Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Multi-view 3D surface reconstruction from SAR images by inverse rendering

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that a surface-based differentiable SAR renderer, trained with a coarse-to-fine MLP, reconstructs digital elevation models from as few as two to five intensity-only SAR images while separating geometry from backscatter.

desk verdict A genuine new renderer for SAR surface reconstruction, with a plausible synthetic validation but an unproven geometry-texture separation claim. read the letter →

arxiv 2502.10492 v1 pith:YEXNBQKB submitted 2025-02-14 cs.CV eess.SP

classification cs.CVeess.SP
keywords SyntheticApertureRadarinverserenderingdigitalelevationmodelneuralradiancefieldsdifferentiablemulti-viewreconstructionbackscatteringcoarse-to-fineoptimization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that SAR intensity images alone can be inverted into a digital elevation model by a differentiable surface renderer, without interferometric phase, coherent baselines, or volume rendering. The method models a scene as a single-valued elevation map plus a backscattering map, renders expected SAR intensities by summing over rectangular terrain patches, and fits an MLP with a coarse-to-fine training strategy to a few views. On synthetic images from a physically based simulator, the recovered elevation has RMSE 36.7 m on a 75 m resolution scene from five views, 52.9 m from two views, and 4.72 m on a 0.75 m resolution island with multiple materials from five views. The authors take these results as showing that geometry and texture can be separated and that multi-view SAR disparity is a viable cue for surface reconstruction.

What carries the argument

The central object is the differentiable surface-based SAR rendering model of Eq. (2): $\hat{I}_m = \delta_a \sum_{k=0}^{K-1} \sigma_k \hat{v}_k \hat{w}_{m,k} l_k$, which replaces volume sampling by sums over rectangular patches of a piecewise-linear heightmap. Three ingredients make it differentiable: a logistic-sigmoid shadow recursion (Eq. 3), a smooth maximum function for range-cell overlaps (Eqs. 5-6), and a Lambertian cosine normalized radar cross-section (Eq. 7). A coarse-to-fine schedule anneals the number of patches K and the smoothness parameter $\mu$, and frequency-band windowing on the positional encoding limits high-frequency overfitting early in training, so the same renderer can output full azimuth rows efficiently.

What would settle it

Run the pipeline on SAR images of a scene containing forest canopy, urban layover, or vertical cliffs, and compare the recovered elevation against lidar ground truth; if the reconstruction tracks the scattering phase center or smears vertical structures far beyond the reported RMSE, the single-surface Lambertian assumption fails.

Watch

Extended reading notes

Core claim

The central result is that SAR image formation can be discretized over a terrain surface rather than over a volume, and that this surface-only discretization is differentiable enough to drive inverse rendering. Expected intensity in a range cell is written as a sum over piecewise-linear patches of the heightmap, with a backscattering coefficient, an iteratively computed shadow indicatrix, and an analytic range-cell overlap fraction; backscattering itself is a Lambertian cosine term. Using this renderer, the paper fits a multilayer perceptron that outputs elevation and backscattering maps, annealing the patch count and smoothness in a coarse-to-fine schedule to avoid local minima and speed training. The reported RMSE values on two simulated scenes are below the SAR pixel resolution, which the authors interpret as evidence that geometry is reconstructed, not copied from texture.

Load-bearing premise

The assumption that carries the method is that each SAR pixel is produced by direct backscattering from exactly one point of a single-valued surface, with no volume scattering, ground penetration, layover, or multipath.

Editorial extensions

If this is right

  • Multi-view SAR surface reconstruction becomes possible from intensity data alone, removing the need for coherent phase or interferometric acquisition constraints.
  • Surface-only rendering cuts the sampling dimension from three (volume) to two (surface), so coarse-to-fine training needs fewer samples and produces noise-free synthetic images.
  • Geometry and backscatter separate during optimization, meaning texture changes across views are absorbed into the learned appearance map rather than corrupting the recovered elevation.
  • Reconstruction degrades gracefully with view count: two ascending/descending views already yield an elevation map finer than the SAR pixel resolution, with RMSE 52.9 m on 75 m cells.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct extension, not tested in the paper, would be applying the same renderer to real multi-angle SAR stacks to see whether the synthetic-level accuracy survives incidence-dependent scattering and calibration errors.
  • The surface-only model could be extended to mixed scenes by adding a volumetric term for vegetation or snow layers, keeping the patch-based efficiency for bare terrain.
  • Because the backscattering map is learned without conditioning on incidence angle, conditioning the appearance network on local incidence is a natural next step and would likely improve geometry/texture separation on steep slopes.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a surface-based differentiable SAR rendering model for multi-view 3D reconstruction from a few intensity SAR images. The scene is represented as an elevation map Z(x,y) and a backscattering map B(x,y); the renderer discretizes the surface into patches, computes differentiable shadows and range-cell contributions, and synthesizes intensity via Eqs. (1)–(7). A multi-layer perceptron with Instant-NGP encoding is optimized with a coarse-to-fine sampling schedule and a negative log-likelihood loss on log-intensities plus total-variation regularization on B (Eq. (8)). Experiments on synthetic EMPRISE data report DSM RMSEs of 36.7 m on the Forez scene with 5 views, 52.9 m with 2 views, and 4.72 m on the Island scene with 5 views (Table I). The conclusion claims successful retrieval of a neural elevation model and correct separation of geometry from texture.

Significance. If the central claims hold, this is a useful contribution to SAR 3D reconstruction: it replaces costly volume rendering with a surface rasterization adapted to heightmaps, which is a principled and potentially more efficient direction than SAR-specific NeRF adaptations. The evaluation against EMPRISE, a physically based simulator independent of the proposed renderer, is a genuine strength and avoids the self-inversion problem present in some prior work. The main weaknesses are that the geometry–texture separation claim is not quantitatively demonstrated, the physical scope is limited to single-valued surfaces without volume scattering or layover, and the experiments lack baselines and error bars.

major comments (3)
  1. [V, Table I, Eqs. (7)–(8)] The central claim that the method 'correctly separates geometry from texture' is not supported by the quantitative evaluation: Table I reports only DSM RMSE, while the recovered backscattering map B is assessed qualitatively in Section IV and Figures 3–4. Because Eq. (7) multiplies B by |u·n| and Eq. (8) penalizes only the intensity mismatch between rendered and observed images, an erroneous elevation field can be partially compensated by a modified B; the reported elevation RMSE does not rule out such compensation. I request a quantitative validation of the recovered B, for example a comparison with the known mangia-model material map on the Island scene or an ablation with B fixed to a reference map, reported with an explicit metric such as RMSE or structural similarity on B.
  2. [II-A, Eq. (1)] The image formation model explicitly assumes negligible vegetation and ground penetration, a single-valued surface Z(x,y), and direct backscattering only. These assumptions are not tested: the three experiments are synthetic terrain scenes without layover, multipath, volume scattering, or vertical structures, so the paper provides no evidence on how the method behaves when Eq. (1) is misspecified. I ask for an explicit statement of the validity domain of the model and at least one controlled experiment with layover or volume scattering, or a physical argument with a concrete test that the single-surface model remains unbiased in such cases.
  3. [IV, Table I] The evaluation is limited to three synthetic scenarios and contains no comparison with existing reconstruction methods such as SAR-NeRF [13], Radar Fields [14], or classical radargrammetry, and no error bars or repeated runs. Since Section V claims advantages over volume rendering in sample efficiency and reconstruction quality, the authors should provide at least one head-to-head comparison or ablation on the same EMPRISE scenes, together with a measure of variability across training runs; without this, the numerical values in Table I are difficult to interpret.
minor comments (5)
  1. [Table I] The header 'RSME' should read 'RMSE'; please also state the units of the reported errors and the exact evaluation area (the text says 'restricted to emerged lands and to the areas of overlap of at least 2 images', but this is not indicated in the table).
  2. [III-B] The coarse-to-fine schedule is not fully specified: the initial subsampling factor β0, target samples per line K_f, target smooth-maximum accuracy μ_f, and shadow sigmoid steepness ξ are free parameters, but only λ=0.1 and M>80 are given. Please report the values used and the exact frequency-windowing rule mapping |X_{k+1}-X_{k-1}|/2 to a maximum frequency level l.
  3. [II-B, Eq. (8)] The log-intensity loss is taken from the authors' previous SAR2SAR paper [19]; since this loss is load-bearing for training on speckle, please state the assumed speckle distribution and derive the loss, or provide a reference that gives the derivation, rather than citing only a despeckling application.
  4. [Abstract and Introduction] The abstract and introduction describe the method as working on 'unconstrained SAR images', while the experiments use stripmap acquisitions with known geometry and no phase information. Please clarify what degree of acquisition freedom is assumed (for example, arbitrary orbit geometries and incidence angles, or only different look directions).
  5. [Figures 3 and 4] The color scales for the DSM and scattering-coefficient maps are not specified, which makes the claimed visual agreement difficult to assess; consider adding colorbars and side-by-side residual maps.

Circularity Check

0 steps flagged · score 2.0 of 10

Not circular: the DSM reconstruction is tested against EMPRISE/SRTM ground truth external to the method's own renderer; the only self-citations (SAR2SAR loss, EMPRISE simulator) are not load-bearing circular inputs.

full rationale

The derivation chain is not circular. The method fits an MLP to predict elevation Z and backscattering B by minimizing the speckle log-likelihood (Eq. 8) under the forward model sigma0 = B * |u.n| (Eq. 7). The reported elevation errors are computed against DSM ground truth that is not an input of the forward model: the Forez scene uses SRTM elevations, and the Island scene uses EMPRISE simulations generated with the mangia backscattering model, which the paper states is 'more complex than our cosine hypothesis.' The proposed renderer is therefore not inverted against itself; EMPRISE is an independent, physically based simulator (co-developed by a co-author, but not built from Eq. 7). The log-intensity loss cited to the authors' SAR2SAR paper [19] is a standard speckle negative log-likelihood and does not encode the target DSM, so citing it is a minor self-citation rather than a circular step. The only substantive weakness is that 'correctly separate geometry from texture' is supported only visually: because B multiplies the incidence term in Eq. 7, a mis-estimated Z could in principle be partially absorbed by B without changing the rendered intensity much (Eq. 8). That is an identifiability/validation gap, not a circular reduction—the DSM RMSE is still measured against external ground truth, and the B map is not presented as a fitted version of the reference. No uniqueness theorem, ansatz, or known result is imported from the authors' prior work to force the choice of representation.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central reconstruction is supported by an external synthetic benchmark, but the physical model rests on several simplifying assumptions about surface scattering and image formation. No new physical entities are introduced. The only fitted quantities are the MLP weights and the manually chosen hyperparameters.

free parameters (6)
  • lambda (TV regularization weight) = 0.1
    Chosen by hand to balance geometry fitting and backscattering regularization; Section III-A. Not fitted to data, but affects the optimization.
  • beta_0 (initial subsampling factor) = not reported
    Defines the coarse-to-fine start in Section III-B; not specified numerically.
  • K_f (target samples per line) = not reported
    Controls the final resolution of the surface discretization; Section III-B.
  • mu_f (target smooth-maximum accuracy) = not reported
    Controls range-cell assignment sharpness; Section III-B.
  • xi (shadow sigmoid steepness) = not reported
    Sets the differentiability and sharpness of shadow boundaries in Eq. 3.
  • M (minibatch lines) = greater than 80
    Number of rows rendered per minibatch; only a lower bound is given, exact value not specified.
assumptions (6)
  • domain assumption Scene is described by a single-valued elevation map Z(x,y) and backscattering map B(x,y) with no volume scattering or ground penetration.
    Section II-A opening: 'Consider a stripmap acquisition in a radar band with negligible vegetation and ground penetration.' This excludes layover, volume, and multi-path effects.
  • domain assumption Observed intensity is a sum of direct backscatter contributions from illuminated patches, each with NRCS sigma0 and contribution indicatrix C.
    Eq. 1 models intensity as an integral of sigma0*C; ignores coherent speckle beyond the loss model and multiple scattering.
  • domain assumption Backscattering follows a Lambertian cosine model sigma0 = B * |u dot n|.
    Eq. 7 and Section II-E; simplified reflectivity used in place of material-dependent models such as mangia in EMPRISE.
  • domain assumption Shadow state at each scatterer is determined only by the nearest previously illuminated scatterer, with a logistic sigmoid approximation.
    Eq. 3 iterative shadow algorithm; approximates the true shadow demarcation and ignores diffraction.
  • domain assumption Range cell assignment uses locally straight isochron lines and a smooth maximum approximation.
    Eqs. 4-6; standard geometric SAR approximation.
  • domain assumption Speckle is distributed so that the negative log-likelihood of log-intensities is an appropriate training loss.
    Eq. 8 borrowed from SAR2SAR [19]; assumes Gamma-like speckle statistics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-view 3D surface reconstruction from SAR images by inverse rendering." pith.science (2026). https://pith.science/paper/YEXNBQKB

@misc{pith2026250210492,
  author       = {Pith},
  title        = {Pith review of: Multi-view 3D surface reconstruction from SAR images by inverse rendering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YEXNBQKB}},
  note         = {Machine review of arXiv:2502.10492}
}
read the original abstract

3D reconstruction of a scene from Synthetic Aperture Radar (SAR) images mainly relies on interferometric measurements, which involve strict constraints on the acquisition process. These last years, progress in deep learning has significantly advanced 3D reconstruction from multiple views in optical imaging, mainly through reconstruction-by-synthesis approaches pioneered by Neural Radiance Fields. In this paper, we propose a new inverse rendering method for 3D reconstruction from unconstrained SAR images, drawing inspiration from optical approaches. First, we introduce a new simplified differentiable SAR rendering model, able to synthesize images from a digital elevation model and a radar backscattering coefficients map. Then, we introduce a coarse-to-fine strategy to train a Multi-Layer Perceptron (MLP) to fit the height and appearance of a given radar scene from a few SAR views. Finally, we demonstrate the surface reconstruction capabilities of our method on synthetic SAR images produced by ONERA's physically-based EMPRISE simulator. Our method showcases the potential of exploiting geometric disparities in SAR images and paves the way for multi-sensor data fusion.

Figures

Figures reproduced from arXiv: 2502.10492 by the authors.

Figure 1
Figure 1. Iterative determination of shadow areas: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the fully-connected network. Input: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Results of the DSM reconstruction on the low resolution Forez scene. Areas not visible in both images are masked out on the DSM reconstructed [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Results of the surface learning on the high resolution island scene [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SH-SAS: An Implicit Neural Representation for Complex Spherical-Harmonic Scattering Fields for 3D Synthetic Aperture Sonar

    cs.GR 2025-09 conditional novelty 6.0 of 10

    Modeling the complex acoustic scattering field as spherical harmonic coefficients inside a hash-encoded neural network improves 3D synthetic aperture sonar reconstruction over isotropic neural baselines.

Reference graph

Works this paper leans on

23 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [13]

    SAR-NeRF: Neural Radiance Fields for Synthetic Aperture Radar Multiview Repre- sentation,

    Z. Lei, F. Xu, J. Wei, F. Cai, F. Wang, and Y .-Q. Jin, “SAR-NeRF: Neural Radiance Fields for Synthetic Aperture Radar Multiview Repre- sentation,” IEEE TGRS, vol. 62, pp. 1–15, 2024

  2. [14]

    Radar Fields: An Extension of Radiance Fields to SAR,

    T. Ehret, R. Mari, D. Derksen, N. Gasnier, and G. Facciolo, “Radar Fields: An Extension of Radiance Fields to SAR,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2024, pp. 564–574

  3. [1]

    Synthetic Aperture Radar interferometry,

    R. Bamler and P. Hartl, “Synthetic Aperture Radar interferometry,” Inverse Problems, vol. 14, no. 4, p. R1, aug 1998

  4. [2]

    Tomographic processing of interferometric SAR data: Developments, applications, and future research perspectives,

    G. Fornaro, F. Lombardini, A. Pauciullo, D. Reale, and F. Viviani, “Tomographic processing of interferometric SAR data: Developments, applications, and future research perspectives,” IEEE Signal Processing Magazine, vol. 31, no. 4, pp. 41–50, 2014

  5. [3]

    Radargrammetric DEM extraction over urban area using circular SAR imagery,

    S. Palm, H. M. Oriot, and H. M. Cantalloube, “Radargrammetric DEM extraction over urban area using circular SAR imagery,” IEEE Transactions on Geoscience and Remote Sensing , vol. 50, no. 11, pp. 4720–4725, 2012

  6. [4]

    Radarclinometry for ERS- 1 data mapping,

    S. Paquerault, H. Maitre, and J.-M. Nicolas, “Radarclinometry for ERS- 1 data mapping,” in IGARSS ’96. 1996 International Geoscience and Remote Sensing Symposium , vol. 1, 1996, pp. 503–505 vol.1

  7. [5]

    Deep-learning-based single-image height reconstruction from very-high-resolution SAR intensity data,

    M. Recla and M. Schmitt, “Deep-learning-based single-image height reconstruction from very-high-resolution SAR intensity data,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 183, pp. 496– 509, 2022

  8. [6]

    NeRF: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing scenes as neural radiance fields for view synthesis,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2020

Show all 23 references
  1. [7]

    Instant Neural Graphics Primitives with a Multiresolution Hash Encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant Neural Graphics Primitives with a Multiresolution Hash Encoding,” ACM Trans. Graph., vol. 41, no. 4, pp. 102:1–102:15, Jul. 2022

  2. [8]

    V olume rendering of neural implicit surfaces,

    L. Yariv, J. Gu, Y . Kasten, and Y . Lipman, “V olume rendering of neural implicit surfaces,” in Thirty-Fifth Conference on Neural Information Processing Systems, 2021

  3. [9]

    NeuS: Learning Neural Implicit Surfaces by V olume Rendering for Multi-view Reconstruction,

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wang, “NeuS: Learning Neural Implicit Surfaces by V olume Rendering for Multi-view Reconstruction,” NeurIPS, 2021

  4. [10]

    Sat-NeRF: Learning multi-view satellite photogrammetry with transient objects and shadow modeling using RPC cameras,

    R. Mar ´ı, G. Facciolo, and T. Ehret, “Sat-NeRF: Learning multi-view satellite photogrammetry with transient objects and shadow modeling using RPC cameras,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , 2022, pp. 1310–1320

  5. [11]

    Multi-date Earth observation NeRF: The detail is in the shadows,

    R. Mar ´ı, G. Facciolo, and T. Ehret, “Multi-date Earth observation NeRF: The detail is in the shadows,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) Workshops , June 2023, pp. 2034–2044

  6. [12]

    Neural V olumetric Reconstruction for Coherent Synthetic Aperture Sonar,

    A. Reed, J. Kim, T. Blanford, A. Pediredla, D. Brown, and S. Jayasuriya, “Neural V olumetric Reconstruction for Coherent Synthetic Aperture Sonar,” ACM Trans. Graph., vol. 42, no. 4, Jul. 2023

  7. [15]

    EMPRISE,

    “EMPRISE,” https://www.emprise-em.fr, accessed: 2025-01-21

  8. [16]

    EMPRISE : Synthetic Environment for sensor design and virtual qualification,

    N. Trouv ´e, C. Cochin, J. Houssay, E. Everaere, X. Husson, K. Ungers, A. Jouad´e, R. Fabbri, T. Houret, B. Talibart, O. L´evˆeque, and X. Dupuis, “EMPRISE : Synthetic Environment for sensor design and virtual qualification,” in 2024 IEEE International Radar Conference (RADAR) , 2024

  9. [17]

    Smooth Maximum Unit: Smooth activation function for deep networks using smoothing maximum technique,

    K. Biswas, S. Kumar, S. Banerjee, and A. K. Pandey, “Smooth Maximum Unit: Smooth activation function for deep networks using smoothing maximum technique,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2022, pp. 784–793

  10. [18]

    SAR image simulation and analysis of alpine terrain,

    M. Gelautz, H. Frick, J. Raggam, J. Burgstaller, and F. Leberl, “SAR image simulation and analysis of alpine terrain,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 53, no. 1, pp. 17–38, 1998

  11. [19]

    SAR2SAR: a semi-supervised despeckling algorithm for SAR images,

    E. Dalsasso, L. Denis, and F. Tupin, “SAR2SAR: a semi-supervised despeckling algorithm for SAR images,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , pp. 1–1, 2021

  12. [20]

    Nerfies: Deformable neural radiance fields,

    K. Park, U. Sinha, J. T. Barron, S. Bouaziz, D. B. Goldman, S. M. Seitz, and R. Martin-Brualla, “Nerfies: Deformable neural radiance fields,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , 2021, pp. 5845–5854

  13. [21]

    BARF: Bundle- Adjusting Neural Radiance Fields,

    C.-H. Lin, W.-C. Ma, A. Torralba, and S. Lucey, “BARF: Bundle- Adjusting Neural Radiance Fields,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , 2021, pp. 5721–5731

  14. [22]

    NASA Shuttle Radar Topography Mission Global 1 arc second [Data set],

    N. JPL, “NASA Shuttle Radar Topography Mission Global 1 arc second [Data set],” 2013

  15. [23]

    Calibration of a radar cross-section model using a surrogate model optimization algorithm,

    T. Houret, O. L ´evˆeque, N. Trouv ´e, R. Bocheux, X. Husson, and A. Jouad´e, “Calibration of a radar cross-section model using a surrogate model optimization algorithm,” in 2023 IEEE International Radar Conference (RADAR), 2023, pp. 1–6

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.