Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Baltimore Atlas: FreqWeaver Adapter for Semi-supervised Ultra-high Spatial Resolution Land Cover Classification

T0 review · 4 major / 5 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 Keeping the SAM2 backbone frozen and adding a frequency-splitting adapter makes 0.3 m urban land cover classification accurate with a fraction of the trainable parameters.

desk verdict A reasonable applied paper with a genuinely new single-city dataset, but the headline gains rest on a self-built benchmark with five test patches and no error bars; the ideas deserve a serious referee, and the evidence does not yet support the margins. read the letter →

arxiv 2506.15565 v2 pith:HRZHQLE7 submitted 2025-06-18 cs.CV

Junhao Wu , Aboagye-Ntow Stephen , Chuyuan Wang , Gang Chen , Xin Huang This is my paper · ORCID
classification cs.CV
keywords ultra-highspatialresolutionlandcoverclassificationparameter-efficientadapterfrequencydecompositionSAM2semi-supervisedsegmentationmeanteacherurbanremotesensing
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 claims that ultra-high spatial resolution land cover mapping can be done accurately without full fine-tuning of a large segmentation foundation model. It keeps the SAM2 backbone frozen, inserts a small FreqWeaver adapter that splits features into low- and high-frequency branches, and adds an entropy-weighted teacher–student loop to exploit unlabeled imagery. On a new 0.3 m Baltimore Atlas dataset, the full framework reaches 0.5778 IoU, 1.78 points above the best parameter-efficient baseline and 3.44 points above a state-of-the-art high-resolution segmentation method, while updating only 5.96% of model parameters. If correct, this is a practical route to sub-meter land cover maps under the severe annotation bottleneck of UHSR data.

What carries the argument

The FreqWeaver Adapter is a parameter-efficient module inserted into transformer stages 3 and 4 of a frozen SAM2 backbone; it is the only trainable part of the model. Its work is to decouple frequencies with a channel-wise 2D DFT and radial masks, process low-frequency global structure with an 11×11 depthwise convolution, high-frequency local detail with a 5×5 depthwise convolution, retain a spatial MLP branch, and fuse the branches with router weights derived from global average pooling. The companion Uncertainty-Aware Teacher–Student (UATS) mechanism uses an EMA-updated teacher to produce pseudo-labels on weakly perturbed inputs and weights the consistency loss by per-pixel confidence $w_j = 1 - H_j/\log C$, so uncertain pixels contribute less supervision.

What would settle it

Have several remote-sensing analysts independently re-annotate the five test patches and compute inter-annotator agreement; if the agreement is low or if FreqWeaver's edge over Conv-LoRA shrinks below the label-noise level when compared against the fresh annotations, the claimed improvement is not established.

Watch

Extended reading notes

Core claim

The central claim is that a frozen foundation model can be adapted to the multi-scale, fine-textured patterns of 0.3 m urban imagery more effectively by frequency-decoupled adapters than by full fine-tuning or existing adapter designs. FreqWeaver performs a 2D discrete Fourier transform, applies radial masks to isolate low- and high-frequency components, models the low-frequency part with 11×11 depthwise convolutions, the high-frequency part with 5×5 depthwise convolutions, preserves an unfiltered spatial branch, and fuses all three through a router. Inserted only in deeper SAM2 transformer stages, this adapter reaches 0.5520 IoU with 4.12M trainable parameters, outperforming full fine-tuning (0.5239 IoU, 69.11M parameters). Adding the Uncertainty-Aware Teacher–Student framework, which weights pseudo-labels pixel-wise by $1 - H/\log C$, raises IoU to 0.5778 on the Baltimore Atlas test set.

Load-bearing premise

Everything hinges on the corrected test labels of the five Baltimore Atlas test patches being accurate enough that IoU differences of 1.8 to 3.4 points are real signal rather than label noise or bias inherited from the automatically generated initial annotations.

Editorial extensions

If this is right

  • Full fine-tuning of the segmentation backbone becomes unnecessary: FreqWeaver delivers higher IoU than full fine-tuning while updating 4.12M instead of 69.11M parameters.
  • Semi-supervised gains are additive: entropy-weighted pseudo-labels add about 2.6 IoU points on top of the adapter, so unlabeled 0.3 m imagery can be converted into usable supervision.
  • Adapter placement is the decisive design choice: adapters in early SAM2 stages hurt performance, and stages 3–4 are the best insertion points; the authors report the same trend for other adapter designs.
  • Frequency decoupling outperforms low-rank and single-branch adapters as a way to transfer foundation-model priors to sub-meter remote sensing.

Reading between the lines

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

  • Beyond the paper: the headline comparisons rest on a single city and five test patches, so the portability of the 1.8–3.4 point gains to other cities, sensors, or seasons is an open question.
  • Beyond the paper: the radial cutoff $\rho$ and kernel sizes are fixed choices; a sweep over these would show how sensitive the gains are and whether the frequency split is the true driver.
  • Beyond the paper: the same frozen-backbone-plus-frequency-adapter recipe could be tested on other UHSR sensors and on instance-level tasks, since the visualizations already suggest the model resolves individual vehicles and pools.
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

4 major / 5 minor

Summary. The paper proposes Baltimore Atlas, a framework for ultra-high spatial resolution (UHSR) land cover classification at 0.3 m resolution. The framework combines a new dataset (Baltimore Atlas Dataset) derived from NAIP imagery of Baltimore City, a parameter-efficient FreqWeaver Adapter that adapts a frozen SAM2 backbone by frequency-domain decomposition and multi-branch convolution, and an Uncertainty-Aware Teacher–Student (UATS) semi-supervised learning scheme that uses entropy-based confidence weighting for pseudo-labels. Experiments report IoU/Dice improvements over PEFT baselines (e.g., +1.78 IoU over Conv-LoRA) and over prior LULC methods (e.g., +3.44 IoU over SAM-RS) using only 5.96% of total model parameters, along with ablations on adapter components and insertion stages.

Significance. If the reported gains are real, the paper makes a useful practical contribution: a lightweight adapter for adapting SAM2 to sub-meter remote sensing imagery, combined with a semi-supervised training scheme that reduces labeled-data dependence. The explicit parameter accounting, the adapter stage-placement ablation, and the detailed architectural description are strengths. However, the headline claims currently rest on a self-created benchmark with five test patches, single-run evaluations, and labels inherited from an automated 1 m product (UrbanWatch) with undocumented manual correction. The margins of 1.78–3.44 IoU are within the range that could be explained by label noise, patch selection, or training stochasticity. The evaluation weaknesses are load-bearing and need to be addressed before the central claims can be considered established.

major comments (4)
  1. [Sec. 4.1 and Tables 1–5] The entire quantitative evaluation is based on only five 1024×1024 test patches, with no error bars, confidence intervals, or multiple seeds reported for any experiment. The reported advantages are 1.78% IoU over Conv-LoRA (Table 1) and 3.44% IoU over SAM-RS (Table 2). With n=5 patches and a single run, these gaps could easily be reversed by patch selection or training stochasticity. Please report per-patch scores, results over at least 3–5 random seeds (mean ± std), and a significance test (e.g., paired bootstrap over patches) for the main comparisons.
  2. [Sec. 4.1] The ground-truth labels for the Baltimore Atlas Dataset are generated by running the UrbanWatch model (trained on 1 m imagery) to produce initial annotations, followed by "manual corrections" with no protocol, no annotator qualifications, no inter-annotator agreement, and no independent audit. Since a co-author of the present paper is also a co-author of UrbanWatch (ref. [45]), this creates a circularity risk: the benchmark may inherit UrbanWatch's systematic errors, and the 0.3 m corrections are undocumented. Please describe the correction protocol in detail, quantify how many pixels were changed, report inter-annotator agreement on a subset, and ideally validate on an existing independent UHSR benchmark (e.g., Vaihingen, Potsdam, or a manually labeled subset from another city).
  3. [Table 3 and Sec. 6] Table 3 reports 0.00% user's and producer's accuracy for the Agriculture class, meaning the model never predicts Agriculture. The paper acknowledges class imbalance (Agriculture is 0.047% of pixels, Fig. 3), but the abstract and conclusion still claim "high-accuracy and robust" mapping. This complete failure on a defined class materially weakens the claim and may distort overall IoU/Dice comparisons, since Agriculture contributes almost nothing to the union/intersection. Please report per-class IoU for all compared methods in Tables 1 and 2, and discuss how the 0.00% Agriculture result affects the headline comparisons. Consider also reporting macro-averaged metrics or excluding Agriculture with clear justification.
  4. [Sec. 4.5 and Table 2] It is unclear whether the LULC baselines in Table 2 (ABCNet, UNetFormer, CMTFNet, RS3Mamba, SAM-RS) are trained fully supervised on the 55 labeled patches, while "Ours" additionally uses unlabeled data via UATS. Table 4 shows the full model (FW+UATS) achieves 0.5778 IoU, while FW alone gives 0.5520 IoU. If the Table 2 baselines are fully supervised, the reported 3.44% IoU gain conflates the contribution of semi-supervised learning with the adapter's contribution and is not a fair comparison. Please state the training protocol for every method, and add semi-supervised baselines (e.g., self-training, FixMatch, UniMatch) or at least compare FW-only against the Table 2 baselines under identical supervised data.
minor comments (5)
  1. [Sec. 3.2.1, Eq. (2)] The frequency cutoff ρ is a free hyperparameter, but its value is never reported. Please provide the chosen value and a sensitivity analysis showing how IoU varies with ρ.
  2. [Sec. 3.3.1] The weak and strong augmentation pipelines are not specified beyond "weakly augmented" and "strongly augmented." Please list the exact augmentation operations (e.g., color jitter, cutout, flip, scale ranges) so the experiments can be reproduced.
  3. [Table 5] The caption and check-mark notation in Table 5 are difficult to parse: the first row '✓ ✓ ✓ ✓' gives IoU 0.5238, but the text states inserting adapters into all stages can harm performance. Please clarify which stages contain adapters in each row, and consider adding rows for all two-stage combinations to support the claim that Stages 3+4 is best.
  4. [General] There are several typos and inconsistencies: 'Relate Work' in the Section 2 heading, 'Uncertainty-A ware' in the abstract and Section 3.3, and 'A d p t' in Figure 1. Please copyedit throughout.
  5. [References] Reference [31] is incomplete ('Clay Foundation: Clay Foundation Model. 2024. Clay.'), and reference [10] lists 'et al.' after the first author; please provide full author lists and page numbers where applicable.

Circularity Check

1 steps flagged · score 4.0 of 10

The method's architecture is independently specified, but every headline IoU gain is measured on a self-created benchmark whose ground truth is initialized by UrbanWatch, a prior model with overlapping authorship, and whose manual refinement is undocumented.

  1. self citation load bearing [Section 4.1 (Dataset) and the abstract's headline comparisons (Tables 1 and 2)]
    "In the first stage, we use the UrbanWatch model, trained on 1 m resolution imagery, to generate the initial annotations as a coarse baseline. In the second stage, these annotations are refined through manual corrections to improve accuracy."

    All reported IoU/Dice numbers are computed against the Baltimore Atlas Dataset, whose ground truth is initialized by UrbanWatch, a prior model with overlapping authorship (Gang Chen is a co-author of both UrbanWatch [45] and this paper). The paper provides no inter-annotator agreement, no independent label audit, and no external benchmark check; the 'manual corrections' protocol is unspecified. Thus the headline claims '1.78% IoU improvement' and '3.44% IoU gain' reduce to scoring higher on labels that originate from the authors' own prior model plus undocumented fixes. This is load-bearing because every headline comparison uses this dataset and there is no independent validation that the labels are accurate enough to support margins of 1.8 to 3.4 IoU points.

full rationale

I walked the paper's derivation chain. The FreqWeaver Adapter (Section 3.2) is specified by explicit equations: Fourier decomposition with radial masks (Eqs. 1-4), branch convolutions (Eqs. 5-7), and a router-based weighted fusion (Eqs. 8-10). The Uncertainty-Aware Teacher-Student framework (Section 3.3) is also explicitly defined through EMA updates, entropy-based confidence weights, and a combined loss (Eqs. 11-16). No parameter is fitted to a target and then renamed as a prediction; no uniqueness theorem is imported from the authors; no ansatz is smuggled in via citation; and no known result is merely renamed. The method itself is not circular. The one genuine circularity is in the evaluation loop: the benchmark ground truth is generated by the UrbanWatch model, a prior work with overlapping authorship, and the manual refinement is undocumented, so the benchmark is not independent of the authors' prior system. This is a load-bearing self-citation for the empirical claims. The paper also acknowledges in Section 6 that agriculture is severely underrepresented, and Table 3 reports 0.00 user's/producer's accuracy for Agriculture; that is a robustness limitation rather than circularity, but it reinforces that the reported margins rest on a fragile, self-created benchmark. Given that the central architecture has independent content and the circularity is confined to benchmark provenance, a score of 4 is appropriate.

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

All experimental claims rest on a self-created, unreleased benchmark whose labels come from a prior model by the same research group. The method adds several hand-set hyperparameters and no external validation, so the ledger is dominated by domain assumptions about label quality and transferability rather than by new physical or theoretical entities.

free parameters (5)
  • Frequency cutoff rho = not reported
    Defines the radial binary masks in Eq. (2) that split low and high frequency features; no value or sensitivity analysis is given.
  • EMA decay alpha = 0.99
    Set in Eq. (11) for the teacher update; no ablation over momentum values is provided.
  • Consistency loss weight lambda = 0.1
    Set in Eq. (16) to balance labeled and unlabeled losses; no sensitivity analysis is reported.
  • Adapter kernel sizes = 11x11 low frequency, 5x5 high frequency
    Depthwise convolution sizes in Section 3.2.2 are chosen by hand; no ablation over kernel sizes is reported.
  • Adapter placement stages = Stage 3 and Stage 4 only
    Selected empirically from Table 5; adding early stages hurts and only the last stage underfits, but the choice was made on the same test set.
assumptions (4)
  • domain assumption The UrbanWatch-derived, manually corrected labels are valid ground truth for 0.3 m land cover mapping.
    Section 4.1 uses UrbanWatch predictions as initial labels; if these labels carry systematic errors, all IoU numbers inherit them.
  • domain assumption Frozen SAM2 features plus an adapter transfer effectively to UHSR remote sensing imagery.
    Sections 3.1 and 4.3 assume SAM2's pretrained knowledge is a useful base for 0.3 m aerial imagery without reporting a comparison on alternate foundations.
  • domain assumption Entropy-based pixel confidence makes pseudo-labels sufficiently reliable for semi-supervised learning.
    Section 3.3.2 defines confidence weights from entropy; the paper does not test this assumption against alternative uncertainty estimates.
  • domain assumption 60 patches (55 train, 5 test) from selected Baltimore areas represent the diversity of UHSR scenes.
    Section 4.1 and Section 6 acknowledge geographic limitations and rare-class imbalance; the assumption underpins all generalization claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Baltimore Atlas: FreqWeaver Adapter for Semi-supervised Ultra-high Spatial Resolution Land Cover Classification." pith.science (2026). https://pith.science/paper/HRZHQLE7

@misc{pith2026250615565,
  author       = {Pith},
  title        = {Pith review of: Baltimore Atlas: FreqWeaver Adapter for Semi-supervised Ultra-high Spatial Resolution Land Cover Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HRZHQLE7}},
  note         = {Machine review of arXiv:2506.15565}
}
read the original abstract

Ultra-high Spatial Resolution (UHSR) Land Cover Classification is increasingly important for urban analysis, enabling fine-scale planning, ecological monitoring, and infrastructure management. It identifies land cover types on sub-meter remote sensing imagery, capturing details such as building outlines, road networks, and distinct boundaries. However, most existing methods focus on 1 m imagery and rely heavily on large-scale annotations, while UHSR data remain scarce and difficult to annotate, limiting practical applicability. To address these challenges, we introduce Baltimore Atlas, a UHSR land cover classification framework that reduces reliance on large-scale training data and delivers high-accuracy results. Baltimore Atlas builds on three key ideas: (1) Baltimore Atlas Dataset, a 0.3 m resolution dataset based on aerial imagery of Baltimore City; (2) FreqWeaver Adapter, a parameter-efficient adapter that transfers SAM2 to this domain, leveraging foundation model knowledge to reduce training data needs while enabling fine-grained detail and structural modeling; (3) Uncertainty-Aware Teacher Student Framework, a semi-supervised framework that exploits unlabeled data to further reduce training dependence and improve generalization across diverse scenes. Using only 5.96% of total model parameters, our approach achieves a 1.78% IoU improvement over existing parameter-efficient tuning strategies and a 3.44% IoU gain compared to state-of-the-art high-resolution remote sensing segmentation methods on the Baltimore Atlas Dataset.

Figures

Figures reproduced from arXiv: 2506.15565 by the authors.

Figure 1
Figure 1. Overview of the proposed framework. Given a remote sensing image, two augmented views are generated: a strongly [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed FreqWeaver Adapter. The FreqWeaver Adapter is inserted into specific Transformer layers [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visualization of land cover class distribution. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of the State-of-the-Art PEFT Method. In the first row, our method shows better capture of textures and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the State-of-the-Art LULC Method. The results show that earlier methods do not work well on [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visualization of Land Cover Classification on the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 46 canonical work pages

  1. [45]

    Myint, Yuyu Zhou, Geoffrey J

    Yindan Zhang, Gang Chen, Soe W. Myint, Yuyu Zhou, Geoffrey J. Hay, Jelena Vukomanovic, and Ross K. Meentemeyer. 2022. UrbanWatch: A 1-Meter Resolu- tion Land Cover and Land Use Database for 22 Major Cities in the United States. Remote Sens. Environ. (2022), 113106

    work page 2022

  2. [1]

    David Berthelot, Nicholas Carlini, Ian Goodfellow, Avital Oliver, Nicolas Papernot, and Colin Raffel. 2019. MixMatch: a holistic approach to semi-supervised learning. NeurIPS (2019), 11

    work page 2019

  3. [2]

    Gang Chen, Qihao Weng, Geoffrey J Hay, and Yinan He. 2018. Geographic object- based image analysis (GEOBIA): Emerging trends and future opportunities.GISci. Remote Sens. (2018), 159–182

    work page 2018

  4. [3]

    Shoufa Chen, Chongjian GE, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. 2022. AdaptFormer: Adapting Vision Transformers for Scalable Visual Recognition. In NeurIPS. 16664–16678

    work page 2022

  5. [4]

    Jiangfan Feng, Hongxin Luo, and Zhujun Gu. 2025. Improving semi-supervised remote sensing scene classification via Multilevel Feature Fusion and pseudo- labeling. Int. J. Appl. Earth Obs. Geoinf. (2025), 104335

    work page 2025

  6. [5]

    Shengxi Gui, Shuang Song, Rongjun Qin, and Yang Tang. 2024. Remote Sensing Object Detection in the Deep Learning Era—A Review. Remote Sens. (2024), 327

    work page 2024

  7. [6]

    Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. 2024. SkySense: A Multi- Modal Remote Sensing Foundation Model Towards Universal Interpretation. In CVPR. 27672–27683

    work page 2024

  8. [7]

    Ruimei Han, Pei Liu, Guangyan Wang, Hanwei Zhang, and Xilong Wu. 2020. Advantage of Combining OBIA and Classifier Ensemble Method for Very High- Resolution Satellite Imagery Classification. J. Sens. (2020), 8855509

    work page 2020

Show all 50 references
  1. [8]

    G. J. Hay and G. Castilla. 2008. Geographic Object-Based Image Analysis (GEO- BIA): A new name for a new discipline. Object-Based Image Analysis (2008), 75–89

    2008

  2. [9]

    Mingyuan He, Jie Zhang, Yang He, Xinjie Zuo, and Zebin Gao. 2024. Anno- tated Dataset for Training Cloud Segmentation Neural Networks Using High- Resolution Satellite Remote Sensing Imagery. Remote Sens. (2024), 3682

    2024

  3. [10]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In ICLR. 3

    2022

  4. [11]

    Leiyi Hu, Wanxuan Lu, Hongfeng Yu, Dongshuo Yin, Xian Sun, and Kun Fu. 2024. Tea: A training-efficient adapting framework for tuning foundation models in remote sensing. IEEE Trans. Geosci. Remote Sens. (2024)

    2024

  5. [12]

    Leiyi Hu, Hongfeng Yu, Wanxuan Lu, Dongshuo Yin, Xian Sun, and Kun Fu. 2024. AiRs: Adapter in Remote Sensing for Parameter-Efficient Transfer Learning.IEEE Trans. Geosci. Remote Sens. (2024), 1–18

    2024

  6. [13]

    Xiaoxing Hu, Ziyang Gong, Yupei Wang, Yuru Jia, Gen Luo, and Xue Yang. 2025. Earth-Adapter: Bridge the Geospatial Domain Gaps with Mixture of Frequency Adaptation. arXiv preprint arXiv:2504.06220 (2025)

    2025

  7. [14]

    Wei Hua, Dingkang Liang, Jingyu Li, Xiaolong Liu, Zhikang Zou, Xiaoqing Ye, and Xiang Bai. 2023. SOOD: Towards Semi-Supervised Oriented Object Detection. In CVPR. 15558–15567

    2023

  8. [15]

    Wubiao Huang, Fei Deng, Haibing Liu, Mingtao Ding, and Qi Yao. 2025. Multiscale Semantic Segmentation of Remote Sensing Images Based on Edge Optimization. IEEE Trans. Geosci. Remote Sens. (2025), 1–13

    2025

  9. [16]

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. 2022. Visual prompt tuning. In ECCV. 709–727

    2022

  10. [17]

    Mehak Khan, Abdul Hanan, Meruyert Kenzhebay, Michele Gazzea, and Reza Arghandeh. 2024. Transformer-based land use and land cover classification with explainability using satellite imagery. Sci. Rep. (2024), 16744

    2024

  11. [18]

    Maja Kucharczyk, Geoffrey J Hay, Salar Ghaffarian, and Chris H Hugenholtz

  12. [19]

    Dong-Hyun Lee et al . 2013. Pseudo-label: The simple and efficient semi- supervised learning method for deep neural networks. In ICML. 896

    2013

  13. [20]

    Rui Li, Shunyi Zheng, Ce Zhang, Chenxi Duan, Libo Wang, and Peter M Atkinson

  14. [21]

    Xiaoxiang Li, Cheng Cheng, Zhiyong Chen, and Manchun Li. 2024. Hierarchical partition of urban land-use units by unsupervised graph learning from high- resolution satellite images. Int. J. Geogr. Inf. Sci. (2024), 2432546

    2024

  15. [22]

    Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. 2024. S2MAE: A Spatial- Spectral Pretraining Foundation Model for Spectral Remote Sensing Data. In CVPR. 24088–24097

    2024

  16. [23]

    Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)

    2021 arXiv

  17. [24]

    Zhuohong Li, Wei He, Jiepan Li, Fangxiao Lu, and Hongyan Zhang. 2024. Learning without Exact Guidance: Updating Large-Scale High-Resolution Land Cover Maps from Low-Resolution Historical Labels. In CVPR. 27717–27727

    2024

  18. [25]

    Yanan Liu and Libao Zhang. 2025. Confidence-Guided Joint Complementary Learning for Weakly Annotated Remote Sensing Object Segmentation. IEEE Trans. Geosci. Remote Sens. (2025), 1–14

    2025

  19. [26]

    Jinna Lv, Qi Shen, Mingzheng Lv, Yiran Li, Lei Shi, and Peiying Zhang. 2023. Deep learning-based semantic segmentation of remote sensing images: a review. Front. Ecol. Evol. (2023), 1201125

    2023

  20. [27]

    Xianping Ma, Qianqian Wu, Xingyu Zhao, Xiaokang Zhang, Man-On Pun, and Bo Huang. 2024. SAM-Assisted Remote Sensing Imagery Semantic Segmentation With Object and Boundary Constraints. IEEE Trans. Geosci. Remote Sens. (2024), 1–16

    2024

  21. [28]

    Xianping Ma, Xiaokang Zhang, and Man-On Pun. 2024. RS3Mamba: Visual State Space Model for Remote Sensing Image Semantic Segmentation. IEEE Geosci. Remote Sens. Lett. (2024), 1–5

    2024

  22. [29]

    Yue Ma, Bowen Dong, Lin Sun, Donghui Zhang, Xiaoyan Wang, Jun Wang, and Xianfang Sun. 2025. Transform Dual-Branch Attention Net: Efficient Semantic Segmentation of Ultra-High-Resolution Remote Sensing Images. Remote Sens. (2025), 540

    2025

  23. [30]

    Marcos Mendieta, Zitu Yang, Yaxing Chen, Ziqi Yang, Chen Chen, and Zheng Wu

  24. [31]

    Clay Foundation: Clay Foundation Model. 2024. Clay

    2024

  25. [32]

    Junlong Qiu, Wei Liu, Xin Zhang, Erzhu Li, Lianpeng Zhang, and Xing Li. 2025. DED-SAM: Adapting Segment Anything Model 2 for Dual Encoder–Decoder Change Detection. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.(2025), 995–1006

    2025

  26. [33]

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. 2024. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    2024 arXiv

  27. [34]

    Caleb Robinson, Le Hou, Kolya Malkin, Rachel Soobitsky, Jacob Czawlytko, Bistra Dilkina, and Nebojsa Jojic. 2019. Large scale high-resolution land cover mapping with multi-resolution data. In CVPR. 12726–12735

    2019

  28. [35]

    Shreya Roy, Revati More, MM Kimothi, S Mamatha, SP Vyas, and SS Ray. 2018. Comparative analysis of object based and pixel based classification for mapping of mango orchards in Sitapur district of Uttar Pradesh. J. Geomatics (2018), 1–8

    2018

  29. [36]

    Kihyuk Sohn and David Berthelot. 2020. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. In NeurIPS. 596–608

    2020

  30. [37]

    Daniela Szwarcman, Sujit Roy, Paolo Fraccaro, Þorsteinn Elí Gíslason, Benedikt Blumenstiel, Rinki Ghosal, Pedro Henrique de Oliveira, Joao Lucas de Sousa Almeida, Rocco Sedona, Yanghui Kang, et al. 2024. Prithvi-eo-2.0: A versatile multi-temporal foundation model for earth obs...

    2024

  31. [38]

    Antti Tarvainen and Harri Valpola. 2017. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS (2017)

    2017

  32. [39]

    Atkinson

    Libo Wang, Rui Li, Ce Zhang, Shenghui Fang, Chenxi Duan, Xiaoliang Meng, and Peter M. Atkinson. 2022. UNetFormer: A UNet-Like Transformer for Effi- cient Semantic Segmentation of Remote Sensing Urban Scene Imagery. ISPRS J. Photogramm. Remote Sens. (2022), 196–214

    2022

  33. [40]

    Shanwen Wang, Xin Sun, Changrui Chen, Danfeng Hong, and Jungong Han

  34. [41]

    Honglin Wu, Peng Huang, Min Zhang, Wenlong Tang, and Xinyu Yu. 2023. CMTFNet: CNN and Multiscale Transformer Fusion Network for Remote-Sensing Image Semantic Segmentation. IEEE Trans. Geosci. Remote Sens. (2023), 1–12

    2023

  35. [42]

    Junde Wu, Ziyue Wang, Mingxuan Hong, Wei Ji, Huazhu Fu, Yanwu Xu, Min Xu, and Yueming Jin. 2025. Medical sam adapter: Adapting segment anything model for medical image segmentation. Med. Image Anal. (2025), 103547

    2025

  36. [43]

    Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. 2020. Self-training with noisy student improves imagenet classification. In CVPR. 10687–10698

    2020

  37. [44]

    Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. 2023. Revisiting Weak-to-Strong Consistency in Semi-Supervised Semantic Segmentation. In CVPR. 7236–7246

    2023

  38. [46]

    Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, and Chun Yuan. 2024. Convolution meets lora: Parameter efficient finetuning for segment anything model. arXiv preprint arXiv:2401.17868 (2024)

    2024 arXiv

  39. [2020]

    Remote Sens

    Geographic object-based image analysis: a primer and future directions. Remote Sens. (2020), 2012

    2020

  40. [2021]

    ABCNet: Attentive bilateral contextual network for efficient semantic segmentation of fine-resolution remotely sensed imagery. ISPRS J. Photogramm. Remote Sens. (2021), 84–98

    2021

  41. [2024]

    In NeurIPS

    Toward Real Ultra Image Segmentation: Leveraging Surrounding Context to Cultivate General Segmentation Model. In NeurIPS. 129227–129249

  42. [2025]

    IEEE Transactions on Geoscience and Remote Sensing (2025)

    Semi-supervised semantic segmentation for remote sensing images via multi-scale uncertainty consistency and cross-teacher-student attention. IEEE Transactions on Geoscience and Remote Sensing (2025)

    2025

Pith tools

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