/13.14 G FLOPs and UNetD at 66.83 mIoU/12.85 G FLOPs.

Methods

  • The model uses a MobileViT backbone as encoder, with the last pooling and fully connected layers removed, and an efficient layer aggregation decoder (ELAD) containing cross-scale feature fusion (CFF) and compact feature aggregation (CFA) modules; the loss combines weighted IoU loss and weighted BCE loss.
  • Evaluation uses three public datasets: AI4Mars-MSL (17,030 images, 4 classes, 1024 × 1024), MSL-Seg (4155 images, 8 classes, 560 × 500), and S5 Mars (6000 images, 9 classes, 1200 × 1200).
  • Training uses MMSegmentation and PyTorch, 400 epochs, mini-batch size 16 over 4 RTX2080Ti GPUs, SGD with initial LR 1e-2 and polynomial LR policy, random crop to 512 × 512, and augmentation; metrics are mIoU, pixelACC, mFscore, and mPrecision, with comparison against eight SOTA methods.

Limitations

  • Stated: there is an enormous gap between high-end GPU and a low-memory spacecraft device; future work will test on realistic hardware and focus on energy consumption and practical performance.
  • Stated: missed detections and error detections remain, and segmentation accuracy needs further improvement; some MSL-Seg classes have low IoU (e.g., Martian soil and bedrock) because their notions are ambiguous and their pixel counts in training data are low.
  • Evident from the text: evaluation is performed on GPUs (RTX2080Ti for training, NVIDIA 3070Ti for FPS) rather than actual on-board satellite hardware; also, the text reports “8.54 M parameters and 5.6 G FLOPs” while the tables list 5.61 M parameters and 8.54 G FLOPs, indicating a reporting inconsistency.

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

Integrity signals

Registry

Crossref · 10.3390/rs14246297

Cites

67 of 67 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Planetary rover on Mars ( a ) and a sample image ( b ) along with its segmentation a for terrain types ( c ). Figure 1. Planetary rover on Mars ( a ) and a sample image ( b ) along with its segmentation annotation for terrain types ( c ).
Figure 1 Planetary rover on Mars ( a ) and a sample image ( b ) along with its segmentation a for terrain types ( c ). Figure 1. Planetary rover on Mars ( a ) and a sample image ( b ) along with its segmentation annotation for terrain types ( c ).CC-BY · From the PDF
Figure 2 Framework overview of the proposed SegMarsViT. Figure 2. Framework overview of the proposed SegMarsViT.
Figure 2 Framework overview of the proposed SegMarsViT. Figure 2. Framework overview of the proposed SegMarsViT.CC-BY · From the PDF
Figure 3 Two types of building blocks in MobileViT backbone: ( a ) MV2 block; ( b ) MViT block. Figure 3. Two types of building blocks in MobileViT backbone: ( a ) MV2 block; ( b ) MViT block.
Figure 3 Two types of building blocks in MobileViT backbone: ( a ) MV2 block; ( b ) MViT block. Figure 3. Two types of building blocks in MobileViT backbone: ( a ) MV2 block; ( b ) MViT block.CC-BY · From the PDF
Figure 4 Illustration of the proposed Efficient Layer Aggregation Decoder. Figure 4. Illustration of the proposed Efficient Layer Aggregation Decoder.
Figure 4 Illustration of the proposed Efficient Layer Aggregation Decoder. Figure 4. Illustration of the proposed Efficient Layer Aggregation Decoder.CC-BY · From the PDF
Figure 5 Qualitative comparison on S 5 Mars test set. ( a -e ) show five experimental samples. Figure 5. Qualitative comparison on S 5 Mars test set. ( a -e ) show five experimental samples.
Figure 5 Qualitative comparison on S 5 Mars test set. ( a -e ) show five experimental samples. Figure 5. Qualitative comparison on S 5 Mars test set. ( a -e ) show five experimental samples.CC-BY · From the PDF
Image 6
CC-BY · From the PDF
Figure 6 Visualization examples on MSL-Seg test set. ( a -e ) represent five experimental samples. Figure 6. Visualization examples on MSL-Seg test set. ( a -e ) represent five experimental samples.
Figure 6 Visualization examples on MSL-Seg test set. ( a -e ) represent five experimental samples. Figure 6. Visualization examples on MSL-Seg test set. ( a -e ) represent five experimental samples.CC-BY · From the PDF
Figure 8 Illustrating the ELAD architecture of the proposed SegMarsViT. Figure 8. Illustrating the ELAD architecture of the proposed SegMarsViT.
Figure 8 Illustrating the ELAD architecture of the proposed SegMarsViT. Figure 8. Illustrating the ELAD architecture of the proposed SegMarsViT.CC-BY · From the PDF
Image 9
CC-BY · From the PDF

Verification

Claims the paper makes

1 Valid · 1 Not valid · 2 Not checked · 67 Own results and interpretations

36 more claims

Create a free account to see all 71 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 12
  • The quote only partly states this claim. 3
  • Probably not a substantial claim. 18

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.

5 of 5

Citations

Cites

Show all 67 references

Cited by

See all 35

Metadata

Authors
Dai, Y., Zheng, T., Xue, C., & Zhou, L.
Affiliations
Chinese Academy of Sciences (CN)
Corresponding author
Xue, Changbin
Venue
Remote Sensing
Publisher
MDPI AG
Published
2022-12-12
Volume
14
Issue
24
Pages
6297
ISSN
2072-4292 (ISSN-L)
OpenAlex
W4311375979
Cited by
35
References
52
Citations by year
2026: 4 · 2025: 15 · 2024: 11 · 2023: 5
FWCI
3.91
Type
journal article
Language
en
Open access
Locations
  • Remote Sensing (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Chinese Academy of Sciences Project (CXJJ-20S017)
Topics
  • Physical Sciences › Earth and Planetary Sciences › Space and Planetary Science › Planetary Science and Exploration
  • 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
Keywords
lightweight vision transformer, cross-scale feature fusion, Mars 2020 Perseverance rover

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 (gpt-6-sol).

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: SegMarsViT: Lightweight Mars Terrain Segmentation Network for Autonomous Driving in Planetary Exploration. https://app.cochat.ai/papers/10.3390/rs14246297
The paper
Dai, Y., Zheng, T., Xue, C., & Zhou, L. (2022). SegMarsViT: Lightweight Mars Terrain Segmentation Network for Autonomous Driving in Planetary Exploration. Remote Sensing. https://doi.org/10.3390/rs14246297

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.