Pith. sign in

REVIEW 4 major objections 5 minor 51 references

Cleaning noisy parcel labels, not scaling models, doubles global field-boundary mapping accuracy.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 13:31 UTC pith:MKLXNQD5

load-bearing objection A serious data-cleaning effort around a stock YOLOv11, with a valuable dataset release and a big claimed gain that currently rests only on a self-made benchmark. the 4 major comments →

arxiv 2607.19069 v1 pith:MKLXNQD5 submitted 2026-07-21 cs.CV

Delineate Anything v2: A Global Foundation Model for Field Delineation

Mykola Lavreniuk , Nataliia Kussul , Andrii Shelestov , Yevhenii Salii , Volodymyr Kuzin , Charlotte Julia Li-Xing Wang , Zoltan Szantoi This is my paper
classification cs.CV
keywords field boundary delineationfoundation modeldata-centric AIlabel noisesatellite imageryinstance segmentationzero-shot generalizationFBIS-73M
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.

Agricultural field boundaries are the spatial unit behind food-security monitoring, subsidy payments, and carbon accounting, but global maps are unreliable because administrative registries often merge several physical fields into one polygon. This paper argues that the binding constraint is label quality, not model size, and builds a 73-million-field dataset across 61 countries with a curation pipeline that fixes this parcel-versus-field mismatch in resolution-specific ways: manual splitting for high-resolution imagery, and an image-space 'pixel homogenization plus edge enhancement' for medium-resolution imagery. On a new manually curated 100-country benchmark, the trained model improves mean average precision at IoU 0.5 from 0.275 to 0.559, a 103.3% relative gain, and the ablation attributes most of that gain to curation rather than to raw data scaling. The model is also fast enough to map all of Ukraine (603,000 km²) in 5.4 hours on a consumer workstation. If correct, this reframes foundation-model progress for geospatial tasks as a data-quality problem that can be solved deliberately rather than by adding parameters.

Core claim

The paper's central claim is that a resolution-aware data curation pipeline, applied to a large but noisy multi-national field-boundary repository, yields a foundation model whose zero-shot generalization is roughly double that of the prior state of the art. The discovery is not a new architecture — the model uses a standard instance-segmentation backbone — but a demonstration that structural label noise (administrative parcels containing several physical fields) is the dominant bottleneck. For high-resolution imagery the fix is manual geometric splitting; for medium-resolution imagery, where manual correction is too costly, the fix is topological image-space adaptation: flagging parcels wit

What carries the argument

FBIS-73M, a 73-million-instance, multi-resolution (0.25–10 m) field-boundary training set spanning 61 countries, plus a two-stage curation pipeline. Automated screening flags parcels with abnormal intra-mask spectral deviation. Flagged high-resolution parcels are manually re-drawn; flagged medium-resolution parcels are repaired in image space: a cyclical bounding transformation (Eq. 1) homogenizes pixel color around the parcel's mode to erase false internal edges, and a Khalimsky-grid edge enhancement step strengthens genuine but visually weak boundaries via soft-light blending. The curated set is trained on 512×512 patches with a standard real-time instance-segmentation backbone, and evalua

Load-bearing premise

The reported doubling of accuracy rests on a benchmark the authors created themselves — exactly four manually selected 'representative' patches per country, with no documented check that these 100 countries are disjoint from the 61 countries in the training set or that the patches were not chosen with the model's strengths in mind.

What would settle it

Run the released model on an independently annotated, pre-registered held-out benchmark (for example, a third-party global field-boundary test set) and compare mAP@0.5. If the roughly 2× gain over the prior state of the art does not reproduce, or if withholding the homogenization and edge-enhancement edits from a random subset of training patches changes accuracy substantially, then either the benchmark or the curation effect is overstated. The single most decisive observation would be a per-country scatter plot of benchmark score versus whether that country appears in the 61-country training

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

If this is right

  • National- and global-scale field maps can be produced at speeds compatible with operational monitoring: the model maps 603,000 km² of Ukraine in 5.4 hours on a consumer workstation, or about 112,000 km² per hour.
  • Scaling data without fixing structural label noise hits a ceiling; the ablation shows raw scaling from 22M to 73M instances adds only +0.086 mAP, while curation adds +0.198, so future dataset efforts should prioritize quality control over volume.
  • Zero-shot transfer to under-represented regions is where the gain is largest: Africa's regional score roughly doubles from 0.251 to 0.584 and Asia & Oceania nearly triples from 0.161 to 0.440, supporting the claim that geographic diversity plus consistent supervision enables planetary generalization.
  • Because the curation pipeline is architecture-agnostic and the data, weights, code, and vector products are publicly released, other geospatial delineation tasks can adopt the same remediation recipe without changing their underlying model.

Where Pith is reading between the lines

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

  • The image-space homogenization and edge-enhancement steps intentionally change the pixel statistics of training patches, so the model may learn to expect a 'cleaned' texture. A testable consequence is that performance on raw, unedited satellite imagery from novel regions could be lower than on the curated benchmark — a distribution gap the paper does not analyze.
  • The same parcel-versus-field mismatch afflicts other geospatial vector layers such as building footprints and land parcels, so the screening-and-remediation recipe could transfer; a natural next experiment is applying it to those domains.
  • Because the benchmark is four hand-picked patches per country with equal country weighting, the reported mAP measures cross-country landscape diversity rather than within-country reliability; per-region and per-season variance is likely larger than a four-patch sample can capture.
  • An independent, pre-registered held-out annotation campaign in countries not represented in the training corpus would settle whether the +103% gain is a property of the method or of the evaluation design, and would control for any overlap between the 100 benchmark countries and the 61 training countries.

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. The paper introduces Delineate Anything v2, a field-boundary delineation foundation model built on a data-centric paradigm rather than architectural innovation. The authors construct FBIS-73M, a 73-million-instance, 61-country multi-resolution training dataset, and propose a resolution-specific curation pipeline that (a) manually splits high-resolution merged parcels and (b) applies image-space pixel homogenization and Khalimsky-grid edge enhancement to medium-resolution imagery. They also manually curate a 100-country evaluation benchmark (4 patches per country) and report that Delineate Anything v2 surpasses Delineate Anything v1 by 0.284 mAP@0.5 (+103.3% relative gain) on this benchmark, with ablations attributing the bulk of the gain (+0.198) to data curation rather than dataset scaling (+0.086). The paper additionally demonstrates nationwide mapping of Ukraine at 5.4 hours on consumer hardware and releases code, weights, dataset, and vector products.

Significance. If the central claims hold, the paper makes a notable empirical contribution: it is the largest public field-boundary training repository, it releases a geographically diverse benchmark, and its ablation provides evidence that supervision quality, not model architecture or raw data volume, is the binding constraint in global field delineation. The release of code, weights, and vector products is a concrete strength that could support reproducible research. The central SOTA claim, however, rests entirely on a self-created, 400-patch benchmark whose independence from the training countries is not demonstrated and whose patch-selection protocol is not reported. The paper also declines quantitative comparison against public external checkpoints, weakening the falsifiability of the SOTA itself. These concerns are load-bearing, but they are addressable within the manuscript's scope.

major comments (4)
  1. [§3.1, §4.2, Table 1] The headline claim of +0.284 mAP@0.5 (+103.3% relative) is measured exclusively on a benchmark the authors constructed: exactly four manually selected 512×512 patches per country. The manuscript calls the benchmark 'independent' and §5 says it evaluates 'outside the training distribution,' but it never states whether the 100 benchmark countries are disjoint from the 61 FBIS-73M training countries, nor does it describe the patch-selection protocol, annotator instructions, or quality-control procedure. If benchmark countries overlap training countries, or if 'representative' patches were chosen in a way that favors the curation pipeline, the zero-shot SOTA claim is an evaluation artifact. Please report an explicit country-overlap analysis, a step-by-step patch-selection protocol, inter-annotator agreement, and multi-seed error bars (or bootstrap confidence intervals) for Table 1. The Europ
  2. [§4.4] The paper states that direct metric benchmarking against FTW, PRUE, NASA Harvest, and Sinergise is 'unfeasible without public weights [1,40]' and therefore provides only visual comparison. This is contradicted by the public availability of FTW and PRUE weights; NASA Harvest static layers are also described as publicly available in the same sentence. As written, the SOTA claim is non-falsifiable because no external baseline is scored on the same benchmark. Please provide quantitative mAP (or an equivalent, explicitly defined metric) for FTW, PRUE, and NASA Harvest on the 100-country benchmark, or justify in detail why such scoring is infeasible despite the released weights.
  3. [§3.3, §3.4, Eq. (1)] The central mechanism of the paper modifies training images rather than labels: pixel homogenization folds large color deviations into a bounded range using a sensor-specific B (Eq. 1), and Khalimsky-grid edge enhancement darkens/lightens pixels around weak boundaries with a soft-light blend of strength α. A model trained on such edited imagery may learn to rely on edit fingerprints (e.g., the triangular-wave quantization pattern or the synthetic contrast edge) that are absent in real, unedited test imagery. The paper provides no analysis of this distribution gap. Please run an inference-only experiment on unedited validation patches (i.e., raw imagery withheld from the pipeline) or an ablation that disables edge enhancement at inference time to confirm that the reported zero-shot gains are not an artifact of train-time image editing.
  4. [§4.3, Table 3] The ablation table is internally consistent, but the gain attributions are presented in a way that can mislead. The text says data curation contributes an additional +0.198 mAP over raw scaling (0.559 − 0.361), which is correct. However, Table 3 labels the 'Full Curation Pipeline' row as '+0.150', which is the incremental gain over the HR-manual-split row (0.409), not the cumulative curation gain. Please relabel the Gain column as incremental or add a separate cumulative-gain column to avoid conflation of the curation contribution with the final increment.
minor comments (5)
  1. [§4.1] Metric notation is inconsistent: 'mAP@0.5:0.95' and 'mAP@[0.5:0.95]' are both used. Please unify.
  2. [§3.2, §3.3, §3.4] Several thresholds are described as 'empirically set' or 'selected conservatively' with no sensitivity analysis: the 15%/10%/5% area-flag thresholds, B=32/48, the mid-range threshold of 64, and the edge-strength and weak-fraction thresholds. Given that the paper's thesis is that curation drives gains, a brief sensitivity study (or at least a statement that results are stable over a range of thresholds) would substantially strengthen the claim.
  3. [Fig. 7 caption] The caption cites 'DelAny [21]' for a static layer, but [21] is the ECAI paper; it would be clearer to cite the specific released checkpoint/static product (e.g., the v1 model repository) rather than the paper.
  4. [§4.2] The sentence 'performance doubles in Africa (+0.333) and nearly triples in Asia & Oceania (+0.279)' is ambiguous: the +0.333 and +0.279 are absolute mAP increases, not relative multiples. Please state both absolute and relative gains explicitly.
  5. [Abstract/§1] Minor wording: 'Delineate Anything v2 surpasses the current state-of-the-art, including the Delineate Anything framework' is circular because the prior state of the art is the authors' own model. Rephrasing as 'previous state of the art' would be more neutral.

Circularity Check

0 steps flagged

No circular derivation found: the headline gain is an empirical measurement on a self-created benchmark, with self-citation baseline but no equation-level reduction to inputs.

full rationale

The central claim (abstract; §4.2) — that Delineate Anything v2 surpasses Delineate Anything by 0.284 mAP@0.5 (+103.3% relative) — is a measured difference on the authors' own 100-country benchmark, not a quantity derived by construction from the curation equations. No equation in §3.3–3.4 is shown to equal the reported metric; Eq. (1) is a data transformation with empirically set B, and the 15/10/5% flags are described as 'selected conservatively' and 'kept fixed across all datasets,' with no statement that they were tuned on the benchmark. Without such a statement, the possibility of benchmark tuning is speculation, not a demonstrable reduction, so the 'fitted input called prediction' pattern is not established. The baseline is the authors' own Delineate Anything [21] and the benchmark is self-created, which weakens external validity, but the comparison is empirical rather than imported by citation; this is a self-citation that is not load-bearing in the derivation. The paper's own limitation in §4.4 ('direct metric benchmarking against baseline frameworks and open mapping pipelines is unfeasible without public weights') and the missing explicit statement in §3.1 that the 100 benchmark countries are disjoint from the 61 training countries are evaluation-validity concerns, not circularity: they affect whether the SOTA claim is falsifiable and independent, but they do not make the result equal to its inputs. The ablation (Table 3) is internally consistent and each increment is measured on the same benchmark, so the conclusion that curation drives the gain is an empirical claim, not a definitional one. Thus no significant circularity; score 2 reflects only the minor self-citation baseline and self-created evaluation, which are treated as non-circular weaknesses.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

No new physical or conceptual entities are postulated; Eq. (1) and the Khalimsky-grid blending are new procedures, not entities. The paper's real ledger is parametric: roughly seven hand-set curation and training thresholds govern the pipeline that produces the reported gains, and the paper provides no evidence they were set on a validation split independent of the 100-country benchmark. The model itself adds no parameters beyond stock YOLOv11.

free parameters (7)
  • Gaussian blur kernel size 13x13 = 13x13
    Used in §3.2 to suppress sensor noise before screening; size chosen without stated criterion.
  • Deviation threshold sets for screening = [10,20,50]% Sentinel-2; [25,35,50]% Planet
    Levels used to construct the intra-mask deviation map in §3.2; chosen by hand and kept fixed across datasets.
  • Multi-field flagging area thresholds = 15% / 10% / 5% of parcel area
    A parcel is flagged as merged if cumulative deviation-level area exceeds these; the text says 'selected conservatively' (§3.2).
  • Sensor-specific bounding range B = B=32 (Sentinel-2), B=48 (Planet)
    Controls the cyclical bounding transform in Eq. (1); 'empirically set' in §3.3.
  • Mid-range threshold and soft-light strength alpha = 64 (8-bit); alpha in [0,1]
    Parameters of the Khalimsky-grid edge enhancement in §3.4; the paper gives no selection procedure.
  • Edge-strength and weak-fraction thresholds = not reported numerically
    Segments are flagged as visually weak below an 'empirical strength threshold' and 'fraction-of-weak-pixels threshold' (§3.4); values not disclosed.
  • AdamW learning rate = 1e-7
    Stated in §4.1 as 'selected empirically'; three orders of magnitude below typical fine-tuning rates, so either a misprint or a load-bearing atypical choice.
axioms (5)
  • domain assumption Multi-field administrative parcels are detectable by intra-mask spectral deviation (flagged when cumulative deviation-area exceeds 15/10/5% of the parcel).
    §3.2. Fails when merged fields have similar crops/phenology (invisible to spectral screening) or when a single field is spectrally heterogeneous (false positives); no validation of the screening recall is provided.
  • domain assumption Editing the training image (pixel homogenization + synthetic edge darkening) yields samples drawn from the same distribution the model will see at inference.
    §3.3–3.4 erase false internal boundaries and inject edge cues at label positions; real test images contain neither edit, so the model must not key on edit fingerprints. The paper relies on the benchmark as an implicit test and never analyzes this distribution gap.
  • standard math Khalimsky-grid digital-topology properties (odd-odd pixels, odd-even/even-odd edges, even-even junctions) as in Khalimsky et al. [15].
    §3.4 uses the grid to localize inter-pixel boundaries; treated as an unproved background result from the cited literature.
  • domain assumption COCO-pretrained YOLOv11 weights are a suitable initialization for satellite field delineation.
    §4.1 initializes from COCO; the natural-image domain is far from satellite imagery, yet this choice is not ablated.
  • domain assumption The manually annotated 100-country benchmark is accurate, its four patches per country are representative, and its countries are (implicitly) not used for training.
    §3.1 states manual curation but gives no annotator count, quality-control procedure, or explicit disjointness statement; the word 'zero-shot' is never formally defined.

pith-pipeline@v1.3.0-alltime-deepseek · 11725 in / 17995 out tokens · 173193 ms · 2026-08-01T13:31:35.308001+00:00 · methodology

0 comments
read the original abstract

Accurate agricultural field boundary delineation at large scale is a foundational task for food security, supply chain transparency, and carbon accounting. While vision foundation models like SAM show remarkable zero-shot capabilities, they frequently fail in geospatial domains due to topological complexity, cropland texturing patterns, and a lack of physical scale awareness. In this work, we introduce Delineate Anything v2, a globally scalable foundation model designed specifically for wide-area field boundary mapping. We construct FBIS-73M, a 73-million-instance multi-resolution dataset spanning 61 countries. To address the pervasive issue of multi-field administrative parcel merging, we introduce a resolution-specific data curation pipeline that leverages topological image-space adaptation to homogenize merged parcels and strengthen weak physical boundaries. Furthermore, we establish a novel, manually curated evaluation benchmark covering 100 countries to assess independent zero-shot generalization. Our results show that Delineate Anything v2 surpasses the current state-of-the-art, including the Delineate Anything framework, by 0.284 mAP@0.5 (+103.3% relative gain), while maintaining execution speeds suitable for rapid national- and global-scale deployment, as demonstrated by nationwide mapping of Ukraine (603,000 km^2) in 5.4 hours on a consumer-grade workstation. Code, pre-trained weights, the FBIS-73M dataset, and ready-to-use national-scale vector boundary products are publicly available at https://github.com/Lavreniuk/Delineate-Anything.

Figures

Figures reproduced from arXiv: 2607.19069 by Andrii Shelestov, Charlotte Julia Li-Xing Wang, Mykola Lavreniuk, Nataliia Kussul, Volodymyr Kuzin, Yevhenii Salii, Zoltan Szantoi.

Figure 1
Figure 1. Figure 1: Overview of the Delineate Anything v2 framework. To overcome structural label noise in global administrative registries, we introduce a resolution-aware data curation pipeline. Raw parcels undergo screening for anomalies. Problematic masks are then resolved via High-Resolution (HR) manual partitioning or Medium-Resolution (MR) topological image-space adaptation (pixel homogenization and edge enhancement). … view at source ↗
Figure 2
Figure 2. Figure 2: Spatial distribution and country-level heterogeneity of the FBIS-73M train￾ing dataset. The global map (top) illustrates geographic sample density, where green shading corresponds to regions containing agricultural field instances (intensity pro￾portional to patch count) and blue shading represents non-agricultural background reference countries. The histogram (bottom) quantifies the long-tailed sample dis… view at source ↗
Figure 3
Figure 3. Figure 3: Spatial and regional distribution of field density within the independent 100- country evaluation benchmark. The choropleth map (top) visualizes the geographic coverage of the test set (exactly 4 patches per country), with shading representing the average number of field instances per image. The distribution (bottom) quantifies the average number of fields per country, color-coded by continent [PITH_FULL_… view at source ↗
Figure 4
Figure 4. Figure 4: Examples of high-resolution manual label refinement. The layout showcases two agricultural scenarios (rows) across three stages (columns, left to right): (1) raw satellite imagery, (2) original overlapping administrative mask with topological errors, and (3) finalized manually split vector boundaries matching the physical field structure. We compute directional gradients by taking the difference between di… view at source ↗
Figure 5
Figure 5. Figure 5: Pipeline for automated medium-resolution topological image-space adaptation. The five panels show (left to right): (1) raw satellite patch, (2) original administrative contours with parcel-versus-field mismatches highlighted in red, (3) processed, homog￾enized fields and enhanced edges extracted via the anomaly mask, (4) modified raster with unaltered (green) and remediated (red) contours indicating where … view at source ↗
Figure 6
Figure 6. Figure 6: Zero-shot qualitative results across major macro-regions. Rows (top to bot￾tom): raw RGB imagery, manual ground truth, and Delineate Anything v2 predictions. Columns represent Europe (Norway), Asia & Oceania (United Arab Emirates), North America (Costa Rica), Latin America & Caribbean (Peru), and Africa (South Sudan). lated Sentinel-2 imagery. Delineate Anything v2 requires only 25 ms to process a single 5… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison against operational mapping products. The field bound￾aries extracted by Delineate Anything v2 and the manual ground truth are contrasted directly with publicly available static layers from Delineate Anything [21], NASA Har￾vest [40], Sinergise Solutions [1], and Fields of the World (FTW) [14, 30]. achieves a new state of the art in global field delineation, significantly outper￾form… view at source ↗

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

51 extracted references · 9 canonical work pages

  1. [1]

    Batič, M., Peressutti, D., Vesel, N., Verbič, S., Lubej, M., Oman Kadunc, N.: EU field boundaries (2024).https://doi.org/10.5281/zenodo.14229033,https: //doi.org/10.5281/zenodo.14229033

    work page doi:10.5281/zenodo.14229033 2024

  2. [2]

    IEEE Transactions on Geo- science and Remote Sensing60(2022).https://doi.org/10.1109/TGRS.2022

    Burgert, T., Ravanbakhsh, M., Demir, B.: On the effects of different types of label noise in multi-label remote sensing image classification. IEEE Transactions on Geo- science and Remote Sensing60(2022).https://doi.org/10.1109/TGRS.2022. 3226371

    doi:10.1109/tgrs.2022 2022

  3. [3]

    org/abs/2511.16719

    Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Suris, D., Ryali, C., Alwala, K.V., Khedr, H., Huang, A., Lei, J., Ma, T., Guo, B., Kalla, A., Marks, M., Greer, J., Wang, M., Sun, P., Rädle, R., Afouras, T., Mavroudi, E., Xu, K., Wu, T.H., Zhou, Y., Momeni, L., Hazra, R., Ding, S., Vaze, S., Porcher, F., Li, F., Li, S., Kamath, A., Cheng, H.K., ...

    Pith/arXiv arXiv 2026

  4. [4]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–17 (2024)

    Chen, K., Liu, C., Chen, H., Zhang, H., Li, W., Zou, Z., Shi, Z.: Rsprompter: Learn- ing to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing62, 1–17 (2024). https://doi.org/10.1109/TGRS.2024.3356074

    arXiv 2024

  5. [5]

    11th International Conference on Learning Representations, ICLR 2023 (5 2022)

    Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision trans- former adapter for dense predictions. 11th International Conference on Learning Representations, ICLR 2023 (5 2022)

    2023

  6. [6]

    D’andrimont, R., Claverie, M., Kempeneers, P., Muraro, D., Yordanov, M., Peres- sutti, D., Batič, M., Waldner, F.: Ai4boundaries: an open ai-ready dataset to map fieldboundarieswithsentinel-2andaerialphotography.EarthSystemScienceData 15, 317–329 (1 2023).https://doi.org/10.5194/ESSD-15-317-2023,

    doi:10.5194/essd-15-317-2023 2023

  7. [7]

    ISPRS Journal of Photogrammetry and Remote Sensing162, 94–114 (4 2020).https://doi.org/ 10.1016/J.ISPRSJPRS.2020.01.013

    Diakogiannis, F.I., Waldner, F., Caccetta, P., Wu, C.: Resunet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing162, 94–114 (4 2020).https://doi.org/ 10.1016/J.ISPRSJPRS.2020.01.013

    doi:10.1016/j.isprsjprs.2020.01.013 2020

  8. [8]

    IEEE Geoscience and Remote Sensing Letters21(2024).https://doi

    Feng, W., Guan, F., Sun, C., Xu, W.: Road-sam: Adapting the segment anything model to road extraction from large very-high-resolution optical remote sensing images. IEEE Geoscience and Remote Sensing Letters21(2024).https://doi. org/10.1109/LGRS.2024.3430900

    arXiv 2024

  9. [9]

    Remote Sensing 2024, Vol

    Hell, M., Brandmeier, M.: Identifying plausible labels from noisy training data for a land use and land cover classification application in amazônia legal. Remote Sensing 2024, Vol. 16, Page 208016, 2080 (6 2024).https://doi.org/10.3390/ RS16122080

    2024

  10. [10]

    Conference on Human Factors in Computing Systems - Proceedings (5 2021).https://doi.org/10.1145/3411764

    Highfll, N.S.S.K.H., Aroyo, D.A.P.P.L., Sambasivan, N., Kapania, S., Highfll, H., Akrong, D., Paritosh, P., Aroyo, L.: Everyone wants to do the model work, not the data work: Data cascades in high-stakes ai. Conference on Human Factors in Computing Systems - Proceedings (5 2021).https://doi.org/10.1145/3411764. 3445518

    doi:10.1145/3411764 2021

  11. [11]

    Remote Sensing 2024, Vol

    Huang, Z., Jing, H., Liu, Y., Yang, X., Wang, Z., Liu, X., Gao, K., Luo, H.: Segment anything model combined with multi-scale segmentation for extracting complex cultivated land parcels in high-resolution remote sensing images. Remote Sensing 2024, Vol. 16, Page 348916, 3489 (9 2024).https://doi.org/10.3390/ RS16183489 16 M. Lavreniuk et al

    2024

  12. [12]

    Jocher, G., Qiu, J., Liu, M., Lyu, S., Akyon, F.C., Kalfaoglu, M.E.: Ultralytics yolo26: Unified real-time end-to-end vision models (2026),https://arxiv.org/ abs/2606.03748

    Pith/arXiv arXiv 2026

  13. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence39, 28151–28159 (9 2024).https://doi.org/10.1609/aaai.v39i27.35034

    Kerner, H., Chaudhari, S., Ghosh, A., Robinson, C., Ahmad, A., Choi, E., Jacobs, N., Holmes, C., Mohr, M., Dodhia, R., Ferres, J.M.L., Marcus, J.: Fields of the world:Amachinelearningbenchmarkdatasetforglobalagriculturalfieldboundary segmentation. Proceedings of the AAAI Conference on Artificial Intelligence39, 28151–28159 (9 2024).https://doi.org/10.1609...

    work page doi:10.1609/aaai.v39i27.35034 2024

  14. [14]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Kerner, H., Chaudhari, S., Ghosh, A., Robinson, C., Ahmad, A., Choi, E., Jacobs, N., Holmes, C., Mohr, M., Dodhia, R., et al.: Fields of the world: A machine learning benchmark dataset for global agricultural field boundary segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 28151–28159 (2025)

    2025

  15. [15]

    Topology and its Applications36, 1–17 (7 1990)

    Khalimsky, E., Kopperman, R., Meyer, P.R.: Computer graphics and connected topologies on finite ordered sets. Topology and its Applications36, 1–17 (7 1990). https://doi.org/10.1016/0166-8641(90)90031-V

    work page doi:10.1016/0166-8641(90)90031-v 1990

  16. [16]

    arXiv preprint arXiv:2410.17725 (10 2024)

    Khanam, R., Hussain, M.: Yolov11: An overview of the key architectural enhance- ments. arXiv preprint arXiv:2410.17725 (10 2024)

    Pith/arXiv arXiv 2024

  17. [17]

    Proceedings of the IEEE International Conference on Computer Vision pp

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. Proceedings of the IEEE International Conference on Computer Vision pp. 3992– 4003 (4 2023).https://doi.org/10.1109/ICCV51070.2023.00371

    arXiv 2023

  18. [18]

    IEEE Geoscience and Remote Sensing Letters14, 778–782 (5 2017).https://doi.org/10.1109/LGRS

    Kussul, N., Lavreniuk, M., Skakun, S., Shelestov, A.: Deep learning classification of land cover and crop types using remote sensing data. IEEE Geoscience and Remote Sensing Letters14, 778–782 (5 2017).https://doi.org/10.1109/LGRS. 2017.2681128

    arXiv 2017

  19. [19]

    IEEE Journal of Selected Topics in Applied Earth Observations and Re- mote Sensing9, 2500–2508 (6 2016).https://doi.org/10.1109/JSTARS.2016

    Kussul, N., Lemoine, G., Gallego, F.J., Skakun, S.V., Lavreniuk, M., Shelestov, A.Y.: Parcel-based crop classification in ukraine using landsat-8 data and sentinel- 1a data. IEEE Journal of Selected Topics in Applied Earth Observations and Re- mote Sensing9, 2500–2508 (6 2016).https://doi.org/10.1109/JSTARS.2016. 2560141

    work page doi:10.1109/jstars.2016 2016

  20. [20]

    In: European Conference on Computer Vision Workshops (ECCVW)

    Lavreniuk, M., Bhat, S.F., Muller, M., Wonka, P.: Evp: Enhanced visual percep- tion using inverse multi-attentive feature refinement and regularized image-text alignment. In: European Conference on Computer Vision Workshops (ECCVW). pp. 206–225 (2024)

    2024

  21. [21]

    In: European Conference on Artificial Intelligence (4 2025)

    Lavreniuk, M., Kussul, N., Shelestov, A., Yailymov, B., Salii, Y., Kuzin, V., Szantoi, Z.: Delineate anything: Resolution-agnostic field boundary delineation on satellite imagery. In: European Conference on Artificial Intelligence (4 2025). https://doi.org/10.48550/arXiv.2504.02534

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.02534 2025

  22. [22]

    Lei, M., Li, S., Wu, Y., Hu, H., Zhou, Y., Zheng, X., Ding, G., Du, S., Wu, Z., Gao, Y.: Yolov13: Real-time object detection with hypergraph-enhanced adaptive visual perception (2025),https://arxiv.org/abs/2506.17733

    Pith/arXiv arXiv 2025

  23. [23]

    ISPRS Journal of Photogrammetry and Remote Sensing200, 24–40 (6 2023).https: //doi.org/10.1016/J.ISPRSJPRS.2023.04.019

    Li, M., Long, J., Stein, A., Wang, X.: Using a semantic edge-aware multi-task neural network to delineate agricultural parcels from remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing200, 24–40 (6 2023).https: //doi.org/10.1016/J.ISPRSJPRS.2023.04.019

    work page doi:10.1016/j.isprsjprs.2023.04.019 2023

  24. [24]

    Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)8693 LNCS, 740–755 (5 2014).https://doi

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)8693 LNCS, 740–755 (5 2014).https://doi. org/10.1007/978-3-319-10602-1_48 Delineate ...

    doi:10.1007/978-3-319-10602-1_48 2014

  25. [25]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–17 (2024).https://doi.org/10.1109/TGRS.2024.3373908

    Liu, C., Albrecht, C.M., Wang, Y., Li, Q., Zhu, X.X.: Aio2: Online correction of object labels for deep learning with incomplete annotation in remote sensing image segmentation. IEEE Transactions on Geoscience and Remote Sensing62, 1–17 (2024).https://doi.org/10.1109/TGRS.2024.3373908

    arXiv 2024

  26. [26]

    Interna- tional Journal of Applied Earth Observation and Geoinformation118, 103265 (4 2023).https://doi.org/10.1016/J.JAG.2023.103265

    Liu, Y., Zhong, Y., Ma, A., Zhao, J., Zhang, L.: Cross-resolution national-scale land-cover mapping based on noisy label learning: A case study of china. Interna- tional Journal of Applied Earth Observation and Geoinformation118, 103265 (4 2023).https://doi.org/10.1016/J.JAG.2023.103265

    arXiv 2023

  27. [27]

    International Conference on Geoinformatics2022-August(2022).https://doi.org/10.1109/ GEOINFORMATICS57846.2022.9963838

    Long, J., Li, M., Wang, X.: Boundary-aware multi-task neural networks for agri- cultural field extraction from very high-resolution satellite images. International Conference on Geoinformatics2022-August(2022).https://doi.org/10.1109/ GEOINFORMATICS57846.2022.9963838

    arXiv 2022

  28. [28]

    IEEE Geoscience and Remote Sensing Letters 21(2024).https://doi.org/10.1109/LGRS.2024.3454263

    Long, J., Zhao, H., Li, M., Wang, X., Lu, C.: Integrating segment anything model derived boundary prior and high-level semantics for cropland extraction from high- resolution remote sensing images. IEEE Geoscience and Remote Sensing Letters 21(2024).https://doi.org/10.1109/LGRS.2024.3454263

    arXiv 2024

  29. [29]

    IEEE Transactions on Geoscience and Remote Sensing62(2024).https://doi.org/10.1109/TGRS

    Luo, M., Zhang, T., Wei, S., Ji, S.: Sam-rsis: Progressively adapting sam with box prompting to remote sensing image instance segmentation. IEEE Transactions on Geoscience and Remote Sensing62(2024).https://doi.org/10.1109/TGRS. 2024.3460085

    arXiv 2024

  30. [30]

    arXiv preprint arXiv:2603.27101 (2026)

    Muhawenayo, G., Robinson, C., Khanal, S., Fang, Z., Corley, I., Wollam, A., Gao, T.,Strnad,L.,Avery,R.,Estes,L.,etal.:Prue:Apracticalrecipeforfieldboundary segmentation at scale. arXiv preprint arXiv:2603.27101 (2026)

    arXiv 2026

  31. [31]

    Ng, A.: Data-centric ai (2021),https://datacentricai.org/, accessed: 2026-06- 23

    2021

  32. [32]

    International Journal of Applied Earth Observation and Geoinformation124(11 2023).https://doi.org/10.1016/j.jag.2023.103540

    Osco, L.P., Wu, Q., de Lemos, E.L., Gonçalves, W.N., Ramos, A.P.M., Li, J., Marcato, J.: The segment anything model (sam) for remote sensing applications: From zero to one shot. International Journal of Applied Earth Observation and Geoinformation124(11 2023).https://doi.org/10.1016/j.jag.2023.103540

    arXiv 2023

  33. [33]

    Pan, C., Fan, X., Tjahjadi, T., Guan, H., Fu, L., Ye, Q., Wang, R.: Vision founda- tion model guided multimodal fusion network for remote sensing semantic segmen- tation.IEEEJournalofSelectedTopicsinAppliedEarthObservationsandRemote Sensing18, 9409–9431 (2025).https://doi.org/10.1109/JSTARS.2025.3547880

    arXiv 2025

  34. [34]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–15 (2024).https://doi.org/10.1109/ TGRS.2024.3367850

    Pan,Y.,Wang,X.,Wang,Y.,Zhong,Y.:Rbp-mtl:Agriculturalparcelvectorization via region-boundary-parcel decoupled multitask learning. IEEE Transactions on Geoscience and Remote Sensing62, 1–15 (2024).https://doi.org/10.1109/ TGRS.2024.3367850

    arXiv 2024

  35. [35]

    ISPRS Journal of Photogrammetry and Remote Sensing203, 246–264 (9 2023).https://doi.org/10.1016/J.ISPRSJPRS.2023.08.001

    Pan, Y., Wang, X., Zhang, L., Zhong, Y.: E2evap: End-to-end vectorization of smallholder agricultural parcel boundaries from high-resolution remote sensing im- agery. ISPRS Journal of Photogrammetry and Remote Sensing203, 246–264 (9 2023).https://doi.org/10.1016/J.ISPRSJPRS.2023.08.001

    work page doi:10.1016/j.isprsjprs.2023.08.001 2023

  36. [36]

    Remote Sensing 2017, Vol

    Pelletier, C., Valero, S., Inglada, J., Champion, N., Sicre, C.M., Dedieu, G.: Effect of training class label noise on classification performances for land cover mapping with satellite image time series. Remote Sensing 2017, Vol. 9, Page 1739, 173 (2 2017).https://doi.org/10.3390/RS9020173

    doi:10.3390/rs9020173 2017

  37. [37]

    IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 20, 2023 (2023).https://doi.org/10.17026/dans-xy6-ngg6 18 M

    Persello, C., Member, S., Grift, J., Fan, X., Paris, C., Hänsch, R., Koeva, M., Nelson, A.: Ai4smallfarms: A dataset for crop field delineation in southeast asian smallholder farms. IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 20, 2023 (2023).https://doi.org/10.17026/dans-xy6-ngg6 18 M. Lavreniuk et al

    work page doi:10.17026/dans-xy6-ngg6 2023

  38. [38]

    13th International Conference on Learning Representations, ICLR 2025 pp

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: Sam 2: Segment anything in images and videos. 13th International Conference on Learning Representations, ICLR 2025 pp. 41175–41218 (8 2024),https:...

    Pith/arXiv arXiv 2025

  39. [39]

    8340–8350 (4 2023).https: //doi.org/10.1109/WACV57701.2024.00817

    Ren, S., Luzi, F., Lahrichi, S., Kassaw, K., Collins, L.M., Bradbury, K., Malof, J.M.: Segment anything, from space? Proceedings - 2024 IEEE Winter Conference on Applications of Computer Vision, WACV 2024 pp. 8340–8350 (4 2023).https: //doi.org/10.1109/WACV57701.2024.00817

    arXiv 2024

  40. [40]

    Scientific Data12(12 2025).https: //doi.org/10.1038/S41597-025-05190-7,

    Sadeh, Y., Wagner, J., Nair, S.S., Oliinyk, O., Belles, E., Bibih, A., D’Harboullé, L., Bendahmane, H., Gupta, M., Becker-Reshef, I.: National-scale in-season field boundaries of ukraine using remote sensing. Scientific Data12(12 2025).https: //doi.org/10.1038/S41597-025-05190-7,

    work page doi:10.1038/s41597-025-05190-7 2025

  41. [41]

    One Earth3, 237–250 (8 2020).https://doi.org/10.1016/j.oneear.2020.07.011

    Scown, M.W., Brady, M.V., Nicholas, K.A.: Billions in misspent eu agricultural subsidies could support the sustainable development goals. One Earth3, 237–250 (8 2020).https://doi.org/10.1016/j.oneear.2020.07.011

    work page doi:10.1016/j.oneear.2020.07.011 2020

  42. [42]

    Frontiers in Earth Science5, 1–10 (2 2017)

    Shelestov, A., Lavreniuk, M., Kussul, N., Novikov, A., Skakun, S.: Exploring google earth engine platform for big data processing: Classification of multi-temporal satellite imagery for crop mapping. Frontiers in Earth Science5, 1–10 (2 2017). https://doi.org/10.3389/FEART.2017.00017/TEXT

    arXiv 2017

  43. [43]

    International Journal of Applied Earth Observation and Geoinformation 133, 104085 (9 2024).https://doi.org/10.1016/J.JAG.2024.104085

    Sun, J., Yan, S., Yao, X., Gao, B., Yang, J.: A segment anything model based weakly supervised learning method for crop mapping using sentinel-2 time series images. International Journal of Applied Earth Observation and Geoinformation 133, 104085 (9 2024).https://doi.org/10.1016/J.JAG.2024.104085

    arXiv 2024

  44. [44]

    International Journal of Digital Earth18(12 2025).https://doi.org/10.1080/ 17538947.2025.2497487

    Tang,Z.,Wang,X.,Jiang,Q.,Pan,H.,Deng,G.,Chen,H.,You,Y.,Li,S.,Hou,H.: Parcel-scale crop planting structure extraction combining time-series of sentinel-1 and sentinel-2 data based on a semantic edge-aware multi-task neural network. International Journal of Digital Earth18(12 2025).https://doi.org/10.1080/ 17538947.2025.2497487

    arXiv 2025

  45. [45]

    Tian, Y., Ye, Q., Doermann, D.: Yolov12: Attention-centric real-time object detec- tors (2025),https://arxiv.org/abs/2502.12524

    Pith/arXiv arXiv 2025

  46. [46]

    ISPRS Journal of Photogrammetry and Re- mote Sensing79, 106–121 (5 2013).https://doi.org/10.1016/J.ISPRSJPRS

    Turker, M., Kok, E.H.: Field-based sub-boundary extraction from remote sensing imagery using perceptual grouping. ISPRS Journal of Photogrammetry and Re- mote Sensing79, 106–121 (5 2013).https://doi.org/10.1016/J.ISPRSJPRS. 2013.02.009

    doi:10.1016/j.isprsjprs 2013

  47. [47]

    Remote Sensing of En- vironment245, 111741 (8 2020).https://doi.org/10.1016/J.RSE.2020.111741

    Waldner, F., Diakogiannis, F.I.: Deep learning on edge: Extracting field boundaries from satellite images with a convolutional neural network. Remote Sensing of En- vironment245, 111741 (8 2020).https://doi.org/10.1016/J.RSE.2020.111741

    arXiv 2020

  48. [48]

    IEEE Geoscience and Remote Sensing Letters22 (2025).https://doi.org/10.1109/LGRS.2025.3567600

    Wang, J., Zhang, X., Ma, X., Yu, W., Ghamisi, P.: Auto-prompting sam for weakly supervised landslide extraction. IEEE Geoscience and Remote Sensing Letters22 (2025).https://doi.org/10.1109/LGRS.2025.3567600

    arXiv 2025

  49. [49]

    IEEE Journal of Selected Topics in Applied Earth Ob- servations and Remote Sensing14, 951–963 (2021).https://doi.org/10.1109/ JSTARS.2020.3043442

    Wang, Y., Ding, W., Zhang, R., Li, H.: Boundary-aware multitask learning for remote sensing imagery. IEEE Journal of Selected Topics in Applied Earth Ob- servations and Remote Sensing14, 951–963 (2021).https://doi.org/10.1109/ JSTARS.2020.3043442

    arXiv 2021

  50. [50]

    ACM Computing Surveys57, 129 (1 2025)

    Zha, D., Bhat, Z.P., Lai, K.H., Yang, F., Jiang, Z., Zhong, S., Hu, X.: Data- centric artificial intelligence: A survey. ACM Computing Surveys57, 129 (1 2025). https://doi.org/10.1145/3711118

    doi:10.1145/3711118 2025

  51. [51]

    Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training (2023),https://github.com/Sense-X/Co-DETR

    2023