Pith. sign in

REVIEW 3 major objections 7 minor 43 references

Toward Seasonal Guidelines for Robust Deep-Learning Sentinel-2 Building Detection in Different Area Types

T0 review · 3 major / 7 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Building detection from Sentinel-2 imagery is strongly seasonal: models trained on summer scenes are far more accurate than winter-trained models, and matching the training month to the acquisition month usually beats any mismatch.

desk verdict Useful and honest regional study of seasonality in Sentinel-2 building detection; the month-by-month numbers rest on one training scene per month, so treat the rankings as preliminary. read the letter →

arxiv 2607.19994 v2 pith:O2ILILEZ submitted 2026-07-22 cs.CV

Micha{l} Romaszewski , Kamil Drejer , Katarzyna Ko{l}odziej , Anna Zawadzka , Stanis{l}aw Lewiński , Przemys{l}aw G{l}omb , Marek Ruciński , Michal Krupiński
show 2 more authors
Krzysztof Gryguc Przemys{l}aw Seku{l}a Szymon Sala
This is my paper
classification cs.CV
keywords Sentinel-2buildingdetectionseasonalitysemanticsegmentationU-NettransferabilitylandcoverclassificationBDOT10k
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

The paper tries to establish that seasonality is a first-order driver of how well deep-learning models can map buildings from 10 m Sentinel-2 imagery, and to turn that into practical model-selection rules. Using a multi-year set of 65 Sentinel-2 scenes over the Warsaw region, the authors fine-tune U-Net and DeepLabV3+ models for each calendar month, then apply every monthly model to every scene. They find a clear seasonal cycle: median IoU is highest from May to September, peaking in August and September, and collapses in December to median IoU 0.174 (L2A) and 0.221 (L1C). In 10 of 12 acquisition months the best accuracy came from the model trained on the same month, and a Collective Summer Model trained on May–August scenes matches the month-specific models, which matters because it gives practitioners a reusable model for routine annual mapping.

What carries the argument

The central instrument is the seasonal transferability matrix — a 12×12 table of median IoU for every pairing of (training/model month) and (acquisition month), built by fine-tuning one U-Net per month on a single representative Sentinel-2 scene and then classifying every scene with every monthly model. The matrix exposes a high-accuracy May–August block and the winter collapse. Its operational distillation is the Collective Summer Model, a U-Net fine-tuned jointly on four L2A scenes from May–August, which is claimed to replicate month-specific accuracy while removing dependence on a single acquisition date. Ground truth comes from rasterising official Polish topographic building footprints

What would settle it

Rerun the cross-temporal matrix with monthly models trained on several scenes per month (or median composites) and check whether the December collapse and the May–August peak persist; if the winter gap disappears, the seasonal claims are artifacts of single-scene training rather than of season.

Watch

Extended reading notes

Core claim

The paper claims that in the Warsaw region Sentinel-2 building detection has a pronounced seasonal cycle: median IoU peaks in August–September around 0.43 and falls to 0.174 (L2A) or 0.221 (L1C) in December. In 10 of 12 acquisition-month rows, the highest median IoU came from the model trained on the same month as the image. The authors further claim that a Collective Summer Model — one U-Net fine-tuned on four May–August L2A scenes — matches the month-specific models, so it can serve as a reusable, year-after-year building-mapping model. Supporting findings: U-Net beats DeepLabV3+; L2A beats L1C by about 2–3%; dense urban/industrial areas reach median IoU ~0.5 versus ~0.3 in suburban/rural;

Load-bearing premise

Each best monthly model is trained on a single Sentinel-2 scene, so every month-level conclusion — which months are optimal and how well models transfer — rests on the premise that one chosen date represents the whole month's acquisition conditions.

Editorial extensions

If this is right

  • Operational building mapping with Sentinel-2 should preferentially train and acquire imagery in May–August; winter acquisitions should be avoided or handled with special models.
  • Month-matched model selection is the default rule: in 10 of 12 months, the best median IoU came from the model trained on the same calendar month as the image.
  • The Collective Summer Model offers a single reusable model for summer scenes, matching month-specific models and simplifying cloud-free image collection over several years.
  • L2A processing gives a small but real advantage (about 2–3% median IoU/F1), so practitioners wanting cross-date stability should prefer atmospherically corrected products.
  • Accuracy expectations must be typology-aware: medians near 0.5 IoU in dense urban/industrial areas and near 0.3 in suburban/rural areas persist across seasons.

Reading between the lines

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

  • If the paper's single-scene explanation is right, then training each monthly model on a temporal composite (median of several scenes) or an ensemble should specifically rescue winter performance: the December failure is attributed to a snow-covered training scene, so composites should remove that artifact.
  • The morphology-based boundary analysis suggests a testable re-reading of the seasonal matrix: if roughly 30% of raw error is confined to a ±1 pixel boundary band, the apparent seasonal gap might shrink when evaluated on stable building cores and far background rather than full pixel overlap.
  • The May–August optimum is plausibly tied to vegetation contrast and high sun elevation, so the same protocol should yield a different local calendar in other latitudes or in evergreen/arid landscapes; the framework is a method for deriving regional seasonal guidelines rather than a universal month rule.
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 / 7 minor

Summary. This paper asks when, in the seasonal cycle, Sentinel-2 imagery is most (and least) reliable for deep-learning building detection, and how that reliability varies across settlement types. The authors build a multi-temporal dataset of 65 Sentinel-2 acquisitions (2019–2025, L1C and L2A) over a Warsaw-region tile, rasterize official BDOT10k building footprints to 10 m for training and to 2.857 m for evaluation, and fine-tune per-scene U-Net and DeepLabV3+ models with a frozen ResNet-50 encoder pretrained via DeCUR. After selecting U-Net as the more accurate and stable architecture, they pick, for each calendar month and processing level, the single fine-tuned model with the best validation IoU (12 'best monthly models'; Table 3), then apply each monthly model to all 65 dates over eight independent test areas (12,480 classifications). Reported findings: median IoU is highest from May to September and collapses in December (month-matched L2A IoU 0.174 vs 0.427 in August); in 10 of 12 acquisition-month rows the best-scoring model is the month-matched one; L2A narrowly outperforms L1C; industrial/high-density areas are easier than suburban/rural ones; a Collective Summer Model trained on one scene per month from May–August matches the individual monthly models; and a substantial share of residual error is concentrated in a ±1-pixel boundary zone. The paper closes with operational recommendations, notably to prefer the Collective Summer Model for May–August mapping and to avoid singl

Significance. The contribution is a systematic, practice-oriented quantification of a spatio-temporal generalization gap that the literature mostly discusses anecdotally. The experimental design has real strengths: the dataset spans 65 dates with a dedicated, non-overlapping validation strip and eight independent 6.25 km² test areas; the transferability matrix is a measured hold-out result, not a fitted one; the snow-December failure is reported as an explicit negative result with a mechanistic explanation; the external Wrocław check is a genuine spatial generalization test; and the morphology-based error decomposition separates localization from detection. The CSM experiment is a useful, decision-relevant mitigation. If the seasonal trends survive scrutiny, the paper delivers directly actionable guidance (which months to acquire, which model library to keep, when to avoid single-scene training) that is currently lacking in operational Sentinel-2 practice. Its significance is bounded by the single-region, single-backbone scope and by the single-scene-per-month training design discussed below; within those bounds it is a solid empirical contribution.

major comments (3)
  1. [Section 4.2 (Table 3)] Each 'best monthly model' is a single model fine-tuned on one acquisition date, selected on the validation strip. All Phase-2 seasonal conclusions—the May–September optimum, the December collapse, the 10/12 diagonal claim (Fig. 7)—are produced by these twelve single-scene models. The paper acknowledges that the December model is a snow-scene artifact (§4.3.2, §4.3.3), but the same design applies to every month: a favorable single summer scene could inflate that month's results. The rectangular summer block is partly reassuring—several independently selected monthly models transfer well across many independent test scenes—but no within-month variability is quantified: no retraining on alternative scenes from the same month, no selection-stability or leave-one-scene-out analysis, no intervals on monthly medians. With 2–9 scenes per month available (Table 2), this analysis is feasible; §5.6
  2. [Section 4.3.1–4.3.2 (Table 4, Figure 7)] Monthly differences are reported without confidence intervals or significance tests, although several cells rest on very few scenes (January: 2, February: 3; Table 2) and adjacent-month differences are often tiny (L2A IoU 0.427 vs 0.425 vs 0.423 for August/June/May; L1C 0.443 vs 0.431 for May/June). The '10 of 12' diagonal-dominance statement is a comparison of medians; in some rows the best off-diagonal cell is within ~0.01 of the matched-month cell. The 65-date corpus already supplies the repeated test scenes needed for bootstrap confidence intervals or a mixed-effects analysis; adding them would establish which seasonal contrasts are real rather than selection noise.
  3. [Abstract and §4.3.1 vs Table 4] The claim that 'the highest median values are observed from May to September, with peak performance in August and September' is not consistent with the reported medians. For L1C the IoU and F1 peaks are in May (0.443 and 0.613, respectively), not August/September; for L2A the September IoU (0.393) is below April (0.398) and February (0.408), and only August (0.427/0.598) is an actual maximum. The broad May–September block is supported, but the specific peak attribution in the abstract and conclusions should be corrected or qualified.
minor comments (7)
  1. [Tables 3 and 4] Several numeric entries are run together ('2022-02-280.426', '0.4270.5980.824'), likely a formatting artifact; please ensure proper spacing and alignment in the final version.
  2. [Section 3.2] Typographical errors: 'approximately.28%' should read 'approximately 28%', and 'approx.12%M' should read 'approx. 12% of the ~27M total parameters'.
  3. [Section 4.3.2] 'The December model resulted in 0.09–0.13 accuracy for images from almost all months' should read 'median IoU of 0.09–0.13', since Figure 7 reports IoU, not accuracy.
  4. [Section 4.4] The boundary-error decomposition (BA 0.7998→0.8318, boundary error share 0.2974) is computed on a single 'selected random image'. The interpretive statement that 'a significant amount of classification error is linked to boundary localisation' and the explanation of the area bias in §4.3.4 would be more robust if computed over the full corpus or at least several scenes per settlement type.
  5. [Section 5.2] The quantitative bracketing against Feng et al. (IoU 0.311–0.499 across Chinese cities) and Dixit et al. is cross-dataset and uses different ground truth and grids; 'our peak IoU of ~0.43 falls within this range' should be labeled as an informal comparison, not a matched benchmark.
  6. [Section 5.3 and Conclusions] L2A's advantage is stated to be ~2–3% and inconsistent across months (better in six months, comparable in six), yet recommendation 2 says to use L2A 'to ensure the classification stability'. The stability benefit is not measured; either add a dispersion comparison or soften the wording.
  7. [Figure 7] Please add per-cell scene counts (rows range from 2 to 9 scenes; Table 2) and optionally IQRs, since the medians in low-N rows are unstable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the seasonal rankings and transferability matrix are measured hold-out results, not derivations from fitted inputs.

full rationale

The paper's central claims are empirical evaluations rather than derived predictions. Phase 1 selects one model per month on a held-out validation split ('The model with the best validation performance was then selected as the representative monthly model for the subsequent cross-temporal analysis'), and Phase 2 computes the reported metrics on separate test regions ('the reported metrics are computed on the independent test areas'). The seasonal optimum (May–September), the December drop, and the 10/12 diagonal-dominance result are medians over these independent test classifications; they are not obtained by fitting a parameter to the same data and then re-predicting it. The only hand-set quantity entering the metrics is the 30% rasterisation occupancy threshold, which is a fixed preprocessing choice ('A 30% threshold was adopted because it provided the most faithful representation of building shapes'), not a value optimised against the reported IoU. The pretrained encoder [35] is an external dependency rather than a self-citation chain. The acknowledged limitation that each monthly model is trained on a single scene (Section 4.3.3) and the stated geographic and scope restrictions (Section 5.6) weaken external validity but do not make any result equivalent to its input by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests primarily on representativeness assumptions: one scene per month stands for the entire month, the 30% occupancy rasterisation is faithful, and eight small test areas stand for their settlement categories. The paper introduces no new physical entities, and its hyperparameters are standard; the hand-set thresholds are inputs rather than fitted outputs.

free parameters (3)
  • Rasterisation occupancy threshold for 10 m ground-truth masks = 30%
    Chosen from 10%-100% options as 'the most faithful representation of building shapes' (Section 2.2); this threshold affects every training mask and therefore every reported metric.
  • Cloud-exclusion thresholds and s2cloudless parameters = scene selection cloud cover <20%; evaluation clouds removed with threshold=0.7, average_over=4, dilation_size=2; classif
    These hand-set values determine which scenes and pixels enter training and evaluation (Sections 2.2 and 3.4). They are not varied or sensitivity-tested.
  • Patch resampling ratio = 64x64 pixels resampled to 224x224, producing 2.857 m outputs
    Design choice that defines the output grid on which all IoU/F1/BA values are computed (Section 3.2); conclusions about detection versus localisation depend on this grid.
assumptions (4)
  • ad hoc to paper The single selected scene for each month is representative of that month's seasonal and radiometric conditions.
    Each 'best monthly model' is fine-tuned on one scene (Section 4.2, Table 3); if the scene is unusual (e.g., snow cover), the monthly model and transfer matrix reflect that scene, not the month.
  • domain assumption Rasterising BDOT10k footprints with a 30% cell-occupancy threshold faithfully represents true building masks at 10 m.
    The threshold was chosen after visual comparison of 10%-100% options (Section 2.2); no independent accuracy assessment of the resulting masks is provided.
  • domain assumption The eight 6.25 km² test areas adequately represent the five settlement typologies.
    Conclusions about industrial, dense urban, suburban, and rural performance are drawn from one or two small test areas per category (Section 2.1).
  • domain assumption The pretrained DeCUR encoder provides a suitable feature space for both L1C and L2A Sentinel-2 data after the paper's resampling and band alignment.
    The encoder is frozen and used as-is (Section 3.1); its suitability is not validated separately from decoder fine-tuning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toward Seasonal Guidelines for Robust Deep-Learning Sentinel-2 Building Detection in Different Area Types." pith.science (2026). https://pith.science/paper/O2ILILEZ

@misc{pith2026260719994,
  author       = {Pith},
  title        = {Pith review of: Toward Seasonal Guidelines for Robust Deep-Learning Sentinel-2 Building Detection in Different Area Types},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O2ILILEZ}},
  note         = {Machine review of arXiv:2607.19994}
}
read the original abstract

Sentinel-2 imagery offers open access, global coverage, and frequent revisit times, making it attractive for practical building mapping at scale; however, its native 10m resolution makes building vs non-building classification challenging, particularly for small or sub-pixel buildings, and performance can vary with both seasonality and the heterogeneity of built-up environments. This paper introduces a Sentinel-2 building-detection framework designed to systematically quantify these effects and to support more formalised, practice-oriented model selection. We construct a dedicated multi-temporal Sentinel-2 dataset over the Warsaw region and derive binary ground-truth masks by rasterising official Polish topographic database (BDOT10k) building footprints onto the Sentinel-2 pixel grid. Using two established convolutional segmentation backbones (U-Net and DeepLabV3+), we first perform scene-specific fine-tuning to select a robust architecture and identify the best monthly models for L1C and L2A products separately. We then conduct cross-temporal inference by applying each best monthly model to all scenes, enabling an assessment of (i) which months provide favourable training and inference conditions, (ii) how performance transfers between seasons, (iii) the impact of processing level, and (iv) how these effects differ across built-up typologies. Based on these results, we provide practical guidance for routine Sentinel-2 building classification under varying acquisition periods and settlement characteristics.

Figures

Figures reproduced from arXiv: 2607.19994 by the authors.

Figure 1
Figure 1. localisation of training, validation and evaluation (test) areas on the landcover basemap with highlighted [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Selected test areas represent different types of built-up environments: 1 – industrial built-up area, 2 – high [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Example of ground truth data. a) vector database derived from BDOT10k database, b) mask rasterized [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: General workflow of experimental pipeline. Phase 1 performs scene-specific fine-tuning of UNet and [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Distribution of performance metrics (IoU, F1-score, and Balanced Accuracy) for UNet and DeepLabV3+ [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Monthly distributions of Phase 2 classification performance for Sentinel-2 L1C and L2A products. For [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Seasonal transferability of the selected monthly models for Sentinel-2 L2A products. Values show the [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Monthly distribution of IoU values across the eight built-up test regions for Sentinel-2 L2A products in [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Results of the classification with the highest mean IoU across all areas. The regions are representative [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Distribution of classified building area per region, based on the matched-month checkpoints from best [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Classification result for the spatially distinct Wrocław test area, located outside the Warsaw training [PITH_FULL_IMAGE:figures/full_fig_p023_11.png]
Figure 12
Figure 12. Figure 12: Comparison of spectral curves of a roof and deciduous forest recorded in the Sentinel-2 images in the [PITH_FULL_IMAGE:figures/full_fig_p026_12.png]
Figure 13
Figure 13. Figure 13: The sun elevation for Warsaw at 9:45 UTC (during Sentinel-2 acquisitions). [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: Comparison of building shadows at Sentinel-2 images recorded at the end of October and July [PITH_FULL_IMAGE:figures/full_fig_p027_14.png]
Figure 15
Figure 15. Figure 15: Comparison of single monthly models and the Collective Summer Model (CSM) for the stable May–August [PITH_FULL_IMAGE:figures/full_fig_p028_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 11 canonical work pages

  1. [1]

    Dixit, K

    M. Dixit, K. Chaurasia, V. K. Mishra, Dilated-resunet: A novel deep learning architecture for building extraction from medium resolution multi-spectral satellite imagery, Expert Systems with Applications 184 (2021) 115530.doi:10.1016/j.eswa.2021.115530

    arXiv 2021

  2. [2]

    Dixit, K

    M. Dixit, K. Chaurasia, V. K. Mishra, D. Singh, H.-N. Lee, 6+: A novel approach for building extraction from a medium resolution multi-spectral satellite, Sustainability 14 (3) (2022) 1615. doi:10.3390/su14031615

    work page doi:10.3390/su14031615 2022

  3. [3]

    Y.Meng, S.Chen, Y.Liu, L.Li, Z.Zhang, T.Ke, X.Hu, Unsupervisedbuildingextractionfrom multimodal aerial data based on accurate vegetation removal and image feature consistency constraint, Remote Sensing 14 (8) (2022) 1912.doi:10.3390/rs14081912

    work page doi:10.3390/rs14081912 2022

  4. [4]

    K. Reda, M. Kedzierski, Detection, classification and boundary regularization of buildings in satellite imagery using faster edge region convolutional neural networks, Remote Sensing 12 (14) (2020) 2240.doi:10.3390/rs12142240

    work page doi:10.3390/rs12142240 2020

  5. [5]

    S. Zou, L. Wang, Mapping individual abandoned houses across cities by integrating vhr re- mote sensing and street view imagery, International Journal of Applied Earth Observation and Geoinformation 113 (2022) 103018.doi:10.1016/j.jag.2022.103018

    arXiv 2022

  6. [6]

    C. Li, L. Fu, Q. Zhu, J. Zhu, Z. Fang, Y. Xie, Y. Guo, Y. Gong, Attention enhanced u-net for building extraction from farmland based on google and worldview-2 remote sensing images, Remote Sensing 13 (21) (2021) 4411.doi:10.3390/rs13214411. 30

    work page doi:10.3390/rs13214411 2021

  7. [7]

    Y. Gao, S. Lang, D. Tiede, G. W. Gella, L. Wendt, Comparing obia-generated labels and manually annotated labels for semantic segmentation in extracting refugee-dwelling footprints, Applied Sciences 12 (21) (2022) 11226

    2022

  8. [8]

    Rawat, P

    A. Rawat, P. K. Gupta, C. Persello, Deep learning for built-up fractional mapping us- ing sentinel-2 images: A case study in delhi, india, preprint (2024).doi:10.20944/ preprints202401.1879.v1

    arXiv 2024

Show all 43 references
  1. [9]

    Q. Li, L. Mou, Y. Sun, Y. Hua, Y. Shi, X. X. Zhu, A review of building extraction from remote sensing imagery: Geometrical structures and semantic attributes, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–15

    2024

  2. [10]

    Sirko, S

    W. Sirko, S. Kashubin, M. Ritter, A. Annkah, Y. S. E. Bouchareb, Y. Dauphin, D. Keysers, M. Neumann, M. Cisse, J. Quinn, Continental-Scale Building Detection from High Resolution Satellite Imagery, arXiv:2107.12283 [cs] (Jul. 2021).doi:10.48550/arXiv.2107.12283. URLhttp://arxi...

    2021 doi arXiv

  3. [11]

    URLhttps://github.com/microsoft/GlobalMLBuildingFootprints?tab= readme-ov-file#will-there-be-more-data-coming-for-other-geographies

    microsoft/GlobalMLBuildingFootprints: Worldwide building footprints derived from satellite imagery, dataset; licensed under CDLA Permissive 2.0 (2025). URLhttps://github.com/microsoft/GlobalMLBuildingFootprints?tab= readme-ov-file#will-there-be-more-data-coming-for-other-geographies

    2025

  4. [12]

    Y. Wu, X. Li, Y. Gong, D. Mwaniki, A flexible framework for identifying urban villages using Sentinel-2 observations and deep learning, International Journal of Applied Earth Observation and Geoinformation 141 (2025) 104631.doi:10.1016/j.jag.2025.104631. URLhttps://www.science...

    2025

  5. [13]

    L. Feng, P. Xu, H. Tang, Z. Liu, P. Hou, National-scale mapping of building footprints us- ing feature super-resolution semantic segmentation of Sentinel-2 images, GIScience & Remote Sensing 60 (1) (2023) 2196154.doi:10.1080/15481603.2023.2196154

    2023

  6. [14]

    P. Xu, H. Tang, J. Ge, L. Feng, ESPC_NASUnet: An End-to-End Super-Resolution Semantic Segmentation Network for Mapping Buildings From Remote Sensing Images, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14 (2021) 5421–5435. doi:10.1109/JSTARS...

    2021

  7. [15]

    P. M. Atkinson, Mapping sub-pixel boundaries from remotely sensed images, in: Innovations In GIS, 1st Edition, CRC Press, 1997, pp. 184–202

    1997

  8. [16]

    Schug, D

    F. Schug, D. Frantz, A. Okujeni, P. Hostert, Sub-pixel building area mapping based on syn- thetic training data and regression-based unmixing using sentinel-1 and -2 data, Remote Sens- ing Letters 13 (8) (2022) 822–832.doi:10.1080/2150704X.2022.2088253. URLhttps://doi.org/10.1...

    2022

  9. [17]

    M. Hao, G. Dou, X. Zhang, H. Lin, W. Huo, A subpixel mapping method for urban land use by reducing shadow effects, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 16 (2023) 2163–2177.doi:10.1109/JSTARS.2023.3243895

    2023

  10. [18]

    M. Debella-Gilo, Relative performance of super-resolved Sentinel-2 and Copernicus VHR im- ages in mapping built-up areas and building footprints using deep learning, European Journal of Remote Sensing 58 (1) (2025) 2517381.doi:10.1080/22797254.2025.2517381. 31

    2025

  11. [19]

    S. Chen, Y. Ogawa, C. Zhao, Y. Sekimoto, Large-scale individual building extraction from open-source satellite imagery via super-resolution-based instance segmentation approach, IS- PRS Journal of Photogrammetry and Remote Sensing 195 (2023) 129–152.doi:10.1016/j. isprsjprs.20...

    2023 doi

  12. [20]

    S. Du, X. Zhang, Y. Lei, X. Huang, W. Tu, B. Liu, Q. Meng, S. Du, Mapping urban functional zones with remote sensing and geospatial big data: a systematic review, GIScience & Remote Sensing 61 (1) (2024) 2404900.doi:10.1080/15481603.2024.2404900. URLhttps://doi.org/10.1080/154...

    2024

  13. [21]

    Z. Li, W. He, J. Li, H. Zhang, Identifying every building’s function in large-scale urban areas with multi-modality remote-sensing data, in: IGARSS 2024-2024 IEEE International Geo- science and Remote Sensing Symposium, IEEE, 2024, pp. 310–314

    2024

  14. [22]

    T. Hu, M. Han, Z. Guo, A seasonal-series lstm network for irregular urban function zone recognition using sentinel-2 images, Applied Soft Computing 172 (2025) 112876

    2025

  15. [23]

    M. S. Chowdhury, GIS based method for mapping actual LULC by combining seasonal LULCs, MethodsX 11 (2023) 102472.doi:10.1016/j.mex.2023.102472. URLhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC10652103/

    2023

  16. [24]

    J. Liu, J. Heiskanen, E. Aynekulu, P. K. E. Pellikka, Seasonal variation of land cover clas- sification accuracy of Landsat 8 images in Burkina Faso, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences XL-7/W3 (2015) 455– 460.doi:1...

    2015 doi

  17. [25]

    D. T. Myers, D. Jones, D. Oviedo-Vargas, J. P. Schmit, D. L. Ficklin, X. Zhang, Seasonal vari- ation in land cover estimates reveals sensitivities and opportunities for environmental models, Hydrology and Earth System Sciences 28 (23) (2024) 5295–5310, publisher: Copernicus Gm...

    2024 doi

  18. [26]

    L. Li, J. Zhu, G. Cheng, B. Zhang, Detecting High-Rise Buildings from Sentinel-2 Data Based on Deep Learning Method, Remote Sensing 13 (20) (Oct. 2021).doi:10.3390/rs13204073

    2021 doi

  19. [27]

    Vincent, J

    E. Vincent, J. Ponce, M. Aubry, Satellite image time series semantic change detection: Novel architecture and analysis of domain shift (2024).arXiv:2407.07616,doi:10.48550/arXiv. 2407.07616. URLhttps://arxiv.org/abs/2407.07616

    2024 doi arXiv

  20. [28]

    A. Roy, A. B. Inamdar, Multi-temporal Land Use Land Cover (LULC) change analysis of a dry semi-arid river basin in western India following a robust multi-sensor satellite image calibration strategy, Heliyon 5 (4) (2019) e01478.doi:10.1016/j.heliyon.2019.e01478. URLhttps://www....

    2019 doi

  21. [29]

    Robinson, A

    C. Robinson, A. Ortiz, J. M. L. Ferres, B. Anderson, D. E. Ho, Temporal Cluster Matching for Change Detection of Structures from Satellite Imagery (Jun. 2021).arXiv:2103.09787, doi:10.48550/arXiv.2103.09787. 32

    2021 doi arXiv

  22. [30]

    Van Etten, D

    A. Van Etten, D. Hogan, J. M. Manso, J. Shermeyer, N. Weir, R. Lewis, The multi-temporal urban development spacenet dataset, in: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2021, pp. 6398–6407

    2021

  23. [31]

    E. Sun, Y. Cui, P. Liu, J. Yan, A Decade of Deep Learning for Remote Sensing Spatiotemporal Fusion: Advances, Challenges, and Opportunities, arXiv:2504.00901 [cs] (Jul. 2025).doi: 10.48550/arXiv.2504.00901. URLhttp://arxiv.org/abs/2504.00901

    2025 doi arXiv

  24. [32]

    Richling, J

    A. Richling, J. Solon, A. Macias, J. Balon, J. Borzyszkowski, M. Kistowski (Eds.), Region- alna geografia fizyczna Polski: praca zbiorowa, Bogucki Wydawnictwo Naukowe ; na zlecenie Generalnej Dyrekcji Ochrony Środowiska, Poznań : [Warszawa], 2021

    2021

  25. [33]

    Gatti, A

    A. Gatti, A. Bertolini, Sentinel-2 products specification document, Rapport technique (2015) 4–7. URLhttps://sentinels.copernicus.eu/documents/247904/349490/S2_MSI_Product_ Specification.pdf

    2015

  26. [34]

    Ginting, K

    D. Ginting, K. Setiawan, N. Anggraini, A. Suardana, M. Nandika, A. Ulfa, K. Aziz, R. Dewanti, Comparison between top and bottom of atmosphere Sentinel-2 image for mangrove mapping in Balikpapan Bay, East Kalimantan, BIO Web of Conferences 89 (2024) 07003.doi:10.1051/ bioconf/2...

    2024

  27. [35]

    Y. Wang, C. M. Albrecht, N. A. A. Braham, C. Liu, Z. Xiong, X. X. Zhu, Decoupling Common and Unique Representations for Multimodal Self-supervised Learning (Jul. 2024).arXiv: 2309.05300,doi:10.48550/arXiv.2309.05300

    2024 doi arXiv

  28. [36]

    Fitness for purpose

    S. Borkowska, E. Bielecka, K. Pokonieczny, OpenStreetMap – building data completeness vi- sualization in terms of “Fitness for purpose”, Advances in Geodesy and Geoinformation (Oct. 2023).doi:10.24425/agg.2022.141922

    2023

  29. [37]

    Polish Parliament (Sejm), Regulation of the Minister of Development, Labour and Technology of 27 july 2021 on the database of topographic objects and the database of general-geographic objects, as well as standard cartographic products, Journal of Laws of the Republic of Polan...

    2021

  30. [38]

    URLhttps://www.geoportal.gov.pl/

    Główny Urząd Geodezji i Kartografii, Geoportal.gov.pl: Geoportal infrastruktury informacji przestrzennej,https://www.geoportal.gov.pl/(2026). URLhttps://www.geoportal.gov.pl/

    2026

  31. [39]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, in: N. Navab, J. Hornegger, W. M. Wells, A. F. Frangi (Eds.), Medical Im- age Computing and Computer-Assisted Intervention – MICCAI 2015, Springer International Publishing, Ch...

    2015 doi

  32. [40]

    L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, H. Adam, Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation (Aug. 2018).arXiv:1802.02611, doi:10.48550/arXiv.1802.02611. 33

    2018 doi arXiv

  33. [41]

    Sentinel-hub/sentinel2-cloud-detector, Sentinel Hub (Feb. 2026). URLhttps://github.com/sentinel-hub/sentinel2-cloud-detector Appendix A. Examples of classification results This appendix presents selected examples of classification outputs used for visual inspection of model be...

    2026

  34. [43]

    The issue is particularly relevant for detecting new service, office, or public buildings, where green roof designs are increasingly common

    This is an important limitation of the method: vegetated rooftops can be spectrally similar to surrounding green areas and may therefore be missed by the classifier. The issue is particularly relevant for detecting new service, office, or public buildings, where green roof des...

  35. [2023]

    The model broadly reconstructs the shape of buildings and even preserves internal courtyards and gaps between structures. This level of spatial detail would be difficult to obtain with traditional methods such as Random Forest, especially given the 10 m Sentinel-2 input resolu...

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.