Pith. sign in

REVIEW 4 major objections 6 minor 72 references

Pan-Arctic Permafrost Landform and Human-built Infrastructure Feature Detection with Vision Transformers and Location Embeddings

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Vision transformers with location embeddings beat CNN baselines on two of three Arctic mapping tasks.

desk verdict Solid, honestly-reported ablation study whose main generalization claim is undercut by unverified site-level split integrity and location-embedding shortcuts. read the letter →

arxiv 2506.02868 v1 pith:2V33ISMV submitted 2025-06-03 cs.CV

classification cs.CV
keywords VisionTransformerslocationembeddingsArcticremotesensingpermafrostlandformsretrogressivethawslumpsice-wedgepolygonssemanticsegmentationinstance
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 pre-trained Vision Transformers (ViTs) become effective detectors for Arctic permafrost landforms and built infrastructure when fused with location embeddings—compact codes that tell the model where each image tile was captured. On held-out test data from three previously published CNN-era datasets, the proposed models outperform the CNN baselines on retrogressive thaw slump segmentation (F1 0.9239 vs. 0.8455) and ice-wedge polygon instance segmentation (mAP50 0.5722 vs. 0.4968), while infrastructure detection does not improve (F1 0.8604 vs. 0.8887). The paper's motivation is pan-Arctic scale: the same semantic class looks spectrally different across regions, and location embeddings supply the geographic context needed to disambiguate such features. If the claim holds, large-area monitoring of thaw disturbance and permafrost-adjacent infrastructure can lean on pretrained transformers rather than task-specific CNN pipelines.

What carries the argument

The central machinery is the ViTDet feature extractor—a plain Vision Transformer that produces one feature map, from which a simple feature pyramid builds multi-scale features—combined with location embeddings from SatCLIP, a contrastively trained model that maps geographic coordinates to vectors at two granularities, L10 and L40. The paper grafts this onto two downstream heads: ViTDet-UNet for semantic segmentation (replacing TransUNet's CNN encoder with the feature pyramid and cascaded upsampler) and ViTDet-Mask-RCNN for instance segmentation. The load-bearing operation is the fusion of image and location embeddings, tested as elementwise addition, concatenation, projected addition, projected concatenation, and cross-attention, before or after the feature pyramid. Across the ablations, fusion placement and embedding granularity change results, with finer L40 embeddings helping when sites are spatially clustered and coarser L10 helping when sites are widely dispersed.

What would settle it

Run a strict site-level or scene-level split—hold out entire islands for the retrogressive thaw slump task or entire countries for ice-wedge polygons—and compare F1 and mAP50 on those never-seen sites; if the location-embedding advantage disappears, the paper's generalization claim is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a plain ViT backbone arranged in the ViTDet style—with a simple feature pyramid replacing a hierarchical CNN—can be adapted for both semantic and instance segmentation of Arctic features, and that adding a location-embedding signal (SatCLIP) improves accuracy beyond both the same ViT without location embeddings and prior CNN baselines on two of the three tasks. The best configurations are task-dependent: a base ViT with post-pyramid L40 concatenation for retrogressive thaw slumps, a large ViT with pre-pyramid L40 cross-attention for ice-wedge polygons, and a base ViT with post-pyramid L10 projection-concatenation for infrastructure. In qualitative examples, the location-embedding models reduce false positives and improve segmentation in sites with very low training representation, such as a Wainwright tile contributing roughly one percent of the infrastructure training data. The paper also compares against Mask2Former, a universal segmentation transformer, and the proposed models score higher than Mask2Former on all three tasks, while the CNN baseline remains ahead on infrastructure.

Load-bearing premise

The reported test gains assume that the held-out test tiles are geographically independent of the training tiles; if tiles from the same scenes or sites appear in both, the improvements could reflect location memorization rather than pan-Arctic generalization.

Editorial extensions

If this is right

  • On the paper's results, pan-Arctic mapping of retrogressive thaw slumps and ice-wedge polygons can be run with pretrained ViT backbones rather than task-specific CNN pipelines, reducing the need for large labeled datasets.
  • Location embeddings act as a spatial prior that improves detection in under-represented areas, such as a site contributing roughly one percent of infrastructure training tiles, where segmentation of roads and buildings improved.
  • The best way to fuse location information depends on the geographic structure of the task: fine-grained L40 embeddings for clustered sites like thaw slumps and ice-wedge polygons, coarser L10 for widely dispersed infrastructure sites.
  • Replacing the CNN encoder in TransUNet with a ViTDet feature pyramid is a workable design for semantic segmentation, since the multiscale feature map supplies the cross-connections for the upsampler.

Reading between the lines

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

  • A question the paper leaves open is whether the test gains reflect true pan-Arctic generalization or memorization of site-specific signals: the train and test tiles for each task come from the same geographic sites, and the SatCLIP t-SNE shows distinct site clusters, so a strict site-level or scene-level split could shrink the reported gaps.
  • A strong control baseline would be a model that receives only the location embedding without image features; since SatCLIP codes are so site-discriminative, such a baseline would reveal how much of the gain comes from location alone versus from genuine image understanding.
  • A natural next step, which the paper lists as future work, is pretraining a ViT with masked autoencoding directly on Arctic satellite imagery; that could close the infrastructure gap, where the CNN baseline still leads, and would test whether ImageNet-pretrained features are the bottleneck.
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 / 6 minor

Summary. The paper proposes a family of Vision Transformer (ViT) models with SatCLIP location embeddings for three Arctic remote sensing tasks: retrogressive thaw slump (RTS) semantic segmentation, ice-wedge polygon (IWP) instance segmentation, and human-built infrastructure detection. The authors build on ViTDet with a simple feature pyramid network, adapt TransUNet for semantic segmentation, and systematically ablate backbone size, fusion placement, merge strategy, and location-embedding granularity on validation sets. On the held-out test sets, the best configurations are reported to outperform prior CNN baselines on RTS (F1 0.9239 vs 0.8455) and IWP (mAP50 0.5722 vs 0.4968), while performing worse on infrastructure (F1 0.8604 vs 0.8887). The paper's central claim is therefore the two-of-three improvement, with the additional claim that location embeddings provide a measurable benefit over image-only ViTs.

Significance. If the test results are trustworthy, this is a useful contribution to pan-Arctic GeoAI: it demonstrates that ViT-based backbones with external location embeddings can be adapted to sub-meter satellite imagery for permafrost landform and infrastructure mapping, and it provides a broad ablation of fusion strategies. The paper also compares against Mask2Former and previously published CNN baselines, and it supplies validation-set error bars for the ablations. The main risk is that the reported test-set gains may be inflated by geographic leakage between training and test tiles, especially because the location embeddings are shown to form per-site clusters (Figure 7); if that is the case, the headline 'pan-Arctic generalization' claim is not supported by the experiments as described.

major comments (4)
  1. [Section IV.A and Table XIII] The paper does not establish that the test tiles are geographically independent of the training tiles. For RTS, all 2,132 tiles come from only two study areas (Banks Island and Eureka Sound Lowlands) and are split into 1705/213/214 tiles with no description of a scene-level or site-level split. For infrastructure, the 5,374 tiles come from 25 scenes over 18 sites, and for IWP, 855 tiles come from a small set of circumpolar sites; again the split is reported only as tile counts. Since the SatCLIP L40 embeddings form distinct per-site clusters (Figure 7), a tile-level split from the same scenes or even overlapping tiles would allow the model to use location embeddings as a near-perfect site-identity shortcut. The authors should either demonstrate scene/site-disjoint splits for all three datasets or report results under such splits; without this, the RTS and IWP gains in Table XIII may reflect spatial autocorrelation and location memorization rather than pan-Arctic generalization.
  2. [Section V.A, Figure 6, and Table XIII] The benefit of location embeddings is demonstrated only on the validation set, not on the test set. Figure 6 compares 'ViTDet' with 'ViTDetLoc' using validation metrics, but Table XIII reports test metrics only for the final location-aware configuration against the CNN baseline and Mask2Former. No test-set number is given for the corresponding ViT model without location embeddings. Therefore the paper's claim that location embeddings improve detection is not directly tested on the held-out test data. The authors should report test metrics for the no-location-embedding ViT baselines for all three tasks, ideally with the same selected configuration and multiple seeds.
  3. [Table XIII and Section V.A] The test-set comparison lacks uncertainty quantification. Table XIII reports single-point metrics with no error bars or significance tests, whereas the validation ablations in Appendix Tables A I-A V include ± ranges. The observed differences (RTS +0.078 F1, IWP +0.075 mAP50) could be within run-to-run variability, especially given the small test sets (214, 123, and 506 tiles). The infrastructure test result actually favors the CNN baseline (F1 0.8604 vs 0.8887), so the 'two of three' claim rests entirely on the two positive differences, which need error bars or a repeated-seed evaluation to be convincing.
  4. [Section V.A, Mask2Former comparison] The Mask2Former baseline is not controlled for comparability. The text states it uses a Swin backbone at 384x384 input, batch size 16, 50 epochs, and was trained 'multiple times with a few basic parameter variations' selecting the best model. This is a reasonable sanity check, but it is not a tuned or comparably trained transformer baseline, so the conclusion that 'transformer-based models with location embeddings outperform prior CNN-based models' should not be framed as a general ViT-vs-CNN verdict based on this comparison. The central comparison should be against the CNN baselines with the same data splits, and the Mask2Former results should be described as preliminary.
minor comments (6)
  1. [Section IV.A] The text says 'Regressive Thaw Slumps' in the dataset description; this should be 'Retrogressive Thaw Slumps'.
  2. [Section V.A and Conclusion] There are typographical errors: 'embedings' and 'embedgings' should be 'embeddings', and 'mloU' should be 'mIoU'.
  3. [Appendix, Figure A1] The caption and in-text references for the appendix boxplots are inconsistent: Section IV.A refers to RTS geometric properties as 'Figure A1(b)', but in the appendix Figure A1(b) is captioned as ice-wedge polygons and A1(c) as RTS. Please correct the cross-references.
  4. [Appendix, Table A V] The row for 'Infra L40 Pre Cross Attention' in Table A V appears corrupted: the values are listed as '0.9657±0.9645 0.8592±0.8475 0.901±0.8903 0.8787±0.8679 0.7904±0.7746', which cannot be correct. This table row should be regenerated from the actual runs.
  5. [References] References [49] and [55] are duplicate entries for the same DeiT paper; one should be removed or replaced with a distinct citation.
  6. [General] The paper does not state whether code, trained models, or the exact train/validation/test tile assignments will be made available; for a results-oriented paper of this type, releasing the split metadata is important for reproducibility, particularly given the geographic-leakage concern raised in the major comments.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the study is an empirical benchmark evaluation whose compared models are externally pre-trained and whose test metrics are reported on a held-out split.

full rationale

This paper contains no derivation chain whose conclusion is equivalent to its premises. The proposed models are built from two externally published components: MAE-pretrained ViTDet backbones [56] and SatCLIP location embeddings [27]; neither is fitted to the test labels or to the CNN baseline numbers. Model configuration is selected on validation and the test set is explicitly reserved: 'We ensure that the validation data sets are used for the empirical model design process through architectural ablations, and the test set is kept independent to report the final metrics for the respective datasets' (Section IV.A), and 'we make all model architecture decisions based on observed results from the validation dataset and hold out the test dataset exclusively for reporting the results' (Section V.A). The reported gains (RTS F1 0.9239 vs 0.8455; IWP mAP50 0.5722 vs 0.4968; infrastructure F1 0.8604 vs 0.8887, Table XIII) are ordinary test-set comparisons against previously published CNN baselines [58],[59],[60]; those baselines are imported as fixed external numbers, not re-derived or fitted here. Self-citations occur only as dataset provenance and baseline references and are not load-bearing for any mathematical claim. The SatCLIP embedding t-SNE (Figure 7) does show per-site clustering, and the paper does not document a scene-disjoint train/test split; that is a potential spatial-autocorrelation threat to external validity, not a circularity of the kind where a prediction reduces by construction to its input. No fitted parameter is relabeled as a prediction, no uniqueness theorem is imported from the authors' prior work, and no known result is merely renamed. Accordingly the appropriate finding is no significant circularity.

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

The central empirical results rest on three task-specific architecture choices selected from validation ablations (Table XII), plus assumptions about label quality, satellite image consistency, and the independence of the geographic splits. No new physical entities are introduced; ViTDet-UNet is an architectural variant, not an invented entity. The main ledger item is the train/test geographic independence assumption, which is not verified in the paper.

free parameters (3)
  • RTS selected configuration = Base ViT, L40, post-pyramid concat
    Chosen from validation metrics in Table X as the best F1; not fitted to the test set. Test F1 in Table XIII is 0.9239.
  • Infrastructure selected configuration = Base ViT, L10, post-pyramid projection concat
    Chosen from validation metrics in Table IX as the best mIoU; test F1 in Table XIII is 0.8604, below the CNN baseline of 0.8887.
  • IWP selected configuration = Large ViT, L40, pre-pyramid cross attention
    Chosen from validation metrics in Table XI as the best mAP50; test mAP50 in Table XIII is 0.5722.
assumptions (4)
  • domain assumption Train/test tile splits are geographically independent within the same study sites.
    RTS train/test both come from Banks Island and ESL (Section IV.A); IWP and infrastructure train/test come from the same circumpolar sites. No scene-level or site-level partitioning is described, so spatial autocorrelation may inflate test scores.
  • domain assumption Human digitized labels are accurate and consistent across the three datasets.
    Datasets from prior studies [58]-[60] were created by on-screen digitization by human analysts (Section IV.A); label noise and inconsistencies are not quantified.
  • domain assumption SatCLIP pretrained location embeddings provide meaningful signal for Arctic regions.
    SatCLIP is trained on global satellite imagery (Section III.B); the paper does not fine-tune it for Arctic data or validate its suitability beyond a t-SNE visualization.
  • domain assumption MAE-pretrained ViT backbones transfer from ImageNet to sub-meter satellite imagery.
    Section III.A uses ImageNet-1K MAE pretrained weights from Detectron2; no Arctic-specific pretraining is performed, and the domain shift is handled only by fine-tuning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pan-Arctic Permafrost Landform and Human-built Infrastructure Feature Detection with Vision Transformers and Location Embeddings." pith.science (2026). https://pith.science/paper/2V33ISMV

@misc{pith2026250602868,
  author       = {Pith},
  title        = {Pith review of: Pan-Arctic Permafrost Landform and Human-built Infrastructure Feature Detection with Vision Transformers and Location Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2V33ISMV}},
  note         = {Machine review of arXiv:2506.02868}
}
read the original abstract

Accurate mapping of permafrost landforms, thaw disturbances, and human-built infrastructure at pan-Arctic scale using sub-meter satellite imagery is increasingly critical. Handling petabyte-scale image data requires high-performance computing and robust feature detection models. While convolutional neural network (CNN)-based deep learning approaches are widely used for remote sensing (RS),similar to the success in transformer based large language models, Vision Transformers (ViTs) offer advantages in capturing long-range dependencies and global context via attention mechanisms. ViTs support pretraining via self-supervised learning-addressing the common limitation of labeled data in Arctic feature detection and outperform CNNs on benchmark datasets. Arctic also poses challenges for model generalization, especially when features with the same semantic class exhibit diverse spectral characteristics. To address these issues for Arctic feature detection, we integrate geospatial location embeddings into ViTs to improve adaptation across regions. This work investigates: (1) the suitability of pre-trained ViTs as feature extractors for high-resolution Arctic remote sensing tasks, and (2) the benefit of combining image and location embeddings. Using previously published datasets for Arctic feature detection, we evaluate our models on three tasks-detecting ice-wedge polygons (IWP), retrogressive thaw slumps (RTS), and human-built infrastructure. We empirically explore multiple configurations to fuse image embeddings and location embeddings. Results show that ViTs with location embeddings outperform prior CNN-based models on two of the three tasks including F1 score increase from 0.84 to 0.92 for RTS detection, demonstrating the potential of transformer-based models with spatial awareness for Arctic RS applications.

Figures

Figures reproduced from arXiv: 2506.02868 by the authors.

Figure 1
Figure 1. A possible application of a ViT for remote sensing a downstream task based on the model introduced by [18]. Each image tile is patched and combined with a positional embedding before being consumed by the transformer block. Features extracted from the Transformer are used by the downstream detection head to complete the detection task. The transformer encoder consists of multiple layers (L) of transformers. D. Locat… view at source ↗
Figure 2
Figure 2. Left hand side represents the original TransUNet architecture proposed by Chen et al. The right-hand side shows our proposed ViTDet-UNet architecture which substitutes the CNN encoder for a feature pyramid in order to create the hierarchical feature maps [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The top half of the figure shows a high-level representation of the Mask-RCNN framework. The bottom half of the figure showcases how the framework can be adapted to work with the feature maps produced by a SFPN instead of relying on a hierarchical backbone [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Pre and Post (SFP) Location Encoder placement with multiple merge strategies. In addition to tiling the 1D location embedding to produce a 2D map that can be merged with the image embeddings as described above, we also introduce the idea of projecting the 1D embedding …
Figure 5
Figure 5. Figure 5: Sampled locations from which target objects for each dataset were labeled from Maxar satellite imagery. Distribution of tundra types based on the Circumpolar Arctic Vegetation Map [66] and permafrost extent based on the Circum-Arctic map of permafrost and ground￾ice co…
Figure 6
Figure 6. Figure 6: shows the comparison of model accuracy of the best model configuration (Table XII) with location embeddings, compared with the model without location embeddings and the existing CNN model [16],[58],[60] for the respective dataset. As can be seen there is an increase in…
Figure 7
Figure 7. Figure 7: t-distributed stochastic neighbor embedding (t-SNE) visualization of the SatCLIP embeddings using the L40 encoder for a subset of the infrastructure training data with independent clusters for each site [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Segmentation model results for four example tiles (a-d), showing base image (Satellite Images © Maxar), annotated ground truth, and outputs for the different models. Green outlines represent buildings, black represent roads, and magenta represents storage tanks. [a] Ti…
Figure 9
Figure 9. Figure 9: Segmentation model results for three example tiles (a-c), showing base image (Satellite Images © Maxar), annotated ground truth, and outputs from the CNN, VitDet, and VitDetLoc models. Yellow outlines represent the annotated retrogressive thaw slumps. [a] Both VitDet a…
Figure 10
Figure 10. Figure 10: Segmentation model results for three example tiles (a-c), showing base image (Satellite Images © Maxar), annotated ground truth, and outputs from the CNN, VitDet, and VitDetLoc models. Blue outlines represent the annotated low centered ice-wedge polygons and the orang…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

72 extracted references · 42 canonical work pages

  1. [1]

    A transdisciplinary, comparative analysis reveals key risks from Arctic permafrost thaw,

    S. Gartler et al., “A transdisciplinary, comparative analysis reveals key risks from Arctic permafrost thaw,” Commun. Earth Environ., vol. 6, no. 1, pp. 1–20, Jan. 2025, doi: 10.1038/s43247-024-01883-w

  2. [2]

    Environmental Impacts from Projected Permafrost Thaw in Alaska: Defining Knowledge Gaps, Data Needs, and Research Priorities,

    T. L. Spero, N. L. Briggs, and L. Boldrick, “Environmental Impacts from Projected Permafrost Thaw in Alaska: Defining Knowledge Gaps, Data Needs, and Research Priorities,” Apr. 2025, doi: 10.1175/WCAS- D-24-0150.1

  3. [3]

    Unmet Needs of Environmentally Threatened Alaska Native Villages: Assessment and Recommendations,

    Alaska Native Tribal Health Consortium, “Unmet Needs of Environmentally Threatened Alaska Native Villages: Assessment and Recommendations,” ANTHC, AL, 2024

  4. [4]

    The Arctic has warmed nearly four times faster than the globe since 1979,

    M. Rantanen et al., “The Arctic has warmed nearly four times faster than the globe since 1979,” Commun. Earth Environ., vol. 3, no. 1, pp. 1–10, Aug. 2022, doi: 10.1038/s43247-022-00498-3

  5. [5]

    Climate-driven thaw of permafrost preserved glacial landscapes, northwestern Canada,

    S. V. Kokelj, T. C. Lantz, J. Tunnicliffe, R. Segal, and D. Lacelle, “Climate-driven thaw of permafrost preserved glacial landscapes, northwestern Canada,” Geology, vol. 45, no. 4, pp. 371–374, Apr. 2017, doi: 10.1130/G38626.1

  6. [6]

    Increase in Arctic coastal erosion and its sensitivity to warming in the twenty-first century,

    D. M. Nielsen et al., “Increase in Arctic coastal erosion and its sensitivity to warming in the twenty-first century,” Nat. Clim. Change, vol. 12, no. 3, pp. 263–270, Mar. 2022, doi: 10.1038/s41558-022- 01281-0

  7. [7]

    Impacts of permafrost degradation on infrastructure,

    J. Hjort, D. Streletskiy, G. Doré, Q. Wu, K. Bjella, and M. Luoto, “Impacts of permafrost degradation on infrastructure,” Nat. Rev. Earth Environ., vol. 3, no. 1, pp. 24–38, Jan. 2022, doi: 10.1038/s43017-021- 00247-8

  8. [8]

    Permafrost and Climate Change: Carbon Cycle Feedbacks From the Warming Arctic,

    E. A. G. Schuur et al., “Permafrost and Climate Change: Carbon Cycle Feedbacks From the Warming Arctic,” 2022, doi: https://doi.org/10.1146/annurev-environ-012220-011847

Show all 72 references
  1. [9]

    Towards long-term records of rain-on-snow events across the Arctic from satellite data,

    A. Bartsch et al., “Towards long-term records of rain-on-snow events across the Arctic from satellite data,” The Cryosphere, vol. 17, no. 2, pp. 889–915, Feb. 2023, doi: 10.5194/tc-17-889-2023

  2. [10]

    The capillaries of the Arctic tundra,

    A. K. Liljedahl, C. Witharana, and E. Manos, “The capillaries of the Arctic tundra,” Nat. Water, vol. 2, no. 7, pp. 611–614, 2024

  3. [11]

    An Optimal GeoAI Workflow for Pan-Arctic Permafrost Feature Detection from High-Resolution Satellite Imagery,

    M. Udawalpola, A. Hasan, A. Liljedahl, A. Soliman, J. Terstriep, and C. Witharana, “An Optimal GeoAI Workflow for Pan-Arctic Permafrost Feature Detection from High-Resolution Satellite Imagery,” Photogramm. Eng. Remote Sens., vol. 88, pp. 181–188, Mar. 2022, doi: 10.14358/PERS...

  4. [12]

    Use of commercial satellite imagery to monitor changing arctic polygonal tundra,

    A. Hasan, M. Udawalpola, A. Liljedahl, and C. Witharana, “Use of commercial satellite imagery to monitor changing arctic polygonal tundra,” Photogramm. Eng. Remote Sens., vol. 88, no. 4, pp. 255–262, 2022

  5. [13]

    An Object-Based Approach for Mapping Tundra Ice-Wedge Polygon Troughs from Very High Spatial Resolution Optical Satellite Imagery,

    C. Witharana et al., “An Object-Based Approach for Mapping Tundra Ice-Wedge Polygon Troughs from Very High Spatial Resolution Optical Satellite Imagery,” Remote Sens., vol. 13, no. 4, Art. no. 4, Jan. 2021, doi: 10.3390/rs13040558

  6. [14]

    Rapid Changes in Retrogressive Thaw Slump Dynamics in the Russian High Arctic Based on Very High-Resolution Remote Sensing,

    S. Barth, I. Nitze, B. Juhls, A. Runge, and G. Grosse, “Rapid Changes in Retrogressive Thaw Slump Dynamics in the Russian High Arctic Based on Very High-Resolution Remote Sensing,” Geophys. Res. Lett., vol. 52, no. 7, p. e2024GL113022, 2025, doi: 10.1029/2024GL113022

  7. [15]

    A Collaborative and Scalable Geospatial Data Set for Arctic Retrogressive Thaw Slumps with Data Standards,

    Y. Yang et al., “A Collaborative and Scalable Geospatial Data Set for Arctic Retrogressive Thaw Slumps with Data Standards,” Sci. Data, vol. 12, no. 1, p. 18, Jan. 2025, doi: 10.1038/s41597-025-04372-7

  8. [16]

    A multi- objective comparison of CNN architectures in Arctic human-built infrastructure mapping from sub-meter resolution satellite imagery,

    E. Manos, C. Witharana, A. S. Perera, and A. K. Liljedahl, “A multi- objective comparison of CNN architectures in Arctic human-built infrastructure mapping from sub-meter resolution satellite imagery,” Int. J. Remote Sens., vol. 44, no. 24, pp. 7670–7705, Dec. 2023, doi: 10.10...

  9. [17]

    Permafrost thaw-related infrastructure damage costs in Alaska are projected to double under medium and high emission scenarios,

    E. Manos, C. Witharana, and A. K. Liljedahl, “Permafrost thaw-related infrastructure damage costs in Alaska are projected to double under medium and high emission scenarios,” Commun. Earth Environ., vol. 6, no. 1, pp. 1–11, Mar. 2025, doi: 10.1038/s43247-025-02191-7

  10. [18]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” in 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3- 7, 2021, OpenReview.net, 2021. Accessed: Nov. 15, 2024. [Online]. Availab...

  11. [19]

    PLG-ViT: Vision Transformer with Parallel Local and Global Self-Attention,

    N. Ebert, D. Stricker, and O. Wasenmüller, “PLG-ViT: Vision Transformer with Parallel Local and Global Self-Attention,” Sensors, vol. 23, no. 7, p. 3447, Mar. 2023, doi: 10.3390/s23073447

  12. [20]

    Self-Supervised Vision Transformers for Joint SAR-Optical Representation Learning,

    Y. Wang, C. M. Albrecht, and X. X. Zhu, “Self-Supervised Vision Transformers for Joint SAR-Optical Representation Learning,” in IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium, Jul. 2022, pp. 139–142. doi: 10.1109/IGARSS46834.2022.9883983

  13. [21]

    SatMAE: pre-training transformers for temporal and multi-spectral satellite imagery,

    Y. Cong et al., “SatMAE: pre-training transformers for temporal and multi-spectral satellite imagery,” in Proceedings of the 36th International Conference on Neural Information Processing Systems, in NIPS ’22. Red Hook, NY, USA: Curran Associates Inc., Nov. 2022, pp. 197–211

  14. [22]

    Morphological Building/Shadow Index for Building Extraction From High-Resolution Imagery Over Urban Areas,

    X. Huang and L. Zhang, “Morphological Building/Shadow Index for Building Extraction From High-Resolution Imagery Over Urban Areas,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 5, no. 1, pp. 161–172, Feb. 2012, doi: 10.1109/JSTARS.2011.2168195

  15. [23]

    DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images,

    I. Demir et al., “DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Salt Lake City, UT, USA: IEEE, Jun. 2018, pp. 172–17209. doi: 10.1109/CVPRW.2018.00031

  16. [24]

    GeoCLIP: Clip-Inspired Alignment between Locations and Images for Effective Worldwide Geo-localization,

    V. V. Cepeda, G. K. Nayak, and M. Shah, “GeoCLIP: Clip-Inspired Alignment between Locations and Images for Effective Worldwide Geo-localization,” Nov. 21, 2023, arXiv: arXiv:2309.16020. doi: 10.48550/arXiv.2309.16020

  17. [25]

    Learning Transferable Visual Models From Natural Language Supervision,

    A. Radford et al., “Learning Transferable Visual Models From Natural Language Supervision,” in Proceedings of the 38th International Conference on Machine Learning, PMLR, Jul. 2021, pp. 8748–8763. Accessed: Jan. 02, 2025. [Online]. Available: https://proceedings.mlr.press/v139...

  18. [26]

    A systematic review of geospatial location embedding approaches in large language models: A path to spatial AI systems,

    S. Tucker, “A systematic review of geospatial location embedding approaches in large language models: A path to spatial AI systems,” Jan. 12, 2024, arXiv: arXiv:2401.10279. doi: 10.48550/arXiv.2401.10279

  19. [27]

    SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery,

    K. Klemmer, E. Rolf, C. Robinson, L. Mackey, and M. Rußwurm, “SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery,” Apr. 12, 2024, arXiv: arXiv:2311.17179. doi: 10.48550/arXiv.2311.17179

  20. [28]

    A relative evaluation of multiclass image classification by support vector machines,

    G. M. Foody and A. Mathur, “A relative evaluation of multiclass image classification by support vector machines,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 6, pp. 1335–1343, Jun. 2004, doi: 10.1109/TGRS.2004.827257

  21. [29]

    Spectral–Spatial Classification of Hyperspectral Imagery Based on Partitional Clustering Techniques,

    Y. Tarabalka, Jó. A. Benediktsson, and J. Chanussot, “Spectral–Spatial Classification of Hyperspectral Imagery Based on Partitional Clustering Techniques,” IEEE Trans. Geosci. Remote Sens., vol. 47, no. 8, pp. 2973–2987, Aug. 2009, doi: 10.1109/TGRS.2009.2016214

  22. [30]

    Locality-Preserving Dimensionality Reduction and Classification for Hyperspectral Image Analysis,

    W. Li, S. Prasad, J. E. Fowler, and L. M. Bruce, “Locality-Preserving Dimensionality Reduction and Classification for Hyperspectral Image Analysis,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 4, pp. 1185– 1198, Apr. 2012, doi: 10.1109/TGRS.2011.2165957

  23. [31]

    Classification of hyperspectral remote sensing images with support vector machines,

    F. Melgani and L. Bruzzone, “Classification of hyperspectral remote sensing images with support vector machines,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 8, pp. 1778–1790, Aug. 2004, doi: 10.1109/TGRS.2004.831865

  24. [32]

    Support vector machines for classification of hyperspectral data,

    J. A. Gualtieri and S. Chettri, “Support vector machines for classification of hyperspectral data,” in IGARSS 2000. IEEE 2000 International Geoscience and Remote Sensing Symposium. Taking the Pulse of the Planet: The Role of Remote Sensing in Managing the Environment. Proceedi...

  25. [33]

    Support vector machines in remote sensing: A review,

    G. Mountrakis, J. Im, and C. Ogole, “Support vector machines in remote sensing: A review,” ISPRS J. Photogramm. Remote Sens., vol. 66, no. 3, pp. 247–259, May 2011, doi: 10.1016/j.isprsjprs.2010.11.001

  26. [34]

    Introduction Neural networks in remote sensing,

    P. M. Atkinson and A. R. L. Tatnall, “Introduction Neural networks in remote sensing,” Int. J. Remote Sens., vol. 18, no. 4, pp. 699–709, Mar. 1997, doi: 10.1080/014311697218700

  27. [35]

    A technique for the selection of kernel- function parameters in RBF neural networks for classification of remote-sensing images,

    L. Bruzzone and D. F. Prieto, “A technique for the selection of kernel- function parameters in RBF neural networks for classification of remote-sensing images,” IEEE Trans. Geosci. Remote Sens., vol. 37, no. 2, pp. 1179–1184, Mar. 1999, doi: 10.1109/36.752239

  28. [36]

    Reducing the Dimensionality of Data with Neural Networks,

    G. E. Hinton and R. R. Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,” Science, vol. 313, no. 5786, pp. 504– 507, Jul. 2006, doi: 10.1126/science.1127647

  29. [37]

    Deep learning in remote sensing applications: A meta-analysis and review,

    L. Ma, Y. Liu, X. Zhang, Y. Ye, G. Yin, and B. A. Johnson, “Deep learning in remote sensing applications: A meta-analysis and review,” ISPRS J. Photogramm. Remote Sens., vol. 152, pp. 166–177, Jun. 2019, doi: 10.1016/j.isprsjprs.2019.04.015

  30. [38]

    Deep Convolutional Neural Networks for Hyperspectral Image Classification,

    W. Hu, Y. Huang, L. Wei, F. Zhang, and H. Li, “Deep Convolutional Neural Networks for Hyperspectral Image Classification,” J. Sens., vol. 2015, no. 1, p. 258619, 2015, doi: 10.1155/2015/258619. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. XX, 202X 19

  31. [39]

    Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources,

    X. X. Zhu et al., “Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources,” IEEE Geosci. Remote Sens. Mag., vol. 5, no. 4, pp. 8–36, Dec. 2017, doi: 10.1109/MGRS.2017.2762307

  32. [41]

    SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2021, pp. 12077–12090. Accessed: Jan. 03, 2025...

  33. [42]

    Semantic texton forests for image categorization and segmentation,

    J. Shotton, M. Johnson, and R. Cipolla, “Semantic texton forests for image categorization and segmentation,” in 2008 IEEE Conference on Computer Vision and Pattern Recognition, Jun. 2008, pp. 1–8. doi: 10.1109/CVPR.2008.4587503

  34. [43]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

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

  35. [44]

    Mask r-cnn,

    K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask r-cnn,” presented at the Proceedings of the IEEE international conference on computer vision, 2017, pp. 2961–2969

  36. [45]

    FCOS: Fully Convolutional One-Stage Object Detection,

    Z. Tian, C. Shen, H. Chen, and T. He, “FCOS: Fully Convolutional One-Stage Object Detection,” in 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2019, pp. 9626–9635. doi: 10.1109/ICCV.2019.00972

  37. [46]

    CenterMask: Real-Time Anchor-Free Instance Segmentation,

    Y. Lee and J. Park, “CenterMask: Real-Time Anchor-Free Instance Segmentation,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2020, pp. 13903–13912. doi: 10.1109/CVPR42600.2020.01392

  38. [47]

    Masked-Attention Mask Transformer for Universal Image Segmentation,

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-Attention Mask Transformer for Universal Image Segmentation,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1290–1299. Accessed: Jan. 03, 202...

  39. [48]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,

    Z. Liu et al., “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2021, pp. 9992–10002. doi: 10.1109/ICCV48922.2021.00986

  40. [50]

    Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation,

    R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation,” in 2014 IEEE Conference on Computer Vision and Pattern Recognition, Jun. 2014, pp. 580–587. doi: 10.1109/CVPR.2014.81

  41. [51]

    ImageNet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition, Jun. 2009, pp. 248–255. doi: 10.1109/CVPR.2009.5206848

  42. [52]

    ImageNet Classification with Deep Convolutional Neural Networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” in Advances in Neural Information Processing Systems, Curran Associates, Inc.,

  43. [53]

    Very Deep Convolutional Networks for Large-Scale Image Recognition,

    K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” Apr. 10, 2015, arXiv: arXiv:1409.1556. doi: 10.48550/arXiv.1409.1556

  44. [54]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2016, pp. 770–778. doi: 10.1109/CVPR.2016.90

  45. [55]

    Training data-efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jegou, “Training data-efficient image transformers & distillation through attention,” in Proceedings of the 38th International Conference on Machine Learning, PMLR, Jul. 2021, pp. 10347–10357. Accessed: Mar. 05, ...

  46. [56]

    Exploring Plain Vision Transformer Backbones for Object Detection,

    Y. Li, H. Mao, R. Girshick, and K. He, “Exploring Plain Vision Transformer Backbones for Object Detection,” Jun. 10, 2022, arXiv: arXiv:2203.16527. doi: 10.48550/arXiv.2203.16527

  47. [57]

    TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers,

    J. Chen et al., “TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers,” Med. Image Anal., vol. 97, p. 103280, Oct. 2024, doi: 10.1016/j.media.2024.103280

  48. [58]

    Automated Detection of Retrogressive Thaw Slumps in the High Arctic Using High-Resolution Satellite Imagery,

    C. Witharana et al., “Automated Detection of Retrogressive Thaw Slumps in the High Arctic Using High-Resolution Satellite Imagery,” Remote Sens., vol. 14, no. 17, Art. no. 17, Jan. 2022, doi: 10.3390/rs14174132

  49. [59]

    Convolutional Neural Networks for Automated Built Infrastructure Detection in the Arctic Using Sub-Meter Spatial Resolution Satellite Imagery,

    E. Manos, C. Witharana, M. R. Udawalpola, A. Hasan, and A. K. Liljedahl, “Convolutional Neural Networks for Automated Built Infrastructure Detection in the Arctic Using Sub-Meter Spatial Resolution Satellite Imagery,” Remote Sens., vol. 14, no. 11, Art. no. 11, Jan. 2022, doi:...

  50. [60]

    Hyperparameter Optimization for Large-Scale Remote Sensing Image Analysis Tasks: A Case Study Based on Permafrost Landform Detection Using Deep Learning,

    A. S. Perera, C. Witharana, E. Manos, and A. K. Liljedahl, “Hyperparameter Optimization for Large-Scale Remote Sensing Image Analysis Tasks: A Case Study Based on Permafrost Landform Detection Using Deep Learning,” IEEE Access, vol. 12, pp. 43062– 43077, 2024

  51. [61]

    Yuxin Wu, Alexander Kirillov, Wan-Yen Lo, and Ross Girshick, Detectron2. (2019). [Online]. Available: https://github.com/facebookresearch/detectron2

  52. [62]

    Frontera: The Evolution of Leadership Computing at the National Science Foundation,

    D. Stanzione, J. West, R. T. Evans, T. Minyard, O. Ghattas, and D. K. Panda, “Frontera: The Evolution of Leadership Computing at the National Science Foundation,” in Practice and Experience in Advanced Research Computing, in PEARC ’20. New York, NY, USA: Association for Comput...

  53. [63]

    Lonestar6

    “Lonestar6.” Accessed: Jan. 07, 2025. [Online]. Available: http://tacc.utexas.edu/systems/lonestar6/

  54. [64]

    Google Cloud Documentation,

    “Google Cloud Documentation,” Google Cloud. Accessed: Jan. 07,

  55. [65]

    MIT Press, 2016

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning. MIT Press, 2016. Accessed: Apr. 04, 2025. [Online]. Available: https://www.deeplearningbook.org/

  56. [66]

    A raster version of the Circumpolar Arctic Vegetation Map (CAVM),

    M. K. Raynolds et al., “A raster version of the Circumpolar Arctic Vegetation Map (CAVM),” Remote Sens. Environ., vol. 232, p. 111297, Oct. 2019, doi: 10.1016/j.rse.2019.111297

  57. [67]

    Circum-Arctic map of permafrost and ground-ice conditions,

    J. L. Brown, O. J. F. Jr, J. A. Heginbottom, and E. S. Melnikov, “Circum-Arctic map of permafrost and ground-ice conditions,” U.S. Geological Survey, 45, 1997. doi: 10.3133/cp45

  58. [68]

    Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation,

    G. Ghiasi et al., “Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation,” Jun. 23, 2021, arXiv: arXiv:2012.07177. doi: 10.48550/arXiv.2012.07177

  59. [69]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2015, pp. 3431–3440. doi: 10.1109/CVPR.2015.7298965

  60. [70]

    Microsoft COCO: Common Objects in Context,

    T.-Y. Lin et al., “Microsoft COCO: Common Objects in Context,” in Computer Vision – ECCV 2014, D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, Eds., Cham: Springer International Publishing, 2014, pp. 740–755. doi: 10.1007/978-3-319-10602-1_48. IEEE TRANSACTIONS ON GEOSCIEN...

  61. [2006]

    He has a PhD in Computer Science

    He is an IEEE member since 2014. He has a PhD in Computer Science. His general areas of interest include data science, data mining, database systems, and software engineering. Currently he is working on developing and deploying deep learning-based computer vision models on Big...

  62. [2012]

    07, 2025

    Accessed: Apr. 07, 2025. [Online]. Available: https://proceedings.neurips.cc/paper_files/paper/2012/hash/c399862d3 b9d6b76c8436e924a68c45b-Abstract.html

  63. [2019]

    Her research interests include cyberinfrastructure, geospatial big data, GeoAI, and their applications in data-intensive environmental and social sciences

    She is currently a Professor of GIScience in the School of Geographical Sciences and Urban Planning at Arizona State University, where she also directs the Spatial Analysis Research Center and the Cyberinfrastructure and Computational Intelligence Lab. Her research interests i...

  64. [2025]

    Available: https://cloud.google.com/docs

    [Online]. Available: https://cloud.google.com/docs

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.