pith. machine review for the scientific record. sign in

arxiv: 2605.08179 · v1 · submitted 2026-05-05 · 📡 eess.SP · astro-ph.IM· cs.LG

Recognition: no theorem link

Neural Posterior Estimation of Terrain Parameters from Radar Sounder Data

Jordy Dal Corso , Annalena Kofler , Marco Cortellazzi , Lorenzo Bruzzone , Bernhard Sch\"olkopf
Authors on Pith no claims yet

Pith reviewed 2026-05-12 01:32 UTC · model grok-4.3

classification 📡 eess.SP astro-ph.IMcs.LG
keywords neural posterior estimationradar sounderterrain parameter inversionsimulation-based inferenceMars radar datauncertainty quantificationsubsurface analysis
0
0 comments X

The pith

A neural density estimator trained on simulated radar observations recovers full posterior distributions over terrain parameters from real Mars sounder data.

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

Radar sounders probe subsurface structure by recording echoes of transmitted electromagnetic waves, yet conventional analysis relies on simplifying assumptions and returns only single-point estimates that ignore noise correlations and parameter trade-offs. The paper replaces those methods with simulation-based inference: a GPU simulator generates large numbers of synthetic observations that train a neural network to output the complete probability distribution over terrain parameters such as permittivity and roughness. The framework conditions the estimator on chosen reference surface values so that the effect of those assumptions on the resulting posteriors can be quantified directly. A reader should care because the approach supplies calibrated uncertainties and explicit robustness checks, enabling more trustworthy maps of subsurface properties when the same trained model is applied to actual planetary radar profiles.

Core claim

Training a neural posterior estimator on synthetic radar sounder data produced by a GPU-based physical simulator yields well-calibrated posterior distributions over terrain parameters; these distributions remain valid when the estimator is applied to real Mars radar profiles once it is conditioned on literature-informed reference surface assumptions.

What carries the argument

Conditional neural density estimator for amortized posterior inference, trained via simulation-based inference to map observed radar echoes to full distributions over terrain parameters while explicitly conditioning on reference surface assumptions.

If this is right

  • The method supplies full posterior distributions that quantify uncertainties and correlations among terrain parameters instead of isolated point estimates.
  • Explicit conditioning on reference surface assumptions permits systematic testing of how posterior inferences change with different surface priors.
  • Once trained, the estimator can be applied directly to real radar profiles without requiring additional real-world training data.
  • Galactic and instrumental noise are folded into the inferred distributions rather than treated as separate post-processing corrections.

Where Pith is reading between the lines

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

  • The same simulation-trained estimator could be retrained for other remote-sensing modalities whenever a sufficiently accurate forward simulator exists.
  • Full posteriors could be propagated forward into higher-level models of subsurface layering or ice detection to produce uncertainty-aware planetary maps.
  • Improvements in simulator fidelity would directly tighten the posteriors obtained on future mission data without changing the inference architecture.

Load-bearing premise

The GPU-based simulator must reproduce the physics, noise statistics, and echo characteristics of real radar sounder instruments closely enough that posteriors learned from its outputs remain valid on actual Mars observations.

What would settle it

If the posterior distributions produced by the trained model on real Mars radar profiles are shown to be inconsistent with independent geological or in-situ measurements of the same terrain parameters, the claim of simulator-to-reality transfer would be refuted.

Figures

Figures reproduced from arXiv: 2605.08179 by Annalena Kofler, Bernhard Sch\"olkopf, Jordy Dal Corso, Lorenzo Bruzzone, Marco Cortellazzi.

Figure 1
Figure 1. Figure 1: Overview of the NPE framework: During data generation, prior samples are passed through the simulator to extract peak [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SHARAD Radargram 186301. We highlight in light pink and blue the CP and ZP areas, respectively. On the right, we [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: Since NPE allows rapid inference without additional [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Inferred posterior distributions for CP and ZP under varying [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
read the original abstract

Radar sounders are electromagnetic instruments that can probe deep into the subsurface of Earth and other planetary bodies by processing the echo of transmitted radar waves. Conventional approaches for analyzing such data rely on approximate assumptions and often produce point estimates that ignore parameter correlations as well as galactic and measurement noise. We propose a simulation-based inference approach to terrain parameter inversion from radar sounder data, where synthetic observations from a GPU-based simulator are used to train a neural network-based density estimator for neural posterior estimation (NPE). By explicitly conditioning on reference surface assumptions, the proposed framework allows systematic evaluation of posterior robustness to reference surface variability. We demonstrate that our NPE model is well calibrated on simulated data and transferable to real Mars radar profiles, where we analyze terrain parameters using literature-informed reference values.

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

2 major / 2 minor

Summary. The manuscript proposes a simulation-based inference method using neural posterior estimation (NPE) to recover terrain parameters from radar sounder data. Synthetic echoes generated by a GPU-based forward simulator train a neural density estimator; the resulting posteriors are asserted to be well-calibrated on simulated data and directly applicable to real Mars SHARAD/MARSIS profiles when conditioned on literature-derived reference surfaces.

Significance. If the simulator faithfully reproduces instrument noise, volume scattering, and echo statistics, the framework would represent a meaningful advance over conventional point-estimate inversions by supplying full posterior distributions, explicit uncertainty quantification, and systematic robustness checks to reference-surface assumptions. Such an approach could improve subsurface characterization in planetary radar sounding.

major comments (2)
  1. [Abstract] Abstract: the central transferability claim—that posteriors trained on GPU-simulated data remain valid for real Mars profiles—rests on the untested assumption that the simulator reproduces the actual echo statistics, noise, and volume scattering of SHARAD/MARSIS. No quantitative cross-validation (e.g., posterior predictive checks, comparison of simulated vs. observed echo histograms, or agreement with independent dielectric/roughness estimates from other sensors) is reported to support this load-bearing step.
  2. [Methods/Results] The manuscript provides no ablation studies, training details, or calibration diagnostics (coverage probabilities, posterior sharpness, or simulation-based calibration plots) that would allow evaluation of the NPE calibration claim on simulated data.
minor comments (2)
  1. Define all acronyms (NPE, SHARAD, MARSIS, etc.) at first use and ensure consistent notation for terrain parameters throughout.
  2. Clarify the precise parameterization of the reference surface (dielectric constant, roughness, layering) and how it is conditioned in the NPE network.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for their insightful comments on our manuscript. Below, we provide detailed responses to the major comments and indicate the revisions made to address them.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central transferability claim—that posteriors trained on GPU-simulated data remain valid for real Mars profiles—rests on the untested assumption that the simulator reproduces the actual echo statistics, noise, and volume scattering of SHARAD/MARSIS. No quantitative cross-validation (e.g., posterior predictive checks, comparison of simulated vs. observed echo histograms, or agreement with independent dielectric/roughness estimates from other sensors) is reported to support this load-bearing step.

    Authors: We acknowledge that the manuscript does not report explicit quantitative cross-validation between the simulator outputs and real SHARAD/MARSIS echo statistics. The transferability demonstration in the original text relies on obtaining terrain parameter posteriors from real profiles that are consistent with literature-derived expectations for Mars. To directly address this concern, the revised manuscript now includes posterior predictive checks (sampling parameters from the inferred posterior and comparing simulated echo distributions to observed real-data histograms), as well as quantitative comparisons of key statistics and discussion of agreement with independent dielectric and roughness estimates from other sensors where available. We also clarify the simulator assumptions regarding noise and volume scattering. revision: yes

  2. Referee: [Methods/Results] The manuscript provides no ablation studies, training details, or calibration diagnostics (coverage probabilities, posterior sharpness, or simulation-based calibration plots) that would allow evaluation of the NPE calibration claim on simulated data.

    Authors: The original manuscript states that the NPE is well calibrated on simulated data and provides some training information, but we agree that the level of detail and diagnostics is insufficient for full evaluation. In the revision, we have expanded the Methods section with complete training hyperparameters, network architecture, and optimization details. We have added ablation studies examining the impact of network choices and conditioning variables. Calibration diagnostics have been augmented with coverage probability plots, posterior sharpness metrics, and simulation-based calibration (SBC) plots in the Results section to allow readers to assess calibration on simulated data. revision: yes

Circularity Check

0 steps flagged

No circularity: simulator-generated training data and real-data application remain independent

full rationale

The derivation relies on an external GPU simulator to produce synthetic observations for training the NPE density estimator, followed by separate calibration on held-out simulations and application to real Mars profiles using literature reference values. No equation or step reduces a reported posterior, calibration metric, or transfer result to a fitted parameter or self-citation by construction; the forward model and real observations are treated as distinct inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the unverified fidelity of the GPU simulator and on the assumption that literature reference surface values are appropriate conditioning inputs.

axioms (1)
  • domain assumption The GPU-based simulator accurately models real radar sounder physics, noise, and echo formation
    Invoked when training data are generated and when claiming transfer to real Mars profiles.

pith-pipeline@v0.9.0 · 5443 in / 1287 out tokens · 57116 ms · 2026-05-12T01:32:23.704071+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

36 extracted references · 36 canonical work pages · 2 internal anchors

  1. [1]

    The lunar radar sounder (lrs) onboard the kaguya (selene) spacecraft,

    T. Ono, A. Kumamoto, Y . Kasahara, Y . Yamaguchi, A. Yamaji, T. Kobayashi, S. Oshigami, H. Nakagawa, Y . Goto, K. Hashimoto et al., “The lunar radar sounder (lrs) onboard the kaguya (selene) spacecraft,”Space science reviews, vol. 154, pp. 145–192, 2010

    work page 2010

  2. [2]

    Lunar radar sounder observations of subsurface layers under the nearside maria of the moon,

    T. Ono, A. Kumamoto, H. Nakagawa, Y . Yamaguchi, S. Os- higami, A. Yamaji, T. Kobayashi, Y . Kasahara, and H. Oya, “Lunar radar sounder observations of subsurface layers under the nearside maria of the moon,”Science, vol. 323, no. 5916, pp. 909–912, 2009

    work page 2009

  3. [3]

    Science results from sixteen years of mro sharad operations,

    N. E. Putzig, R. Seu, G. A. Morgan, I. B. Smith, B. A. Campbell, M. R. Perry, M. Mastrogiuseppeet al., “Science results from sixteen years of mro sharad operations,”Icarus, vol. 419, p. 115715, 2024

    work page 2024

  4. [4]

    Quantitative analysis of mars surface radar reflectivity at 20mhz,

    C. Grima, W. Kofman, A. Herique, R. Orosei, and R. Seu, “Quantitative analysis of mars surface radar reflectivity at 20mhz,”Icarus, vol. 220, no. 1, pp. 84–99, 2012. [Online]. Available: https://www.sciencedirect.com/science/ article/pii/S0019103512001558

    work page 2012

  5. [5]

    Rime: Radar for icy moon exploration,

    L. Bruzzone, J. J. Plaut, G. Alberti, D. D. Blankenship, F. Bovolo, B. A. Campbell, A. Ferro, Y . Gim, W. Kofman, G. Komatsu et al., “Rime: Radar for icy moon exploration,” in2013 IEEE international geoscience and remote sensing symposium-IGARSS. IEEE, 2013, pp. 3907–3910

    work page 2013

  6. [6]

    Envision mission to venus: Subsurface radar sounding,

    L. Bruzzone, F. Bovolo, S. Thakur, L. Carrer, E. Donini, C. Gerekos, S. Paterna, M. Santoni, and E. Sbalchiero, “Envision mission to venus: Subsurface radar sounding,” inIGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2020, pp. 5960–5963

    work page 2020

  7. [7]

    Radar evidence for ice in lobate debris aprons in the mid-northern latitudes of mars,

    J. J. Plaut, A. Safaeinili, J. W. Holt, R. J. Phillips, J. W. Head III, R. Seu, N. E. Putzig, and A. Frigeri, “Radar evidence for ice in lobate debris aprons in the mid-northern latitudes of mars,” Geophysical research letters, vol. 36, no. 2, 2009

    work page 2009

  8. [8]

    Dielectric map of the martian northern hemisphere and the nature of plain filling materials,

    J. Mouginot, A. Pommerol, P. Beck, W. Kofman, and S. M. Clifford, “Dielectric map of the martian northern hemisphere and the nature of plain filling materials,”Geophysical research letters, vol. 39, no. 2, 2012

    work page 2012

  9. [9]

    A coherent multilayer simulator of radargrams acquired by radar sounder instruments,

    C. Gerekos, A. Tamponi, L. Carrer, D. Castelletti, M. Santoni, and L. Bruzzone, “A coherent multilayer simulator of radargrams acquired by radar sounder instruments,”IEEE Transactions on Geoscience and Remote Sensing, vol. 56, no. 12, pp. 7388–7404, 2018

    work page 2018

  10. [10]

    Radar sounding using the cassini altimeter: Waveform modeling and monte carlo approach for data inversion of observations of titan’s seas,

    M. Mastrogiuseppe, A. Hayes, V . Poggiali, R. Seu, J. I. Lunine, and J. Hofgartner, “Radar sounding using the cassini altimeter: Waveform modeling and monte carlo approach for data inversion of observations of titan’s seas,”IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 10, pp. 5646–5656, 2016

    work page 2016

  11. [11]

    A dictionary-based integrated simulation approach to model large- and small-scale coherent surface scattering phenomena in radar sounder data,

    E. Sbalchiero, M. Cortellazzi, S. Thakur, and L. Bruzzone, “A dictionary-based integrated simulation approach to model large- and small-scale coherent surface scattering phenomena in radar sounder data,”IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–19, 2023

    work page 2023

  12. [12]

    The frontier of simulation-based inference,

    K. Cranmer, J. Brehmer, and G. Louppe, “The frontier of simulation-based inference,”Proceedings of the National Academy of Sciences, vol. 117, no. 48, pp. 30 055–30 062, 2020

    work page 2020

  13. [13]

    Normalizing flows for probabilistic modeling and inference,

    G. Papamakarios, E. Nalisnick, D. J. Rezende, S. Mohamed, and B. Lakshminarayanan, “Normalizing flows for probabilistic modeling and inference,”Journal of Machine Learning Research, vol. 22, no. 57, pp. 1–64, 2021

    work page 2021

  14. [14]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inThe Eleventh International Conference on Learning Representations,

    work page

  15. [15]

    Available: https://openreview.net/forum?id= XVjTT1nw5z

    [Online]. Available: https://openreview.net/forum?id= XVjTT1nw5z

    work page

  16. [16]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inThe Eleventh International Conference on Learning Representations,

    work page

  17. [17]

    Available: https://openreview.net/forum?id= PqvMRDCJT9t

    [Online]. Available: https://openreview.net/forum?id= PqvMRDCJT9t

    work page

  18. [18]

    Building Normalizing Flows with Stochastic Interpolants

    M. S. Albergo and E. Vanden-Eijnden, “Building normalizing flows with stochastic interpolants,” inThe Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://arxiv.org/abs/2209.15571

    work page internal anchor Pith review arXiv 2023

  19. [19]

    Stochastic interpolants: A unifying framework for flows and diffusions,

    M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden, “Stochastic interpolants: A unifying framework for flows and diffusions,”

    work page

  20. [20]

    Stochastic Interpolants: A Unifying Framework for Flows and Diffusions

    [Online]. Available: https://arxiv.org/abs/2303.08797

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    Flow matching for scalable simulation-based inference,

    J. Wildberger, M. Dax, S. Buchholz, S. R. Green, J. H. Macke, and B. Sch ¨olkopf, “Flow matching for scalable simulation-based inference,”NeurIPS 2023, dec 2023

    work page 2023

  22. [22]

    Simulation-based inference: A practical guide,

    M. Deistler, J. Boelts, P. Steinbach, G. Moss, T. Moreau, M. Gloeckler, P. L. C. Rodrigues, J. Linhart, J. K. Lappalainen, B. K. Miller, P. J. Gon c ¸alves, J.-M. Lueckmann, C. Schr ¨oder, and J. H. Macke, “Simulation-based inference: A practical guide,”

    work page

  23. [23]

    Deistler, J

    [Online]. Available: https://arxiv.org/abs/2508.12939

    work page arXiv

  24. [24]

    Long and F

    D. Long and F. Ulaby,Microwave radar and radiometric remote sensing. Artech, 2015

    work page 2015

  25. [25]

    Radio science computing low-frequency radar surface echoes for planetary radar using huygens-fresnel’s principle,

    Y . Berquin, A. Herique, W. Kofman, and E. Heggy, “Radio science computing low-frequency radar surface echoes for planetary radar using huygens-fresnel’s principle,”Radio Science, vol. 50, 09 2015

    work page 2015

  26. [26]

    Cal- ibration of low-frequency radio telescopes using the galactic background radiation,

    G. Dulk, W. Erickson, R. Manning, and J.-L. Bougeret, “Cal- ibration of low-frequency radio telescopes using the galactic background radiation,”Astronomy & Astrophysics, vol. 365, no. 2, pp. 294–300, 2001

    work page 2001

  27. [27]

    Global permittivity mapping of the martian surface from sharad,

    L. Castaldo, D. M `ege, J. Gurgurewicz, R. Orosei, and G. Alberti, “Global permittivity mapping of the martian surface from sharad,” Earth and Planetary Science Letters, vol. 462, pp. 55–65, 2017

    work page 2017

  28. [28]

    Geometric power fall-off in radar sounding,

    M. S. Haynes, E. Chapin, and D. M. Schroeder, “Geometric power fall-off in radar sounding,”IEEE Transactions on Geo- science and Remote Sensing, vol. 56, no. 11, pp. 6571–6585, 2018

    work page 2018

  29. [29]

    Surface and subsurface radar equations for radar sounders,

    M. S. Haynes, “Surface and subsurface radar equations for radar sounders,”Annals of Glaciology, vol. 61, no. 81, pp. 135–142, 2020

    work page 2020

  30. [30]

    Permittivity estimation over mars by using sharad data: the cerberus palus area,

    G. Alberti, L. Castaldo, R. Orosei, A. Frigeri, and G. Cirillo, “Permittivity estimation over mars by using sharad data: the cerberus palus area,”Journal of Geophysical Research: Planets, vol. 117, no. E9, 2012. [Online]. Available: https://agupubs. onlinelibrary.wiley.com/doi/abs/10.1029/2012JE004047

    work page doi:10.1029/2012je004047 2012

  31. [31]

    Neural spline flows,

    C. Durkan, A. Bekasov, I. Murray, and G. Papamakarios, “Neural spline flows,”Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  32. [32]

    sbi reloaded: a toolkit for simulation-based inference workflows,

    J. Boelts, M. Deistler, M. Gloeckler, ´Alvaro Tejero-Cantero, J.- M. Lueckmann, G. Moss, P. Steinbach, T. Moreau, F. Muratore, J. Linhart, C. Durkan, J. Vetter, B. K. Miller, M. Herold, A. Ziaeemehr, M. Pals, T. Gruner, S. Bischoff, N. Krouglova, R. Gao, J. K. Lappalainen, B. Mucs ´anyi, F. Pei, A. Schulz, Z. Stefanidi, P. Rodrigues, C. Schr ¨oder, F. A. ...

    work page

  33. [33]

    Available: https://doi.org/10.21105/joss.07754

    [Online]. Available: https://doi.org/10.21105/joss.07754

    work page doi:10.21105/joss.07754

  34. [34]

    Validation of software for bayesian models using posterior quantiles,

    S. R. Cook, A. Gelman, and D. B. Rubin, “Validation of software for bayesian models using posterior quantiles,” Journal of Computational and Graphical Statistics, vol. 15, no. 3, pp. 675–692, 2006. [Online]. Available: https: //doi.org/10.1198/106186006X136976

    work page doi:10.1198/106186006x136976 2006

  35. [35]

    Scuola Medica Salernitana,

    S. Talts, M. Betancourt, D. Simpson, A. Vehtari, and A. Gelman, “Validating bayesian inference algorithms with simulation-based calibration,”arXiv preprint arXiv:1804.06788, 2018

    work page arXiv 2018

  36. [36]

    Revisiting Classifier Two-Sample Tests

    D. Lopez-Paz and M. Oquab, “Revisiting classifier two-sample tests,”arXiv preprint arXiv:1610.06545, 2016

    work page Pith review arXiv 2016