Pith. sign in

REVIEW 3 major objections 5 minor 26 references

Generative Algorithms for Wildfire Progression Reconstruction from Multi-Modal Satellite Active Fire Measurements and Terrain Height

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

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A conditional GAN trained on physics-based fire simulations reconstructs wildfire arrival times from sparse satellite detections, matching aircraft-measured perimeters with an average Dice coefficient of 0.81.

desk verdict A credible cWGAN extension with real WRF-SFIRE training and independent validation on five fires, but the hand-built observation operator needs direct validation to make the 0.81 Dice claim stick. read the letter →

arxiv 2506.10404 v1 pith:R6TOWFK4 submitted 2025-06-12 cs.LG

Bryan Shaddy , Brianna Binder , Agnimitra Dasgupta , Haitong Qin , James Haley , Angel Farguell , Kyle Hilburn , Derek V. Mallia
show 3 more authors
Adam Kochanski Jan Mandel Assad Oberai
This is my paper
classification cs.LG
keywords wildfirespreadreconstructionfirearrivaltimeconditionalWassersteinGANdataassimilationforVIIRSactiveGOESignitionWRF-SFIREsimulationuncertaintyquantification
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 argues that sparse, noisy satellite fire detections can be turned into a complete, high-resolution map of when the fire front arrived at each point. It does this with a conditional Wasserstein generative adversarial network trained on simulations from the coupled atmosphere–wildfire model WRF-SFIRE, so the physics of fire spread is baked into the estimates. On five Pacific US wildfires, the reconstructed perimeters match aircraft-measured infrared perimeters with an average Sørensen-Dice coefficient of 0.81. The paper also reports that adding terrain height as a conditioning input changes predictions very little once satellite measurements are given, with 97% of pixels differing by under 30 minutes.

What carries the argument

The central object is the observation operator $M$ that maps a fine-grained fire arrival time field $\tau$ to a coarse, sparse, noisy measurement field $\bar{\tau}$ mimicking VIIRS and GOES products: coarsening to 375 m, keeping or dropping pixels with probability 1/2, sampling four measurement times, assigning interval-binned arrival times, adding a GOES ignition-time error, erasing two 3 km patches, and upsampling back to 25 m. Training tuples $(\tau,\bar{\tau},h)$ are produced by applying this operator to WRF-SFIRE solutions, and a cWGAN with a U-Net generator (using conditional instance normalization and a latent vector $z\sim P_Z$) and a critic with Wasserstein-1 loss learns to sample from $P_{T|\bar{T},H}$. This makes the physics of the simulation model the prior, while the latent vector provides an ensemble of plausible reconstructions and hence pixel-wise uncertainty.

What would settle it

Compare the statistical distribution of gap sizes, patch obstructions, and ignition-time errors in the synthetic measurements against real VIIRS and GOES data over many fires; if the distributions do not match, the agreement on the five validation fires cannot be attributed to a faithful observation model.

Watch

Extended reading notes

Core claim

The central claim is that a conditional Wasserstein generative adversarial network, trained exclusively on synthetic measurement–arrival-time–terrain tuples derived from WRF-SFIRE simulations of 140 real 2023 CONUS wildfires, learns a conditional distribution of fire arrival times given VIIRS active fire detections, GOES-derived ignition times, and terrain height. When applied to five real fires in California and Washington, the generator's mean output yields fire perimeters at the time of aircraft IR overflight that agree with measured high-resolution perimeters at an average Dice coefficient of 0.81, a probability of detection of 0.83, and a false alarm ratio of 0.20. The paper treats this as evidence that physics-informed generative reconstruction can serve as the initial state for coupled atmosphere–wildfire forecast initialization, and that the conditional distribution provides uncertainty estimates through the spread of generated samples.

Load-bearing premise

The approximate observation operator used to generate synthetic VIIRS and GOES measurements during training must faithfully reproduce the real measurement process; if real satellite data contain systematic biases or noise patterns not captured by those hand-designed steps, the trained generator's performance on real fires is not guaranteed.

Editorial extensions

If this is right

  • Reconstructed arrival time maps can initialize coupled atmosphere–wildfire simulations, giving a matching atmospheric state via the spin-up procedure.
  • The ensemble of generated arrival times yields pixel-wise standard deviations that quantify uncertainty, with higher values in measurement-sparse and sharp-gradient regions.
  • The method extends naturally to other conditioning fields; the terrain analysis suggests that fuel maps may be more informative than terrain height.
  • With higher-resolution or more frequent measurements, finer peripheral detail in fire perimeters could be captured.
  • The approach eliminates the need for satellite data during training, relying only on simulations plus an approximate observation operator.

Reading between the lines

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

  • If the method scales operationally, it could turn the current retrospective validation into a near-real-time pipeline for initializing forecast models from incoming satellite detections.
  • The finding that terrain adds little after satellite conditioning suggests the model is essentially learning smooth, physics-guided interpolation; a controlled experiment that withholds measurements entirely could test whether terrain alone carries meaningful spread information.
  • The same conditional-generative pipeline could be retrained with fuel maps, soil moisture, or wind fields as additional conditioning variables, with the hand-tuned observation operator replaced by a learned forward model.
  • Validating against aircraft IR perimeters at only one time per fire leaves open whether the arrival-time field is correct in shape as well as instantaneous extent; sequences of IR perimeters over time would test that more strictly.
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

3 major / 5 minor

Summary. The paper develops a conditional Wasserstein GAN (cWGAN) that reconstructs wildfire fire-arrival-time maps from VIIRS active-fire detections, GOES-derived ignition times, and terrain height. Training data are generated by applying an approximate observation operator (Section 3b) to 140 WRF-SFIRE simulations of 2023 CONUS wildfires, augmented by rotation/translation/cropping. The trained generator is applied to five real Pacific US wildfires (2019-2022) and validated against NIROPS aircraft IR perimeters, reporting a mean Sørensen-Dice coefficient of 0.81 (0.85 excluding Tennant), mean POD 0.83, and mean FAR 0.20. The paper also reports that conditioning on terrain height has minimal effect on the reconstructions when satellite measurements are available.

Significance. If the result holds, the paper offers a practical pipeline for initializing coupled atmosphere-wildfire forecasts from sparse satellite active-fire data, with built-in uncertainty quantification. Strengths include the use of realistic WRF-SFIRE physics in the training distribution, an independent external validation against high-resolution IR perimeters, transparent discussion of the Tennant fire failure, and a reproducible-looking data-generation protocol. The method is clearly motivated and the architecture follows established cWGAN practice. However, the central empirical claim rests on a small validation set and on an observation operator whose fidelity to real VIIRS/GOES measurements is not demonstrated; these issues need to be addressed before the results can be considered robust.

major comments (3)
  1. [§3b, §5c] The approximate observation operator M, defined in Section 3b, is central to the method because all training data are generated by applying M to WRF-SFIRE arrival times, yet the paper provides no quantitative evidence that M reproduces the statistics of real VIIRS/GOES active-fire measurements. The hand-chosen steps (50% per-pixel dropout, burning-interval delta ~ U(6,12), two random 3 km x 3 km gaps, ignition-time error U(0,2), and four overpasses) condense cloud/smoke obscuration, fire-radiative-power-dependent detectability, viewing geometry, and actual overpass schedules into a few uniform choices. If M is systematically unrepresentative, the generator learns the conditional distribution for a synthetic process, and the five real-fire validations (one with SC=0.67) may not detect the mismatch. I recommend validating M directly by comparing synthetic measurement statistics (detection density, gap-length distributions, temporal sampling, and persistence of obstructions) against real VIIRS data for the validation fires or a larger fire sample, and performing a sensitivity analysis of the key hyperparameters (delta, dropout probability, number of overpasses).
  2. [Table 3, §5e] The validation set of five fires is too small to support the headline 'average Sørensen-Dice coefficient of 0.81' without uncertainty quantification. No confidence intervals are provided for SC, POD, or FAR, and the mean is heavily influenced by the Tennant outlier (SC=0.67). In addition, no baseline method is compared on the same fires; the SVM-based approach of Farguell et al. (2021), which the paper cites as prior work, is the natural benchmark and would contextualize the 0.81 figure. I recommend reporting bootstrap or per-fire confidence intervals for the metrics and including at least one baseline reconstruction on the same five fires.
  3. [§5f] The claim that terrain height has minimal influence on arrival time inference is based solely on 200 synthetic validation cases, not on the five real fires, and is stated as a general finding in the abstract. This experiment inherits any biases of the synthetic observation operator: if the measurements already constrain arrival times strongly, the terrain may be redundant in the training distribution even if it is informative for real fires with different detection densities or error characteristics. Additionally, the histogram in Figure 7 excludes pixels with arrival time exceeding 47 hours, but the excluded fraction is not reported, and the mean difference of -3 minutes is an order of magnitude smaller than the 30-minute threshold used to define agreement. I recommend either restricting the terrain conclusion to the synthetic validation setting or re-running the comparison on the real-fire predictions, and reporting the fraction of excluded pixels.
minor comments (5)
  1. [§4b] The Wasserstein objective in Eq. (1) requires a 1-Lipschitz critic, and the theoretical guarantee of Eq. (3) depends on this condition, but the training section does not specify how Lipschitz continuity is enforced (e.g., gradient penalty, weight clipping, or spectral normalization). Please state the exact penalty term and its weight.
  2. [Table 3] The table caption contains a typo: 'False Alarm Ration' should be 'False Alarm Ratio'.
  3. [§5f and Figure 7] The figure caption should state the exclusion rule (pixels with arrival time >47 h in both predictions are removed) and the total number of pixels included, since the histogram's interpretation depends on this filtering.
  4. [§3a] The data augmentation counts are consistent (140 fires x 25 augmentations x 5 observation realizations = 17,500 training tuples), but it would be helpful to state explicitly whether the 12 validation simulations are disjoint from the 140 training simulations and whether any of the five real validation fires overlap with fires used in training (they are from 2019-2022, while training is 2023, so this appears to be the case, but it should be stated for clarity).
  5. [§5b] The GOES ignition time estimate procedure in steps 1-6 selects the earliest high-confidence detection in the 12.8 km domain, but the paper does not discuss potential contamination from neighboring fires or false detections; a sentence on how such cases were handled would strengthen the description.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the WRF-SFIRE/synthetic-observation training is independent from the five-fire NIROPS validation, and self-citations are not load-bearing.

full rationale

The paper's derivation chain is self-contained with respect to its central claim. Training tuples (τ, τ̄, h) are generated by applying the approximate observation operator M of Section 3b to WRF-SFIRE arrival-time maps; the cWGAN learns the conditional distribution P_{T|T̄,H}. At inference, real VIIRS detections and GOES ignition times are processed into measurement maps and passed through the frozen generator. Validation uses NIROPS aircraft IR perimeters, which were never used during training and do not enter the construction of M. Thus the reported Sørensen-Dice coefficients are predictions on held-out real observations, not fits to those observations. The cited prior work (Shaddy et al. 2024; Ray et al. 2023) supplies architecture choices and a WGAN approximation theorem; it does not define the validation metric, constrain the comparison, or supply the five perimeters. The fidelity of the hand-built observation operator is a legitimate generalizability risk, but that is a correctness/robustness concern, not circularity: no equation in the paper makes the validation result an identity or a renamed input.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on three pillars: WRF-SFIRE simulations as physics-based training targets, a hand-crafted observation operator that approximates VIIRS and GOES data, and the cWGAN's ability to generalize from simulated to real measurements. The observation operator alone contributes six or more hand-chosen parameters that are not fitted to real satellite data. No new physical entities are introduced.

free parameters (7)
  • Number of VIIRS overpasses per training sample = 4
    The observation operator in Section 3b creates four copies of each coarsened arrival time map to simulate multiple satellite overpasses during the first 48 hours; this number is chosen by hand and directly shapes the training measurement distribution.
  • Per-pixel elimination probability = 1/2
    Step 3 of the observation operator randomly drops half the coarse pixels to simulate independent noise between VIIRS measurements.
  • Measurement time sampling bounds = U(2, max(tau)-0.1) hours
    Step 4 samples four observation times uniformly between 2 hours and just before the maximum arrival time; these bounds are ad hoc choices.
  • Burning interval delta = U(6,12) hours
    Step 5 assigns arrival times within (t_j - delta, t_j) to t_j to mimic pixels that ignited earlier and are still burning; the 6 to 12 hour window is chosen by hand.
  • Ignition time error = U(0,2) hours
    Step 8 subtracts a random 0 to 2 hour error from arrival times to account for GOES ignition time uncertainty; the range is assumed, not measured.
  • Persistent obstruction patches = two 3 km x 3 km patches
    Step 9 randomly removes two 3 km by 3 km regions from each measurement to simulate persistent cloud or smoke obstructions; size and number are hand chosen.
  • Normalization constants = arrival time divided by 48 h; terrain shifted and divided by 3000 m
    These scaling choices in Section 3b affect the loss landscape and could influence which features the generator learns, though their effect on the final result is not evaluated.
assumptions (5)
  • domain assumption WRF-SFIRE (Rothermel-based) simulations provide realistic fire arrival time fields for training.
    The paper relies on 140 WRF-SFIRE simulations as ground truth for fire progression; if the fire model systematically misrepresents real spread, the learned conditional distribution is biased. Section 3a.
  • domain assumption The hand-crafted observation operator in Section 3b generates synthetic measurements statistically similar to real VIIRS and GOES products.
    The cWGAN is trained entirely on synthetic measurements; domain shift between these and real satellite data would degrade real-fire performance. Section 3b.
  • standard math The cWGAN, with sufficient capacity and training, approximates the true conditional distribution P(T | measured arrival time, terrain).
    The theoretical guarantee in Eq. 3 is cited from Ray et al. 2023 and relies on the critic being 1-Lipschitz and the generator having enough parameters; these conditions are assumed, not verified. Section 4.
  • domain assumption NIROPS aircraft IR perimeters are accurate ground truth for fire extent at the measurement time.
    Predicted perimeters are validated by comparing to NIROPS perimeters; any bias or timing error in these perimeters propagates into the reported SC, POD, and FAR values. Section 5e.
  • domain assumption Validation fires selected by the criteria in Section 5a are representative of the target deployment population.
    Fires are chosen only if their first 48 hour extent fits in the 12.8 km domain and IR perimeters are available, which may exclude fast-growing or remote fires and bias the average scores. Section 5a.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Algorithms for Wildfire Progression Reconstruction from Multi-Modal Satellite Active Fire Measurements and Terrain Height." pith.science (2026). https://pith.science/paper/R6TOWFK4

@misc{pith2026250610404,
  author       = {Pith},
  title        = {Pith review of: Generative Algorithms for Wildfire Progression Reconstruction from Multi-Modal Satellite Active Fire Measurements and Terrain Height},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6TOWFK4}},
  note         = {Machine review of arXiv:2506.10404}
}
read the original abstract

Increasing wildfire occurrence has spurred growing interest in wildfire spread prediction. However, even the most complex wildfire models diverge from observed progression during multi-day simulations, motivating need for data assimilation. A useful approach to assimilating measurement data into complex coupled atmosphere-wildfire models is to estimate wildfire progression from measurements and use this progression to develop a matching atmospheric state. In this study, an approach is developed for estimating fire progression from VIIRS active fire measurements, GOES-derived ignition times, and terrain height data. A conditional Generative Adversarial Network is trained with simulations of historic wildfires from the atmosphere-wildfire model WRF-SFIRE, thus allowing incorporation of WRF-SFIRE physics into estimates. Fire progression is succinctly represented by fire arrival time, and measurements for training are obtained by applying an approximate observation operator to WRF-SFIRE solutions, eliminating need for satellite data during training. The model is trained on tuples of fire arrival times, measurements, and terrain, and once trained leverages measurements of real fires and corresponding terrain data to generate samples of fire arrival times. The approach is validated on five Pacific US wildfires, with results compared against high-resolution perimeters measured via aircraft, finding an average Sorensen-Dice coefficient of 0.81. The influence of terrain height on the arrival time inference is also evaluated and it is observed that terrain has minimal influence when the inference is conditioned on satellite measurements.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 19 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]

    \"O ktem, 2018: Deep bayesian inversion

    Adler, J., and O. \"O ktem, 2018: Deep bayesian inversion. arXiv preprint arXiv:1811.05910, doi:10.48550/arXiv.1811.05910

    arXiv doi:10.48550/arxiv.1811.05910 2018

  4. [4]

    A., 1976: Estimating wildfire behavior and effects

    Albini, F. A., 1976: Estimating wildfire behavior and effects. General technical report, INT-30. USDA Forest Service, Intermountain Forest and Range Experiment Station

    work page 1976

  5. [5]

    Bakhshaii, A., and E. A. Johnson, 2019: A review of a new generation of wildfire--atmosphere modeling. Can.\ J.\ For.\ Res., 49 (6), 565--574, doi:10.1139/cjfr-2018-0138

    doi:10.1139/cjfr-2018-0138 2019

  6. [6]

    Driscoll, S

    Burke, M., A. Driscoll, S. Heft-Neal, J. Xue, J. Burney, and M. Wara, 2021: The changing risk and burden of wildfire in the united states. Proceedings of the National Academy of Sciences, 118 (2), e2011048\,118

    work page 2021

  7. [7]

    AccuWeather, ://www.accuweather.com/en/weather-news/accuweather-estimates-more-than-250-billion-in-damages-and-economic-loss-from-la-wildfires/1733821

    Danielle, M., 2025: Accuweather estimates more than \ 250 billion in damages and economic loss from la wildfires. AccuWeather, ://www.accuweather.com/en/weather-news/accuweather-estimates-more-than-250-billion-in-damages-and-economic-loss-from-la-wildfires/1733821

    work page arXiv 2025

  8. [8]

    Dennison, P. E., S. C. Brewer, J. D. Arnold, and M. A. Moritz, 2014: Large wildfire trends in the western United States , 1984--2011. Geophysical Research Letters, 41 (8), 2928--2933, doi:10.1002/2014GL059576

    work page doi:10.1002/2014gl059576 2014

Show all 26 references
  1. [9]

    Shlens, and M

    Dumoulin, V., J. Shlens, and M. Kudlur, 2016: A learned representation for artistic style. arXiv preprint arXiv:1610.07629, doi:10.48550/arXiv.1610.07629

    2016 doi arXiv

  2. [10]

    Mandel, J

    Farguell, A., J. Mandel, J. Haley, D. V. Mallia, A. Kochanski, and K. Hilburn, 2021: Machine learning estimation of fire arrival time from Level-2 Active Fires satellite data. Remote Sens., 13 (11), 2203, doi:10.3390/rs13112203

    2021 doi

  3. [11]

    Flannigan, M., A. S. Cantin, W. J. De Groot, M. Wotton, A. Newbery, and L. M. Gowman, 2013: Global wildland fire season severity in the 21st century. Forest Ecology and Management, 294, 54--61

    2013

  4. [12]

    Greenfield, P. H., W. Smith, and D. C. Chamberlain, 2003: Phoenix-the new forest service airborne infrared fire detection and mapping system. 2nd Int. Wildland Fire Ecology and Fire Management Congress and the 5th Symposium on Fire and Forest Meteorology

    2003

  5. [13]

    Huang, G., Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, 2017: Densely connected convolutional networks. Proceedings of the IEEE conference on computer vision and pattern recognition, 4700--4708, doi:10.48550/arXiv.1608.06993

    2017 doi arXiv

  6. [14]

    P., and C

    Lareau, N. P., and C. B. Clements, 2016: Environmental controls on pyrocumulus and pyrocumulonimbus initiation and development. Atmospheric Chemistry and Physics, 16 (6), 4005--4022

    2016

  7. [15]

    P., and C

    Lareau, N. P., and C. B. Clements, 2017: The mean and turbulent properties of a wildfire convective plume. J.\ Appl.\ Meteor.\ Climatol., 56 (8), 2289--2299, doi:10.1175/JAMC-D-16-0384.1

    2017 doi

  8. [16]

    Beezley, and A

    Mandel, J., J. Beezley, and A. Kochanski, 2011: Coupled atmosphere-wildland fire modeling with wrf-fire version 3.3. Geoscientific Model Development Discussions, 4 (1), 497--545

    2011

  9. [17]

    Mandel, J., J. D. Beezley, A. K. Kochanski, V. Y. Kondratenko, and M. Kim, 2012: Assimilation of perimeter data and coupling with fuel moisture in a wildland fire--atmosphere dddas. Procedia Computer Science, 9, 1100--1109, doi:10.1016/j.procs.2012.04.119

    2012 doi

  10. [18]

    Natural Hazards and Earth System Sciences, 14 (10), 2829--2845, doi:10.5194/nhess-14-2829-2014

    Mandel, J., and Coauthors, 2014: Recent advances and applications of WRF--SFIRE . Natural Hazards and Earth System Sciences, 14 (10), 2829--2845, doi:10.5194/nhess-14-2829-2014

    2014 doi

  11. [19]

    Z \'a rate, E

    Pastor, E., L. Z \'a rate, E. Planas, and J. Arnaldos, 2003: Mathematical models and calculation systems for the study of wildland fire behaviour. Progress in Energy and Combustion Science, 29 (2), 139--153

    2003

  12. [20]

    Murgoitio-Esandi, A

    Ray, D., J. Murgoitio-Esandi, A. Dasgupta, and A. A. Oberai, 2023: Solution of physics-based inverse problems using conditional generative adversarial networks with full gradient penalty. arXiv preprint arXiv:2306.04895, doi:10.48550/arXiv.2306.04895

    2023 doi arXiv

  13. [21]

    C., 1972: A mathematical model for predicting fire spread in wildland fuels

    Rothermel, R. C., 1972: A mathematical model for predicting fire spread in wildland fuels. Research Paper, INT-115. US Department of Agriculture, Intermountain Forest and Range Experiment Station

    1972

  14. [22]

    Hoffman, E

    Schmidt, C., J. Hoffman, E. Prins, and S. Lindstrom, 2010: Goes-r advanced baseline imager (abi) algorithm theoretical basis document for fire/hot spot characterization, version 2.0, noaa, silver spring, md. NOAA NESDIS, Cent. Satell. Appl. Res

    2010

  15. [23]

    Giglio, and J

    Schroeder, W., L. Giglio, and J. Hall, 2024: Collection 2 visible infrared imaging radiometer suite (viirs) 375-m active fire product user’s guide version 1.0

    2024

  16. [24]

    Artificial Intelligence for the Earth Systems

    Shaddy, B., and Coauthors, 2024: Generative algorithms for fusion of physics-based wildfire spread models with satellite data for initializing wildfire forecasts. Artificial Intelligence for the Earth Systems

    2024

  17. [25]

    Nat.\ Sustain., 4 (3), 252--260, doi:10.1038/s41893-020-00646-7

    Wang, D., and Coauthors, 2021: Economic footprint of California wildfires in 2018. Nat.\ Sustain., 4 (3), 252--260, doi:10.1038/s41893-020-00646-7

    2021 doi

  18. [26]

    Westerling, A. L., H. G. Hidalgo, D. R. Cayan, and T. W. Swetnam, 2006: Warming and earlier spring increase western US forest wildfire activity. science, 313 (5789), 940--943, doi:10.1126/science.1128834

    2006 doi

Pith tools

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