Pith. sign in

REVIEW 36 references

Remote sensing colour image semantic segmentation of trails created by large herbivorous Mammals

T0 review · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read On a new 100-image dataset of Spanish mountain pastures, U-Net plus MambaOut outperformed 69 other segmentation model-encoder pairs at pixel-level mapping of grazing trails.

desk verdict New dataset and broad benchmark for pixel-level grazing-trail segmentation, but the ground-truth section as written cannot produce the reported masks and model selection uses the same folds; worth peer review after major revision. read the letter →

arxiv 2504.12121 v3 pith:E67X6YHJ submitted 2025-04-16 cs.CV

Jose Francisco Diez-Pastor , Francisco Javier Gonzalez-Moya , Pedro Latorre-Carmona , Francisco Javier Perez-Barbería , Ludmila I.Kuncheva , Antonio Canepa-Oneto , Alvar Arnaiz-González , Cesar Garcia-Osorio This is my paper · ORCID
classification cs.CV
keywords trailsbasisencodergrazingsegmentationsemanticaimingassessed
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

Grazing trails are bare-soil paths created when cattle and wild ungulates repeatedly walk the same routes. Knowing where these trails are helps ecologists locate heavy grazing pressure, which can threaten biodiversity. This paper tests whether deep-learning image segmentation, the same family of methods used to find roads or buildings in satellite photos, can draw these trails automatically.

The authors collected 100 aerial images of five Spanish mountain ranges. An expert ecologist hand-drew the centerline of every visible trail in each image, and those drawings were converted into soft target masks. Then they trained 70 combinations: five segmentation architectures, namely U-Net, FPNet, UperNet, Segformer, and PSPNet, each with fourteen pretrained encoder networks. Every combination was evaluated with 10-fold cross-validation. The best average Intersection over Union was about 0.42, achieved by U-Net with the MambaOut encoder; the same pair also had the best F1 score at about 0.58. The authors report that the model generally distinguished trails from dirt roads and other linear features.

The result is a benchmark more than a validation. The ground truth comes from a single expert's interpretation, so the scores measure agreement with that one person's labels. There is no comparison against the earlier patch-level method, no final test set held out after model selection, and no standard deviations reported for the scores. The dataset and code are promised on public repositories, but the paper does not give a commit hash or training hyperparameters. So the practical claim, that this can become a monitoring tool, rests on further independent evaluation.

Extended reading notes

Core claim

The paper's central claim is that semantic segmentation can map grazing trails at pixel level, and specifically that "The UNet architecture with the MambaOut encoder achieved the highest overall performance across combinations of architectures and encoders" (Section 3), with mean IoU of about 0.417 and F1 of about 0.582. The abstract adds that this is "the first instance of achieving competitive image segmentation performance for the detection and precise outlining of large herbivore trail systems." If these claims are correct, a U-Net model with a MambaOut encoder trained on expert-labeled aerial images can outline grazing trails as continuous pixel maps, advancing beyond patch-level presence detection.

Load-bearing premise

The load-bearing premise is that the ground-truth masks are valid. A single expert visually identified every trail from Google Earth Pro orthoimages and drew centerlines in ImageJ (Section 2.2), and the masks were then smoothed with a hand-chosen sigma of 16. There is no field survey, no second annotator, and no quantitative labeling protocol. All IoU and F1 scores, and therefore the ranking of models, measure agreement with this one subjective interpretation. If the labels misrepresent actual trails, the benchmark's conclusions do not transfer to real monitoring applications.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The benchmark rests on expert-generated labels, a hand-set Gaussian smoothing width, and standard transfer-learning assumptions. No new physical or mathematical entities are introduced, so the ledger is mostly about label validity and preprocessing choices.

free parameters (4)
  • sigma (ground-truth smoothing) = 16
    Section 2.2 sets sigma to 16, visually chosen to match half the average width of paths in the training set; it defines the width of the target masks used in every evaluation metric.
  • Th (HSI label threshold) = 3
    Section 2.2 sets the threshold for converting hand-drawn yellow pencil traces into binary centerline masks; this digitization threshold affects the binary mask before Gaussian smoothing.
  • Image downscale factor = 8
    Section 2.2 says images were resized by a factor of 8 to reduce computational cost; factors 2 and 4 were tested without showing quantitative results, so the claimed equivalence is not independently supported.
  • ROPE width for Bayesian tests = 0.01
    Section 2.4 sets the region of practical equivalence to 0.01 in the Bayesian comparison; this analysis choice affects whether differences between models are called equivalent.
assumptions (5)
  • domain assumption Single-expert visual identification of trails from Google Earth Pro orthoimages is an adequate ground truth.
    Section 2.2; there is no field verification and no second annotator, and the Discussion acknowledges that observer subjectivity is an unavoidable source of error.
  • domain assumption The yellow-pencil centerlines and HSI threshold procedure faithfully represent trail locations.
    Section 2.2; the conversion from hand-drawn lines to binary masks is assumed to be accurate for generating training targets.
  • domain assumption The Gaussian distance transform with sigma equal to 16 represents true trail extent.
    Equation 1 and the following paragraph; the width of the target masks is set by this visually chosen parameter.
  • domain assumption ImageNet-pretrained encoders transfer to aerial grazing-trail segmentation.
    Section 2.3 and Table 1; all 14 encoders are used with ImageNet pretraining, which is standard practice but not verified for this domain.
  • domain assumption Google Earth Pro orthoimages are geometrically and radiometrically adequate for trail delineation.
    Section 2.1; images are taken from a consumer mapping service rather than a calibrated remote-sensing source, and the paper provides no validation against ground measurements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Remote sensing colour image semantic segmentation of trails created by large herbivorous Mammals." pith.science (2026). https://pith.science/paper/E67X6YHJ

@misc{pith2026250412121,
  author       = {Pith},
  title        = {Pith review of: Remote sensing colour image semantic segmentation of trails created by large herbivorous Mammals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E67X6YHJ}},
  note         = {Machine review of arXiv:2504.12121}
}
read the original abstract

Identifying spatial regions where biodiversity is threatened is crucial for effective ecosystem conservation and monitoring. In this stydy, we assessed varios machine learning methods to detect grazing trails automatically. We tested five semantic segmentation models combined with 14 different encoder networks. The best combination was UNet with MambaOut encoder. The solution proposed could be used as the basis for tools aiming at mapping and tracking changes in grazing trails on a continuous temporal basis.

Figures

Figures reproduced from arXiv: 2504.12121 by the authors.

Figure 1
Figure 1. Study area. Map of Spain showing the areas (red triangles) images were obtained from. The selected mountain systems correspond to: (i) “Cantabrian” mountain range, (ii) “Palencia” mountain range, (iii) “Pyrenees”, (iv) “Sierra de la Demanda”, and (v) “Sierra de Béjar”. J.F. Díez-Pastor et al.: Preprint submitted to Elsevier Page 19 of 17 [PITH_FULL_IMAGE:figures/full_fig_p019_1.png] view at source ↗
Figure 2
Figure 2. Examples of different grazing trails networks in the “Cantabrian” mountain range. (A) Parallel trails on heather, their layout dominant direction is perpendicular to the slope, 43° 1’37.14" N, 5°30’4.25" O WGS84; (B) As (A) but on a rocky slope, coordinates 43°11’17.56" N, 4°45’24.56" O; (C) Trails on mountain grassland, no dominant layout direction, coordinates 43° 2’25.66" N, 6°13’20.43" O; (D) Trails on heather, … view at source ↗
Figure 3
Figure 3. Example of the original RGB groundtruth image (left columns) and the result of the automatic trail segmentation (right column). The trails are highlighted in orange. Best viewed in colour. J.F. Díez-Pastor et al.: Preprint submitted to Elsevier Page 21 of 17 [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Heatmap of Intersection over Union (IoU) performance across all encoder-architecture combinations. Each cell displays the mean IoU for a specific pair, with underlined values indicating the best encoder for each architecture and bold values showing the overall top comb…
Figure 5
Figure 5. Figure 5: Heatmap of the 𝐹1 performance measure across all encoder-architecture combinations. Each cell displays the mean IoU for a specific pair, with underlined values indicating the best encoder for each architecture and bold values showing the overall top combination. The ro…
Figure 6
Figure 6. Figure 6: Bayesian tests heatmap, comparing the performance of the five architectures with their best encoder. Each cell contains the probability that the method in the row is better than the method in the column (𝑝𝑙𝑒𝑓 𝑡) on top and the probability that the opposite occurs (𝑝𝑟𝑖𝑔…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 15 canonical work pages

  1. [1]

    Potentialoflinearfeaturesdetectioninamediterraneanlandscapefrom3dvhropticaldata: Applicationtoterrace walls, in: 2012 IEEE International Geoscience and Remote Sensing Symposium, pp

    Bailly,J.S.,Levavasseur,F.,2012. Potentialoflinearfeaturesdetectioninamediterraneanlandscapefrom3dvhropticaldata: Applicationtoterrace walls, in: 2012 IEEE International Geoscience and Remote Sensing Symposium, pp. 7110–7113. doi:10.1109/IGARSS.2012.6352024. Benavoli,A.,Corani,G.,Demšar,J.,Zaffalon,M.,2017. Timeforachange: atutorialforcomparingmultiplecla...

    work page arXiv 2012

  2. [7]

    Hao, S.,Zhou, Y., Guo,Y., 2020a

    doi:10.1007/s13735-017-0141-z. Hao, S.,Zhou, Y., Guo,Y., 2020a. Abriefsurveyonsemanticsegmentationwithdeeplearning. Neurocomputing406, 302–321. doi: 10.1016/j. neucom.2019.11.118. Hao, S., Zhou, Y., Guo, Y., 2020b. A brief survey on semantic segmentation with deep learning. Neurocomputing 406, 302–321. doi:https: //doi.org/10.1016/j.neucom.2019.11.118. J....

    doi:10.1007/s13735-017-0141-z 2019

  3. [10]

    Earth Surface Processes and Landforms 45, 3368–3380

    Detection of terracettes in semi-arid rangelands using fourier-based image anal- ysis of very-high-resolution satellite imagery. Earth Surface Processes and Landforms 45, 3368–3380. doi: 10.1002/esp.4971, arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/esp.4971. Higgins, C.G.,

    work page doi:10.1002/esp.4971

  4. [12]

    Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.,

    doi:10.3390/rs12020221. Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.,

    work page doi:10.3390/rs12020221

  5. [14]

    Sustainable Cities and Society 79, 103630

    Citysurfaces: City-scale semantic segmentation of sidewalk materials. Sustainable Cities and Society 79, 103630. doi:10.1016/j.scs.2021.103630. Howard,A.,Sandler,M.,Chu,G.,Chen,L.C.,Chen,B.,Tan,M.,Wang,W.,Zhu,Y.,Pang,R.,Vasudevan,V.,Le,Q.V.,Adam,H.,2019. Searching for mobilenetv3. URL:https://arxiv.org/abs/1905.02244, arXiv:1905.02244. Huang, G., Liu, Z.,...

    arXiv 2021

  6. [15]

    Chollet, F.,

    doi:10.3390/rs15061536. Chollet, F.,

    doi:10.3390/rs15061536

  7. [16]

    Mathematical Models in Engineering 4, 112–137

    Automatic semantic segmentation and classification of remote sensing data for agriculture. Mathematical Models in Engineering 4, 112–137. doi:10.21595/mme.2018.19840. Jamali, A., Roy, S.K., Li, J., Ghamisi, P.,

    work page arXiv 2018

  8. [17]

    CATENA 216, 106380

    Multi-parallel structure and a generalized conceptual model of livestock track network. CATENA 216, 106380. URL: https://www.sciencedirect.com/science/article/pii/S0341816222003666, doi:https://doi. org/10.1016/j.catena.2022.106380. Jin, B., Sun, G., Zhang, Y., Zou, M., Ni, X., Luo, K., Zhang, X., Cheng, H., Li, F., Wu, X.B.,

    arXiv 2022

Show all 36 references
  1. [19]

    Feature pyramid networks for object detection, in: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 936–944. doi:10.1109/CVPR.2017.106. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.,

    2017 doi

  2. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Luo,Z.,Yang,W.,Yuan,Y.,Gou,R.,Li,X.,2023. Semanticsegmentationofagriculturalimages: Asurvey. InformationProcessinginAgriculture J.F. Díez-Pastor et al.:Preprint ...

    2023 doi

  3. [21]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 3523–3542

    Image segmentation using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 3523–3542. doi:10.1109/TPAMI.2021.3059968. Mnih, V., Hinton, G.E.,

    2021

  4. [22]

    1933–1936

    Patch based land cover classification: A comparison of deep learning, svm and nn classifiers, in: IGARSS 2020 - 2020 IEEE International Geoscience and Remote Sensing Symposium, pp. 1933–1936. doi:10.1109/IGARSS39084.2020. 9323755. Parajuli,J.,Fernandez-Beltran,R.,Kang,J.,Pla,F...

    2020

  5. [23]

    International Journal of Applied Earth Observation and Geoinformation 84, 101977

    Terra: Terrain extraction from elevation rasters through repetitive anisotropic filtering. International Journal of Applied Earth Observation and Geoinformation 84, 101977. doi:10.1016/j.jag. 2019.101977. Powell, R., Matzke, N., de Souza, C., Clark, M., Numata, I., Hess, L., R...

    2019

  6. [25]

    Detection of old agricultural terraces in steep, vegetated terrain using airborne lidar: Case studies from hong kong, in: 33rd Asian Conference on Remote Sensing 2012, p. 408–417. URL:https://api.semanticscholar. org/CorpusID:15246203. Schneider, C.A., Rasband, W.S., Eliceiri, K.W.,

    2012

  7. [26]

    Nature Methods 9, 671–675

    NIH image to ImageJ: 25 years of image analysis. Nature Methods 9, 671–675. doi:10. 1038/nmeth.2089. Simonyan,K.,Zisserman,A.,2015. Verydeepconvolutionalnetworksforlarge-scaleimagerecognition. URL: https://arxiv.org/abs/1409. 1556, arXiv:1409.1556. Sofia, G., Marinello, F., Ta...

    2015 arXiv

  8. [29]

    Ecology and Evolution 11, 13684–13691

    Positive impacts of livestock and wild ungulate routes on functioning of dryland ecosystems. Ecology and Evolution 11, 13684–13691. doi: 10.1002/ece3.8147, arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/ece3.8147. J.F. Díez-Pastor et al.:Preprint submitted to Elsevier P...

    doi

  9. [30]

    URL: https://arxiv.org/abs/1602.07261, arXiv:1602.07261

    Inception-v4, inception-resnet and the impact of residual connections on learning. URL: https://arxiv.org/abs/1602.07261, arXiv:1602.07261. Tan,M.,Le,Q.V.,2020.Efficientnet: Rethinkingmodelscalingforconvolutionalneuralnetworks.URL: https://arxiv.org/abs/1905.11946, arXiv:1905....

    2020 arXiv

  10. [31]

    arXiv preprint arXiv:2206.04040

    An improved one millisecond mobile backbone. arXiv preprint arXiv:2206.04040 . Washington-Allen, R.A., Niel, T.G.V., Ramsey, R.D., West, N.E.,

    arXiv

  11. [32]

    Remote Sensing of Environment 287, 113452

    Semantic segmentation of water bodies in very high-resolution satellite and aerial images. Remote Sensing of Environment 287, 113452. doi:10.1016/j.rse.2023.113452. Wieren, S.E., Bakker, J.P.,

    2023

  12. [34]

    International Journal of Applied Earth Observation and Geoinformation 112, 102885

    Semantic image segmentation for sea ice parameters recognition using deep convolutional neural networks. International Journal of Applied Earth Observation and Geoinformation 112, 102885. doi:10.1016/j.jag.2022.102885. Zhang,X.,Jin,J.,Lan,Z.,Li,C.,Fan,M.,Wang,Y.,Yu,X.,Zhang,Y....

    2022

  13. [36]

    Cantabrian

    Pyramid scene parsing network, in: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE Computer Society, Los Alamitos, CA, USA. pp. 6230–6239. doi:10.1109/CVPR.2017.660. J.F. Díez-Pastor et al.:Preprint submitted to Elsevier Page 17 of 17 Remote sensin...

    2017 doi

  14. [1982]

    Zeitschrift fur Geomorphologie 26, 459–472

    Grazing-step terracettes and their significanceGrazing-step terracettes and their significance. Zeitschrift fur Geomorphologie 26, 459–472. doi:10.1127/zfg/26/1982/459. ADS Bibcode: 1982ZGm....26..459H. Hiltbrunner, D., Schulze, S., Hagedorn, F., Schmidt, M.W., Zimmmermann, S.,

    1982 doi

  15. [1992]

    Springer, pp

    Individual comparisons by ranking methods, in: Breakthroughs in statistics: Methodology and distribution. Springer, pp. 196–202. Wurm,M.,Stark,T.,Zhu,X.X.,Weigand,M.,Taubenböck,H.,2019. Semanticsegmentationofslumsinsatelliteimagesusingtransferlearningon fully convolutional neu...

    2019 doi

  16. [2004]

    Remote Sensing of Environment 90, 221–234

    Sources of error in accuracy assessment of thematic land-cover maps in the brazilian amazon. Remote Sensing of Environment 90, 221–234. doi:10.1016/j.rse.2003.12.007. Robbins, C.,

    2003 doi

  17. [2008]

    CATENA 75, 288–296

    Grazing-induced spatial variability of soil bulk density and content of moisture, organic carbon and calcium carbonate in a semi-arid rangeland. CATENA 75, 288–296. doi:10.1016/j.catena.2008.07.007. Stavi, I., Yizhaq, H., Osem, Y., Argaman, E.,

    2008 doi

  18. [2012]

    Geoderma 170, 369–377

    Cattle trampling alters soil properties and changes soil microbial communities in a swiss sub-alpine pasture. Geoderma 170, 369–377. doi:10.1016/j.geoderma.2011.11.026. Hong, D., Zhang, B., Li, H., Li, Y., Yao, J., Li, C., Werner, M., Chanussot, J., Zipf, A., Zhu, X.X.,

    2011 doi

  19. [2014]

    ISPRS Journal of Photogrammetry and Remote Sensing 96, 123–133

    A new landscape metric for the identification of terraced sites: The slope local length of auto-correlation (sllac). ISPRS Journal of Photogrammetry and Remote Sensing 96, 123–133. doi:10.1016/j.isprsjprs.2014.06.018. Stavi, I., Ungar, E.D., Lavee, H., Sarah, P.,

    2014 doi

  20. [2016]

    Ecosphere 7, e01337

    Livestock tracks transform resource distribution on terracette landscapes of the loess plateau. Ecosphere 7, e01337. doi: 10.1002/ecs2.1337, arXiv:https://esajournals.onlinelibrary.wiley.com/doi/pdf/10.1002/ecs2.1337. Jones,C.G.,Lawton,J.H.,Shachak,M.,1994. Organismsasecosyste...

    doi

  21. [2017]

    URL:https://arxiv.org/abs/1610.02357, arXiv:1610.02357

    Xception: Deep learning with depthwise separable convolutionsfefficient. URL:https://arxiv.org/abs/1610.02357, arXiv:1610.02357. da Costa, L.B., de Carvalho, O.L.F., de Albuquerque, A.O., Gomes, R.A.T., Guimarães, R.F., de Carvalho Júnior, O.A.,

    arXiv

  22. [2018]

    URL:https://arxiv.org/abs/ 1608.06993, arXiv:1608.06993

    Densely connected convolutional networks. URL:https://arxiv.org/abs/ 1608.06993, arXiv:1608.06993. Iakubovskii, P.,

    arXiv

  23. [2020]

    Ecology Letters 23, 1298–1309

    The effects of livestock grazing on biodiversity are multi-trophic: a meta-analysis. Ecology Letters 23, 1298–1309. doi: 10.1111/ele.13527, arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/ele.13527. Gordon, I.J., Prins, H.H.T.,

    doi

  24. [2021]

    Ecological Informatics 62, 101279

    Amazon forest cover change mapping based on semantic segmentation by u-nets. Ecological Informatics 62, 101279. doi:10.1016/j.ecoinf.2021.101279. Cai, H., Li, J., Hu, M., Gan, C., Han, S.,

    2021

  25. [2022]

    Geocarto International 37, 6538–6550

    Deep semantic segmentation for detecting eucalyptus planted forests in the brazilian territory using sentinel-2 imagery. Geocarto International 37, 6538–6550. doi:10.1080/10106049.2021.1943009, arXiv:https://doi.org/10.1080/10106049.2021.1943009. DaniloGodone,D.G.,Baldo,M.,201...

    2021

  26. [2023]

    Remote Sensing of Environment 299, 113856

    Cross-city matters: A multimodal remote sensing benchmark dataset for cross-city semantic segmentation using high-resolution domain adaptation networks. Remote Sensing of Environment 299, 113856. doi:10.1016/j.rse.2023.113856. Hosseini, M., Miranda, F., Lin, J., Silva, C.T.,

    2023

  27. [2024]

    arXiv preprint arXiv:2312.00752

    Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 . Guo,Y.,Liu,Y.,Georgiou,T.,Lew,M.,2018. Areviewofsemanticsegmentationusingdeepneuralnetworks. InternationalJournalofMultimedia Information Retrieval

    2018 arXiv

  28. [2025]

    Phileobench: Evaluatinggeo-spatialfoundationmodels

    Fibaek,C.,Camilleri,L.,Luyts,A.,Dionelis,N.,LeSaux,B.,2024. Phileobench: Evaluatinggeo-spatialfoundationmodels. arXiv:2401.04464. Filazzola, A., Brown, C., Dettlaff, M.A., Batbaatar, A., Grenke, J., Bao, T., Peetoom Heida, I., Cahill Jr, J.F.,

    2024 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.