Pith. sign in

REVIEW 4 major objections 8 minor 56 references

EcoVision: AI-Powered Drone Imaging for Salt Marsh Vegetation Monitoring and Dominance Mapping

T0 review · 4 major / 8 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Drone pipeline maps salt marsh dominance within 8% of field surveys

desk verdict EcoVision integrates SegFormer + ConvNeXt + grid-based dominance scoring for salt marsh UAV monitoring. The pipeline is sensible and the ecological framing is genuine, but the validation has real gaps that need addressing before publication. read the letter →

arxiv 2607.06105 v1 pith:DOETWLRU submitted 2026-07-07 cs.CV cs.AI

Innocent Onyenonachi , Peter J. Lawerance , Nadia Kanwal This is my paper
classification cs.CVcs.AI
keywords saltmarshmonitoringUAVremotesensingsemanticsegmentationvegetationdominancemappingdeeplearningecologyConvNeXtSegFormerspeciesclassification
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 prove that a modular computer-vision pipeline can turn drone photos of salt marshes into ecologically meaningful dominance maps — not just species detection, but quantitative estimates of how much of each species covers each 2×2 metre patch, directly comparable to what a field ecologist would record with a quadrat. The pipeline chains four stages: a SegFormer-B5 transformer segments vegetation from background at the pixel level; connected-component analysis extracts contiguous vegetation patches (called blobs); a ConvNeXt classifier labels each blob as one of two target grasses (Spartina maritima or Puccinellia maritima); and a spatial aggregation step computes proportional areal coverage per grid cell. The key claim is that this chain, validated against field quadrat surveys at a North Wales salt marsh, achieves 97.6% agreement on dominant-species identification per grid patch and a mean absolute dominance error below 8%. The central object carrying the argument is the dominance score — a proportional areal coverage metric computed within fixed 2×2 m grid cells — which bridges pixel-level model outputs and the spatial units ecologists actually use. If the paper is right, the pipeline demonstrates that modern deep learning can produce repeatable, spatially dense vegetation dominance estimates at scales and resolutions that manual surveys cannot practically achieve, without replacing fieldwork as a calibration tool.

What carries the argument

SegFormer-B5 for semantic segmentation, ConvNeXt-Base for blob classification, connected-component analysis for vegetation blob extraction, and grid-based dominance scoring (proportional areal coverage per 2×2 m cell) as the aggregation mechanism.

What would settle it

If dominance scores from the pipeline diverged from field quadrat estimates by more than ~15-20% MAE, or if dominant-species agreement per grid patch dropped below ~85%, the claim that pixel-level predictions translate into ecologically interpretable metrics at field-survey quality would be undermined.

Watch

Extended reading notes

Core claim

The paper's core discovery is that an end-to-end chain of transformer-based segmentation, blob extraction, ConvNeXt classification, and grid-based aggregation can translate raw UAV RGB imagery into species dominance scores at 2×2 m resolution that match field quadrat surveys within 8% mean absolute error and 97.6% dominant-species agreement. The aggregation step — computing proportional areal coverage per grid cell — is what converts pixel-level predictions into an ecologically interpretable metric, and the paper shows that spatial aggregation at this scale absorbs localised segmentation and classification errors, preserving system-level accuracy even where individual blob boundaries are im.

Load-bearing premise

The ConvNeXt classifier was trained on close-range photos from public biodiversity repositories (iNaturalist, GBIF) but applied to top-down drone imagery. The paper acknowledges this as a known domain-shift challenge. If the 99% classification accuracy does not hold across viewpoints, lighting, and canopy perspectives typical of UAV surveys, the downstream dominance scores would inherit that error — though the grid aggregation step may partially absorb it.

Editorial extensions

If this is right

  • Salt marsh monitoring programmes could replace or supplement labour-intensive quadrat transects with repeatable drone surveys, increasing spatial coverage and temporal frequency at lower cost.
  • The modular pipeline structure means individual components — segmentation model, classifier, grid resolution — can be swapped or upgraded without redesigning the system, making it adaptable to other ecosystems and species sets.
  • Dominance maps at 2×2 m resolution could enable early detection of invasive species expansion or habitat fragmentation that coarse field transects would miss.
  • If RGB-only imagery suffices for two-species dominance mapping, conservation programmes with limited budgets could deploy the approach without multispectral or LiDAR sensor investments.
  • Longitudinal drone surveys using this pipeline could track interannual dominance shifts relevant to sea-level rise monitoring in vulnerable coastal habitats.

Reading between the lines

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

  • The dominance aggregation step may be doing significant error-correction work: if pixel-level IoU is only 0.56 but grid-level dominance error is under 8%, the spatial binning is absorbing substantial segmentation noise. This suggests the approach could tolerate even weaker segmentation models for coarser ecological questions, but also raises the question of whether finer grid resolutions (e.g., 1×
  • The two-species case is relatively favourable for classification because Spartina and Puccinellia have contrasting growth forms (erect vs. prostrate). Adding species with similar canopy architectures — as the authors propose for Salicornia and Atriplex — may degrade ConvNeXt classification accuracy more than the 99% figure suggests, and the dominance pipeline's robustness would then depend more he
  • If the classifier was trained on close-range public repository images but validated on top-down UAV blobs, the 99% accuracy may partly reflect the morphological distinctness of the two species rather than genuine cross-domain generalization. Testing on species with subtler visual differences would be a stronger validation of the domain-transfer claim.
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 / 8 minor

Summary. This manuscript presents EcoVision, an end-to-end pipeline that combines UAV-acquired RGB imagery, SegFormer-based semantic segmentation, ConvNeXt-based species classification, and grid-based dominance scoring to monitor salt marsh vegetation (Spartina maritima and Puccinellia maritima) at 2×2 m resolution. The authors report pixel-level segmentation accuracy of ~0.96, classification F1 of ~0.99, and end-to-end dominance agreement with field surveys of 97.6% with <8% MAE. The work targets a genuine gap: translating pixel-level predictions into ecologically interpretable, quadrat-scale dominance metrics.

Significance. The integration of segmentation, classification, and spatially explicit dominance scoring into a single pipeline validated against field data is a meaningful contribution to UAV-based ecological monitoring. The dominance score (Eq. 1) is a straightforward, parameter-free proportional area calculation, which is appropriate. The modular architecture and alignment with JNCC survey guidelines increase practical relevance. However, the significance of the end-to-end validation results is substantially undermined by missing methodological detail (see Major Comments).

major comments (4)
  1. §4.6, Table 6: The central end-to-end validation claims (97.6% agreement, <8% MAE) are reported without sample size, spatial distribution of quadrats, field survey protocol, or confidence intervals. The manuscript does not state how many quadrats were surveyed, how field dominance was estimated, or over what spatial extent. Without these, the statistical reliability of the agreement figure is impossible to assess. This is load-bearing for the paper's central claim and must be addressed.
  2. §4.2, Table 5: The ConvNeXt classification F1 of 0.99 is computed on a held-out test split drawn from the same iNaturalist/GBIF distribution as the training data (§3.3, Table 2). The UAV-derived blobs that actually pass through the pipeline are never independently classified and evaluated against manual labels. The only evidence that the classifier works on UAV data is the end-to-end agreement number, which itself lacks methodological transparency. An independent classification evaluation on UAV-derived blobs is needed to close this validation gap.
  3. §4.1, Table 4: The reported mIoU of 0.5571 does not match the mean of the two per-class IoUs ((0.9041 + 0.7670)/2 = 0.836). The footnote indicates background is excluded from mean calculations, but the discrepancy is unexplained. If background IoU is included in the mIoU, its value should be reported; if not, the mIoU should equal 0.836. This inconsistency in a primary segmentation metric must be resolved.
  4. §3.5.1 vs. §3.3: There is an internal contradiction regarding training data. §3.5.1 states SegFormer was 'fine-tuned on 512×512 UAV imagery,' but §3.3 and Table 2 indicate all training data came from iNaturalist/GBIF, with UAV data 'reserved exclusively for independent testing.' Clarification is needed on whether the segmentation model was trained on UAV imagery, public repository imagery, or both, as this directly affects the interpretation of the segmentation results in Table 4.
minor comments (8)
  1. Abstract: 'mean IoU = 0.56' is reported, but given the discrepancy in Table 4 (see major comment), this value may need correction.
  2. §3.4.2: The augmentation pipeline states each original image generates five augmented variants, yielding ~2,300 samples from 460 images. However, 460 × 5 = 2,300, which implies the original images are not retained in the training set. Confirm whether originals were included alongside augmentations.
  3. Table 2: The 'After Manual QC' column shows 460 images total, but the per-source breakdown (80 + 160 + 78 + 142 = 460) is consistent. The 'Raw images' column (1,012) vs. 'After Manual QC' (460) indicates substantial filtering (~55%); briefly justify this retention rate.
  4. Figure 8 caption states '307 samples' in the test set, but Table 5 also states 307. However, §4.2 text mentions '390 test samples.' Reconcile this discrepancy.
  5. §3.5.4: The area threshold for blob extraction (100 pixels) is described as 'empirically defined.' A brief sensitivity analysis or justification for this specific value would strengthen the reproducibility of the post-processing step.
  6. References [17] and [24] appear to cite the same work (Moreno et al., Remote Sensing 17(12), 2025) with slightly different author lists. Consolidate.
  7. §3.6.3, Eq. (9): The Pearson correlation formula is rendered with formatting issues in the denominator (square root symbols). Ensure correct typesetting in the final version.
  8. Keywords list is excessively long and includes generic terms ('method,' 'dataset,' 'pipeline'). Consider a focused keyword set for discoverability.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee correctly identifies that the end-to-end validation claims need substantially more methodological transparency, and that there are internal inconsistencies in the manuscript that must be resolved. We address each major comment below. We agree with three of the four comments in full and partially agree with the fourth; the manuscript will be revised accordingly.

read point-by-point responses

  1. Referee: §4.6, Table 6: The central end-to-end validation claims (97.6% agreement, <8% MAE) are reported without sample size, spatial distribution of quadrats, field survey protocol, or confidence intervals. The manuscript does not state how many quadrats were surveyed, how field dominance was estimated, or over what spatial extent. Without these, the statistical reliability of the agreement figure is impossible to assess. This is load-bearing for the paper's central claim and must be addressed.

    Authors: The referee is correct. The end-to-end validation in §4.6 and Table 6 omits critical methodological detail, including the number of quadrats surveyed, their spatial distribution across the Foryd Bay site, the field survey protocol used to estimate dominance (e.g., visual percentage cover estimation, point-intercept, or Braun-Blanquet scales), and any confidence intervals or uncertainty estimates on the reported agreement metrics. This is a genuine gap that undermines the interpretability of the central validation claim. We will revise §4.6 (and add corresponding detail to §3.2 or a new methods subsection) to include: (1) the total number of field quadrats surveyed and their spatial layout relative to the 2×2 m grid cells; (2) the field survey protocol, including quadrat size, observer methodology, and how percentage dominance was estimated in the field; (3) the spatial extent and sampling design (e.g., stratified random, transect-based); and (4) confidence intervals or bootstrap uncertainty estimates on both the 97.6% agreement figure and the <8% MAE. We will also report the full distribution of per-quadrat errors rather than only the aggregate MAE. If the sample size is small enough that confidence intervals are wide, we will state this explicitly and temper the claims accordingly. revision: yes

  2. Referee: §4.2, Table 5: The ConvNeXt classification F1 of 0.99 is computed on a held-out test split drawn from the same iNaturalist/GBIF distribution as the training data (§3.3, Table 2). The UAV-derived blobs that actually pass through the pipeline are never independently classified and evaluated against manual labels. The only evidence that the classifier works on UAV data is the end-to-end agreement number, which itself lacks methodological transparency. An independent classification evaluation on UAV-derived blobs is needed to close this validation gap.

    Authors: The referee correctly identifies a validation gap. The ConvNeXt classifier F1 of 0.99 (Table 5) is computed on a held-out test split from the iNaturalist/GBIF distribution, not on UAV-derived blobs. The manuscript does not currently report an independent evaluation of classification accuracy on UAV-extracted vegetation blobs against manual species labels. We acknowledge that the end-to-end dominance agreement (§4.6) is the only current evidence that the classifier generalises to UAV data, and as the referee notes in the previous comment, that metric itself lacks sufficient methodological transparency. We will address this in revision by: (1) adding a classification evaluation on UAV-derived blobs manually labelled by a domain expert, reporting per-class precision, recall, F1, and a confusion matrix on UAV data specifically; (2) clarifying in §4.2 that the Table 5 results reflect performance on the public-repository test split only, and that UAV-domain classification performance is assessed separately; and (3) cross-referencing this evaluation with the end-to-end dominance validation. If the number of manually labelled UAV blobs is limited, we will report this honestly and discuss the implications for interpreting the end-to-end results. revision: yes

  3. Referee: §4.1, Table 4: The reported mIoU of 0.5571 does not match the mean of the two per-class IoUs ((0.9041 + 0.7670)/2 = 0.836). The footnote indicates background is excluded from mean calculations, but the discrepancy is unexplained. If background IoU is included in the mIoU, its value should be reported; if not, the mIoU should equal 0.836. This inconsistency in a primary segmentation metric must be resolved.

    Authors: The referee is correct that there is an inconsistency. If background is excluded from the mean calculation, as the footnote and §3.5.3 state, then the mIoU should be the arithmetic mean of the two vegetation class IoUs: (0.9041 + 0.7670)/2 = 0.8356, not 0.5571. The reported value of 0.5571 is inconsistent with the stated methodology. We will investigate the source of this discrepancy—it likely arises from background IoU being inadvertently included in the mean computation despite the footnote stating otherwise, or from a different averaging convention being applied during evaluation. In the revised manuscript we will: (1) correct Table 4 so that the reported mIoU is consistent with the stated methodology (background excluded, mean of the two vegetation class IoUs); (2) if background IoU was in fact included, report it explicitly and present both values; and (3) ensure the metric definitions in §3.6.3, the loss function description in §3.5.3, and the results in Table 4 are internally consistent. We thank the referee for catching this. revision: yes

  4. Referee: §3.5.1 vs. §3.3: There is an internal contradiction regarding training data. §3.5.1 states SegFormer was 'fine-tuned on 512×512 UAV imagery,' but §3.3 and Table 2 indicate all training data came from iNaturalist/GBIF, with UAV data 'reserved exclusively for independent testing.' Clarification is needed on whether the segmentation model was trained on UAV imagery, public repository imagery, or both, as this directly affects the interpretation of the segmentation results in Table 4.

    Authors: The referee identifies a genuine internal contradiction. §3.5.1 states that SegFormer-B5 was 'fine-tuned on 512×512 UAV imagery,' while §3.3 and Table 2 state that training data came from iNaturalist/GBIF and that UAV data was 'reserved exclusively for independent testing.' These two statements cannot both be correct as written. We will resolve this in revision. The intended data strategy, as described in §3.3 and §3.4.3, is that training was conducted on annotated images from iNaturalist and GBIF (resized to 512×512), while UAV imagery was held out for independent testing. The statement in §3.5.1 that the model was 'fine-tuned on 512×512 UAV imagery' is erroneous and will be corrected to 'fine-tuned on 512×512 imagery from iNaturalist and GBIF.' However, we must also honestly flag a related issue: if the segmentation model in Table 4 was evaluated on UAV imagery (as §4.1 states: 'evaluated on the held-out test set comprising annotated UAV imagery'), but was trained only on iNaturalist/GBIF imagery, then the domain gap between close-up repository photos and overhead UAV imagery is substantial, and the reported per-class IoU values (0.90 and 0.77) would need to be interpreted in that cross-domain context. We will clarify in revision exactly which data the segmentation model was trained and evaluated on, and if there is a domain mismatch between training and test data, we will discuss its implications explicitly. If it turns out that some UAV imagery was used in training (contradicting §3.3), we will correct §3.3 and Table 2 accordingly and restate the data partitioning transparently. revision: partial

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: dominance formula is parameter-free, models trained on external data, end-to-end validation against independent field surveys.

full rationale

The paper's central derivation chain is not circular. Equation 1 defines dominance as D_s(P) = A_s(P) / sum(A_i(P)) * 100, a parameter-free proportional area calculation with no fitted constants. The ConvNeXt classifier is trained on iNaturalist/GBIF images (Table 2, Section 3.3) and evaluated on a held-out test split from the same distribution (Table 5), while UAV imagery is reserved for independent end-to-end validation (Section 3.3: 'UAV imagery was stored separately as independent validation data'). The 97.6% agreement and <8% MAE (Table 6) are computed against field-derived ground-truth quadrat surveys (Section 3.6.3, Equation 9-10), which are external to the model's training pipeline. No self-citation chain is load-bearing: the paper cites SegFormer [15], ConvNeXt [33], and standard ecological methods [34] as external architectural and methodological foundations. The reader's concern about domain shift (close-range training images vs. top-down UAV imagery) and missing sample sizes for field validation are correctness and transparency risks, not circularity. The end-to-end agreement number is not forced by construction; it depends on whether the trained models genuinely generalize to UAV data, which is an empirical question the paper attempts to answer. No equation reduces to its own inputs, no prediction is a renamed fit, and no self-citation is load-bearing for the central claim.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The pipeline relies on two empirically chosen parameters (blob area threshold, grid size) and three core domain assumptions about the transferability and error tolerance of the chosen models. No new physical entities or forces are introduced.

free parameters (2)
  • Area threshold for blob extraction = 100 pixels
    Stated in Section 3.5.4 as an empirically defined threshold to balance noise suppression with retention of ecologically relevant patches.
  • Grid cell size = 2x2 m
    Chosen to align with JNCC guidelines and standard quadrat sizes, but acts as a fixed spatial parameter in the aggregation.
assumptions (3)
  • domain assumption SegFormer-B5 can effectively segment vegetation from background in high-resolution UAV imagery.
    The entire pipeline depends on the segmentation model producing accurate vegetation masks. The paper reports mIoU=0.56, which is moderate, but assumes this is sufficient for downstream tasks.
  • domain assumption ConvNeXt classification on public repository images generalizes to UAV-derived blob crops.
    The classifier is trained on iNaturalist/GBIF images and applied to UAV blobs. The paper assumes the morphological features learned transfer to the top-down UAV perspective.
  • domain assumption Spatial aggregation at 2x2m mitigates pixel-level segmentation and classification errors.
    Section 4.7 states that spatial aggregation reduced the impact of pixel and blob-level errors, preserving robustness at the ecological scale. This is assumed to hold for the reported dominance accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EcoVision: AI-Powered Drone Imaging for Salt Marsh Vegetation Monitoring and Dominance Mapping." pith.science (2026). https://pith.science/paper/DOETWLRU

@misc{pith2026260706105,
  author       = {Pith},
  title        = {Pith review of: EcoVision: AI-Powered Drone Imaging for Salt Marsh Vegetation Monitoring and Dominance Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DOETWLRU}},
  note         = {Machine review of arXiv:2607.06105}
}
read the original abstract

High-resolution RGB imagery acquired from low-altitude UAV surveys was processed through a modular pipeline incorporating transformer-based semantic segmentation, connected-component vegetation extraction, fine-grained species classification using a ConvNeXt architecture, and grid-based dominance scoring at 2x2m resolution. The framework targeted two ecologically significant halophytic grasses, Spartina maritima and Puccinellia maritima, and was trained using a curated and manually annotated UAV imagery, along with biodiversity imagery sourced from publicly accessible datasets. In order to identify these plants from the imagery, our segmentation yielded reliable species masks (mean IoU = 0.56; pixel-level accuracy = 0.96), while object-level classification achieved very good discrimination (F1 = 0.99). Dominance estimates closely matched quadrat-based field surveys, with mean absolute differences below 8%, preserving fine-scale spatial structure under realistic survey conditions. The developed system, named EcoVision, establishes a practical foundation for scalable, high-resolution salt marsh monitoring, demonstrating how AI-driven workflows can translate pixel-level predictions into ecologically interpretable metrics.

Figures

Figures reproduced from arXiv: 2607.06105 by the authors.

Figure 1
Figure 1. End-to-end flow of the EcoVision pipeline. [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Conceptual architecture of the EcoVision framework. The system is organised as [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Representative field photographs of the two target species. (a) Spartina mar [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Study area location. (a) Regional context showing the survey site on the North [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: SegFormer-B5 architecture used for semantic segmentation of UAV salt marsh [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Representative dominance map produced by the EcoVision pipeline. Each [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: ConvNeXt-Base training accuracy and macro-averaged F1-score across 15 [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Confusion matrix for ConvNeXt-Base species classification on the held-out test [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages

  1. [1]

    E. B. Barbier, S. D. Hacker, C. Kennedy, E. W. Koch, A. C. Stier, B. R. Silliman, The value of estuarine and coastal ecosystem services, Ecological monographs 81 (2) (2011) 169–193

    work page 2011

  2. [2]

    Wäldchen, P

    J. Wäldchen, P. Mäder, Plant species identification using com- puter vision techniques: A systematic literature review, Archives of Computational Methods in Engineering 25 (2) (2017) 507–543. doi:10.1007/s11831-016-9206-z

    work page doi:10.1007/s11831-016-9206-z 2017

  3. [3]

    secretariat, Global assessment report on biodiversity and ecosystem services, accessed: 2026-02-05 (2024)

    I. secretariat, Global assessment report on biodiversity and ecosystem services, accessed: 2026-02-05 (2024). URLhttps://www.ipbes.net/global-assessment

    work page 2026

  4. [4]

    Bertacchi, V

    A. Bertacchi, V. Giannini, C. Di Franco, N. Silvestri, Using unmanned aerial vehicles for vegetation mapping and identification of botanical species in wetlands, Landscape and Ecological Engineering 15 (2) (2019) 231–240. doi:10.1007/s11355-018-00368-1. 31

    work page doi:10.1007/s11355-018-00368-1 2019

  5. [5]

    C. L. Elzinga, D. W. Salzer, J. W. Willoughby, J. P. Gibbs, Monitoring plant and animal populations, Pacific Conservation Biology 8 (3) (2002)

    work page 2002

  6. [6]

    doi:10.1071/pc020219

    work page doi:10.1071/pc020219

  7. [7]

    R. A. Stevens, S. Shull, J. Carter, E. Bishop, N. Herold, C. A. Riley, et al., Marsh migration and beyond: A scalable framework to assess tidal wetland resilience and support strategic management, PLOS ONE 18 (11) (2023) e0293177. doi:10.1371/journal.pone.0293177

    work page doi:10.1371/journal.pone.0293177 2023

  8. [8]

    C. S. Reddy, Remote sensing of biodiversity: what to measure and mon- itor from space to species?, Biodiversity and Conservation 30 (2021) 2617–2631. doi:10.1007/s10531-021-02216-5

    work page doi:10.1007/s10531-021-02216-5 2021

Show all 56 references
  1. [9]

    Z. Yang, T. Blount, C. Lester, N. Blackford, A. D’Alpaos, M. Marani, B. Murray, S. Silvestri, Unmixing marsh vegetation species across mul- tiple sensors and spatial scales, GIScience & Remote Sensing 62 (1) (2025). doi:10.1080/15481603.2025.2481689

    2025 doi

  2. [10]

    Chang, P

    M. Chang, P. Li, Z. Li, H. Wang, Time series satellite remote sensing re- veals annual distribution, composition, and trajectory of tidal wetlands in the yellow river delta, Estuarine Coastal and Shelf Science 319 (2025) 109264. doi:10.1016/j.ecss.2025.109264

    2025 doi

  3. [11]

    Kattenborn, J

    T. Kattenborn, J. Eichel, F. E. Fassnacht, Convolutional neural net- works enable efficient, accurate and fine-grained segmentation of plant species and communities from high-resolution uav imagery, Scientific Reports 9 (1) (2019) 17656. doi:10.1038/s41598-019-53797-9

    2019 doi

  4. [12]

    A. I. De Castro, Y. Shi, J. M. Maja, J. M. Peña, Uavs for vegetation monitoring: overview and recent scientific contributions, Remote Sens- ing 13 (11) (2021) 2139. doi:10.3390/rs13112139

    2021 doi

  5. [13]

    C. Cruz, K. McGuinness, P. M. Perrin, J. O’Connell, J. R. Martin, J. Connolly, Improving the mapping of coastal invasive species using uav imagery and deep learning, International Journal of Remote Sensing 44 (18) (2023) 5713–5735. doi:10.1080/01431161.2023.2251186

    2023 doi

  6. [14]

    I. L. Turner, M. D. Harley, C. D. Drummond, Uavs for coastal surveying, Coastal Engineering 114 (2016) 19–24. doi:10.1016/j.coastaleng.2016.03.011. 32

    2016 doi

  7. [15]

    Aasen, E

    H. Aasen, E. Honkavaara, A. Lucieer, P. J. Zarco-Tejada, Quantitative remote sensing at ultra-high resolution with uav spectroscopy: A re- view of sensor technology, measurement procedures, and data correction workflows, Remote Sensing 10 (7) (2018) 1091. doi:10.3390/rs10071091

    2018 doi

  8. [16]

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, P. Luo, Seg- former: Simple and efficient design for semantic segmentation with transformers, Advances in neural information processing systems 34 (2021) 12077–12090

    2021

  9. [17]

    URLhttps://data.jncc.gov.uk/data/a407ebfc-2859-49cf- 9710-1bde9c8e28c7/JNCC-NVC-UsersHandbook-2006.pdf

    Joint Nature Conservation Committee, National Vegetation Classifica- tion: Users’ Handbook, JNCC, accessed: 2024-01-01 (2006). URLhttps://data.jncc.gov.uk/data/a407ebfc-2859-49cf- 9710-1bde9c8e28c7/JNCC-NVC-UsersHandbook-2006.pdf

    2024

  10. [19]

    G. R. Morgan, M. E. Hodgson, C. Wang, S. R. Schill, Deep learning of high-resolution aerial imagery for coastal marsh change detection: A comparative study, ISPRS International Journal of Geo-Information 11 (2) (2022) 100. doi:10.3390/ijgi11020100

    2022 doi

  11. [20]

    Pineau, P

    J. Pineau, P. Vincent-Lamarre, K. Sinha, V. Larivière, A. Beygelzimer, F. d’Alché Buc, E. Fox, H. Larochelle, Improving reproducibility in ma- chine learning research: A report from the neurips 2019 reproducibility program, Journal of Machine Learning Research 22 (2021) 1–20

    2019

  12. [21]

    Van Alphen, et al., Uav-based wetland monitoring: Multispectral and lidar fusion with random forest classification, Drones 8 (3) (2024)

    R. Van Alphen, et al., Uav-based wetland monitoring: Multispectral and lidar fusion with random forest classification, Drones 8 (3) (2024)

    2024

  13. [22]

    doi:10.3390/drones8030113

    doi

  14. [23]

    P. C. Gray, G. D. Larsen, D. W. Johnston, Drones address an observa- tional blind spot for biological oceanography, Frontiers in Ecology and the Environment 20 (7) (2022) 413–421. doi:10.1002/fee.2472. URLhttps://doi.org/10.1002/fee.2472

    2022 doi

  15. [24]

    Gonzalez-Perez, A

    A. Gonzalez-Perez, A. Perez, et al., Deep and machine learning image classification of coastal wetlands using unpiloted aircraft system multi- spectral images and lidar datasets, Remote Sensing 14 (15) (2022) 3582. 33 doi:10.3390/rs14153582. URLhttps://pubs.usgs.gov/publicatio...

    2022 doi

  16. [25]

    X. Bai, C. Yang, et al., Identification of salt marsh vegetation in the yellow river delta using uav multispectral imagery and deep learning, Drones 9 (4) (2025) 235. doi:10.3390/drones9040235. URLhttps://www.mdpi.com/2504-446X/9/4/235

    2025 doi

  17. [26]

    Moreno, S

    M. Moreno, S. Dalai, G. Cott, B. Bartlett, M. Santos, T. Dorian, J. Ri- ordan, C. McGonigle, F. Sacchetti, G. Dooly, Multi-camera machine learning for salt marsh species classification and mapping, Remote Sens- ing 17 (12) (2025) 1964. doi:10.3390/rs17121964. URLhttps://www.md...

    2025 doi

  18. [27]

    Agate, et al., Mapping and monitoring heterogeneous plant commu- nities in restored and established salt marshes using uavs and machine learning, Remote Sensing 18 (6) (2026) 866

    J. Agate, et al., Mapping and monitoring heterogeneous plant commu- nities in restored and established salt marshes using uavs and machine learning, Remote Sensing 18 (6) (2026) 866. doi:10.3390/rs18060866. URLhttps://www.mdpi.com/2072-4292/18/6/866

    2026 doi

  19. [28]

    S. Liu, et al., Integrating drone multispectral and lidar data with machine learning: Toward species-level biomass estimation in heteroge- neous salt marsh vegetation, Geo-spatial Information Science (2025). doi:10.1080/10095020.2025.2574921. URLhttps://www.tandfonline.com/doi...

    2025 doi

  20. [29]

    A. C. Curcio, L. Barbero, G. Peralta, Enhancing salt marshes moni- toring: Estimating biomass with drone-derived habitat-specific models, Remote Sensing Applications: Society and Environment 35 (2024) 101216. doi:10.1016/j.rsase.2024.101216. URLhttps://www.sciencedirect.com/sc...

    2024 doi

  21. [30]

    Pöttker, K

    M. Pöttker, K. Kiehl, T. Jarmer, D. Trautz, Convolutional neural net- work maps plant communities in semi-natural grasslands using multi- spectral unmanned aerial vehicle imagery, Remote Sensing 15 (7) (2023) 1945

    2023

  22. [31]

    Gao, X.Du, P.Xu, E.Gao, Y

    N. Gao, X.Du, P.Xu, E.Gao, Y. Yang, High-resolutionmappingcoastal wetland vegetation using frequency-augmented deep learning method, Remote Sensing 18 (2) (2026) 247. doi:10.3390/rs18020247. 34

    2026 doi

  23. [32]

    Gautam, et al., Detection of invasive species (siam weed) using drone- based imaging and yolo deep learning model, Remote Sensing 17 (1) (2025) 120

    D. Gautam, et al., Detection of invasive species (siam weed) using drone- based imaging and yolo deep learning model, Remote Sensing 17 (1) (2025) 120. doi:10.3390/rs17010120

    2025 doi

  24. [33]

    T. Deng, et al., Comparison of multi-class and fusion of multiple single- class segnet model for mapping karst wetland vegetation using uav im- ages, Scientific Reports 12 (1) (2022) 13270. doi:10.1038/s41598-022- 17620-2

    2022 doi

  25. [34]

    Milioto, P

    A. Milioto, P. Lottes, C. Stachniss, Real-time semantic segmen- tation of crop and weed for precision agriculture robots leverag- ing background knowledge in cnns, in: IEEE International Con- ference on Robotics and Automation (ICRA), 2018, pp. 2229–2235. doi:10.1109/icra.2018.8460962

    2018 doi

  26. [35]

    Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, S. Xie, A con- vnet for the 2020s, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 11976– 11986

    2022

  27. [36]

    Kent, Vegetation Description and Data Analysis: A Practical ap- proach, 2nd Edition, John Wiley & Sons, Ltd, 2012

    M. Kent, Vegetation Description and Data Analysis: A Practical ap- proach, 2nd Edition, John Wiley & Sons, Ltd, 2012

    2012

  28. [37]

    L. Ma, Y. Liu, X. Zhang, Y. Ye, G. Yin, B. A. Johnson, Deep learn- ing in remote sensing applications: A meta-analysis and review, ISPRS Journal of Photogrammetry and Remote Sensing 152 (2019) 166–177. doi:10.1016/j.isprsjprs.2019.04.015

    2019 doi

  29. [38]

    J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for se- mantic segmentation, in: Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2015, pp. 3431–3440. doi:10.1109/cvpr.2015.7298965

    2015 doi

  30. [39]

    R. M. Haralick, et al., Computer and Robot Vision, Addison-Wesley, 1992. URLhttps://scispace.com/pdf/computer-and-robot-vision-1upftzg46q.pdf

    1992

  31. [40]

    R. C. Gonzalez, R. E. Woods, Digital Image Processing, 4th Edition, Pearson Education Limited, 2018. URLhttps://api.pageplace.de/preview/DT0400.9781292223070_ A37747583/preview-9781292223070_A37747583.pdf 35

    2018

  32. [41]

    Anderson, K

    K. Anderson, K. J. Gaston, Lightweight unmanned aerial vehicles will revolutionize spatial ecology, Frontiers in Ecology and the Environment 11 (3) (2013) 138–146. doi:10.1890/120150

    2013 doi

  33. [42]

    Adam, Saltmarsh Ecology, Cambridge University Press, 1990

    P. Adam, Saltmarsh Ecology, Cambridge University Press, 1990. doi:10.1017/cbo9780511565328

    1990 doi

  34. [43]

    M. D. Bertness, Zonation of spartina patens and spartina alterni- flora in new england salt marsh, Ecology 72 (1) (1991) 138–148. doi:10.2307/1938909

    1991 doi

  35. [44]

    Iglhaut, et al., Structure from motion photogrammetry in forestry: a review, Current Forestry Reports 5 (3) (2019) 155–168

    J. Iglhaut, et al., Structure from motion photogrammetry in forestry: a review, Current Forestry Reports 5 (3) (2019) 155–168. doi:10.1007/s40725-019-00094-3

    2019 doi

  36. [45]

    Colomina, P

    I. Colomina, P. Molina, Unmanned aerial systems for photogrammetry and remote sensing: A review, ISPRS Journal of Photogrammetry and Remote Sensing 92 (2014) 79–97. doi:10.1016/j.isprsjprs.2014.02.013

    2014 doi

  37. [46]

    doi:10.15468/ab3s5x

    iNaturalist contributors, iNaturalist, iNaturalist Research-grade Ob- servations, Available at:https://www.inaturalist.org, occurrence dataset (2023). doi:10.15468/ab3s5x. URLhttps://www.inaturalist.org

    2023 doi

  38. [47]

    URLhttps://www.gbif.org/what-is-gbif

    GBIF: The Global Biodiversity Information Facility, Free and Open Access Biodiversity Data, Available at: https://www.gbif.org/what-is-gbif, accessed: 17 May 2025 (2023). URLhttps://www.gbif.org/what-is-gbif

    2025

  39. [48]

    Ac- cessed via GBIF occurrence search API on [May 25, 2025] (2025)

    GBIF.org, GBIF Occurrence Images forSpartina maritimaand Puccinellia maritimafor EcoVision Study, derived dataset. Ac- cessed via GBIF occurrence search API on [May 25, 2025] (2025). doi:110.15468/dd.f6b6ng. URLhttps://doi.org/10.15468/DD.F6B6NG

    2025 doi

  40. [49]

    B. C. Russell, et al., Labelme: a database and web-based tool for image annotation, International Journal of Computer Vision 77 (1–3) (2007) 157–173. doi:10.1007/s11263-007-0090-8. 36

    2007 doi

  41. [50]

    Buslaev, V

    A. Buslaev, V. I. Iglovikov, E. Khvedchenya, A. Parinov, M. Druzhinin, A. A. Kalinin, Albumentations: Fast and flexible image augmentations, Information 11 (2) (2020) 125. doi:10.3390/info11020125. URLhttp://dx.doi.org/10.3390/info11020125

    2020 doi

  42. [51]

    Shorten, T

    C. Shorten, T. M. Khoshgoftaar, A survey on image data augmentation for deep learning, Journal of Big Data 6 (1) (2019). doi:10.1186/s40537- 019-0197-0

    2019 doi

  43. [52]

    Hugging Face, Transformers documentation, https://huggingface.co/docs/transformers/en/index, accessed: 2026-03-09 (2023)

    2026

  44. [53]

    L. Lei, Q. Yang, L. Yang, T. Shen, R. Wang, C. Fu, Deep learn- ing implementation of image segmentation in agricultural applications: a comprehensive review, Artificial Intelligence Review 57 (6) (2024). doi:10.1007/s10462-024-10775-6

    2024 doi

  45. [54]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, arXiv preprint arXiv:1711.05101 (2017). doi:10.48550/arxiv.1711.05101. URLhttps://doi.org/10.48550/arxiv.1711.05101

    2017 doi arXiv

  46. [55]

    Micikevicius, et al., Mixed precision training, arXiv preprint arXiv:1710.03740 (2017)

    P. Micikevicius, et al., Mixed precision training, arXiv preprint arXiv:1710.03740 (2017). doi:10.48550/arxiv.1710.03740

    2017 doi arXiv

  47. [56]

    Tejani, S

    A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, S. Chintala, Py- torch: An imperative style, high-performance deep learning library, arXiv preprint arXiv:1912.01703 (2019). doi:10.48550/arxiv.1912.01703

    1912 doi arXiv

  48. [57]

    Legendre, L

    P. Legendre, L. Legendre, Numerical Ecology, 3rd Edition, Develop- ments in Environmental Modelling, Elsevier, Amsterdam, 2012. URLhttps://www.scirp.org/reference/referencespapers? referenceid=3732179 37

    2012

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.