Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

PrediTree is presented as the first openly available dataset pairing sub-meter multi-temporal RGB-NIR imagery with later LiDAR-derived canopy height maps, letting models predict tree height from past observations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A sub-meter, multi-temporal, multi-spectral canopy height dataset for training tree height prediction models, released openly with benchmark results.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection PrediTree's dataset claim (first open sub-meter multi-temporal multi-spectral CHM dataset) holds up on inspection, but the paper's experimental reporting is sloppy and the target CHM year is an unanalyzed tile-level mean that could bias the growth labels. the 4 major comments →

arxiv 2509.01202 v2 pith:ZZR73HSO submitted 2025-09-01 cs.CV

PrediTree: A Multi-Temporal Sub-meter Dataset of Multi-Spectral Imagery Aligned With Canopy Height Maps

Hiyam Debary , Mustansar Fiaz , Levente Klein This is my paper
classification cs.CV
keywords tree growth predictioncanopy height mapmulti-temporal remote sensingLiDARmulti-spectral imagerysub-meter resolutiondeep learning benchmarkforest monitoring
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 introduces PrediTree, a large open dataset built to train models that predict tree height from historical imagery. It aligns three yearly snapshots of 0.5 m RGB, near-infrared, and NDVI imagery across French forests with a later LiDAR-derived canopy height map at the same resolution, yielding over three million 256 by 256 tiles. The authors argue this is the first openly available dataset combining sub-meter resolution, multi-spectral imagery, and multiple temporal snapshots with a target height map, a combination existing datasets lack. To use it, they propose an encoder-decoder framework that takes the three image timestamps plus the relative time differences in years to the target height map and outputs a predicted canopy height. In their experiments, a U-Net-based model reaches the lowest masked error, and ablations show all three spectral bands and all three timestamps help.

Core claim

The central claim is that the task of tree growth prediction—forecasting future canopy height from past observations—can be supported by a public dataset in which multi-spectral aerial imagery at three past times is spatially aligned with a later 0.5 m LiDAR canopy height map. The paper constructs this dataset from French national LiDAR and orthophoto data, derives canopy height by subtracting a smoothed terrain model from the surface model, merges overlapping tiles, and keeps three random timestamps for departments with more than three. The companion model encodes each timestamp's RGB-NIR-NDVI stem features, fuses them across time, injects the relative year gaps between each image and the t

What carries the argument

The load-bearing object is the temporal pairing itself: each sample consists of three 256 by 256, 0.5 m images with five bands (RGB, NIR, NDVI) acquired at times t1, t2, and t3, plus a spatially aligned canopy height map at a later time ty, together with the relative time differences Δti = ty − ti. The model is an encoder-decoder: per-timestamp stem encoders produce spatial features, a spatio-temporal fusion block combines them, and embeddings of the Δt values are fused in to condition the prediction on how much growth time has elapsed; a weighted masked MSE loss concentrates learning on vegetation pixels. The Δt conditioning is what turns static height estimation into growth prediction.

Load-bearing premise

The model's notion of growth time is only as accurate as the year attached to the target canopy height map, which is computed as the mean acquisition year of merged LiDAR tiles; if those tiles were captured in different years, the time gaps Δt are wrong and the learned 'growth' signal is corrupted.

What would settle it

For any PrediTree tile, look up the original LiDAR tile acquisition years instead of the merged mean; recompute the model's Δt inputs with the true dates and compare predictions. If replacing the mean-year label with exact acquisition years produces materially different height predictions, the dataset's temporal reference does not support the claimed growth modeling; if results are unchanged, the approximation is harmless.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Tree height at a future time can be predicted without future LiDAR, using only historical multi-spectral imagery and elapsed time, if the dataset's pairing is learned.
  • The relative-time conditioning makes the same trained model applicable to sites whose snapshots are separated by 1–3 years, not just the exact dates in the dataset.
  • The public processing pipeline lets users regenerate the dataset at 0.15–0.5 m resolutions and extend it to other French departments, so the benchmark can grow beyond its current 33 departments.
  • The ablation results imply that adding spectral bands (RGB, NIR, NDVI) and more timestamps improves prediction, giving a concrete recipe for future growth models.
  • If it is truly the first open sub-meter multi-temporal LiDAR-CHM dataset, PrediTree becomes the natural training and evaluation ground for crown-level growth forecasting.

Where Pith is reading between the lines

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

  • Because the target year is the mean acquisition year of merged LiDAR tiles, the effective Δt values are uncertain to about half the spread of the merged tiles; growth measured over 1–3 years may be confounded unless the actual per-tile years are used or validated.
  • The reported model results come from 30k training samples of 785k; the architecture ranking could shift on the full dataset, and the dataset's quality as a benchmark will depend on standardized train/test splits that use exact tile metadata.
  • One testable extension is to feed the model a future image and ask it to output the height change between two historical dates, using the same Δt encoding; the dataset could then be evaluated for change detection as well as forward prediction.
  • Linking PrediTree-style imagery to process-based growth models like the 3-PG lineage mentioned in the paper may let machine-learned growth predictions be constrained by physiological plausibility, reducing the need for massive paired LiDAR coverage.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. PrediTree introduces a new open dataset pairing sub-meter (0.5 m) LiDAR-derived canopy height maps with multi-temporal, multi-spectral (RGB, NIR, NDVI) aerial imagery from 33 French departments, totaling over 3.1M aligned 256x256 tiles. The paper also proposes an encoder-decoder framework that ingests three past image timestamps plus relative year differences Δt_i = t_y − t_i and predicts a canopy height map at a later target time t_y. Experiments compare U-Net and several DeepLabv3-based encoders on a 30K-sample training subset and 393-sample test set, and ablations study spectral-band and temporal-input contributions. The central claim is that PrediTree is the first openly available dataset enabling sub-meter, multi-temporal tree-height prediction from historical imagery.

Significance. If the dataset construction is reliable, PrediTree would fill a real gap: existing high-resolution CHM datasets are mostly single-timestamp, while multi-temporal products are typically at 10–30 m resolution. The dataset is large, openly released, and accompanied by processing and training code, which are concrete strengths for reproducibility. The proposed temporal-conditioning formulation is natural and the ablation design is informative. However, the paper's central value depends on the accuracy of the target CHM timestamp t_y: the model's temporal supervision signal is the relative time difference between each image and the CHM, so any error in t_y directly propagates into the growth labels. The manuscript does not currently quantify or mitigate this error, and the experimental reporting contains errors that undermine confidence in the quantitative claims.

major comments (4)
  1. The target timestamp t_y is defined as the mean acquisition year across all merged CHM tiles, while overlapping tiles are merged by 'taking the most recent data at the overlap location.' This means different pixels within one CHM tile can originate from LiDAR flights in different years, and the scalar t_y cannot represent all of them. Since the model uses Δt_i = t_y − t_i as temporal conditioning, per-pixel errors in t_y directly corrupt the growth signal the dataset is designed to provide. The paper gives no distribution of LiDAR acquisition years within merged tiles, no statistics on within-tile timestamp spread, and no sensitivity analysis. Please either provide per-pixel acquisition timestamps or metadata that allows users to construct them, restrict the dataset to tiles with sufficiently homogeneous acquisition years, or otherwise demonstrate that the mean-year approximation does no
  2. The abstract states that U-Net 'provides the highest masked mean squared error of 11.78%,' but Table 2 reports U-Net's MSE as 6.47% and MAE as 11.78%. The number 11.78 is the MAE, not the MSE, and 'highest' should be 'lowest' since lower errors are better. This is not a typo in isolation: the conclusion section repeats the framing, and readers cannot tell which metric is being claimed. The abstract must be corrected to report MSE and MAE separately with the correct direction of comparison.
  3. All model comparisons and ablations are based on a single training run on a 393-sample test set, with no repeated seeds, confidence intervals, or statistical tests. Differences such as U-Net MSE 6.47 vs. ResNet-50 MSE 7.37 may be real, but with a small test set and no variance estimate the ranking is not established. Please add at least three seeds with mean±std, or a statistical test, and report test-set size and selection procedure. This is load-bearing for the claim that U-Net 'outperforms' the alternatives.
  4. Training uses a weighted masked MSE with constants k=10 and θ=0.5 m, while inference reports unweighted masked MSE and MAE restricted to pixels with y > θ. The relationship between the training objective and the reported metrics is clear, but the paper does not state whether θ is applied consistently in all reported metrics or how much of the 393 test tiles consists of tree pixels. Without the valid-pixel fraction, the absolute error values are hard to interpret. Please report the effective number of evaluated pixels per test sample and the fraction of masked pixels.
minor comments (5)
  1. The phrase 'cutting the error of the same experiments but on fewer bands (red, green, blue only), by around 30%' is ambiguous. Table 3 shows MAE reduction from 13.47 to 11.78 (about 12.5%) and MSE reduction from 9.23 to 6.47 (about 30%). Specify which metric the 30% refers to.
  2. Typo: 'encode the the relative time difference' should read 'encode the relative time difference.' Also 'passed to a header' should likely be 'passed to a head.'
  3. The captions describe only 'the first row' of each figure, but the figures contain multiple rows or panels. Please clarify what each row/panel shows, especially the color-coded detection row in Figure 6.
  4. The ablation table denotes timestamp sets with subscripts but the text sometimes says 't 1, t2, t3' without subscript formatting. Use consistent notation, and define whether 'Timestamps' refers to the number of input images or which specific acquisition years are used.
  5. The reference for the IGN LiDAR HD data is cited as a URL accessed March 2025. Consider adding a formal data citation with version or DOI if available.

Circularity Check

0 steps flagged

No circularity: the dataset is constructed from external IGN data and the model is evaluated on held-out tiles.

full rationale

This paper is a dataset-and-benchmark contribution, not a derivation whose conclusion is equivalent to its inputs. The multi-spectral imagery and LiDAR CHMs are sourced from external IGN data; the CHM is produced by subtracting a smoothed DTM from a DSM, independent of the imagery used as model inputs. The prediction target is the CHM at t_y, and inputs are past images plus relative time offsets; there is no equation in which the target is defined in terms of the inputs. The models are trained on a 30K subset and tested on 393 held-out samples, so the reported MSE/MAE are genuine out-of-sample numbers. No fitted parameter is renamed as a prediction, no self-citation carries an argument, and no uniqueness or ansatz is imported from prior work by the authors. The skeptical concern about using a mean LiDAR acquisition year as the CHM timestamp affects the accuracy of the Δt conditioning and is a data-quality limitation, not circularity: even if the timestamp approximation is crude, the evaluation remains a comparison against held-out CHM values. Therefore the paper has no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No invented entities. The dataset and model rely on standard remote sensing assumptions plus a few hand-chosen parameters. The most fragile choices are the 10m DTM smoothing and the mean-year timestamp approximation, both explicitly introduced for computational convenience.

free parameters (3)
  • loss weighting constants k and theta = k=10, theta=0.5m
    Chosen by hand to amplify loss where target height exceeds 0.5m (Section 4). These affect training but are not fitted to the test set.
  • DTM smoothing filter side length = 10m
    Square filter of side 10m chosen 'for computational efficiency purposes' (Section 3, Figure 1). Alters CHM values near data gaps such as water bodies.
  • target resolution = 0.5m
    Selected as the study resolution; configurable in the pipeline. A design choice affecting alignment with CHM pixels.
axioms (4)
  • domain assumption IGN LiDAR point cloud classification into ground and vegetation points is accurate.
    The entire CHM pipeline (Section 3) separates ground from vegetation points using PDAL; inaccuracies propagate to all CHM labels.
  • domain assumption The orthophoto and LiDAR tiles are precisely georeferenced and sub-meter alignment is achieved by resampling.
    Pixel-level pairing of RGB/NIR/NDVI and CHM assumes zero registration error (Section 3).
  • domain assumption CHM derived as DSM minus smoothed DTM represents true canopy height.
    Used as ground truth Y in Eq. 1; no validation against field measurements or independent LiDAR is reported (Section 3).
  • ad hoc to paper The mean acquisition year across merged CHM tiles is a valid target timestamp ty.
    Section 3 merges overlapping LiDAR tiles and takes the mean year as ty; if tiles span different years, the relative time differences Δt become inaccurate.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of PrediTree: A Multi-Temporal Sub-meter Dataset of Multi-Spectral Imagery Aligned With Canopy Height Maps." pith.science (2026). https://pith.science/paper/ZZR73HSO

@misc{pith2026250901202,
  author       = {Pith},
  title        = {Pith review of: PrediTree: A Multi-Temporal Sub-meter Dataset of Multi-Spectral Imagery Aligned With Canopy Height Maps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZR73HSO}},
  note         = {Machine review of arXiv:2509.01202}
}
Share X Bluesky LinkedIn Reddit HN
abstract

We present PrediTree, the first comprehensive open-source dataset designed for training and evaluating tree height prediction models at sub-meter resolution. This dataset combines very high-resolution (0.5m) LiDAR-derived canopy height maps, spatially aligned with multi-temporal and multi-spectral imagery, across diverse forest ecosystems in France, totaling 3,141,568 images. PrediTree addresses a critical gap in forest monitoring capabilities by enabling the training of deep learning methods that can predict tree growth based on multiple past observations. To make use of this PrediTree dataset, we propose an encoder-decoder framework that requires the multi-temporal multi-spectral imagery and the relative time differences in years between the canopy height map timestamp (target) and each image acquisition date for which this framework predicts the canopy height. The conducted experiments demonstrate that a U-Net architecture trained on the PrediTree dataset provides the highest masked mean squared error of $11.78\%$, outperforming the next-best architecture, ResNet-50, by around $12\%$, and cutting the error of the same experiments but on fewer bands (red, green, blue only), by around $30\%$. This dataset is publicly available on https://huggingface.co/datasets/hiyam-d/PrediTree, and both processing and training codebases are available on {GitHub}.

Figures

Figures reproduced from arXiv: 2509.01202 by Hiyam Debary, Levente Klein, Mustansar Fiaz.

Figure 1
Figure 1. Figure 1: Effect of DTM smoothing on CHM quality near water features. Left: RGB aerial imagery showing a forest area in￾tersected by a water body (center of the image). Center: Unprocessed DTM with￾out smoothing, exhibiting data gaps (black areas), particularly above the water body. Right: Improved DTM produced using a square filter of side 10m, effectively filling gaps where canopy information would oth￾erwise be t… view at source ↗
Figure 2
Figure 2. Figure 2: Flowchart of the three-phase method￾ology for dataset creation. Phase 1 (purple boxes) shows the RGB and NIR-RG data ac￾quisition process, from building an index to ex￾tracting coordinates. Phase 2 (green boxes) de￾picts the LiDAR data acquisition workflow, in￾cluding downloading LAZ files and generating CHM tiles. Phase 3 (orange boxes) represents the processing, pairing, and filtering steps that integrat… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the obtained products through the described pipeline. From left to right: RGB orthophotography, fake color representation (using NIR, R and G bands), DTM, DSM, and CHM. All products displayed at the same spatial resolution of 0.5m. The elevation products use a grayscale representation with lighter tones indicating higher values. extent of the RGB-NIR-NDVI products. This implementation merg… view at source ↗
Figure 4
Figure 4. Figure 4: Flowchart of the proposed framework, taking three timestamp input images (Ii) where i ∈ [1, 2, 3]. Each Ii input consists of RGB, NIR, and NDVI bands. These input images are input to the stem encoders to encode spatial stem features (Xi) for each timestamp. These spatial stem features are fused using the Spatio-Temporal Fusion block to obtain fused spatial-temporal embeddings (Xe). We also input the relati… view at source ↗
Figure 5
Figure 5. Figure 5: A sample with visible tree growth, outlined in red, from PrediTree. We collect multi-spectral (RGB+NIR+NDVI) data at three timestamps t1, t2, t3 and their corre￾sponding spatially aligned CHM tile at ty. spatio-temporal embeddings Xe to obtain X¯, which encodes both spatio-temporal and relative time difference embeddings. These embeddings are realized using a projector layer, which yields Xˆ, and forwarded… view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the CHM predictions for the sample in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Biomazon: A Multimodal Dataset for 3D Forest Structure and Biomass Modeling in the Amazon Basin

    cs.CV 2026-06 unverdicted novelty 7.0

    Biomazon is a multimodal benchmark dataset pairing GEDI RH profiles and AGBD targets with Sentinel, ALOS, DEM, and other predictors for joint 3D forest structure and biomass modeling in the Amazon.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Scientific Data12(1), 322 (2025)

    Allred, B.W., McCord, S.E., Morford, S.L.: Canopy height model and naip imagery pairs across conus. Scientific Data12(1), 322 (2025)

    work page 2025

  2. [2]

    Global change biology20(10), 3177–3190 (2014)

    Chave, J., et al.: Improved allometric models to estimate the aboveground biomass of tropical trees. Global change biology20(10), 3177–3190 (2014)

    work page 2014

  3. [3]

    arXiv preprint arXiv:1706.05587 (2017)

    Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017)

    Pith/arXiv arXiv 2017

  4. [4]

    In: Proceedings of the IEEE CVPR

    Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE CVPR. pp. 1251–1258 (2017)

    work page 2017

  5. [5]

    Remote Sensing of Environment270, 112845 (2022)

    Duncanson,L.,Kellner,J.R.,etal.:Abovegroundbiomassdensitymodelsfornasa’s global ecosystem dynamics investigation (gedi) lidar mission. Remote Sensing of Environment270, 112845 (2022)

    work page 2022

  6. [6]

    Open-Canopy: Towards Very High Resolution Forest Monitoring

    Fogel, F., et al.: Open-canopy: A country-scale benchmark for canopy height esti- mation at very high resolution. arXiv preprint arXiv:2407.09392 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    In: Proceedings of the IEEE CVPR

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE CVPR. pp. 770–778 (2016)

    work page 2016

  8. [8]

    IGN: Diffusion Lidar HD IGN.https://diffusion-lidarhd.ign.fr/(2025), [On- line; accessed 3-March-2025]

    work page 2025

  9. [9]

    Environmental Research: Ecology 3(4), 045002 (2024)

    Kwon, T., Shibata, H., Takemi, T., Takagi, K.: Canopy height damage by typhoon songda in northern hokkaido, japan, in 2004. Environmental Research: Ecology 3(4), 045002 (2024)

    work page 2004

  10. [10]

    Forest ecology and management95(3), 209–228 (1997)

    Landsberg, J., Waring, R.: A generalised model of forest productivity using sim- plified concepts of radiation-use efficiency, carbon balance and partitioning. Forest ecology and management95(3), 209–228 (1997)

    work page 1997

  11. [11]

    Nature Ecology & Evolution7(11), 1778–1789 (2023) PrediTree: A Multi-Temporal Sub-meter Dataset 11

    Lang, N., Jetz, W., Schindler, K., Wegner, J.D.: A high-resolution canopy height model of the earth. Nature Ecology & Evolution7(11), 1778–1789 (2023) PrediTree: A Multi-Temporal Sub-meter Dataset 11

    work page 2023

  12. [12]

    Geophysical research letters32(22) (2005)

    Lefsky, M.A., Harding, D.J., Keller, M., Cohen, W.B., Carabajal, C.C., Others: Es- timates of forest canopy height and aboveground biomass using icesat. Geophysical research letters32(22) (2005)

    work page 2005

  13. [13]

    International Journal of Applied Earth Observation and Geoinformation92, 102163 (2020)

    Li, W., Niu, Z., Shang, R., Qin, Y., Wang, L., Chen, H.: High-resolution mapping of forest canopy height using machine learning by coupling icesat-2 lidar with sentinel-1, sentinel-2 and landsat-8 data. International Journal of Applied Earth Observation and Geoinformation92, 102163 (2020)

    work page 2020

  14. [14]

    Science Advances9(37), eadh4097 (2023)

    Liu, S., Brandt, M., et al.: The overlooked contribution of trees outside forests to tree cover and woody biomass across europe. Science Advances9(37), eadh4097 (2023)

    work page 2023

  15. [15]

    Global Bio- geochemical Cycles30(11), 1639–1660 (2016)

    Longo, M., Keller, M., dos Santos, M.N., Leitold, V., Pinagé, E.R., Baccini, A., Saatchi, S., Nogueira, E.M., Batistella, M., Morton, D.C.: Aboveground biomass variability across intact and degraded forests in the brazilian amazon. Global Bio- geochemical Cycles30(11), 1639–1660 (2016)

    work page 2016

  16. [16]

    Remote Sensing of Environ- ment253, 112165 (2021)

    Potapov, P., Li, X., Hernandez-Serna, A., et al.: Mapping global forest canopy height through integration of gedi and landsat data. Remote Sensing of Environ- ment253, 112165 (2021)

    work page 2021

  17. [17]

    European Journal of Forest Research 137, 29–44 (2018)

    Rohner, B., Waldner, P., Lischke, H., Ferretti, M., Thürig, E.: Predicting individual-tree growth of central european tree species as a function of site, stand, management, nutrient, and climate effects. European Journal of Forest Research 137, 29–44 (2018)

    work page 2018

  18. [18]

    In: MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. pp. 234–241. Springer (2015)

    work page 2015

  19. [19]

    High-resolution canopy height map in the Landes forest (France) based on GEDI, Sentinel-1, and Sentinel-2 data with a deep learning approach

    Schwartz,M.,etal.:High-resolutioncanopyheightmapinthelandesforest(france) based on gedi, sentinel-1, and sentinel-2 data with a deep learning approach. arXiv preprint arXiv:2212.10265 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    Journal of Geophysical Research: Biogeosciences 116(G4) (2011)

    Simard, M., Pinto, N., Fisher, J.B., Baccini, A.: Mapping forest canopy height globally with spaceborne lidar. Journal of Geophysical Research: Biogeosciences 116(G4) (2011)

    work page 2011

  21. [21]

    In: 2019 IEEE 10th annual ubiquitous computing, electronics & mobile communi- cation conference (UEMCON)

    Sinha, D., El-Sharkawy, M.: Thin mobilenet: An enhanced mobilenet architecture. In: 2019 IEEE 10th annual ubiquitous computing, electronics & mobile communi- cation conference (UEMCON). pp. 0280–0285. IEEE (2019)

    work page 2019

  22. [22]

    In: International conference on machine learning

    Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105–6114. PMLR (2019)

    2019

  23. [23]

    Remote Sensing of Environment300, 113888 (2024)

    Tolan, J., et al.: Very high resolution canopy height maps from rgb imagery using self-supervised vision transformer and convolutional decoder trained on aerial lidar. Remote Sensing of Environment300, 113888 (2024)

    work page 2024

  24. [24]

    Remote Sensing of Environment298, 113797 (2023)

    Turubanova, S., Potapov, P., Hansen, M.C., Li, X., Tyukavina, A., Pickens, A.H., Hernandez-Serna, A., Arranz, A.P., Guerra-Hernandez, J., Senf, C., et al.: Tree canopy extent and height change in europe, 2001–2021, quantified using landsat data archive. Remote Sensing of Environment298, 113797 (2023)

    work page 2001

  25. [25]

    Remote Sensing of Environment305, 114099 (2024)

    Wagner, F.H., Roberts, S., Ritz, A.L., Carter, G., Dalagnol, R., Favrichon, S., Hirye, M.C., Brandt, M., Ciais, P., Saatchi, S.: Sub-meter tree height mapping of california using aerial images and lidar-informed u-net model. Remote Sensing of Environment305, 114099 (2024)

    work page 2024

  26. [26]

    Webb, N.P., McCord, S.E., Edwards, B.L., Herrick, J.E., Kachergis, E., Okin, G.S., Van Zee, J.W.: Vegetation canopy gap size and height: Critical indicators for wind erosionmonitoringandmanagement.RangelandEcology&Management76,78–83 (2021)

    work page 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.