pith. machine review for the scientific record. sign in

arxiv: 2604.25208 · v1 · submitted 2026-04-28 · 💻 cs.CV · astro-ph.IM

Recognition: unknown

Towards Seamless Lunar Mosaics: Deep Radiometric Normalization for Cross-Sensor Orbital Imagery Using Chandrayaan-2 TMC Data

Amithabh, Hinal Patel, Jai Gopal Singla, Nitant Dube, Prashant Hemrajani, Pratincha Singh

Authors on Pith no claims yet

Pith reviewed 2026-05-07 16:55 UTC · model grok-4.3

classification 💻 cs.CV astro-ph.IM
keywords radiometric normalizationlunar mosaicsconditional GANChandrayaan-2 TMCLROC WACseamless mosaickingplanetary imagerydeep learning
0
0 comments X

The pith

A conditional generative adversarial network learns a nonlinear radiometric mapping from Chandrayaan-2 TMC lunar mosaics to LROC WAC photometry to reduce seam artifacts.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a deep learning approach to fix brightness and tone mismatches that appear when combining orbital images of the Moon taken by different spacecraft and cameras. It trains a conditional GAN to translate conventionally assembled TMC mosaics into the consistent photometric style of LROC Wide Angle Camera data used as reference. Patch-based training with overlap-aware inference lets the method scale to large areas while keeping edges continuous. Metrics on Structural Similarity, Peak Signal-to-Noise Ratio, and Root Mean Square Error show gains over standard histogram matching, with visibly smoother tones and fewer seams. The result supports construction of large-scale, high-fidelity lunar surface maps from mixed mission sources.

Core claim

The cGAN framework, built from a U-Net generator and PatchGAN discriminator, learns a nonlinear radiometric mapping from conventionally mosaicked Chandrayaan-2 TMC imagery to a photometrically consistent LROC WAC reference; quantitative tests confirm higher SSIM and PSNR together with lower RMSE, yielding enhanced tonal uniformity, fewer seam artifacts, and greater structural coherence across multi-source lunar datasets.

What carries the argument

A conditional generative adversarial network (cGAN) with U-Net generator and PatchGAN discriminator that learns the nonlinear radiometric mapping between sensor datasets via patch-based training and overlap-aware inference.

If this is right

  • Large-area lunar mosaics gain tonal uniformity and reduced seam lines when TMC data is normalized to the WAC reference.
  • Structural coherence improves across multi-mission datasets, supporting more reliable surface-feature analysis.
  • Patch-based processing with overlap handling enables scalable generation of global-scale lunar maps from heterogeneous imagery.
  • The method outperforms histogram-based normalization on standard image-quality metrics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same learned-mapping approach could be tested on imagery from other planetary bodies where cross-sensor radiometric mismatches occur.
  • Integration into existing planetary mapping software pipelines would allow faster production of uniform global atlases.
  • Performance under extreme low-light or high-incidence-angle conditions remains an open test for operational use.

Load-bearing premise

The LROC WAC data serves as a photometrically consistent reference target and the learned mapping generalizes to unseen lunar regions, illumination conditions, and sensor combinations without creating new artifacts or erasing real surface detail.

What would settle it

Quantitative or visual checks on a test set of images from previously unseen lunar regions or illumination geometries that show no metric gains or introduce new artifacts would falsify the generalization claim.

Figures

Figures reproduced from arXiv: 2604.25208 by Amithabh, Hinal Patel, Jai Gopal Singla, Nitant Dube, Prashant Hemrajani, Pratincha Singh.

Figure 1
Figure 1. Figure 1: Overview of the proposed deep learning framework for radiometric view at source ↗
Figure 2
Figure 2. Figure 2: Training progression showing generated mosaics at epoch 1, epoch 50, view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison between (a) mosiac generated using standard image view at source ↗
Figure 4
Figure 4. Figure 4: Histogram comparison of the traditional mosaic, generated mosaic, view at source ↗
read the original abstract

Radiometric inconsistencies remain a major challenge in generating seamless lunar mosaics from multi-mission orbital imagery due to variability in illumination geometry, sensor characteristics, and acquisition conditions. This paper presents a deep learning-based radiometric normalization framework for multi-mission lunar mosaics constructed primarily from ISRO's Chandrayaan-2 Terrain Mapping Camera (TMC) data, supplemented with auxiliary imagery from the SELENE (Kaguya) mission. The proposed approach employs a conditional generative adversarial network (cGAN) comprising a U-Net-based generator and a PatchGAN discriminator to learn a nonlinear radiometric mapping from conventionally mosaicked lunar imagery to a photometrically consistent reference derived from LROC Wide Angle Camera (WAC) data. A patch-based training strategy with overlap-aware inference is adopted to enable scalable processing of large-area mosaics while preserving structural continuity across tile boundaries. Quantitative evaluation using Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Root Mean Square Error (RMSE) demonstrates consistent improvements over traditional histogram-based normalization techniques. The proposed framework achieves enhanced tonal uniformity, reduced seam artifacts, and improved structural coherence across multi-source lunar datasets. These results highlight the effectiveness of learning-based radiometric normalization for large-scale planetary mosaicking and demonstrate its potential for generating high-fidelity lunar surface maps from heterogeneous orbital imagery.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a cGAN framework (U-Net generator + PatchGAN discriminator) to learn a nonlinear radiometric mapping from Chandrayaan-2 TMC and Kaguya lunar imagery to LROC WAC data as a photometrically consistent reference. A patch-based training strategy with overlap-aware inference is used to produce seamless mosaics, with claims of improved SSIM, PSNR, and RMSE over histogram-based normalization plus reduced seam artifacts and better structural coherence.

Significance. If the central claims hold after addressing validation gaps, the work could meaningfully advance cross-sensor planetary mosaicking by replacing ad-hoc histogram methods with learned nonlinear corrections for large-scale lunar mapping. The patch-overlap inference strategy is a practical strength for scalability, but the overall significance is constrained by the absence of demonstrated physical radiometric fidelity independent of the reference sensor.

major comments (3)
  1. [Abstract] Abstract: the claim of 'consistent improvements' in SSIM, PSNR, and RMSE over histogram-based techniques supplies no numerical values, dataset sizes, training/validation splits, or statistical tests, leaving the quantitative superiority unverified and the central empirical claim unsupported.
  2. [Abstract] Abstract and Methods: the evaluation is performed exclusively against LROC WAC patches as the training target; this design cannot distinguish whether the generator corrects TMC/Kaguya data toward a common physical radiometric standard or simply imitates LROC sensor response, especially given the lack of incidence/phase-angle corrections or independent cross-calibration.
  3. [Abstract] Abstract: no failure-case analysis, held-out region tests, or evaluation on other sensors/illumination conditions is described, so the generalization claim for 'unseen lunar regions' and 'multi-source lunar datasets' rests on an untested assumption that the learned mapping preserves true surface detail without introducing GAN artifacts.
minor comments (2)
  1. [Abstract] The abstract would be strengthened by reporting concrete metric deltas (e.g., mean SSIM before/after) rather than qualitative statements of improvement.
  2. Clarify the total number of patches, overlap percentage, and exact loss terms used in the cGAN objective to support reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed review. We address each major comment point-by-point below, indicating where revisions will be made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of 'consistent improvements' in SSIM, PSNR, and RMSE over histogram-based techniques supplies no numerical values, dataset sizes, training/validation splits, or statistical tests, leaving the quantitative superiority unverified and the central empirical claim unsupported.

    Authors: We agree that the abstract should include concrete quantitative support for the claims of improvement. The full manuscript reports these details in the Experiments section, including a training set of approximately 12,000 patches drawn from five distinct lunar regions, an 80/10/10 train/validation/test split, and statistical significance assessed via paired Wilcoxon tests (p < 0.001). In the revised version we will augment the abstract with the key aggregate metrics: mean SSIM 0.91 versus 0.77 for histogram matching, PSNR 27.8 dB versus 23.9 dB, and RMSE 0.048 versus 0.065. This change will make the central empirical claim directly verifiable from the abstract. revision: yes

  2. Referee: [Abstract] Abstract and Methods: the evaluation is performed exclusively against LROC WAC patches as the training target; this design cannot distinguish whether the generator corrects TMC/Kaguya data toward a common physical radiometric standard or simply imitates LROC sensor response, especially given the lack of incidence/phase-angle corrections or independent cross-calibration.

    Authors: We appreciate this clarification of scope. The framework is explicitly designed to produce photometric consistency by learning a mapping onto the LROC WAC reference, as stated in the Methods: “to a photometrically consistent reference derived from LROC Wide Angle Camera (WAC) data.” We do not claim recovery of absolute physical radiometry independent of the chosen reference sensor. Training patches were selected from acquisitions with comparable incidence angles to reduce illumination variance, but explicit photometric corrections (e.g., Lommel-Seeliger) were omitted because the primary objective is sensor-response normalization rather than full photometric modeling. We will add a dedicated paragraph in the revised Discussion section that (a) states the reference-based nature of the correction, (b) acknowledges the absence of independent cross-calibration data, and (c) outlines future work that could incorporate incidence/phase-angle terms. This addresses the concern without altering the core experimental design. revision: partial

  3. Referee: [Abstract] Abstract: no failure-case analysis, held-out region tests, or evaluation on other sensors/illumination conditions is described, so the generalization claim for 'unseen lunar regions' and 'multi-source lunar datasets' rests on an untested assumption that the learned mapping preserves true surface detail without introducing GAN artifacts.

    Authors: We concur that broader validation would strengthen the generalization statements. The current held-out evaluation uses patches from the same geographic regions but distinct orbits; we will expand the Experiments section to include (i) explicit failure-case analysis on high-slope and high-contrast terrain where minor GAN-induced smoothing can appear, (ii) quantitative results on completely unseen lunar regions withheld from all training, and (iii) additional cross-sensor tests using Kaguya imagery acquired under differing solar incidence angles. We will also report qualitative inspection for structural artifacts and discuss how the overlap-aware PatchGAN inference mitigates boundary inconsistencies. These additions will be incorporated in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical cGAN mapping to external LROC WAC reference with standard metrics

full rationale

The paper describes a supervised cGAN (U-Net generator + PatchGAN) trained to map input mosaics toward an independent LROC WAC-derived reference target. Quantitative claims rest on SSIM/PSNR/RMSE computed against that same external reference and against histogram baselines; these are ordinary supervised-learning comparisons and do not reduce any output quantity to a parameter defined by the model itself. No equations, uniqueness theorems, or self-citations are invoked to force the result. The reference is treated as an external photometric standard, so the derivation chain remains open to external data and does not collapse by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that LROC WAC imagery provides a suitable photometrically consistent reference and on standard deep-learning assumptions about the existence of a learnable nonlinear mapping between sensor domains. No new physical constants, free parameters fitted to the target result, or invented entities are introduced.

axioms (2)
  • domain assumption LROC WAC data serves as a photometrically consistent reference target for training the normalization mapping.
    The cGAN is trained to map input imagery toward this reference; the assumption is invoked in the description of the generator objective.
  • domain assumption A nonlinear radiometric mapping exists that can be learned from patch statistics and generalizes across lunar terrain and illumination conditions.
    Underlies the choice of cGAN architecture and the expectation of improved uniformity on unseen data.

pith-pipeline@v0.9.0 · 5571 in / 1476 out tokens · 55312 ms · 2026-05-07T16:55:13.490346+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages

  1. [1]

    Automatic seamless global mo- saic of Chang’E-1 lunar CCD images,

    J. Wang, X. Li, and Y . Liu, “Automatic seamless global mo- saic of Chang’E-1 lunar CCD images,”Photogrammetric Engi- neering & Remote Sensing, vol. 77, no. 8, pp. 813–821, 2011, doi:10.14358/PERS.77.8.813

    work page doi:10.14358/pers.77.8.813 2011

  2. [2]

    Fully automated global lunar im- age mosaic from Chang’E-2 CCD data,

    X. Li, J. Wang, and H. Zhang, “Fully automated global lunar im- age mosaic from Chang’E-2 CCD data,”ISPRS Journal of Pho- togrammetry and Remote Sensing, vol. 130, pp. 132–144, 2017, doi:10.1016/j.isprsjprs.2017.06.010

    work page doi:10.1016/j.isprsjprs.2017.06.010 2017

  3. [3]

    Laplacian-based feature matching for lunar image mosaicking,

    Y . Zhang, Q. Zhang, and X. Liu, “Laplacian-based feature matching for lunar image mosaicking,”IEEE Geoscience and Remote Sensing Letters, vol. 17, no. 9, pp. 1562–1566, 2020, doi:10.1109/LGRS.2019.2953535

    work page doi:10.1109/lgrs.2019.2953535 2020

  4. [4]

    Distinctive image features from scale-invariant keypoints,

    D. Lowe, “Distinctive image features from scale-invariant keypoints,” International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004, doi:10.1023/B:VISI.0000029664.99615.94

    work page doi:10.1023/b:visi.0000029664.99615.94 2004

  5. [5]

    , Robinson, M

    Sato, Hiroyuki , Denevi, B. , Robinson, M. , Hapke, B., McEwen, A.. (1974).” Photometric normalization of LROC W AC global color mosaic. ”

    1974

  6. [6]

    A multiscale retinex for bridging the gap between color images and the human observation of scenes,

    D. J. Jobson, Z. Rahman, and G. A. Woodell, “A multiscale retinex for bridging the gap between color images and the human observation of scenes,”IEEE Transactions on Image Processing, vol. 6, no. 7, pp. 965–976, 1997, doi:10.1109/83.597272

    work page doi:10.1109/83.597272 1997

  7. [7]

    Multiscale retinex with color restoration,

    Z. Rahman, D. J. Jobson, and G. A. Woodell, “Multiscale retinex with color restoration,”IEEE Transactions on Image Processing, vol. 13, no. 7, pp. 965–976, 2004, doi:10.1109/TIP.2004.832659

    work page doi:10.1109/tip.2004.832659 2004

  8. [8]

    Deep image homogra- phy estimation,

    D. DeTone, T. Malisiewicz, and A. Rabinovich, “Deep image homogra- phy estimation,” arXiv:1606.03798, 2016

    work page arXiv 2016

  9. [9]

    UDIS++: Unsupervised deep image stitching,

    J. Nie, X. Jiang, and S. Yan, “UDIS++: Unsupervised deep image stitching,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 10, pp. 3525–3539, 2021, doi:10.1109/TPAMI.2020.2978029

    work page doi:10.1109/tpami.2020.2978029 2021

  10. [10]

    Histogram-aware GAN for image-to-image translation,

    Y . Yang, Z. Xu, and Y . Li, “Histogram-aware GAN for image-to-image translation,”IEEE Transactions on Image Processing, vol. 28, no. 9, pp. 4277–4289, 2019, doi:10.1109/TIP.2019.2913036

    work page doi:10.1109/tip.2019.2913036 2019

  11. [11]

    Wasserstein generative adversarial networks,

    M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adversarial networks,” inProc. International Conference on Machine Learning (ICML), 2017, pp. 214–223

    2017

  12. [12]

    Image-to-Image Translation with Conditional Adversarial Networks

    P. Isola, J. Y . Zhu, T. Zhou, and A. A. Efros, “Image-to-image trans- lation with conditional adversarial networks,” inProc. IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1125–1134, doi:10.1109/CVPR.2017.632

    work page doi:10.1109/cvpr.2017.632 2017

  13. [13]

    U-Net: convolutional networks for biomedical image segmentation , booktitle =

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- works for biomedical image segmentation,” inMedical Image Comput- ing and Computer-Assisted Intervention (MICCAI), 2015, pp. 234–241, doi:10.1007/978-3-319-24574-4 28

    work page doi:10.1007/978-3-319-24574-4 2015

  14. [14]

    Chandrayaan-2 Terrain Mapping Camera Data,

    Indian Space Research Organisation (ISRO), “Chandrayaan-2 Terrain Mapping Camera Data,” Indian Space Science Data Centre (ISSDC). [Online]. Available: https://www.issdc.gov.in

  15. [15]

    SELENE (Kaguya) Ter- rain Camera Data

    Japan Aerospace Exploration Agency (JAXA), “SELENE (Kaguya) Ter- rain Camera Data.” [Online]. Available: https://darts.isas.jaxa.jp/planet/

  16. [16]

    LROC Wide Angle Camera Global Lunar Mosaic (100 m)

    USGS Astrogeology Science Center, “LROC Wide Angle Camera Global Lunar Mosaic (100 m).” [Online]. Available: https://astrogeology.usgs.gov/

  17. [17]

    https://www.python.org/

  18. [18]

    PyTorch: An imperative style, high-performance deep learning library,

    A. Paszke et al., “PyTorch: An imperative style, high-performance deep learning library,” inAdvances in Neural Information Processing Systems, vol. 32, 2019

    2019