pith. sign in

arxiv: 2604.27247 · v2 · pith:LEHNWRHAnew · submitted 2026-04-29 · 💻 cs.CV

Towards Generalizable Mapping of Hedges and Linear Woody Features from Earth Observation Data: a national Product for Germany

Thorsten Hoeser , Verena Huber-Garcia , Sarah Asam , Ursula Gessner , Claudia Kuenzer This is my paper

Pith reviewed 2026-05-25 06:20 UTC · model grok-4.3

classification 💻 cs.CV
keywords linear woody featureshedgesEarth observationdeep neural networknational mappingbinary maskgeneralizable workflowGermany
0
0 comments X

The pith

A single neural network separates linear woody features from binary masks created from varied Earth observation inputs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a modular workflow that first converts heterogeneous Earth observation data into a binary woody vegetation mask. A deep neural network then classifies shapes within the mask as linear or non-linear. The same trained model produces national maps of hedges and similar features for all of Germany from three different input resolutions without retraining. Evaluation on refined reference data from four federal states shows competitive performance against existing maps. The design focuses on reusability across sensor types and landscape conditions.

Core claim

The workflow uses two independently optimizable parts: a flexible input interface that consolidates diverse Earth observation data into a binary woody vegetation mask, and a deep neural network trained to separate linear from non-linear shapes in the mask. A single model trained once derives three national-scale linear woody feature maps for Germany from 0.73 m, 1 m, and 3 m resolution sources, with results competitive against refined biotope mapping references and two prior maps across evaluation sites.

What carries the argument

The modular workflow consisting of a flexible input data interface that produces a binary woody vegetation mask, followed by a deep neural network that separates linear from non-linear shapes within the mask.

If this is right

  • National-scale linear woody feature maps can be generated from multiple spatial resolutions using one model without retraining.
  • The approach yields results competitive with refined reference data from multiple federal state biotope campaigns and with two existing maps.
  • The modular structure supports extension to other regions with comparable data diversity and landscape variability.
  • Such mapping directly informs management and preservation of hedges for ecosystem services in agricultural areas.

Where Pith is reading between the lines

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

  • The binary mask intermediate step likely reduces sensor-specific variations to aid cross-resolution generalization.
  • The same structure could be adapted to map other linear landscape elements such as field boundaries or watercourses.
  • Testing the model on data from countries outside Germany with similar agricultural patterns would check broader transferability.
  • Adding multi-temporal inputs to the mask creation step might further stabilize performance under changing acquisition conditions.

Load-bearing premise

That consolidating heterogeneous Earth observation data into a binary woody vegetation mask preserves the shape information needed for a neural network to accurately separate linear from non-linear features across different landscapes and conditions without retraining.

What would settle it

If applying the single trained model to a new input resolution or unseen landscape region produces maps whose accuracy falls substantially below the reported competitive level when checked against independent reference data from biotope surveys.

Figures

Figures reproduced from arXiv: 2604.27247 by Claudia Kuenzer, Sarah Asam, Thorsten Hoeser, Ursula Gessner, Verena Huber-Garcia.

Figure 1
Figure 1. Figure 1: Comparison of different sensor and elevation data products across different sites in Germany, ranging view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the three different input sources separated into a set of heterogeneous data layers provided by the view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the acquisition dates of the DOP20 BKG mosaic tiles and their spatial distribution across view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the spatial distribution of tiles of the BKG DOP20 mosaic and their leaf on / off condition, where view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the spatial distribution of the 5 different reference data sets. The reference data of North view at source ↗
Figure 6
Figure 6. Figure 6: Conceptual overview of the proposed workflow for linear woody feature mapping. The workflow consists view at source ↗
Figure 7
Figure 7. Figure 7: Conditional workflow used to create binary woody vegetation masks from heterogeneous input data of the view at source ↗
Figure 8
Figure 8. Figure 8: Example DOP20 tiles under leaf-on conditions, the respective NDVI distributions with detected peaks and view at source ↗
Figure 9
Figure 9. Figure 9: Eight examples of synthetic scenes composed of the scene elements view at source ↗
Figure 10
Figure 10. Figure 10: Conceptual overview of the modified U-Net architecture with a three-channel input (binary woody vegetation view at source ↗
Figure 11
Figure 11. Figure 11: Training loss and validation F1 scores over training steps, where each step corresponds to one processed view at source ↗
Figure 11
Figure 11. Figure 11: 3.4 Large-scale inference For efficient and seamless inference over the nationwide binary woody vegetation masks organized as GTI, the GTI’s raster extent is used to virtually divide the mosaic into chips of 1,024 × 1,024 pixels with a horizontal and vertical overlap of 50 %. For each virtual chip bounds, raster values are accessed directly via the GTI and passed to the model for inference. On predictions… view at source ↗
Figure 12
Figure 12. Figure 12: Example prediction results derived from BKG-based woody vegetation masks across Germany, showing a view at source ↗
Figure 13
Figure 13. Figure 13: Visual comparison of predictions and reference data across all evaluation sites. The predictions correspond view at source ↗
Figure 14
Figure 14. Figure 14: Quantitative evaluation results across all evaluation sites and predictions. Pixel-wise metrics are reported view at source ↗
read the original abstract

Hedges and other linear woody features provide valuable ecosystem services, particularly within intensively managed agricultural landscapes. They are key elements for climate adaptation and biodiversity amongst others not only due to a largely varying flora, but also as a feeding-, resting-, and nesting place for many animals and insects including valuable pollinators. Therefore, they require dedicated management, preservation, and attention. Thus, systematic and large-scale mapping of these features from Earth observation data is of high importance. However, transferable and reusable workflows for linear woody feature mapping remain a key methodological challenge, given the diversity of sensor types, spatial resolutions, data acquisition conditions, and complex landscape variability encountered across study areas. We introduce a modular workflow built around two independently optimizable components. Firstly, a flexible input data interface that consolidates heterogeneous Earth observation data into a binary woody vegetation mask, and secondly, a deep neural network trained to separate linear from non-linear shapes within these masks. We demonstrate the workflow by deriving three national-scale linear woody feature maps for all of Germany from three input sources with 0.73 m, 1 m and 3 m spatial resolution, respectively, by using a single trained model without retraining. Evaluation against refined reference data from four federal state biotope mapping campaigns and comparison with two existing linear woody feature maps demonstrate that the workflow produces competitive results across all evaluation sites on a national level. The modular design and its demonstrated applicability at national scale provide a foundation for scalable and generalizable linear woody feature mapping beyond Germany.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims to introduce a modular workflow for mapping hedges and linear woody features from Earth observation data at national scale. The workflow has a flexible interface to create binary woody vegetation masks from heterogeneous inputs and a DNN to separate linear from non-linear shapes. A single trained model is used to generate three national maps for Germany from 0.73 m, 1 m, and 3 m resolution data without retraining, with competitive results against biotope mapping reference data.

Significance. If the results hold, the modular workflow could be significant for enabling scalable, generalizable mapping of linear woody features across diverse data sources and resolutions, supporting ecosystem service assessments and biodiversity monitoring at national levels. The use of independent reference data is a positive aspect.

major comments (2)
  1. The central claim of cross-resolution generalization without retraining (Abstract) depends on the binary woody vegetation mask preserving linear topology sufficiently for the fixed DNN. At 3 m resolution, features narrower than one pixel may reduce to isolated or disconnected pixels, qualitatively altering connectivity and aspect ratio relative to 0.73 m inputs. The manuscript provides no ablation isolating this effect, such as comparing native 3 m masks versus downsampled 0.73 m masks fed to the identical network.
  2. The abstract reports competitive national-level results against independent reference data from four federal state biotope mapping campaigns but provides no quantitative metrics, error analysis, or details on training/evaluation protocols. This absence makes it impossible to verify the strength of the cross-site and cross-resolution claims.
minor comments (1)
  1. The abstract would be strengthened by including at least one key quantitative result (e.g., an F1 score or IoU value) to support the competitiveness claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our cross-resolution generalization results. We address each major comment below and will revise the manuscript to strengthen the supporting evidence where needed.

read point-by-point responses

  1. Referee: The central claim of cross-resolution generalization without retraining (Abstract) depends on the binary woody vegetation mask preserving linear topology sufficiently for the fixed DNN. At 3 m resolution, features narrower than one pixel may reduce to isolated or disconnected pixels, qualitatively altering connectivity and aspect ratio relative to 0.73 m inputs. The manuscript provides no ablation isolating this effect, such as comparing native 3 m masks versus downsampled 0.73 m masks fed to the identical network.

    Authors: We agree that an explicit ablation isolating the contribution of the binary mask at native 3 m resolution versus downsampled higher-resolution inputs would provide stronger support for the generalization claim. The current workflow applies the same mask-generation procedure across sources to maintain consistency, and the DNN was trained on diverse linear and non-linear shapes extracted from multiple resolutions; however, the suggested ablation was not performed. We will add this analysis in the revised manuscript, including quantitative comparison of topology metrics (e.g., connectivity and aspect ratio) before and after the DNN stage. revision: yes

  2. Referee: The abstract reports competitive national-level results against independent reference data from four federal state biotope mapping campaigns but provides no quantitative metrics, error analysis, or details on training/evaluation protocols. This absence makes it impossible to verify the strength of the cross-site and cross-resolution claims.

    Authors: The abstract follows the conventional high-level summary style and therefore omits specific numbers. All quantitative metrics (precision, recall, F1, and spatial agreement), error analysis, training protocols, and evaluation details against the four independent biotope-mapping datasets are reported in the results and methods sections of the full manuscript, along with comparisons to two existing maps. To address the referee's concern about verifiability from the abstract alone, we will incorporate the key national-scale performance figures into the abstract in the revision. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical ML workflow with external validation

full rationale

The paper describes a modular empirical pipeline (binary mask generation from heterogeneous EO inputs followed by a fixed DNN for linear/non-linear separation) trained once and applied to three resolutions. All performance claims rest on evaluation against independent biotope mapping reference data from four federal states plus comparison to two external maps. No equations, fitted parameters, or self-citations are invoked as load-bearing derivations; the workflow is self-contained against external benchmarks and does not reduce any result to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard remote sensing and deep learning assumptions rather than new fitted parameters or invented entities; full paper would be needed to audit training details.

axioms (2)
  • domain assumption Heterogeneous Earth observation data from different sensors and resolutions can be consolidated into a reliable binary woody vegetation mask without loss of critical shape information.
    This is the first independently optimizable component of the workflow described in the abstract.
  • domain assumption A deep neural network trained on vegetation masks can separate linear from non-linear shapes in a manner that generalizes across landscapes, acquisition conditions, and input resolutions without retraining.
    This is the second component and the basis for using a single model at national scale.

pith-pipeline@v0.9.0 · 5823 in / 1394 out tokens · 40670 ms · 2026-05-25T06:20:24.256292+00:00 · methodology

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages · 1 internal anchor

  1. [1]

    URL https://www.sciencedirect.com/science/ article/pii/S0167880999001231

    doi:https://doi.org/10.1016/S0167-8809(99)00123-1. URL https://www.sciencedirect.com/science/ article/pii/S0167880999001231. J. R. Brandle, L. Hodges, and X. H. Zhou. Windbreaks in north american agricultural systems.Agroforestry Systems, 61(1):65–78, 7 2004. ISSN 1572-9680. doi:10.1023/B:AGFO.0000028990.31801.62. URL https: //doi.org/10.1023/B:AGFO.00000...

    work page doi:10.1016/s0167-8809(99)00123-1 2004

  2. [2]

    Julie Betbeder, Jean Nabucet, Eric Pottier, Jacques Baudry, Samuel Corgne, and Laurence Hubert-Moy

    doi:10.1109/IGARSS.2013.6723533. Julie Betbeder, Jean Nabucet, Eric Pottier, Jacques Baudry, Samuel Corgne, and Laurence Hubert-Moy. Detection and characterization of hedgerows using terrasar-x imagery.Remote Sensing, 6(5):3752–3769, 2014. ISSN 2072-4292. doi:10.3390/rs6053752. URLhttps://www.mdpi.com/2072-4292/6/5/3752. Kevin Tansey, Ian Chambers, Andrew...

    work page doi:10.1109/igarss.2013.6723533 2013

  3. [3]

    doi:https://doi.org/10.1016/j.isprsjprs.2015.09.007

    ISSN 0924-2716. doi:https://doi.org/10.1016/j.isprsjprs.2015.09.007. URL https://www.sciencedirect. com/science/article/pii/S0924271615002087. Verena Huber-García, Jennifer Kriese, Sarah Asam, Mariel Dirscherl, Michael Stellmach, Johanna Buchner, Kris- tel Kerler, and Ursula Gessner. Hedgerow map of bavaria, germany, based on orthophotos and convolutional...

    work page doi:10.1016/j.isprsjprs.2015.09.007 2015

  4. [4]

    URL https://zslpublications.onlinelibrary.wiley.com/ doi/abs/10.1002/rse2.432

    doi:https://doi.org/10.1002/rse2.432. URL https://zslpublications.onlinelibrary.wiley.com/ doi/abs/10.1002/rse2.432. Richard K. Broughton, Rich J. Burkmar, Morag E. McCracken, Nadine Mitschunas, Lisa R. Norton, Denise W. Pallett, Justine Patton, John W. Redhead, Jo T. Staley, Claire M. Wood, and Richard F. Pywell. A lidar based national model of hedgerows...

    work page doi:10.1002/rse2.432

  5. [5]

    doi:https://doi.org/10.1016/j.jenvman.2025.126705

    ISSN 0301-4797. doi:https://doi.org/10.1016/j.jenvman.2025.126705. URL https://www.sciencedirect. com/science/article/pii/S0301479725026817. Michelangelo Conserva, Alex Wilson, Charlotte Stanton, Vishal Batchu, and Varun Gulshan. Mapping farmed landscapes from remote sensing.arXiv preprint, abs/2506.13993, 2025. URL https://arxiv.org/abs/2506. 13993. Anna...

    work page doi:10.1016/j.jenvman.2025.126705 2025

  6. [6]

    doi:https://doi.org/10.1016/j.rse.2023.113888

    ISSN 0034-4257. doi:https://doi.org/10.1016/j.rse.2023.113888. URL https://www.sciencedirect.com/ science/article/pii/S003442572300439X. Oriane Siméoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan...

    work page doi:10.1016/j.rse.2023.113888 2023

  7. [7]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library

    URLhttps://arxiv.org/abs/1912.01703. Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2019. URLhttps://arxiv.org/abs/1711.05101. Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2017. URLhttps://arxiv.org/abs/1608.03983. Zengqiang...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/tmi.2017.2778748 1912

  8. [8]

    URL https://www.bdvi.de/application/files/8716/9565/1129/FORUM_3-2023_Digitaler_ Zwilling.pdf. 31

    work page