CoChat

The DOI matches its Crossref record; 20 of the 20 references it cites were found in a registry or an index; 4 of the 7 sources its claims cite were checked for support.

Trustworthy Celestial Eye: Calibrated and Robust Planetary Classification via Self-Supervised Vision Transformers

Open access

Abstract

Automated recognition of celestial bodies from observational imagery is a cornerstone of autonomous space exploration. However, deploying deep learning models in space environments entails rigorous requirements not only for accuracy but also for reliability (calibration) and safety (anomaly rejection). Traditional Convolutional Neural Networks (CNNs) trained on small-scale astronomical datasets often suffer from overfitting and overconfidence on Out-of-Distribution (OOD) artifacts. In this work, we present a robust classification framework based on DINOv2, a Vision Transformer pre-trained via discriminative self-supervised learning. We curate a high-fidelity dataset of seven planetary classes sourced from NASA archives and propose a two-stage domain adaptation strategy to transfer large-scale foundation model features to this fine-grained task. Extensive experiments show that our method reaches 100% Top-1 accuracy on the canonical split, and remains highly stable under split variation, achieving 99.43% ± 0.85% Top-1 accuracy across R = 5 repeated stratified splits. More importantly, we address the critical issue of model trustworthiness. Through post hoc temperature scaling, our model achieves a state-of-the-art Expected Calibration Error (ECE) of 0.08%, representing a 36-fold improvement over ResNet50 (2.90%) and a 4.5-fold improvement over the EfficientNet-B3 baseline (0.36%). Furthermore, by integrating Energy-based OOD detection, the system effectively rejects non-planetary artifacts with an AUROC of 93.7%. Qualitative analysis using Grad-CAM reveals that self-supervised attention mechanisms naturally focus on intrinsic planetary features (e.g., surface textures and rings) while ignoring background noise, confirming the superior robustness of vision foundation models in astronomical vision tasks.

Overview

CoChat Summary

TL;DR

The paper proposes a two-stage DINOv2-based framework for classifying seven planetary classes from NASA imagery, adding post hoc temperature scaling and energy-based OOD detection for trustworthiness. It reports 100% Top-1 accuracy on the canonical split, 0.08% ECE after calibration, and 93.7% OOD AUROC, while stress tests identify Gaussian noise as the main robustness bottleneck.

Key findings

  • DINOv2 achieved 100% Top-1 accuracy on the held-out test set, versus 96.2% for ResNet50; across 5 repeated stratified splits, a frozen-feature linear probe reached 99.43% ± 0.85% Acc@1.
  • After temperature scaling, DINOv2 reached 0.08% ECE, compared with 0.36% for EfficientNet-B3 and 2.90% for ResNet50—reported as a 4.5-fold and 36-fold improvement, respectively.
  • Energy-based OOD detection gave DINOv2 an AUROC of 93.7% (Table 1: 93.68%), versus ResNet50 at 86.8% (Table 1: 86.77%) and EfficientNet-B3 at 94.0% (Table 1: 93.96%).
  • Stress tests on N = 105 showed clean accuracy of 100%; Gaussian noise reduced accuracy to 87.62% (σ = 5), 77.14% (σ = 10), and 68.57% (σ = 20), while blur, JPEG compression, downsampling, and cropping mostly stayed within 94.29–100%. RandAugment was critical: weak augmentation averaged 97.4% versus 100.0% with RandAugment.

Methods

  • Dataset: manually curated 700 images across 7 classes (Earth, Jupiter, Mars, Moon, Neptune, Saturn, Uranus), 100 images per class, sourced from NASA archives; split 69:16:15 into 483 training, 112 validation, and 105 test images; OOD set of approximately 200 Google Image Search images; pHash verification found 0 exact or near-duplicates between training and test sets.
  • Model and training: DINOv2 ViT-Base/14 via timm, with a two-stage transfer-learning strategy—Stage 1 linear probing (default 5 epochs) then Stage 2 selective fine-tuning of the last 4 blocks (default 20 epochs); AdamW with lr = 3 × 10⁻⁴ for Stage 1 and 3 × 10⁻⁵ for Stage 2, weight decay 1 × 10⁻⁴, early stopping patience 5, and RandAugment; baselines included ResNet50, EfficientNet-B3, ConvNeXt V2-Tiny, standard ViT-Base, and Swin Transformer, with OpenCLIP ViT-B-16 zero-shot accuracy of 84.76%.
  • Evaluation: Top-1/Top-5 accuracy, macro-F1, balanced accuracy, Wilson and bootstrap confidence intervals, temperature scaling fitted on validation logits, ECE with 15 bins, TTA via flips, energy-based OOD AUROC and TPR at fixed FPR, Grad-CAM visualization, and controlled stress-test perturbations.

Limitations

  • Stated: the curated dataset is small and balanced (700 images, 7 classes), so clean-test accuracy saturates for several modern architectures; the authors note this is not a fully realistic model of long-tailed operational astronomy data.
  • Stated: the OOD set is sourced from web images and does not fully cover mission-specific sensor artifacts such as cosmic-ray hits, saturation spikes, and instrument-dependent noise; reported AUROC is therefore a web-image OOD benchmark rather than a definitive mission-sensor estimate.
  • Stated: DINOv2 has deployment costs—ViT-Base/14 has 86M parameters, about 344 MB in FP32, and at S = 518 uses 1369 tokens versus 256 tokens at S = 224, giving roughly 5.35× more tokens and 28.6× larger quadratic attention cost; no learning-rate scheduler is used, and the small test split (15 images per class) limits stability, with repeated-split frozen-feature OOD AUROC at only 61.50% ± 3.49%.
  • Evident from the text: some reported comparisons are inconsistent—DINOv2’s OOD AUROC is described as 93.7% and “statistically comparable” to EfficientNet, while Table 1 lists EfficientNet-B3 slightly higher at 93.96%—and saturated clean accuracy limits discrimination among modern architectures.

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

Integrity signals

Registry

Crossref · 10.3390/aerospace13030222

Cites

20 of 20 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Two-stage transfer learning schedule (The symbol '*' denotes a placeholder for the model name in the saved weight filenames).
Figure 1 Two-stage transfer learning schedule (The symbol '*' denotes a placeholder for the model name in the saved weight filenames).CC-BY · From the PDF
Figure 2 High-fidelity Celestial Body Dataset from NASA's official mission (class = Earth).
Figure 2 High-fidelity Celestial Body Dataset from NASA's official mission (class = Earth).CC-BY · From the PDF
Figure 3 The samples of the OOD dataset.
Figure 3 The samples of the OOD dataset.CC-BY · From the PDF
Figure 4 Validation accuracy convergence.
Figure 4 Validation accuracy convergence.CC-BY · From the PDF
Figure 5 Reliability diagrams (post-calibration). The plots compare the alignment between confidence and accuracy. ( Left ) ResNet50 shows visible gaps from the diagonal ideal. ( Right ) DINOv2 demonstrates near-perfect alignment (ECE = 0.08%), confirming its superior trustworthiness.
Figure 5 Reliability diagrams (post-calibration). The plots compare the alignment between confidence and accuracy. ( Left ) ResNet50 shows visible gaps from the diagonal ideal. ( Right ) DINOv2 demonstrates near-perfect alignment (ECE = 0.08%), confirming its superior trustworthiness.CC-BY · From the PDF
Figure 6 Energy-based OOD detection distributions. Comparison of energy score distributions for ResNet50 ( Left ) vs. DINOv2 ( Right ). Blue histograms represent in-distribution (planets), and orange histograms represent OOD (artifacts). DINOv2 exhibits a clearer separation between the two peaks, facilitating a safer rejection threshold.
Figure 6 Energy-based OOD detection distributions. Comparison of energy score distributions for ResNet50 ( Left ) vs. DINOv2 ( Right ). Blue histograms represent in-distribution (planets), and orange histograms represent OOD (artifacts). DINOv2 exhibits a clearer separation between the two peaks, facilitating a safer rejection threshold.CC-BY · From the PDF
Figure 7 Grad-CAM visual explanation. Comparison of attention maps for a sample image of Saturn. ( a ) ResNet50: attention is scattered, often focusing on the black background or edges (indicating shortcut learning). ( b ) EfficientNet-B3: attention covers the planet but is diffuse. ( c ) DINOv2: attention is tightly constrained to the planetary disk and rings, demonstrating object-centric semantic understanding.
Figure 7 Grad-CAM visual explanation. Comparison of attention maps for a sample image of Saturn. ( a ) ResNet50: attention is scattered, often focusing on the black background or edges (indicating shortcut learning). ( b ) EfficientNet-B3: attention covers the planet but is diffuse. ( c ) DINOv2: attention is tightly constrained to the planetary disk and rings, demonstrating object-centric semantic understanding.CC-BY · From the PDF

Verification

Claims the paper makes

2 Valid · 1 Not valid · 1 Not checked · 93 Own results and interpretations

48 more claims

Create a free account to see all 97 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 24
  • It may state two separate claims as one. 25
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Citations

Cites

Show all 17 references

Research works printed but not linked: 1

  1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., and Gelly, S. (2021, January 3–7). An image is worth 16×16 words: Transformers for image recognition at scale. Proceedings of the 9th International Conference on Learning Representations (ICLR), Virtual Event. Available online: https://openreview.net/forum?id=YicbFdNTTy.

Metadata

Authors
Xu, Z., Choi, Y., Yi, C., Park, C., Park, J., Park, H., & Song, S.
Affiliations
  • Kyungpook National University (KR)
  • Earth Turbine, 36, Dongdeok-ro 40-gil, Jung-gu, Daegu 41905, Republic of Korea
  • Intelligent Construction Automation Center, Kyungpook National University, 80 Daehak-ro, Buk-gu, Daegu 41566, Republic of Korea
Corresponding author
  • Choi, Young
  • Yi, Changyong
Venue
Aerospace
Publisher
MDPI AG
Published
2026-02-27
Volume
13
Issue
3
Pages
222
ISSN
2226-4310 (ISSN-L)
OpenAlex
W7131846472
Cited by
0
References
17
FWCI
0.00
Type
journal article
Language
en
Open access
Locations
  • Aerospace (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Ministry of Education (RS-2023-00249407)
  • Korea government (RS-2024-00336025, RS-2025-00558871)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
Keywords
planetary classification, self-supervised vision transformers, DINOv2, out-of-distribution detection, temperature scaling, calibration, domain adaptation, foundation models, celestial imagery, Grad-CAM

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: Trustworthy Celestial Eye: Calibrated and Robust Planetary Classification via Self-Supervised Vision Transformers. https://app.cochat.ai/papers/10.3390/aerospace13030222
The paper
Xu, Z., Choi, Y., Yi, C., Park, C., Park, J., Park, H., & Song, S. (2026). Trustworthy Celestial Eye: Calibrated and Robust Planetary Classification via Self-Supervised Vision Transformers. Aerospace. https://doi.org/10.3390/aerospace13030222

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.