Pith. sign in

REVIEW 4 major objections 4 minor 32 references

Probabilistic Spatial Interpolation of Sparse Data using Diffusion Models

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that KrigSCD, a kriging-smoothed conditioned diffusion model, reconstructs full temperature fields from as little as 1% observational coverage with better perceptual fidelity than classical interpolators and base diffusion.

desk verdict KrigSCD is a legitimate diffusion-inpainting variant for sparse climate fields, but the '1% coverage' claim is inflated because the kriging step expands the conditioning mask to ~6% and the baselines don't get that advantage. read the letter →

arxiv 2506.00033 v1 pith:3YQMP27R submitted 2025-05-26 stat.AP cs.LG

classification stat.APcs.LG MSC 62P1286A3268T07
keywords spatialinterpolationdiffusionmodelskrigingdataassimilationimageinpaintingtemperaturefieldssparseobservationsconditionalgeneration
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

This paper tries to establish that a generative diffusion model can serve as a spatial interpolator for climate fields under extreme data sparsity. Its proposed framework, KrigSCD, first smooths a binary observation mask with ordinary kriging and then runs a denoising diffusion probabilistic model with mask conditioning and resampling, producing full 2-meter temperature fields over the Southern Great Plains from as little as 1% observed pixels. The paper reports that KrigSCD achieves lower LPIPS—a perceptual similarity metric—than inverse distance weighting, conditional Gaussian simulations, kriging, and base diffusion at every tested known percentage from 1% to 30%. If this holds, the method offers a concrete way to fill data gaps in historical reanalysis and real-time forecasting pipelines, and it extends in proof-of-concept form to 3D inputs such as weather balloons, plane swaths, and satellite tracks.

What carries the argument

The load-bearing mechanism is the KrigSCD conditioning pipeline, which combines three components. A binary mask simulates realistic observations: isolated pixels for in-situ sensors and linear segments of fixed width for satellite swaths. Ordinary kriging is then applied to estimate every unobserved pixel together with its kriging variance, and pixels whose variance falls in the lowest 5th percentile are promoted to known status with their kriged values, so the diffusion model sees coherent regions of plausibility rather than isolated high-frequency spikes. The reverse diffusion process follows the mask-conditioned sampling scheme of Lugmayr et al. (2022): known pixels are re-noised from ground truth, unknown pixels are denoised by the UNet, and the two are recombined with periodic resampling every $j=10$ steps, allowing the generated and conditioned regions to harmonize. The kriging smoother is what converts point observations into usable conditioning information; without it, the authors observe that the UNet's stride-2 pooling attenuates the isolated spikes and most conditioning is lost.

What would settle it

Rerun the 1% covered experiment with the kriging smoothing disabled—conditioning only on the original 1% of true observed pixels—and compare LPIPS and RMSE against IDW, CGS, and KrigSCD; if the gap mostly closes, the reported advantage comes from the kriging head start rather than the diffusion prior. A second check is to report the effective known fraction after Algorithm 1 and give the baselines the same effective fraction before computing improvements.

Watch

Extended reading notes

Core claim

The central discovery is that a diffusion model conditioned on a kriging-smoothed mask reconstructs perceptually faithful temperature fields where classical interpolators and unconditioned diffusion fail. At 1% known pixels, KrigSCD improves LPIPS by 33.77% over inverse distance weighting, 23.28% over kriging, and 8.52% over base diffusion; at 20% known the improvements rise to 74.82%, 74.43%, and 7.48% respectively, and the LPIPS scores reach a level the paper calls virtually indistinguishable from the ground truth. The reconstructions preserve localized features such as urban cool islands that IDW and conditional Gaussian simulations miss, and they do so while the known pixels exactly match the observed values. The authors frame the result as evidence that generative priors can complement classical geostatistics at the extreme end of sparsity.

Load-bearing premise

The load-bearing premise is that feeding the diffusion model kriged pseudo-observations at low-variance pixels is a fair part of the method rather than hidden extra information, so the comparison against baselines that see only the original mask remains apples-to-apples.

Editorial extensions

If this is right

  • At every tested mask percentage (1%, 5%, 10%, 20%, 30%), KrigSCD yields lower LPIPS than IDW, kriging, CGS, and base diffusion on afternoon summer temperature fields over the Southern Great Plains.
  • The framework handles realistic observation geometries—all-swath, all-in-situ, and mixed configurations—and its error decreases monotonically as the fraction of in-situ points grows.
  • At 20% known coverage or above, LPIPS scores approach the range the paper describes as virtually indistinguishable from ground truth, so the perceptual benefit of additional observations saturates around that density.
  • The same pipeline can be applied to 3D inputs (weather balloons, plane swaths, satellites) with modest architectural changes, as demonstrated on a 64x64x64 cube.
  • The method is a candidate initialization strategy for data assimilation and earth system models, offering a probabilistic full-field state estimate from sparse observations.

Reading between the lines

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

  • Algorithm 1 replaces unknown pixels with kriged estimates when their kriging variance is in the lowest 5th percentile, so a nominal 1% mask actually conditions the diffusion model on roughly 6% of pixels; the paper never reports this effective known fraction, so its comparisons with IDW and CGS do not fully isolate method quality from information quantity.
  • The LPIPS advantage over classical methods is partly a consequence of metric choice: LPIPS rewards perceptually plausible texture, which diffusion priors are designed to produce, while kriging-based methods are designed to minimize pointwise error and will look smoother; downstream users should ask whether their task needs perceptual fidelity or pointwise accuracy.
  • A direct testable extension would be to vary the 5th-percentile kriging threshold and the resampling parameters ($r=10$, $j=10$) for variables with different spatial correlation lengths, such as wind or precipitation, to see whether the current settings are transferable or need per-variable tuning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes KrigSCD, a diffusion-model-based inpainting framework for reconstructing 2-meter temperature fields from very sparse observations, with a kriging-smoothed mask as conditioning. The model is trained on 64x64 HRRR temperature images over the Southern Great Plains and is evaluated at 1%-30% known pixels under in-situ and swath mask configurations. The authors compare KrigSCD against IDW, conditional Gaussian simulations (CGS), and base diffusion, reporting lower LPIPS for KrigSCD, while acknowledging higher RMSE/MAE/MRE at low coverage. The headline claim is that full fields can be reconstructed from as little as 1% observational coverage.

Significance. If the central claim were fully supported, the method would be a useful step toward generative data assimilation and gap-filling in reanalysis products, especially under extreme sparsity. The paper ships a public data link and a GitHub code link, and it evaluates a realistic mix of in-situ and swath observation patterns. The main strength is the systematic use of multiple metrics and the finding that KrigSCD yields consistently lower LPIPS, suggesting better perceptual fidelity. However, the headline 1% claim is clouded by the kriging-mask expansion, and the abstract's 'strong reconstruction accuracy' is contradicted by the paper's own pixel-error numbers. The potential is real, but the evaluation needs substantial rework before the claims can be accepted.

major comments (4)
  1. [Algorithm 1, Section 3e] The kriging-smoothing step (Algorithm 1, lines 10-14) converts unknown pixels into known pseudo-observations: any unknown pixel whose kriging variance falls in the lowest 5th percentile is assigned its kriged estimate and added to the known set. At a nominal 1% mask, this adds roughly 5% of the remaining 99% of pixels, so the diffusion model conditions on about 6% known pixels, not 1%. The comparison baselines IDW and CGS (Appendix A2-A3) are evaluated only on the original 1% mask. Consequently, the abstract's 'as little as 1% observational coverage' is not supported, and the comparisons in Table 2 and Figure 10 are not apples-to-apples. The authors should either report the effective coverage for each threshold, provide the same expanded mask to all baselines, or clearly state that the diffusion model sees kriged pseudo-observations in addition to the original 1%.
  2. [Table 2, Section 4b] The abstract claims 'strong reconstruction accuracy,' but the reported pixel-error metrics at 1% known do not support that wording. From Table 2, at 1% known, KrigSCD has RMSE about 83 and MRE about 0.41 (on a 0-255 scale), which are considerably worse than IDW (RMSE about 32, MRE about 0.26) and CGS (RMSE about 36, MRE about 0.30). The LPIPS advantage is real and worth reporting, but the abstract's phrasing should be qualified to reflect that KrigSCD improves perceptual similarity while having larger pixel-wise errors at very low coverage, or a metric that justifies 'strong' accuracy should be presented.
  3. [Section 2 vs Section 4b] The test set size is inconsistently reported. Section 2 states that 150 samples were randomly selected as the test set, while Section 4b says '100 test images' are used for KrigSCD and Figure 10 mentions 100 reconstructions; Figure 8 also states '100 reconstruction samples per method.' Please clarify which number is correct and ensure that all reported statistics are computed on the same test set.
  4. [Algorithm 2, line 9] The known-region update in Algorithm 2 line 9 deviates from the RePaint formulation. RePaint samples the known region at timestep t-1 from q(x_{t-1} | x0), namely sqrt(alpha_bar_{t-1}) x0 + sqrt(1 - alpha_bar_{t-1}) epsilon. The paper instead uses sqrt(alpha_bar_t) x_t + sqrt(1 - alpha_bar_t) epsilon_t, which does not correspond to the forward noising of the ground-truth image and may weaken the conditioning signal. Please justify this modification or correct the formula to match RePaint.
minor comments (4)
  1. [Section 3f, Eq. (11)] The MRE formula lacks an absolute value, although the text describes an 'average discrepancy.' Without an absolute value, positive and negative errors can cancel; please clarify whether the metric is signed or absolute and adjust the formula accordingly.
  2. [Section 3b, Section 4a] There are several typos, including 'distrubitions' and 'Intuituively' in Section 3b. Also, Section 4a refers to models trained at 100k, 150k, and 200k epochs after earlier stating the model was trained for 100k epochs; clarify whether these are separate training runs or checkpoints.
  3. [Section 3g] The ensemble-size argument says 'with n=500 samples' but then uses n=10 ensemble members; the connection is confusing. Please rewrite so that the CLT-based reasoning for the choice of n=10 is explicit and not conflated with the 500-sample histogram.
  4. [Appendix A1] The exponential variogram parameters c and tau are said to be fitted, but the paper does not describe how they are estimated in the main experiments (e.g., globally versus per mask). Please provide the fitting procedure and, if applicable, the values used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KrigSCD's reconstruction claims are empirically tested on held-out data with external baselines; the kriging preprocessing issue is a comparison-fairness concern, not a circular derivation.

full rationale

The paper's derivation chain is self-contained with respect to circularity. The diffusion model is trained on full-coverage HRRR fields and evaluated on a held-out set of 150 samples (Section 2a), with quantitative comparisons against IDW, CGS, and base diffusion on test images not used in training. No parameter is fitted to the test-set targets: the kriging step (Algorithm 1) uses only the observed pixels in the current mask, and the 5th-percentile threshold is a preprocessing hyperparameter, not a fitted quantity that determines the reported errors. The conditioning machinery is imported from external references (Ho et al. 2020; Lugmayr et al. 2022; Dhariwal and Nichol 2021), and there are no load-bearing self-citations or imported uniqueness claims. A caveat outside circularity: the abstract's '1% observational coverage' is imprecise because Algorithm 1 expands the diffusion model's conditioning mask with low-variance kriged estimates, so the model effectively sees more than 1% of pixels filled with estimated values; this affects the fairness of the comparison with baselines but does not make the reconstruction equivalent to its inputs by construction. Hence the circularity score is 0.

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

The central claim rests on the proxy assumption that HRRR fields plus synthetic masks stand for real observations, and on the ad hoc kriging threshold that expands the conditioning mask. No new entities are postulated.

free parameters (6)
  • Kriging variance threshold percentile = 5th percentile
    Chosen in Section 3e; controls how many kriged pseudo-observations are added to the mask. Not sensitivity-tested.
  • Exponential variogram sill c and range tau = Not reported
    Fitted to empirical semivariogram per mask (Appendix A); values affect kriged estimates and therefore the expanded mask.
  • Diffusion resampling parameters r, j, and timestep respacing t_T = r=10, j=10, t_T=150
    Set in Section 3g; known to affect inpainting quality in RePaint-style methods.
  • Pixel normalization bounds = Not reported
    Temperatures are mapped to 0-255 greyscale for the OpenAI guided-diffusion code; the mapping is never specified, so reported RMSE and MRE units are ambiguous.
  • Ensemble size = n=10
    Chosen from an ensemble study (Figure 4); the statistical argument uses n=500 for CLT and then n=10 for final runs.
  • Training iterations = 100k (called epochs)
    KID scores at 100k and 200k are used to select 100k; 'epochs' is likely a misnomer for iterations.
assumptions (6)
  • domain assumption HRRR reanalysis output is treated as ground truth for training and evaluation.
    Section 2a sources data from the HRRR model; Section 6b acknowledges reanalysis is synthetic, so no independent observations validate the reconstructions.
  • domain assumption Randomly generated binary masks simulate real in-situ and swath observation patterns.
    Section 3c generates masks as random pixels and linear segments; there is no test on real sparse observation geometries.
  • domain assumption Observation error is ignored during conditioning (R unspecified, eta not sampled).
    Section 3a defines y = Hx + eta, but Algorithm 2 treats known pixels as exact, assigning them ground truth values.
  • domain assumption Temperature fields are stationary with an exponential semivariogram.
    Appendix A uses exponential variogram fits; stationarity is assumed throughout kriging steps.
  • standard math Diffusion model theory (DDPM, score matching, resampling) is correct.
    The method builds on Ho et al. (2020) and Lugmayr et al. (2022).
  • ad hoc to paper The 5th-percentile kriging variance threshold is a good fixed hyperparameter.
    Section 3e sets it without sensitivity analysis; it determines the amount of extra conditioning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic Spatial Interpolation of Sparse Data using Diffusion Models." pith.science (2026). https://pith.science/paper/3YQMP27R

@misc{pith2026250600033,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Spatial Interpolation of Sparse Data using Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3YQMP27R}},
  note         = {Machine review of arXiv:2506.00033}
}
read the original abstract

The large underlying assumption of climate models today relies on the basis of a "confident" initial condition, a reasonably plausible snapshot of the Earth for which all future predictions depend on. However, given the inherently chaotic nature of our system, this assumption is complicated by sensitive dependence, where small uncertainties in initial conditions can lead to exponentially diverging outcomes over time. This challenge is particularly salient at global spatial scales and over centennial timescales, where data gaps are not just common but expected. The source of uncertainty is two-fold: (1) sparse, noisy observations from satellites and ground stations, and (2) internal variability stemming from the simplifying approximations within the models themselves. In practice, data assimilation methods are used to reconcile this missing information by conditioning model states on partial observations. Our work builds on this idea but operates at the extreme end of sparsity. We propose a conditional data imputation framework that reconstructs full temperature fields from as little as 1% observational coverage. The method leverages a diffusion model guided by a prekriged mask, effectively inferring the full-state fields from minimal data points. We validate our framework over the Southern Great Plains, focusing on afternoon (12:00-6:00 PM) temperature fields during the summer months of 2018-2020. Across varying observational densities--from swath data to isolated in-situ sensors--our model achieves strong reconstruction accuracy, highlighting its potential to fill in critical data gaps in both historical reanalysis and real-time forecasting pipelines.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 16 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archive author booktitle chapter doi edition editor eid eprint howpublished department institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 '...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION if.digit duplicate "0" = swap duplicate "1" = swap duplicate "2" = swap duplicate "3" = swap duplicate "4" = swap duplicate "5" = swap duplicate "6" = swap duplicate "7" = swap duplicate "8" = swap "9" = or or or or or or or or or FUNCTION n.separate 't := "" #0 'numnames := t empty not t #-1 #1 subs...

  3. [3]

    Shaharbany, E

    Amit, T., T. Shaharbany, E. Nachmani, and L. Wolf, 2022: Segdiff: Image segmentation with diffusion probabilistic models. ://arxiv.org/abs/2112.00390, 2112.00390

  4. [4]

    Bińkowski, M., D. J. Sutherland, M. Arbel, and A. Gretton, 2021: Demystifying mmd gans. 1801.01401

  5. [5]

    Chen, S., P. Sun, Y. Song, and P. Luo, 2023: Diffusiondet: Diffusion model for object detection. ://arxiv.org/abs/2211.09788, 2211.09788

  6. [6]

    Choi, H., J. Choi, J. Hwang, K. Lee, D. E. Lee, and N. Park, 2023: Climate modeling with neural advection–diffusion equation. Knowledge and Information Systems, 65, 2403--2427, doi:10.1007/s10115-023-01829-2

  7. [7]

    P., 1991: Simplification of the kalman filter for meteorological data assimilation

    Dee, D. P., 1991: Simplification of the kalman filter for meteorological data assimilation. Quarterly Journal of the Royal Meteorological Society, 117 (498), 365--384, doi:https://doi.org/10.1002/qj.49711749806, https://rmets.onlinelibrary.wiley.com/doi/pdf/10.1002/qj.49711749806

  8. [8]

    Nichol, 2021: Diffusion models beat gans on image synthesis

    Dhariwal, P., and A. Nichol, 2021: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y. Dauphin, P. Liang, and J. W. Vaughan, Eds., Curran Associates, Inc., Vol. 34, 8780--8794, ://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf

Show all 32 references
  1. [9]

    S., and M

    Fotheringham, A. S., and M. E. O'Kelly, 1989: Spatial interaction models: formulations and applications, Vol. 1. Kluwer academic publishers Dordrecht

  2. [10]

    Goodfellow, I. J., J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, 2014: Generative adversarial networks. 1406.2661

  3. [11]

    Jain, and P

    Ho, J., A. Jain, and P. Abbeel, 2020: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., Curran Associates, Inc., Vol. 33, 6840--6851, ://proceedings.neurips.cc/paper_...

  4. [12]

    Gianinazzi, Y

    Huang, L., L. Gianinazzi, Y. Yu, P. D. Dueben, and T. Hoefler, 2024: Diffda: a diffusion model for weather-scale data assimilation. ://arxiv.org/abs/2401.05932, 2401.05932

  5. [13]

    K., and Coauthors, 2023: Overview: Understanding risks, impacts, and responses, book section 1

    Jay, A. K., and Coauthors, 2023: Overview: Understanding risks, impacts, and responses, book section 1. U.S. Global Change Research Program, Washington, DC, USA, doi:10.7930/NCA5.2023.CH1

  6. [14]

    P., and M

    Kingma, D. P., and M. Welling, 2015: Auto-encoding variational bayes. 1312.6114

  7. [15]

    G., 1951: A statistical approach to some basic mine valuation problems on the witwatersand

    Krige, D. G., 1951: A statistical approach to some basic mine valuation problems on the witwatersand. Journal of the Chemical, Metallurgical and Mining Society of South Africa, 52, 119--139

  8. [16]

    Dong, and P

    Li, Z., B. Dong, and P. Zhang, 2023: Latent assimilation with implicit neural representations for unknown dynamics. doi:10.2139/ssrn.4591077

  9. [17]

    Danelljan, A

    Lugmayr, A., M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool, 2022: Repaint: Inpainting using denoising diffusion probabilistic models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11\,461--11\,471

  10. [18]

    Dhariwal, A

    Nichol, A., P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. McGrew, I. Sutskever, and M. Chen, 2022: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ://arxiv.org/abs/2112.10741, 2112.10741

  11. [19]

    McBratney, and D

    Odeh, I., A. McBratney, and D. Chittleborough, 1995: Further results on prediction of soil properties from terrain attributes: heterotopic cokriging and regression-kriging. Geoderma, 67 (3), 215--226, doi:https://doi.org/10.1016/0016-7061(95)00007-B

  12. [20]

    J., and S

    Rezende, D. J., and S. Mohamed, 2016: Variational inference with normalizing flows. 1505.05770

  13. [21]

    S., Gregory J

    Robert Tardif, C. S., Gregory J. Hakim, 2014: Coupled atmosphere – ocean data assimilation experiments with a low-order climate model. Climate Dynamics, 43, 1631 -- 1643, doi:https://doi.org/10.1007/s00382-013-1989-0

  14. [22]

    Blattmann, D

    Rombach, R., A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, 2022: High-resolution image synthesis with latent diffusion models. ://arxiv.org/abs/2112.10752, 2112.10752

  15. [23]

    J., 1987: Climate of the great plains region of the united states

    Rosenberg, N. J., 1987: Climate of the great plains region of the united states. Great Plains Quarterly, 22--32

  16. [24]

    Saharia, C., J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, 2023: Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 (4), 4713--4726, doi:10.1109/TPAMI.2022.3204461

  17. [25]

    Sisterson, D. L., R. A. Peppler, T. S. Cress, P. J. Lamb, and D. D. Turner, 2016: The arm southern great plains (sgp) site. Meteorological Monographs, 57, 6.1 -- 6.14, doi:10.1175/AMSMONOGRAPHS-D-16-0004.1

  18. [26]

    Weiss, N

    Sohl-Dickstein, J., E. Weiss, N. Maheswaranathan, and S. Ganguli, 2015: Deep unsupervised learning using nonequilibrium thermodynamics. Proceedings of the 32nd International Conference on Machine Learning, F. Bach, and D. Blei, Eds., PMLR, Lille, France, Proceedings of Machine...

  19. [27]

    Ermon, 2019: Generative modeling by estimating gradients of the data distribution

    Song, Y., and S. Ermon, 2019: Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alch\' e -Buc, E. Fox, and R. Garnett, Eds., Curran Associates, Inc., Vol. 32,...

  20. [28]

    Awaji, S

    Sugiura, N., T. Awaji, S. Masuda, T. Mochizuki, T. Toyoda, T. Miyama, H. Igarashi, and Y. Ishikawa, 2008: Development of a four-dimensional variational coupled data assimilation system for enhanced analysis and prediction of seasonal to interannual climate variations. Journal ...

  21. [29]

    Thomas Milewski, M. S. B., 2013: Potential of an ensemble kalman smoother for stratospheric chemical-dynamical data assimilation. Tellus A: Dynamic Meteorology and Oceanography, 65, doi:https://doi.org/10.3402/tellusa.v65i0.18541

  22. [30]

    Heinonen, and V

    Verma, Y., M. Heinonen, and V. Garg, 2024: Climode: Climate and weather forecasting with physics-informed neural odes. arXiv preprint arXiv:2404.10024

  23. [31]

    Zou, and J

    Wang, B., X. Zou, and J. Zhu, 2000: Data assimilation and its applications. Proceedings of the National Academy of Sciences, 97, 11\,143--11\,144, doi:10.1073/pnas.97.21.11143

  24. [32]

    Isola, A

    Zhang, R., P. Isola, A. A. Efros, E. Shechtman, and O. Wang, 2018: The unreasonable effectiveness of deep features as a perceptual metric. 1801.03924

Pith tools

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