Pith. sign in

REVIEW 2 major objections 7 minor 29 references

High-Resolution Live Fuel Moisture Content (LFMC) Maps for Wildfire Risk from Multimodal Earth Observation Data

T0 review · 2 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning a pretrained multimodal Earth-observation model on 41,214 field samples yields accurate wall-to-wall live fuel moisture maps at 10 m resolution across the U.S., with roughly 20 percent lower error than a randomly initialized…

desk verdict A useful, honest application paper whose relative claim (pretrained > random init) holds, but whose absolute accuracy for unseen locations is not yet established because the random split leaks site information. read the letter →

arxiv 2506.20132 v2 pith:K2634PIO submitted 2025-06-25 cs.LG

Patrick Alan Johnson , Gabriel Tseng , Yawen Zhang , Heather Heward , Virginia Sjahli , Favyen Bastani , Joseph Redmon , Patrick Beukema This is my paper
classification cs.LG
keywords LiveFuelMoistureContentwildfireriskremotesensinggeospatialfoundationmodelsmultimodalEarthobservationGalileomodel10-metermappingGlobe-LFMC2.0
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

Live Fuel Moisture Content (LFMC) — the share of water in live vegetation — is a strong predictor of wildfire ignition and spread, but field sampling is slow and sparse. The paper claims that fine-tuning a pretrained, highly multimodal Earth-observation model (Galileo) on the public Globe-LFMC 2.0 dataset yields accurate, spatially complete LFMC maps at 10-meter resolution across the continental United States, with a test-set root mean squared error of 18.91 percentage points and an $R^2$ of 0.72. The key comparison is against a randomly initialized model of the same architecture, which gives an RMSE of 23.61; the pretrained model's roughly 20 percent lower error is offered as evidence that geospatial pretraining transfers to this regression task. If correct, this makes a practical pipeline for generating on-demand, wall-to-wall fuel-moisture maps for fire research, prescribed-burn planning, and incident response.

What carries the argument

The load-bearing object is the Galileo model (Galileo-Tiny, 5.3M parameters), a vision-transformer-based geospatial foundation model pretrained on ten remote-sensing products — Sentinel-2 optical bands plus NDVI, Sentinel-1 radar, VIIRS night lights, ERA5-Land weather, TerraClimate water-balance variables, SRTM elevation and slope, and latitude/longitude — organized along spatial and temporal axes. Its pretraining exposes the encoder to a broad distribution of modalities and timesteps; fine-tuning under mean squared error on the filtered Globe-LFMC 2.0 samples adapts that representation to LFMC regression. The machinery does two jobs: it supplies spatial and temporal context that a small randomly initialized model lacks, and it keeps predictions stable when modalities are missing at fine-tuning or inference time, which matters for cloud-covered or data-poor scenes.

What would settle it

Retrain the model with a spatial leave-one-site-out or block split that keeps all samples from a given area or site in one fold; if the held-out RMSE rises substantially above 18.91, or the geographic pattern of errors changes, the random-split numbers are not representative of new-region accuracy.

Watch

Extended reading notes

Core claim

The paper's central claim is that a pretrained multimodal remote-sensing transformer, fine-tuned on 41,214 CONUS ground samples from Globe-LFMC 2.0 (2017–2023), can predict LFMC at 10 m resolution with an overall MAE of 12.58, an RMSE of 18.91, and an $R^2$ of 0.72 on a randomly split test set. This outperforms the same architecture trained from random weights (RMSE 23.61, $R^2$ 0.57) and a monthly-average baseline (RMSE 33.66, $R^2$ 0.11). The authors argue that pretraining is the cause of the gain, and they support that interpretation by showing that removing any one input modality degrades the pretrained model only slightly, while the randomly initialized model loses up to 8 percent of RMSE performance when TerraClimate is removed. They further claim the model generalizes across seasons and land cover classes, and that case-study maps for the 2025 Eaton and Palisades fire areas match expert expectations, even though those areas had no labels.

Load-bearing premise

The evaluation assumes that a random split of the 41,214 samples into training and test sets measures true generalization; if nearby locations in the test set are correlated with training locations, the reported 18.91 RMSE and 0.72 $R^2$ may overstate how well the model predicts in regions it has never seen.

Editorial extensions

If this is right

  • Users can generate on-demand LFMC maps for arbitrary spatiotemporal bounding boxes across the CONUS domain, including areas with no ground labels.
  • The 10-meter resolution is a step change over prior LFMC products at 250–500 meters, enabling fuel-moisture patterns to be resolved at the scale of individual land parcels and fire perimeters.
  • The fine-tuned model stays within an $R^2$ of 0.68–0.72 even when spatial context is reduced to a single pixel and temporal context to three timesteps, so mapping can be run for small or data-poor areas.
  • Dropping any single input, including weather or topography, changes pretrained-model RMSE by less than 1.3 percentage points, so operational maps can still be produced when a data source is unavailable.

Reading between the lines

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

  • If the random split leaks spatial information, a site-disjoint evaluation would likely raise the reported RMSE; that test is the natural next experiment and would also tell users how far the 10 m maps can be trusted in regions far from the training sites.
  • Because pretraining appears to stabilize predictions when inputs are missing, a weekly or daily version of the pipeline—rather than monthly averages—could be feasible for operational forecasting, a step the paper itself flags as future work.
  • The same recipe could be transferred to other regions by fine-tuning on non-CONUS Globe-LFMC samples, which would probe whether the pretrained representation generalizes globally or only within the U.S. distribution.
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

2 major / 7 minor

Summary. The paper introduces a pipeline for generating 10 m resolution wall-to-wall Live Fuel Moisture Content (LFMC) maps over the conterminous U.S. by fine-tuning Galileo-Tiny, a multimodal pretrained geospatial foundation model, on the Globe-LFMC 2.0 dataset. On a random 70/15/15 split of 41,214 samples, the authors report an RMSE of 18.91, MAE of 12.58, and R2 of 0.72, representing about a 20% RMSE improvement over a randomly initialized model of the same architecture and a large gain over a monthly-average baseline. The paper also includes an input-ablation study and two qualitative case studies for the 2025 Eaton and Palisades fires.

Significance. The central empirical claim—that fine-tuning a pretrained multimodal geospatial model substantially outperforms a randomly initialized model on LFMC regression—is supported by the experimental design, and the relative comparison is not threatened by the split issue. The open pipeline and the high-resolution mapping capability are timely and useful. The evidence includes a random-init control, an input-shape sensitivity analysis, and a missing-modality ablation; these are appropriate internal controls. However, the absolute accuracy figures are weakened by the acknowledged spatial autocorrelation in residuals (Moran's I 0.057, p=0.001) arising from the random split, so the paper's claim of accurate maps for previously unseen locations requires either a spatial holdout evaluation or a more cautious interpretation.

major comments (2)
  1. [3.3.1 and 4.1.1] The random split of 41,214 samples from 1,031 sites likely places the same locations in both training and test sets, since multiple observations per site are collected across seasons and years. The paper's own residual analysis reports Moran's I = 0.057 with p = 0.001, and the authors acknowledge that this 'suggests potential information leakage from the random split' (§4.1.1). Because the model includes latitude/longitude and static SRTM features (§3.2.1), the test set can reward memorizing site-specific conditions rather than learning transferable relationships. The reported RMSE 18.91 and R2 0.72 are thus not reliable estimates of accuracy on previously unseen locations, which is precisely the target of the wall-to-wall mapping application. Please add a spatially stratified or site-holdout evaluation, or explicitly reframe the headline numbers as random-split performance and discuss the expected degradation for new regions.
  2. [4.1.3, Table 6] The text states that removing TerraClimate from the randomly initialized model yields 'a significant degradation in performance (8% reduction in RMSE)', but Table 6 shows that RMSE increases from 23.61 to 25.57, i.e., an 8.3% increase in RMSE, not a reduction; the 14% degradation is in R2 (0.57 to 0.49). Please correct this sentence and clarify which metric is being reported. This matters because the robustness claim for pretrained models is based on a correct reading of Table 6.
minor comments (7)
  1. [Abstract] The abstract says '20 reduction in RMSE' and should read '20% reduction in RMSE'.
  2. [3.2.1] In the bullet list, 'Location DataWe provide' is missing a space; it should be 'Location Data: We provide'.
  3. [Table 6] The header 'W/O IN' is ambiguous; use 'Removed Input' or 'Missing Input' instead.
  4. [4.1.1] The claim that results are 'comparable to, and in some cases exceed, the performance of existing LFMC models' is not supported by a direct comparison on identical data; consider removing or softening this statement, since prior models were evaluated at coarser resolutions and with different train/test splits.
  5. [4.1.1] Please report the number of nearest neighbors k used for the K-NN spatial weights matrix and the distance metric, to make the spatial autocorrelation test reproducible.
  6. [3.3.2] Fine-tuning details such as learning rate, batch size, optimizer, and the number of epochs until early stopping are missing; include them for reproducibility.
  7. [4.1.2, Table 5] The sensitivity analysis to input shape would benefit from confidence intervals or repeated runs to confirm that the differences are not noise, especially since only one random seed appears to be used.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central improvement claim is empirically evaluated against a random-initialized control; only a minor non-load-bearing self-citation of the Galileo model is present.

full rationale

The paper's central quantitative claim—that fine-tuning the pretrained Galileo model achieves an RMSE of 18.91, MAE of 12.58, R2 of 0.72, and about a 20% RMSE reduction over a randomly initialized model of the same architecture—is tested on a held-out 15% test split against both a random-initialized control and a monthly-average baseline (Table 1, Section 4.1.1). This comparison does not reduce by construction: the pretrained weights are not fitted to the test labels, and the random-weight baseline provides an independent empirical benchmark for the value of pretraining. The LFMC labels come from the external Globe-LFMC 2.0 dataset (Yebra et al., 2024), and the remote-sensing inputs are standard Earth observation products, so the prediction target is not defined in terms of the model's outputs. The only circularity-adjacent point is that the Galileo model is cited from Tseng et al. (2025) with overlapping authors; however, the paper does not rely on that citation for its quantitative conclusion, which is established by the direct pretrained-versus-randomized experiment. The acknowledged Moran's I = 0.057 (p = 0.001) spatial autocorrelation and possible information leakage from the random split are evaluation-protocol concerns about spatial generalization, not circular derivation; the paper explicitly defers spatial partitioning to future work. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is smuggled in via self-citation. The score of 2 reflects only the minor, non-load-bearing self-citation rather than any actual circularity in the derivation chain.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The analysis rests externally on the Galileo pretrained model (same group, overlapping authors), the Globe-LFMC 2.0 dataset, and public satellite products. The model itself contributes fine-tuned weights as fitted parameters, plus hand-chosen preprocessing values (302% cap, 1 km window, 32x32x12 input, 70/15/15 split). No novel physical entities are introduced. The paper does not derive LFMC from first principles; it fits a supervised regressor.

free parameters (4)
  • Fine-tuned Galileo-Tiny weights = 5.3M parameters, not enumerated
    The regression model's weights are fit to the 70% training split under MSE loss; this is the central fitting of the empirical claim.
  • LFMC capping threshold = 302%
    Section 3.1: labels capped at 302%, rounded up from the 99.9th percentile, to limit outlier influence; changes the target distribution.
  • Default input shape = 32x32 pixels, 12 monthly timesteps
    Section 4.1.2 Table 5: the default spatial/temporal window; performance varies by up to ~1.4 RMSE across shapes, so this hand-chosen value affects results modestly.
  • Random split ratio = 70/15/15
    Section 3.3.1: train/validation/test proportions; evaluation outcomes depend on this split, and the paper's Moran's I analysis suggests potential spatial leakage.
assumptions (4)
  • domain assumption Galileo pretraining provides transferable representations for LFMC estimation.
    Section 3.2: the method assumes the self-supervised pretraining of Tseng et al. (2025) on multimodal EO data yields features relevant to vegetation moisture; the random-init baseline tests this but the pretrained model itself is taken as an artifact.
  • domain assumption Globe-LFMC 2.0 labels are accurate measurements of LFMC at their stated locations and dates.
    Section 3.1: all training and test targets come from this dataset; no account is taken of sampling or laboratory error.
  • domain assumption The random split gives unbiased generalization estimates.
    Section 3.3.1: the split is random rather than spatial; the paper's own Moran's I result (Section 4.1.1) indicates weak positive spatial autocorrelation, so this assumption is partially violated.
  • domain assumption A 1 km x 1 km bounding box around each sample captures the conditions represented by the sample.
    Section 3.3.1: satellite data are averaged over the box; mismatch between exact field sampling footprints and the window could bias labels, though Table 5 probes window-size sensitivity only through the model input shape, not label extraction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Resolution Live Fuel Moisture Content (LFMC) Maps for Wildfire Risk from Multimodal Earth Observation Data." pith.science (2026). https://pith.science/paper/K2634PIO

@misc{pith2026250620132,
  author       = {Pith},
  title        = {Pith review of: High-Resolution Live Fuel Moisture Content (LFMC) Maps for Wildfire Risk from Multimodal Earth Observation Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2634PIO}},
  note         = {Machine review of arXiv:2506.20132}
}
read the original abstract

Wildfires are increasing in intensity and severity at an alarming rate. Recent advances in AI and publicly available satellite data enable monitoring critical wildfire risk factors globally, at high resolution and low latency. Live Fuel Moisture Content (LFMC) is a critical wildfire risk factor and is valuable for both wildfire research and operational response. However, ground-based LFMC samples are both labor intensive and costly to acquire, resulting in sparse and infrequent updates. In this work, we explore the use of a pretrained, highly-multimodal earth-observation model for generating large-scale spatially complete (wall-to-wall) LFMC maps. Our approach achieves significant improvements over previous methods using randomly initialized models (20 reduction in RMSE). We provide an automated pipeline that enables rapid generation of these LFMC maps across the United States, and demonstrate its effectiveness in two regions recently impacted by wildfire (Eaton and Palisades).

Figures

Figures reproduced from arXiv: 2506.20132 by the authors.

Figure 1
Figure 1. Meteorological season breakdown of the Globe [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Location breakdown of Globe-LFMC 2.0 dataset for CONUS samples 2017-2023 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Coverage of the Globe-LFMC 2.0 dataset for [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Elevation range breakdown of the Globe-LFMC [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 6
Figure 6. Figure 6: Percentage error for the test dataset. Galileo [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: LFMC predictions for the San Gabriel Moun [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: LFMC predictions for the Santa Monica Moun [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 8
Figure 8. Figure 8: Monthly LFMC average predictions for two wild [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 21 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    T., Dobrowski, S

    Abatzoglou, J. T., Dobrowski, S. Z., Parks, S. A., and Hegewisch, K. C. Terraclimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015. Scientific Data, 5, 2018

    work page 1958

  3. [3]

    Anysat: An earth observation model for any resolutions, scales, and modalities

    Astruc, G., Gonthier, N., Mallet, C., and Landrieu, L. Anysat: An earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024

    arXiv 2024

  4. [4]

    Conversion of fuel moisture content values to ignition potential for integrated fire danger assessment

    Chuvieco, E., Aguado, I., and Dimitrakopoulos, A. Conversion of fuel moisture content values to ignition potential for integrated fire danger assessment. Canadian Journal of Forest Research, 34: 0 2284--2293, 2004

    work page 2004

  5. [5]

    Dionelis, N., Fibaek, C., Camilleri, L., Luyts, A., Bosmans, J., and Saux, B. L. Evaluating and benchmarking foundation models for earth observation and geospatial ai. ArXiv, abs/2406.18295, 2024

    arXiv 2024

  6. [6]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

  7. [7]

    Farr, T. G. and Kobrick, M. Shuttle radar topography mission produces a wealth of data. Eos, Transactions American Geophysical Union, 81: 0 583--585, 2000

    work page 2000

  8. [8]

    D., McNorton, J., Lombardi, A., and Wetterhall, F

    Giuseppe, F. D., McNorton, J., Lombardi, A., and Wetterhall, F. Global data-driven prediction of fire activity. Nature Communications, 16, 2025

    work page 2025

Show all 29 references
  1. [9]

    Google earth engine: Planetary-scale geospatial analysis for everyone

    Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., and Moore, R. Google earth engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202: 0 18--27, 2017

    2017

  2. [10]

    Hosten, P. E. and Stringham, T. K. Prescribed fire basics: Fire weather. https://extension.oregonstate.edu/catalog/pub/em-9385-prescribed-fire-basics-fire-weather, 2020. EM 9385

    2020

  3. [11]

    Jakubik, J., Roy, S., Phillips, C., Fraccaro, P., Godwin, D., Zadrozny, B., Szwarcman, D., Gomes, C., Nyirjesy, G., Edwards, B., Kimura, D., Simumba, N., Chu, L., Mukkavilli, S. K., Lambhate, D., Das, K., Bangalore, R., Oliveira, D., Muszynski, M., Ankur, K., Ramasubramanian, ...

    arXiv

  4. [12]

    M., Freeborn, P

    Jolly, W. M., Freeborn, P. H., Bradshaw, L. S., Wallace, J., and Brittain, S. Modernizing the us national fire danger rating system (version 4): Simplified fuel models and improved live and dead fuel moisture calculations. Environ. Model. Softw., 181: 0 106181, 2024

    2024

  5. [13]

    I., Bistinas, I., Whitley, R., Burton, C

    Kelley, D. I., Bistinas, I., Whitley, R., Burton, C. A., Marthews, T. R., and Dong, N. How contemporary bioclimatic and human controls change global fire regimes. Nature Climate Change, pp.\ 1--7, 2019

    2019

  6. [14]

    L., Madrigal, J., Guijarro, M., and Hernando, C

    Marino, E., Yebra, M., Guill \'e n-Climent, M., Algeet, N., Tom \'e , J. L., Madrigal, J., Guijarro, M., and Hernando, C. Investigating live fuel moisture content estimation in fire-prone shrubland from remote sensing using empirical modelling and rtm simulations. Remote Sensi...

    2020

  7. [15]

    Miller, L., Zhu, L., Yebra, M., R \"u diger, C., and Webb, G. I. Projecting live fuel moisture content via deep learning. International Journal of Wildland Fire, 32 0 (5): 0 709--727, 2023

    2023

  8. [16]

    Notes on continuous stochastic phenomena

    Moran, P. Notes on continuous stochastic phenomena. Biometrika, 37 1-2: 0 17--23, 1950

    1950

  9. [17]

    Era5-land monthly averaged data from 1950 to present

    Muñoz Sabater, J. Era5-land monthly averaged data from 1950 to present. https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-land-monthly-means, 2019. Accessed 2025

    1950

  10. [18]

    Wfigs current interagency fire perimeters

    National Interagency Fire Center . Wfigs current interagency fire perimeters. https://data-nifc.opendata.arcgis.com/datasets/nifc::wfigs-current-interagency-fire-perimeters/about, 2025. Accessed 13 Jan 2025

    2025

  11. [19]

    P., Flefil, J

    Rao, K., Williams, A. P., Flefil, J. F., and Konings, A. G. Sar-enhanced mapping of live fuel moisture content. Remote Sensing of Environment, 245: 0 111797, 2020

    2020

  12. [20]

    P., Diffenbaugh, N

    Rao, K., Williams, A. P., Diffenbaugh, N. S., Yebra, M., Bryant, C., and Konings, A. G. Dry live fuels increase the likelihood of lightning-caused fires. Geophysical Research Letters, 50 0 (15): 0 e2022GL100975, 2023

    2023

  13. [21]

    Rossa, C. G. The effect of fuel moisture content on the spread rate of forest fires in the absence of wind or slope. International Journal of Wildland Fire, 26: 0 24--31, 2017

    2017

  14. [22]

    E., Blumenstiel, B., Ghosal, R., de Oliveira, P

    Szwarcman, D., Roy, S., Fraccaro, P., G \'i slason, T. E., Blumenstiel, B., Ghosal, R., de Oliveira, P. H., de Sousa Almeida, J. L., Sedona, R., Kang, Y., Chakraborty, S., Wang, S., Kumar, A., Truong, M., Godwin, D., Lee, H., Hsu, C.-Y., Asanjan, A. A., Mujeci, B., Keenan, T.,...

    2024

  15. [23]

    Lightweight, pre-trained transformers for remote sensing timeseries

    Tseng, G., Cartuyvels, R., Zvonkov, I., Purohit, M., Rolnick, D., and Kerner, H. Lightweight, pre-trained transformers for remote sensing timeseries. arXiv preprint arXiv:2304.14065, 2023

    2023 arXiv

  16. [24]

    R., Shelhamer, E., Kerner, H., and Rolnick, D

    Tseng, G., Fuller, A., Reil, M., Herzog, H., Beukema, P., Bastani, F., Green, J. R., Shelhamer, E., Kerner, H., and Rolnick, D. Galileo: Learning global and local features in pretrained remote sensing models. ArXiv, abs/2502.09356, 2025

    2025 arXiv

  17. [25]

    Tucker, C. J. Red and photographic infrared linear combinations for monitoring vegetation. Remote sensing of Environment, 8 0 (2): 0 127--150, 1979

    1979

  18. [26]

    Vatsavai, R. R. Geospatial foundation models: Recent advances and applications. Proceedings of the 12th ACM SIGSPATIAL International Workshop on Analytics for Big Geospatial Data, 2024

    2024

  19. [27]

    Yebra, M., Scortechini, G., Adeline, K. R. M., Aktepe, N., Almoustafa, T. A., Bar‐Massada, A., Beget, M. E., Boer, M. M., Bradstock, R., Brown, T., Castro, F. X., Chen, R., Chuvieco, E., Danson, M. F., Değirmenci, C. \"U ., Delgado-D \'a vila, R., Dennison, P. E., Bella, C. M....

    2024

  20. [28]

    Zanaga, D., Kerchove, R. V. D., de Keersmaecker, W., Souverijns, N., Brockmann, C., Quast, R., Wevers, J., Grosu, A. C., Paccini, A., Vergnaud, S., Cartus, O., Santoro, M., Fritz, S., Georgieva, I., Lesiv, M., Carter, S., Herold, M., Li, L., Tsendbazar, N.-E., Ramoino, F., and...

    2020

  21. [29]

    I., Yebra, M., Scortechini, G., Miller, L., and Petitjean, F

    Zhu, L., Webb, G. I., Yebra, M., Scortechini, G., Miller, L., and Petitjean, F. Live fuel moisture content estimation from modis: A deep learning approach. ISPRS Journal of Photogrammetry and Remote Sensing, 179: 0 81--91, 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.