Pith. sign in

REVIEW 3 major objections 6 minor 23 references

Segment Anything for Satellite Imagery: A Strong Baseline and a Regional Dataset for Automatic Field Delineation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LoRA fine-tuning of SAM is a strong baseline for automatic field delineation from satellite imagery, beating Mask R-CNN on every dataset tested.

desk verdict The new ERAS dataset and a clean LoRA-finetuned SAM recipe make this a useful empirical paper; the central baseline claim holds, but the temporal split confounds year and region and only one baseline is compared. read the letter →

arxiv 2506.16318 v2 pith:AWKMBHDM submitted 2025-06-19 cs.CV cs.AI

Carmelo Scribano , Elena Govi , Paolo Bertellini , Simone Parisi , Giorgia Franchini , Marko Bertogna This is my paper · ORCID
classification cs.CVcs.AI
keywords fielddelineationSegmentAnythingModelLow-RankAdaptationsatelliteimageryagriculturalparcelsegmentationSentinel-2instanceERASdataset
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 establish that the Segment Anything Model, adapted with parameter-efficient Low-Rank Adaptation (LoRA) fine-tuning, is a strong baseline for automatically drawing agricultural field boundaries from satellite imagery. The adaptation matters because zero-shot SAM is far from usable on this domain, reaching at most 10.5 mAP50 on high-resolution ERAS tiles and less than 7 on Sentinel-2 data. After LoRA fine-tuning with rank 32 on the image encoder and mask decoder, SAM reports mAP50 27.0 on AI4B, 17.1 on the ERAS Sentinel-2 version, and 19.3 on the ERAS high-resolution version, while Mask R-CNN under the same evaluation reaches 14.6, 10.9, and 11.8 respectively. The paper also introduces ERAS, a publicly released dataset covering the Emilia-Romagna region of Italy, and reports that the fine-tuned model generalizes to 2024 imagery with only a 0.9-point drop. A sympathetic reader would care because it offers a low-cost recipe for mapping fields over large areas without expensive ground surveys or training from scratch.

What carries the argument

LoRA, or Low-Rank Adaptation, is the mechanism carrying the argument: it represents the weight update $\Delta W$ as a product $BA$ of two low-rank matrices, so the forward pass becomes $y = W_0 x + BA x$ and only the small matrices $A$ and $B$ are trained. The paper applies LoRA to the attention projection layers of SAM's ViT image encoder and, in the best configuration, to the self-attention and cross-attention layers of the mask decoder, while freezing the prompt encoder. This restricts the number of trainable parameters to about 1.4 million at rank 32, making adaptation to field delineation feasible with a limited number of labeled tiles.

What would settle it

Take a random sample of ERAS tiles, have an independent annotator draw parcel boundaries from the same season's imagery or from a GPS field survey, recompute mAP50 of the LoRA-SAM model against these verified masks, and compare with the reported 17.1 on Sentinel-2 and 19.3 on high-resolution tiles; a large drop would show the central claim is an artifact of label noise.

Watch

Extended reading notes

Core claim

The paper's central claim is that parameter-efficient LoRA fine-tuning adapts SAM into a reliable field-delineation model, providing a strong baseline without training from scratch. On the AI4B Sentinel-2 dataset, fine-tuned SAM with LoRA rank 32, decoder LoRA, and a single-point prompt reaches mAP50 27.0, compared with 14.6 for Mask R-CNN; on the new ERAS dataset it reaches 17.1 on the Sentinel-2 version and 19.3 on the 2.5-meter high-resolution version, versus 10.9 and 11.8 for Mask R-CNN. The paper also reports that zero-shot SAM is far below usable accuracy, with mAP50 between 2.3 and 10.5 across datasets, and that fine-tuning first on AI4B and then on ERAS gives no improvement over training on ERAS alone. Finally, it claims that the ERAS dataset supports geographic and temporal generalization tests, with only a 0.9-point drop when moving from 2023 to 2024 imagery.

Load-bearing premise

The ground-truth parcel boundaries in ERAS come from farmer self-declarations submitted to the administration, and the paper provides no independent verification that those declarations are accurate or aligned with the satellite images; if they are not, the reported mAP50 numbers overstate real field-delineation accuracy.

Editorial extensions

If this is right

  • A parameter-efficient fine-tuned SAM is a practical, low-cost option for automated field-boundary mapping, since only about 1.4 million trainable parameters are needed as opposed to full fine-tuning.
  • On all three benchmarks, fine-tuned SAM beats Mask R-CNN by roughly 60 to 85 percent relative mAP50, so future field-delineation studies should take it as a baseline for comparison.
  • The small 2023-to-2024 drop of 0.9 mAP50 points implies the model can be deployed across growing seasons without immediate retraining, at least in similar agricultural landscapes.
  • Training only on the small ERAS Sentinel-2 set matches two-step AI4B-to-ERAS training, so new regions can be covered from SAM's pretrained weights with a limited number of labeled tiles.
  • The release of ERAS gives the community a public benchmark for Italian agricultural parcels at two resolutions, enabling direct comparisons in future work.

Reading between the lines

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

  • The paper leaves implicit that the 2024 temporal test is partly a test of administrative label stability: because ERAS masks come from farmer self-declarations, year-over-year consistency may reflect unchanged paperwork rather than unchanged fields, and an independent record of actual parcel changes would separate the two.
  • The authors do not search over LoRA ranks beyond 8 and 32, nor do they apply LoRA to the ViT MLP blocks; their own observation that trainable parameters are the bottleneck suggests intermediate ranks or broader encoder coverage could push accuracy further.
  • The modest gap between 10-meter Sentinel-2 imagery (17.1 mAP50) and 2.5-meter commercial imagery (19.3 mAP50) hints that, for large and regular Emilia-Romagna fields, extra resolution buys little; the same comparison on smallholder or fragmented landscapes could change that conclusion.
  • A direct extension of the released dataset is to use ERAS as an out-of-distribution test set for models trained elsewhere, which is exactly the geographic-generalization role the paper demonstrates but does not fully explore for architectures other than Mask R-CNN.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a fine-tuning pipeline for the Segment Anything Model (SAM) based on Low-Rank Adaptation (LoRA) for agricultural field delineation from satellite imagery. It introduces ERAgriSeg (ERAS), a new regional dataset for Emilia-Romagna, Italy, with Sentinel-2 and high-resolution Maxar modalities, and evaluates zero-shot performance, fine-tuning on AI4B and ERAS, comparison against Mask R-CNN, and temporal generalization to 2024. The central claim is that the proposed LoRA fine-tuned SAM provides a robust baseline for automated field delineation, supported by mAP50 improvements over Mask R-CNN in Table 5.

Significance. If the results hold, the paper makes a useful contribution: it shows that parameter-efficient fine-tuning of a foundation model can serve as a strong baseline for field delineation, and it releases a new regional dataset that extends geographic coverage beyond existing benchmarks. The temporal generalization experiment is a valuable addition, as are the comparisons across Sentinel-2 and high-resolution imagery. The main strengths are the public dataset release, the systematic exploration of LoRA configurations, and the clear experimental structure. However, the validity of the central claim depends on label quality and on the soundness of the evaluation protocol, both of which need strengthening.

major comments (3)
  1. [Section 3.2, Tables 4-6] The ERAS ground-truth masks are derived from farmer self-declarations submitted to the administration in 2023, with no independent verification or error analysis reported. Because the same labels are used for training and evaluation, the reported mAP numbers measure agreement with these administrative declarations rather than with true field boundaries. If the declarations simplify boundaries, omit fields, or are temporally misaligned with the imagery, the results overstate real-world delineation accuracy. The paper should report label-quality statistics (e.g., boundary displacement estimates, omission rates, or a manually verified subset) and discuss the limitations of self-declared labels as ground truth.
  2. [Section 5.4, Table 6] The temporal generalization evaluation is confounded by a change in validation-set composition. The 2023 evaluation uses only the Reggio Emilia province, while the 2024 Q1 evaluation uses all provinces, including provinces that appear in the training set. The reported drop of 0.9 mAP50 could therefore reflect a shift in province distribution rather than temporal robustness. The authors should evaluate the 2024 data on the same Reggio Emilia split as 2023, or report per-province results, to isolate the temporal effect.
  3. [Section 5.3, Table 5] The central 'robust baseline' claim rests on a comparison against a single baseline, Mask R-CNN, with no error bars, confidence intervals, or repeated runs. The differences in Table 5 (e.g., 27.0 vs 14.6 on AI4B) are large, but the paper does not establish statistical significance, and the conclusions about specific design choices in Table 2 (e.g., decoder LoRA at r=32) rely on differences of about 1.5 mAP points that may be within run-to-run variability. Reporting multiple seeds or significance tests would substantially strengthen the claims.
minor comments (6)
  1. [Section 4.1] The matrix dimensions given for LoRA are inconsistent: with W0 in R^{d x k}, A in R^{r x k}, and B in R^{d x r}, the product ABx is not defined. The usual convention is A in R^{d x r} and B in R^{r x k}; please correct the notation.
  2. [Table 3] The mAR150 values for ERAS (HR) are reported as 1.66, 1.91, and 2.16, which are far lower than the corresponding Sentinel-2 values (4.8, 5.5, 6.1). This is counterintuitive and suggests a decimal-placement or unit issue; please verify the reported values.
  3. [Section 3.2] The description of the L3 mosaic aggregation as the 'average of the first quartile' is ambiguous; please specify whether the mean is computed over values below the first quartile, or over the lower quartile bin, and clarify the formula.
  4. [Section 5.1] The three pretrained SAM models should be cited to the Segment Anything paper (reference [13]) rather than to the ViT paper (reference [5]).
  5. [Table 1] The Eurocrops row leaves the n-Samples cell empty; using a dash or '—' would make the table easier to read.
  6. [Section 3.2] The rule for choosing the number of tile centroids K ('110% of the total area of the fields, divided by the extent tile extent') is not self-explanatory; please provide an explicit formula and define all quantities.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central baseline claim rests on held-out evaluation, not on fitted inputs or self-citation.

full rationale

The paper's central empirical claim is that LoRA fine-tuning of SAM provides a strong baseline for field delineation, and this claim is supported by evaluations on held-out data that are not used to set the reported metric. Training is performed on AI4B and on ERAS province splits, while validation and temporal generalization use Reggio Emilia (2023) or all-province Q1 2024 data; Table 5 compares SAM LoRA against Mask R-CNN on the same datasets. The LoRA rank, decoder option, and prompting choice are selected through the experimental protocol and then evaluated on separate test splits, so the reported mAP50 values are not forced by construction. The only self-citation, reference [2], appears in a related-work sentence about crop classification methods and is not load-bearing for the paper's derivation or results. The ERAS ground-truth masks are derived from farmer self-declarations, which is a label-quality and external-validity concern rather than a circularity in the derivation chain; it does not make the model's predictions equivalent to its inputs. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and the proposed dataset and fine-tuning pipeline are independently evaluated against published benchmarks. Thus the paper is self-contained with respect to circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central empirical results depend on dataset-label quality, image alignment, resizing choices, and a fixed prompting protocol. No new physical or theoretical entities are introduced. Hyperparameters such as rank, learning rate, batch size, and prompt design are chosen by hand or by small ablations and are treated as free parameters here.

free parameters (6)
  • LoRA rank r = 32
    Chosen from Table 2 comparison; r=8 gave mAP50 25.5 and r=32 gave 27.0 on AI4B. The result depends on this choice.
  • Learning rate = 5e-5 with 256-step linear warmup
    Set in Section 5.2 without a reported sweep; affects convergence and final mAP.
  • Batch size = 4 across 2 GPUs
    Set in Section 5.2; not swept.
  • Training prompt type = single positive point per instance
    Table 2 shows multi-point prompting gives no improvement; single-point prompting is used for all later experiments.
  • Tile centroid number K = 110% of field area divided by tile extent
    Heuristic in Section 3.2 determines ERAS tile composition and affects dataset distribution and results.
  • Inference prompt grid = 32x32 points with NMS
    Automatic SAM pipeline; metrics depend on this fixed grid and post-processing rather than on a deployed prompting strategy.
assumptions (4)
  • domain assumption Farmer self-declarations (LPIS) used to build ERAS ground truth are accurate and current enough for segmentation training and evaluation.
    Section 3.2: masks are derived from self-reported statements submitted by farmers in 2023; no independent verification or error analysis is provided.
  • domain assumption ESA Sentinel-2 Level 3 quarterly mosaics are cloud-free composites that align with the parcel boundaries.
    Section 3.2: pixel values are set to the first quartile of valid observations; no co-registration check with parcel boundaries is reported.
  • domain assumption Rescaling Sentinel-2 tiles from 256x256 to SAM's native 1024x1024 preserves the information needed for field delineation.
    Section 4: required because SAM's ViT encoder does not accept variable input sizes.
  • domain assumption Single-class mAP50 measured with a fixed 32x32 prompt grid is a meaningful proxy for field boundary quality.
    Section 5: metrics are mAP50 and mAR150; boundary accuracy per se is not measured, only mask IoU.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Segment Anything for Satellite Imagery: A Strong Baseline and a Regional Dataset for Automatic Field Delineation." pith.science (2026). https://pith.science/paper/AWKMBHDM

@misc{pith2026250616318,
  author       = {Pith},
  title        = {Pith review of: Segment Anything for Satellite Imagery: A Strong Baseline and a Regional Dataset for Automatic Field Delineation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AWKMBHDM}},
  note         = {Machine review of arXiv:2506.16318}
}
read the original abstract

Accurate mapping of agricultural field boundaries is essential for the efficient operation of agriculture. Automatic extraction from high-resolution satellite imagery, supported by computer vision techniques, can avoid costly ground surveys. In this paper, we present a pipeline for field delineation based on the Segment Anything Model (SAM), introducing a fine-tuning strategy to adapt SAM to this task. In addition to using published datasets, we describe a method for acquiring a complementary regional dataset that covers areas beyond current sources. Extensive experiments assess segmentation accuracy and evaluate the generalization capabilities. Our approach provides a robust baseline for automated field delineation. The new regional dataset, known as ERAS, is now publicly available.

Figures

Figures reproduced from arXiv: 2506.16318 by the authors.

Figure 1
Figure 1. Example predictions from finetuned SAM on the proposed dataset. Additional visualizations are provided in the supplementary material. Accurately mapping the geographical boundaries of cultivated fields is es￾sential for agricultural operations in several ways: enhancing land planning and 4 https://github.com/cscribano/ERAS-dataset arXiv:2506.16318v2 [cs.CV] 23 Jun 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Schematic of SAM architecture in fine-tuning phase. The snowflake indicates that the layer is being frozen, and the flame indicates that it is being trained. For the mask decoder, we experiment with three options (a): frozen, (b): full finetuning, and (c): frozen and adding LoRA layers. 4 Proposed Methodology Data Preparation The tiles obtained from Sentinel-2 have a resolution of (256 × 256) pixels, unlike the CNNs… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 16 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2020)

    Aung, H.L., Uzkent, B., Burke, M., Lobell, D., Ermon, S.: Farm parcel delineation using spatio-temporal convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2020)

    work page 2020

  2. [2]

    In: 2023 18th Conference on Computer Science and Intelligence Systems (FedCSIS)

    Bertellini, P., D’Addese, G., Franchini, G., Parisi, S., Scribano, C., Zanirato, D., Bertogna, M.: Binary classification of agricultural crops using sentinel satellite data and machine learning techniques. In: 2023 18th Conference on Computer Science and Intelligence Systems (FedCSIS). pp. 859–864. IEEE (2023)

    work page 2023

  3. [3]

    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(1), 317–329 (2023)

    work page 2023

  4. [4]

    ISPRS Journal of Photogrammetry and Remote Sensing162, 94–114 (2020)

    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 (2020)

    2020

  5. [5]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    arXiv 2020

  6. [6]

    IEEE Transactions on Geoscience and Remote Sensing 40(5), 1054–1064 (2002)

    Evans, C., Jones, R., Svalbe, I., Berman, M.: Segmenting multispectral landsat tm images into field units. IEEE Transactions on Geoscience and Remote Sensing 40(5), 1054–1064 (2002)

    work page 2002

  7. [7]

    International journal of remote sensing38(7), 1809–1819 (2017)

    Garcia-Pedrero, A., Gonzalo-Martin, C., Lillo-Saavedra, M.: A machine learning approach for agricultural parcel delineation through agglomerative segmentation. International journal of remote sensing38(7), 1809–1819 (2017)

    work page 2017

  8. [8]

    IEEE access7, 158223–158236 (2019) 12 C

    Garcia-Pedrero, A., Lillo-Saavedra, M., Rodriguez-Esparragon, D., Gonzalo- Martin, C.: Deep learning for automatic outlining agricultural parcels: Exploiting the land parcel identification system. IEEE access7, 158223–158236 (2019) 12 C. Scribano, E. Govi, et al

    work page 2019

Show all 23 references
  1. [9]

    In: IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium

    Garnot, V.S.F., Landrieu, L., Giordano, S., Chehata, N.: Time-space tradeoff in deep learning models for crop classification on satellite multi-spectral image time series. In: IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium. pp. 6247–6250. IEEE (2019)

    2019

  2. [10]

    In: Proceedings of the IEEE international conference on computer vision

    He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961–2969 (2017)

    2017

  3. [11]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing12(7), 2217– 2226 (2019)

    Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing12(7), 2217– 2226 (2019)

    2019

  4. [12]

    arXiv preprint arXiv:2106.09685 (2021)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

    2021 arXiv

  5. [13]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026 (2023)

    2023

  6. [14]

    Pattern recognition37(8), 1619–1628 (2004)

    Mueller, M., Segl, K., Kaufmann, H.: Edge-and region-based segmentation tech- nique for the extraction of large, man-made objects in high-resolution satellite imagery. Pattern recognition37(8), 1619–1628 (2004)

    2004

  7. [15]

    International Journal of Applied Earth Observation and Geoinformation 124, 103540 (2023)

    Osco, L.P., Wu, Q., de Lemos, E.L., Gonçalves, W.N., Ramos, A.P.M., Li, J., Junior, J.M.: The segment anything model (sam) for remote sensing applications: From zero to one shot. International Journal of Applied Earth Observation and Geoinformation 124, 103540 (2023)

    2023

  8. [16]

    IEEE Geoscience and Remote Sensing Letters (2023)

    Persello, C., Grift, J., Fan, X., Paris, C., Hänsch, R., Koeva, M., Nelson, A.: Ai4smallfarms: A data set for crop field delineation in southeast asian smallholder farms. IEEE Geoscience and Remote Sensing Letters (2023)

    2023

  9. [17]

    Scientific Data10(1), 612 (2023)

    Schneider, M., Schelte, T., Schmitz, F., Körner, M.: Eurocrops: The largest har- monized open crop dataset across the european union. Scientific Data10(1), 612 (2023)

    2023

  10. [18]

    Remote Sensing 16(9), 1505 (2024)

    Sun, J., Yan, S., Alexandridis, T., Yao, X., Zhou, H., Gao, B., Huang, J., Yang, J., Li, Y.: Enhancing crop mapping through automated sample generation based on segment anything model with medium-resolution satellite imagery. Remote Sensing 16(9), 1505 (2024)

    2024

  11. [19]

    Remote sensing of environment 245, 111741 (2020)

    Waldner, F., Diakogiannis, F.I.: Deep learning on edge: Extracting field bound- aries from satellite images with a convolutional neural network. Remote sensing of environment 245, 111741 (2020)

    2020

  12. [20]

    Remote sensing13(11), 2197 (2021)

    Waldner, F., Diakogiannis, F.I., Batchelor, K., Ciccotosto-Camp, M., Cooper- Williams, E., Herrmann, C., Mata, G., Toovey, A.: Detect, consolidate, delineate: Scalable mapping of field boundaries using satellite images. Remote sensing13(11), 2197 (2021)

    2021

  13. [21]

    Advances in Neural Information Processing Systems36 (2024)

    Wang, D., Zhang, J., Du, B., Xu, M., Liu, L., Tao, D., Zhang, L.: Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems36 (2024)

    2024

  14. [22]

    arXiv preprint arXiv:2304.12620 (2023)

    Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)

    2023 arXiv

  15. [23]

    IEEE Geoscience and Remote Sensing Letters (2024)

    Xue, B., Cheng, H., Yang, Q., Wang, Y., He, X.: Adapting segment anything model to aerial land cover classification with low rank adaptation. IEEE Geoscience and Remote Sensing Letters (2024)

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.