Pith. sign in

REVIEW 3 major objections 7 minor 68 references

Fine-Scale Soil Mapping in Alaska with Multimodal Machine Learning

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MiSo, a vision-based model, produces the first 10-meter/pixel maps of near-surface permafrost and soil taxonomy across Alaska, and under spatial holdout it detects permafrost with higher recall than Random Forest, at the cost of lower…

desk verdict A genuinely new 10 m statewide permafrost/taxonomy map for Alaska with an honest recall-vs-accuracy trade-off, undercut by an unaddressed 1952–2023 label temporal mismatch and no released code or comparison to existing products. read the letter →

arxiv 2506.17302 v1 pith:BZMJPB5P submitted 2025-06-17 cs.CV cs.LG

Yijun Lin , Theresa Chen , Colby Brungard , Grunwald Sabine , Sue Ives , Matt Macander , Timm Nawrocki , Yao-Yi Chiang
show 1 more author
Nic Jelinski
This is my paper
classification cs.CVcs.LG
keywords Fine-ScaleSoilMappingAlaskaNear-SurfacePermafrostTaxonomyMachineLearningGeospatialFoundationModelImplicitNeuralRepresentationContrastive
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 aims to show that a vision-based machine learning model can map near-surface permafrost and soil taxonomy across all of Alaska at 10 meters per pixel, a resolution and coverage no existing soil product achieves. The proposed model, MiSo, combines a pretrained geospatial vision backbone with implicit neural functions that let it make predictions at any query coordinate directly from sparse field observations. Compared against the Random Forest baseline that soil scientists commonly use, MiSo is claimed to generalize better to remote, unseen locations, especially under spatial holdout splits where training and test sites are separated by 1 or 10 kilometers. The central trade-off is that MiSo achieves higher recall for permafrost presence but lower precision and overall accuracy than Random Forest, meaning it flags more potentially frozen ground in unvisited terrain at the expense of more false alarms. A sympathetic reader cares because permafrost thaw threatens infrastructure and carbon storage, and a statewide high-resolution permafrost probability surface would give planners and field scientists a better starting point than the current coarse products.

What carries the argument

The central mechanism is a local implicit image function $f_\theta$ applied to multi-scale features from two Swin Transformer encoders; for a query coordinate $x_q$, it interpolates embeddings from the four nearest encoder features $z_t$ at centers $v_t$ using area weights $S_t/S$, giving $G^{(i)}(x_q)=\sum_{t\in\{00,01,10,11\}} (S_t/S)\, f_\theta(z_t, x_q-v_t)$. One encoder is initialized from a pretrained geospatial foundation model for the satellite imagery, and the other encodes DEM derivatives plus climate normals from scratch. Contrastive learning with an InfoNCE loss aligns the two modality embeddings and aligns a Fourier-based positional encoding of the query coordinates with the image-derived representation, adding geo-location awareness. This machinery lets MiSo train directly on point observations without rasterizing labels to a fixed grid, and at inference time it queries any user-defined grid, such as 10-meter cells, with overlapping crops and Gaussian merging to smooth tile boundaries.

What would settle it

Restrict the test set to field observations sampled in 2020 or later, run the same 1-km and 10-km spatial holdout splits, and check whether MiSo's permafrost-recall advantage over Random Forest persists; if the advantage shrinks or reverses when labels are contemporaneous with the imagery, the generalization claim is an artifact of outdated labels rather than a property of the model.

Watch

Extended reading notes

Core claim

The paper claims to produce the first statewide 10-meter/pixel maps of near-surface permafrost presence-absence and soil taxonomy for Alaska using MiSo, and to demonstrate that MiSo generalizes better to remote, unseen locations than Random Forest. On spatial holdout splits, MiSo consistently achieves higher recall for permafrost presence than RF, which the authors attribute to RF's tendency to favor the dominant non-permafrost class, while MiSo produces more conservative probability estimates that are useful for early-warning and field sampling. For the seven-class soil taxonomy task, MiSo outperforms RF across random, 1-km spatial holdout, and 10-km spatial holdout splits in weighted precision, recall, and F1, with the largest gains on Entisols, Gelisols, and Histosols. The paper also argues that MiSo better captures minority soil classes in ecoregions with skewed class distributions, whereas RF collapses toward the majority class.

Load-bearing premise

The load-bearing premise is that field observations gathered between 1952 and 2023 can be treated as valid current labels for models built on 2019 satellite imagery and 1981-2010 climate normals, an assumption the paper itself flags by noting that broad temporal ranges may misalign observations with current conditions.

Editorial extensions

If this is right

  • Statewide 10-meter maps of near-surface permafrost and seven soil orders become available for Alaska, replacing a coarser product whose 10-meter resolution is largely inherited from 1:1,000,000 STATSGO2 data.
  • In unvisited or under-sampled terrain, MiSo flags more potential permafrost than Random Forest, making it better suited for early-warning assessments and for guiding where new field surveys should go.
  • For soil taxonomy, MiSo improves prediction of minority classes such as Histosols and Gelisols in regions with a dominant soil type, giving a more balanced picture of the soil profile than RF.
  • Because MiSo predicts at arbitrary query coordinates, end users can request probability estimates for any point or custom grid without retraining a fixed-resolution model.
  • The probability surfaces produced by MiSo do not strictly match existing permafrost-zone definitions; users who want zonal alignment can threshold the probability map themselves.

Reading between the lines

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

  • Editorial inference: a temporal holdout test, where training observations are pre-2019 and test observations are post-2019, would directly test whether the reported recall advantage reflects true spatial generalization or simply label drift between sampling dates and the 2019 imagery.
  • Editorial inference: removing the contrastive geo-location alignment in an ablation would isolate how much of MiSo's remote-region recall gain comes from coordinates rather than from visual or environmental features, which matters for interpreting the model in regions far from any training points.
  • Editorial inference: the same architecture could be adapted for forecasting by making the climate covariates time-indexed and treating the observation year as an input, turning the acknowledged label-date mismatch into a permafrost-thaw prediction task.
  • Editorial inference: the model's sharp bimodal probability distributions suggest that threshold-based calibration, rather than raw class-0.5 decisions, would let users trade recall against precision deliberately for different applications such as infrastructure risk versus carbon accounting.
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. The paper introduces MiSo, a multimodal vision-based model for predicting near-surface permafrost (NSP) presence-absence and soil taxonomy across Alaska at 10 m resolution. The architecture combines a pretrained SWIN-transformer geospatial foundation model (SATLASNET) for satellite imagery, a second encoder for DEM-derived and climate covariates, local implicit image functions for continuous predictions at arbitrary coordinates, and contrastive learning objectives for cross-modal and geographic alignment. The model is compared against a Random Forest baseline under three cross-validation schemes: random splits, spatial holdouts with 1 km clusters, and spatial holdouts with 10 km clusters. The paper reports that MiSo achieves higher presence recall than RF in all splits and higher soil taxonomy F1 in all splits, while acknowledging lower overall accuracy and lower presence precision in the spatial holdout scenarios. The authors claim to produce the first 10 m/pixel soil maps for these targets across Alaska and discuss regional analyses by Permafrost Zones and Major Land Resource Areas.

Significance. If the claims are supported, the paper makes a useful applied contribution: it demonstrates a nontrivial deep-learning architecture for sparse-point soil mapping, provides a spatial cross-validation comparison with a standard baseline, and releases code. The use of a geospatial foundation model with implicit neural representations is well matched to the task of predicting from sparse field observations at arbitrary locations. The paper also provides practical guidance on the precision-recall tradeoff, which matters for field sampling and infrastructure planning. However, the headline claim that MiSo 'generalizes better to remote, unseen locations and achieves higher recall than RF' is only partially supported by the reported metrics: in the two spatial holdout scenarios, MiSo has lower overall accuracy and substantially lower presence precision than RF. The main risk to the central claim is the acknowledged temporal mismatch between 1952-2023 field labels and 2019 satellite/1981-2010 climate covariates, which could inflate MiSo's recall advantage if stale presence labels reward predictions of recently thawed permafrost.

major comments (3)
  1. [§5.3.1, Table 1, Abstract] The temporal mismatch between labels and covariates directly affects the interpretation of the central recall claim. Field observations span 1952-2023, while covariates are fixed to 2019 Sentinel-2 imagery and 1981-2010 PRISM normals. Section 3.1 explicitly states that 'the broad temporal range may cause misalignment between the observations and current environmental conditions,' and Section 5.3.1 concedes that some MiSo false positives are areas 'historically underlain by permafrost but recently thawed.' Under spatial holdout, older presence labels that record permafrost that has since thawed will mechanically favor a model that predicts more permafrost, such as MiSo, in its recall score, while RF's lower recall may reflect better alignment with current surface conditions. The conclusion (§6) also acknowledges that the study 'does not explicitly account for the time at which points were sampled.' This is not a minor caveat: the abstract's 'generalizes better to remote, unseen locations and achieves higher recall than RF' is the paper's principal claim. I ask the authors to rerun the spatial holdout experiments on temporally restricted subsets (e.g., observations from 2000 onward or 2010 onward) or to include sample date as a covariate, and to report how the R1 gap between MiSo and RF changes. Without this analysis, the reported recall advantage could be an artifact of label drift rather than true mapping skill.
  2. [Table 1, §5.3.1] The statement that MiSo 'generalizes better to remote, unseen locations' is overstated relative to the evidence in Table 1. In the SH-1km split, MiSo has lower overall accuracy than RF (90.48 vs 92.20) and much lower presence precision (76.17 vs 87.12); in SH-10km, MiSo again has lower overall accuracy (81.06 vs 83.44) and lower presence precision (58.62 vs 69.29). The only metric favoring MiSo in these spatial holdouts is presence recall (84.87 vs 77.19 and 75.93 vs 63.07). 'Generalizes better' should therefore be qualified as 'achieves higher recall at a substantial cost in precision and overall accuracy.' This is not merely phrasing: it changes the practical recommendation, since a monitoring product with high recall but low precision will produce many false-positive alerts. The paper should report a threshold-independent summary (e.g., area under the precision-recall curve) or an explicit discussion of the operating point at which MiSo is preferable, rather than presenting recall as the sole evidence for the generalization claim.
  3. [Tables 1 and 2, §5.2.2] The comparison relies on averages across five folds without any measure of variance or statistical significance. This is particularly important because the Task 2 differences are small: MiSo's F1 advantage over RF is about 2 points in SH-1km (62.77 vs 60.83) and about 0.2 points in SH-10km (55.11 vs 54.88), and the class-level differences in Figure 4 could easily fall within fold-to-fold variability. For the headline NSP recall difference, the paper should report per-fold results, standard deviations, or a paired significance test across the five folds. The current presentation does not allow a reader to assess whether the reported advantages are stable or driven by a single fold.
minor comments (7)
  1. [Throughout] The model name is inconsistently spelled as 'MiSo' in the body and 'MISO' in the abstract and reader-facing text; please unify the spelling.
  2. [§3.4] The text says 'the temporal span of most field observations in MSP dataset' but the dataset is called AKSDB elsewhere; please correct the acronym.
  3. [Figure 8 caption] The caption refers to the 'Cook Islands Lowlands,' but the correct geographic name elsewhere in the paper is 'Cook Inlet Lowlands.'
  4. [§4.1] The claim that the SWIN-T-based model 'consistently outperforming Vision Transformer-based pretraining models including SatMAE [12] and Prithvi [22] in soil prediction tasks' is not supported by any experiment or table in the paper; please add the comparison or remove the claim.
  5. [§4.2, Eq. (1)] The notation '2D coordinate' should be clarified as '2-D coordinate' to avoid ambiguity with '2 times D,' and the dimension of the position encoding should be stated explicitly.
  6. [§5.3.4 and §6] The paper claims to produce the first 10 m/pixel statewide soil maps, but only a 50 km square visualization is shown and the project release link points to code, not to the maps. Including the statewide map products or a clear data-availability statement would strengthen this contribution claim.
  7. [§4.4] The training description does not state how the model is selected across the 30 fine-tuning epochs (e.g., early stopping on a validation subset), which is relevant for interpreting the five-fold results as unbiased estimates.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: MISO's reported gains are empirical comparisons against a baseline on held-out spatial folds, not reductions to fitted inputs or self-citation chains.

full rationale

I walked the claimed derivation chain: MISO is trained on ~38,000 point field observations from AKSDB with Sentinel-2, DEM derivatives, and PRISM climate covariates, and Random Forest is trained on the same labels and evaluated under Random, SH-1km, and SH-10km splits. The reported recall, precision, and accuracy numbers are measurements of fitted models on held-out folds, not quantities defined by the training targets or by the model's own components. The contrastive pretraining uses unlabeled locations and external SATLASNET weights, so it is independent support rather than a self-imported premise. The only author self-citation I found (Ref. [23], Jelinski et al. 2024, in the introduction) supports the importance of NSP/soil-taxonomy data products and is not load-bearing in any derivation. The paper explicitly acknowledges a temporal mismatch between 1952-2023 field observations and 2019/1981-2010 covariates (Section 3.1), and in Section 5.3.1 concedes that some false positives may be 'historically underlain by permafrost but recently thawed.' That is a data-validity threat that could affect which model is better in practice, but it does not make any prediction equivalent to its inputs by construction. No equation is fitted to the quantity it later 'predicts'; no cited theorem is invoked to force the choice of MISO. Therefore the central comparison has independent content and is not circular.

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

The core method is empirical and does not rest on a formal derivation. The main dependencies are label validity, covariate informativeness, and spatial split design. The most fragile dependency is temporal alignment of labels and covariates. No new physical entities are introduced.

free parameters (4)
  • Random Forest buffer distance = 50 m
    Selected empirically as the best-performing buffer among 0, 25, 50, and 100 m (Section 5.2.3); it changes the RF baseline that MISO is judged against.
  • Random Forest hyperparameters = 253 trees, min sample split 5, max depth 16
    Chosen by a 100-iteration random search (Section 5.2.3); the comparison depends on this baseline configuration.
  • MISO training schedule = 50 pretrain epochs, 30 finetune epochs, LR 1e-4 pretrain / 5e-5 finetune, batch 32
    Fixed by the authors without a reported search (Section 5.2.3); these choices affect the reported metrics.
  • Implicit feature dimension and positional encoding length = D = 1024; L frequencies not specified
    The model uses D=1024 embeddings and a Fourier coordinate encoding with L frequencies (Sections 4.2, 4.3.2); L and MLP details are not reported but determine geo awareness.
assumptions (4)
  • domain assumption Historical field observations from 1952 to 2023 are valid current labels despite covariates being fixed to 2019 Sentinel-2 and 1981-2010 climate normals.
    Invoked in Section 3.1, where the paper acknowledges possible misalignment but includes all observations to improve spatial coverage; no temporal correction or sensitivity test is performed.
  • domain assumption Near-surface permafrost presence and soil order can be inferred from surface Sentinel-2 reflectance, DEM derivatives, and climate normals.
    The models use only surface covariates with no subsurface measurements (Sections 3.2-3.4, 4); if surface signals do not track subsurface conditions, the labels cannot be learned.
  • domain assumption DBSCAN spatial clustering with 1 km and 10 km buffers removes spatial autocorrelation leakage between training and test folds.
    Section 5.2.1 defines SH-1km and SH-10km splits; if permafrost or soil spatial autocorrelation extends beyond 10 km, the holdout results are optimistic.
  • domain assumption PRISM climate normals at 800 m resolution can be interpolated to 10 m without introducing meaningful error.
    Section 3.4 states climate data are interpolated to 10 m to align with other modalities; the interpolation adds smoothing that may affect fine-scale predictions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fine-Scale Soil Mapping in Alaska with Multimodal Machine Learning." pith.science (2026). https://pith.science/paper/BZMJPB5P

@misc{pith2026250617302,
  author       = {Pith},
  title        = {Pith review of: Fine-Scale Soil Mapping in Alaska with Multimodal Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZMJPB5P}},
  note         = {Machine review of arXiv:2506.17302}
}
read the original abstract

Fine-scale soil mapping in Alaska, traditionally relying on fieldwork and localized simulations, remains a critical yet underdeveloped task, despite the region's ecological importance and extensive permafrost coverage. As permafrost thaw accelerates due to climate change, it threatens infrastructure stability and key ecosystem services, such as soil carbon storage. High-resolution soil maps are essential for characterizing permafrost distribution, identifying vulnerable areas, and informing adaptation strategies. We present MISO, a vision-based machine learning (ML) model to produce statewide fine-scale soil maps for near-surface permafrost and soil taxonomy. The model integrates a geospatial foundation model for visual feature extraction, implicit neural representations for continuous spatial prediction, and contrastive learning for multimodal alignment and geo-location awareness. We compare MISO with Random Forest (RF), a traditional ML model that has been widely used in soil mapping applications. Spatial cross-validation and regional analysis across Permafrost Zones and Major Land Resource Areas (MLRAs) show that MISO generalizes better to remote, unseen locations and achieves higher recall than RF, which is critical for monitoring permafrost thaw and related environmental processes. These findings demonstrate the potential of advanced ML approaches for fine-scale soil mapping and provide practical guidance for future soil sampling and infrastructure planning in permafrost-affected landscapes. The project will be released at https://github.com/knowledge-computing/Peatland-permafrost.

Figures

Figures reproduced from arXiv: 2506.17302 by the authors.

Figure 1
Figure 1. Field observations of NSP presence-absence across [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The model architecture of MiSo. The model con￾tains two vision encoders based on SWIN Transformers to handle input modalities and uses implicit image functions to generate continuous spatial predictions for a given geo￾coordinate. During training, the model employs contrastive learning (CL) objectives to align both multimodal inputs and visual-geospatial representations (indicated by dash arrows). The final embeddin… view at source ↗
Figure 3
Figure 3. illustrates the learning process. Given an 𝑖-level fea￾ture map 𝑍 (𝑖) ∈ R 𝐻 ×𝑊 ×𝐷 from the encoder, and a query loca￾tion 𝑥𝑞 at within the image, MiSo aims to learn a continuous feature space 𝐺 (𝑖) to represent the context of locations. The embedding at the coordinate 𝑥𝑞 is defined as, 𝐺 (𝑖) (𝑥𝑞) = ∑︁ 𝑡 ∈ {00,01,10,11} 𝑆𝑡 𝑆 · 𝑓𝜃 (𝑧𝑡 , 𝑥𝑞 − 𝑣𝑡 ) (1) Here 𝑥𝑞 is a 2𝐷 coordinate in the continuous image domain. The varia… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Comparison between RF and MiSo across soil tax￾onomy and data splits with probabilities heavily concentrated in the 0-0.1 and 0.9–1.0 ranges. Specifically, MiSo produces a sharper distribution than RF with fewer probabilities in the intermediate bins (0.2—0.8). This ph…
Figure 6
Figure 6. Figure 6: Predicted permafrost presence percentage by Per [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Accuracy comparison between RF and MiSo in the SH-1km split experiment for selected MLRAs. model generalizes well for Spodosols across the two geographically distant regions. In contrast, for Gelisols, both models show similar performance in IAH, but MiSo performs at 5…
Figure 8
Figure 8. Figure 8: Distribution of the ground truth points, RF pre [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 10
Figure 10. Figure 10: Visual example of fine-scale prediction for soil [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 9
Figure 9. Figure 9: Visual example of fine-scale prediction for NSP [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 60 canonical work pages

  1. [1]

    2014.GlobalSoilMap: basis of the global spatial soil information system

    Dominique Arrouays, Neil McKenzie, Jon Hempel, Anne Richer de Forges, and Alex B McBratney. 2014.GlobalSoilMap: basis of the global spatial soil information system. CRC press

    work page 2014

  2. [2]

    Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. 2021. Geography-aware self-supervised learn- ing. InProceedings of the IEEE/CVF International Conference on Computer Vision. 10181–10190

    work page 2021

  3. [3]

    Michael S Balshi, A David McGuire, Paul Duffy, Michael Flannigan, David W Kicklighter, and Jerry Melillo. 2009. Vulnerability of carbon storage in North American boreal forests to wildfires during the 21st century.Global change biology15, 6 (2009), 1491–1510

    work page 2009

  4. [4]

    Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, and Aniruddha Kembhavi. 2023. Satlaspretrain: A large-scale dataset for remote sensing im- age understanding. InProceedings of the IEEE/CVF International Conference on Computer Vision. 16772–16782

    2023

  5. [5]

    Sharon A Billings, Kate Lajtha, Avni Malhotra, Asmeret Asefaw Berhe, M-A de Graaff, Stevan Earl, Jennifer Fraterrigo, Katerina Georgiou, S Grandy, Sarah E Hobbie, et al. 2021. Soil organic carbon is not just for soil scientists: measurement Fine-Scale Soil Mapping in Alaska with Multimodal Machine Learning recommendations for diverse practitioners.Ecologi...

    work page 2021

  6. [6]

    Maurice Blackmon, Byron Boville, Frank Bryan, Robert Dickinson, Peter Gent, Jeffrey Kiehl, Richard Moritz, David Randall, Jagadish Shukla, Susan Solomon, et al. 2001. The community climate system model.Bulletin of the American Meteorological Society82, 11 (2001), 2357–2376

    work page 2001

  7. [7]

    NB Bliss, SW Waltman, and GW Petersen. 1995. Preparing a soil carbon inventory for the United States using geographic information systems. (1995)

    work page 1995

  8. [8]

    FS Chapin Iii, A D Mcguire, J Randerson, R Pielke, Dennis Baldocchi, SE Hobbie, Nigel Roulet, W Eugster, E Kasischke, EB Rastetter, et al. 2000. Arctic and boreal ecosystems of western North America as components of the climate system. Global Change Biology6, S1 (2000), 211–223

    work page 2000

Show all 68 references
  1. [9]

    Vincent Chaplot, Christian Walter, and Pierre Curmi. 2000. Improving soil hydromorphy prediction according to DEM resolution and available pedological data.Geoderma97, 3-4 (2000), 405–422

    2000

  2. [10]

    Yueli Chen, Shile Li, Lingxiao Wang, Magdalena Mittermeier, Monique Bernier, and Ralf Ludwig. 2024. Retrieving freeze-thaw states using deep learning with remote sensing data in permafrost landscapes.International Journal of Applied Earth Observation and Geoinformation126 (202...

    2024

  3. [11]

    Yinbo Chen, Sifei Liu, and Xiaolong Wang. 2021. Learning continuous image representation with local implicit image function. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8628–8638

    2021

  4. [12]

    Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. 2022. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery.Advances in Neural Information Processing Systems35 (2022), 197–211

    2022

  5. [13]

    Christopher Daly, Joseph Smith, and Michael Halbleib. 2018. 1981–2010 High- Resolution Temperature and Precipitation Maps for Alaska Final Report.PRISM Climate Group, Oregon State University: Corvallis, OR, USA(2018)

    2018

  6. [14]

    Nicola Deluigi, Christophe Lambiel, and Mikhail Kanevski. 2017. Data-driven mapping of the potential mountain permafrost distribution.Science of the total environment590 (2017), 370–380

    2017

  7. [15]

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. 1996. A density- based algorithm for discovering clusters in large spatial databases with noise. In kdd, Vol. 96. 226–231

    1996

  8. [16]

    Chandana Gangodagamage, Joel C Rowland, Susan S Hubbard, Steven P Brumby, Anna K Liljedahl, Haruko Wainwright, Cathy J Wilson, Garrett L Altmann, Baptiste Dafflon, John Peterson, et al. 2014. Extrapolating active layer thickness measurements across Arctic polygonal terrain usi...

    2014

  9. [17]

    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 for earth observation imagery. InProceedings of the IEE...

    2024

  10. [18]

    Tomislav Hengl, Jorge Mendes de Jesus, Gerard BM Heuvelink, Maria Ruiperez Gonzalez, Milan Kilibarda, Aleksandar Blagotić, Wei Shangguan, Mar- vin N Wright, Xiaoyuan Geng, Bernhard Bauer-Marschallinger, et al. 2017. Soil- Grids250m: Global gridded soil information based on mac...

    2017

  11. [19]

    Jan Hjort, Dmitry Streletskiy, Guy Doré, Qingbai Wu, Kevin Bjella, and Miska Luoto. 2022. Impacts of permafrost degradation on infrastructure.Nature Reviews Earth & Environment3, 1 (2022), 24–38

    2022

  12. [20]

    Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Naoto Yokoya, Hao Li, Pedram Ghamisi, Xiuping Jia, et al. 2024. SpectralGPT: Spectral remote sensing foundation model.IEEE Transactions on Pattern Analysis and Machine Intelligence(2024)

    2024

  13. [21]

    Elchin E Jafarov, Sergey S Marchenko, and VE Romanovsky. 2012. Numerical modeling of permafrost dynamics in Alaska using a high spatial resolution dataset.The Cryosphere6, 3 (2012), 613–624

    2012

  14. [22]

    Johannes Jakubik, Sujit Roy, CE Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Gabby Nyirjesy, Blair Edwards, et al. 2023. Foundation models for generalist geospatial artificial intelligence. arXiv preprint arXiv:2310.18660(2023)

    2023 arXiv

  15. [23]

    Nicolas A Jelinski, NJ Pastick, AL Kholodov, MJ Sousa, and JM Galbraith. 2024. Estimates of soil taxonomic change due to near-surface permafrost loss in Alaska. Soil Science Society of America Journal88, 5 (2024), 1626–1646

    2024

  16. [24]

    M Torre Jorgenson, Jennifer Harden, Mikhail Kanevskiy, Jonathan O’Donnell, Kim Wickland, Stephanie Ewing, Kristen Manies, Qianlai Zhuang, Yuri Shur, Robert Striegl, et al . 2013. Reorganization of vegetation, hydrology and soil carbon after permafrost degradation across hetero...

    2013

  17. [25]

    M Torre Jorgenson, Yuri L Shur, and Erik R Pullman. 2006. Abrupt increase in permafrost degradation in Arctic Alaska.Geophysical Research Letters33, 2 (2006)

    2006

  18. [26]

    Nafiseh Kakhani, Moien Rangzan, Ali Jamali, Sara Attarchi, Seyed Kazem Alavipanah, Michael Mommert, Nikolaos Tziolas, and Thomas Scholten. 2024. SSL-SoilNet: A Hybrid Transformer-Based Framework with Self-Supervised Learning for Large-Scale Soil Organic Carbon Prediction.IEEE ...

    2024

  19. [27]

    Paahuni Khandelwal, Sangmi Lee Pallickara, and Shrideep Pallickara. 2024. Deep- soil: A science-guided framework for generating high precision soil moisture maps by reconciling measurement profiles across in-situ and remote sensing data. InProceedings of the 32nd ACM Internati...

    2024

  20. [28]

    Rattan Lal, Johan Bouma, Eric Brevik, Lorna Dawson, Damien J Field, Bruno Glaser, Ryusuke Hatano, Alfred E Hartemink, Takashi Kosaki, Bruce Lascelles, et al. 2021. Soils and sustainable development goals of the United Nations: An International Union of Soil Sciences perspectiv...

    2021

  21. [29]

    Rattan Lal, Rabi H Mohtar, Amjad T Assi, Ram Ray, Haimanote Baybil, and Molly Jahn. 2017. Soil as a basic nexus tool: soils at the center of the food–energy–water nexus.Current Sustainable/Renewable Energy Reports4 (2017), 117–129

    2017

  22. [30]

    Zachary L Langford, Jitendra Kumar, Forrest M Hoffman, Amy L Breen, and Colleen M Iversen. 2019. Arctic vegetation mapping using unsupervised training datasets and convolutional neural networks.Remote Sensing11, 1 (2019), 69

    2019

  23. [31]

    David M Lawrence, Rosie A Fisher, Charles D Koven, Keith W Oleson, Sean C Swenson, Gordon Bonan, Nathan Collier, Bardan Ghimire, Léo Van Kampenhout, Daniel Kennedy, et al. 2019. The Community Land Model version 5: Description of new features, benchmarking, and impact of forcin...

    2019

  24. [32]

    David M Lawrence and Andrew G Slater. 2005. A projection of severe near- surface permafrost degradation during the 21st century.Geophysical research letters32, 24 (2005)

    2005

  25. [33]

    David M Lawrence, Andrew G Slater, Vladimir E Romanovsky, and Dmitry J Nicolsky. 2008. Sensitivity of a model projection of near-surface permafrost degradation to soil column depth and representation of soil organic matter. Journal of Geophysical Research: Earth Surface113, F2 (2008)

    2008

  26. [34]

    David M Lawrence, Andrew G Slater, and Sean C Swenson. 2012. Simulation of present-day and future permafrost and seasonally frozen ground conditions in CCSM4.Journal of Climate25, 7 (2012), 2207–2225

    2012

  27. [35]

    John B Lindsay. 2016. Whitebox GAT: A case study in geomorphometric analysis. Computers & Geosciences95 (2016), 75–84

    2016

  28. [36]

    Qi Liu, Jie Niu, Ping Lu, Feifei Dong, Fujun Zhou, Xianglian Meng, Wei Xu, Shan Li, and Bill X Hu. 2022. Interannual and seasonal variations of permafrost thaw depth on the Qinghai-Tibetan plateau: A comparative study using long short-term memory, convolutional neural networks...

    2022

  29. [37]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF international conference on computer vision. 10012–10022

    2021

  30. [38]

    Gengchen Mai, Krzysztof Janowicz, Bo Yan, Rui Zhu, Ling Cai, and Ni Lao. 2020. Multi-scale representation learning for spatial feature distributions using grid cells.arXiv preprint arXiv:2003.00824(2020)

    2020 arXiv

  31. [39]

    Gengchen Mai, Ni Lao, Yutong He, Jiaming Song, and Stefano Ermon. 2023. Csp: Self-supervised contrastive spatial pre-training for geospatial-visual representa- tions. InInternational Conference on Machine Learning. PMLR, 23498–23515

    2023

  32. [40]

    SS Marchenko. 2001. A model of permafrost formation and occurrences in the intracontinental mountains.Norsk Geografisk Tidsskrift-Norwegian Journal of Geography55, 4 (2001), 230–234

    2001

  33. [41]

    Sergei Marchenko, Vladimir Romanovsky, and Gennady Tipenko. 2008. Numeri- cal modeling of spatial permafrost dynamics in Alaska. InProceedings of the ninth international conference on permafrost, Vol. 29. Institute of Northern Engineering, University of Alaska Fairbanks, 1125–1130

    2008

  34. [42]

    Travis W Nauman, Suzann Kienast-Brown, Stephen M Roecker, Colby Brungard, David White, Jessica Philippe, and James A Thompson. 2024. Soil landscapes of the United States (SOLUS): Developing predictive soil property maps of the conterminous United States using hybrid training s...

    2024

  35. [43]

    Vishal Nedungadi, Ankit Kariryaa, Stefan Oehmcke, Serge Belongie, Christian Igel, and Nico Lang. 2024. MMEarth: Exploring multi-modal pretext tasks for geospatial representation learning. InEuropean Conference on Computer Vision. Springer, 164–182

    2024

  36. [44]

    2003.Climate change, permafrost, and impacts on civil infrastructure

    Frederick E Nelson. 2003.Climate change, permafrost, and impacts on civil infrastructure. United States Arctic Research Commission

    2003

  37. [45]

    Ingmar Nitze, Guido Grosse, Benjamin M Jones, Vladimir E Romanovsky, and Julia Boike. 2018. Remote sensing quantifies widespread abundance of permafrost region disturbances across the Arctic and Subarctic.Nature communications9, 1 (2018), 5423

    2018

  38. [46]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748(2018)

    2018 arXiv

  39. [47]

    José Padarian, Budiman Minasny, and Alex B McBratney. 2019. Using deep learning for digital soil mapping.Soil5, 1 (2019), 79–89

    2019

  40. [48]

    SK Panda, SS Marchenko, and VE Romanovsky. 2014. High-resolution permafrost modeling in Wrangell-St. Elias National Park and Preserve.Natural Resource Yijun Lin*, Theresa Chen *, and et al. Report NPS/CAKN/NRTR—2014/861. Fort Collins, Colorado: National Park Service. https://i...

    2014

  41. [49]

    SK Panda, A Prakash, MT Jorgenson, and DN Solie. 2012. Near-surface permafrost distribution mapping using logistic regression and remote sensing in Interior Alaska.GIScience & Remote Sensing49, 3 (2012), 346–363

    2012

  42. [51]

    Santosh K Panda, Sergey S Marchenko, and Vladimir E Romanovsky. 2014. High- resolution permafrost modeling in Denali National Park and Preserve.Nat. Resour. Techn. Rep(2014), 1–44

    2014

  43. [52]

    Santosh K Panda, Anupma Prakash, Diana N Solie, Vladimir E Romanovsky, and M Torre Jorgenson. 2010. Remote sensing and field-based mapping of permafrost distribution along the Alaska Highway corridor, interior Alaska.Permafrost and Periglacial Processes21, 3 (2010), 271–281

    2010

  44. [53]

    Neal J Pastick, M Torre Jorgenson, Bruce K Wylie, Shawn J Nield, Kristofer D Johnson, and Andrew O Finley. 2015. Distribution of near-surface permafrost in Alaska: Estimates of present and future conditions.Remote Sensing of Environ- ment168 (2015), 301–315

    2015

  45. [54]

    Amanda Ramcharan, Tomislav Hengl, Travis Nauman, Colby Brungard, Sharon Waltman, Skye Wills, and James Thompson. 2018. Soil property and class maps of the conterminous United States at 100-meter spatial resolution.Soil Science Society of America Journal82, 1 (2018), 186–201

    2018

  46. [55]

    William U Reybold and Gale W TeSelle. 1989. Soil geographic data bases.Journal of Soil and Water Conservation44, 1 (1989), 28–29

    1989

  47. [56]

    Daniel Riseborough, Nikolay Shiklomanov, Bernd Etzelmüller, Stephan Gru- ber, and Sergei Marchenko. 2008. Recent advances in permafrost modelling. Permafrost and Periglacial Processes19, 2 (2008), 137–156

    2008

  48. [57]

    Pedro A Sanchez, Sonya Ahamed, Florence Carré, Alfred E Hartemink, Jonathan Hempel, Jeroen Huising, Philippe Lagacherie, Alex B McBratney, Neil J McKenzie, Maria De Lourdes Mendonça-Santos, et al. 2009. Digital soil map of the world. Science325, 5941 (2009), 680–681

    2009

  49. [58]

    Xun Shi, A-Xing Zhu, James E Burt, Feng Qi, and Duane Simonson. 2004. A case-based reasoning approach to fuzzy soil mapping.Soil Science Society of America Journal68, 3 (2004), 885–894

    2004

  50. [59]

    Mostafa A Shirazi, Colleen Burch Johnson, James M Omernik, Denis White, Patricia K Haggerty, and Glenn E Griffith. 2003. Quantitative soil descriptions for ecoregions of the United States.Journal of environmental quality32, 2 (2003), 550–561

    2003

  51. [60]

    Yu L Shur and M Torre Jorgenson. 2007. Patterns of permafrost formation and degradation in relation to climate and ecosystems.Permafrost and Periglacial Processes18, 1 (2007), 7–19

    2007

  52. [61]

    Soil Survey Staff. 2022. Gridded national soil survey geographic (gNATSGO) database for the conterminous United States. (2022)

    2022

  53. [62]

    Soil Survey Staff. 2022. State soil geographic database (Statsgo2) for Alaska. (2022)

    2022

  54. [63]

    Evan Austin Thaler, Sebastian Uhleman, Joel C Rowland, J Schwenk, Chen Wang, Baptiste Dafflon, and Katrina Eleanor Bennett. 2023. High-resolution maps of near-surface permafrost for three watersheds on the Seward Peninsula, Alaska derived from machine learning.Earth and Space ...

    2023

  55. [64]

    JA Thompson, TW Nauman, NP Odgers, Z Libohova, and JW Hempel. 2012. Har- monization of legacy soil maps in North America: status, trends, and implications for digital soil mapping efforts. InDigital Soil Assessments and Beyond: Proceed- ings of the Fifth Global Workshop on Dig...

    2012

  56. [65]

    2022.Land Resource Regions and Major Land Resource Areas of the United States, the Caribbean, and the Pacific Basin

    United States Department of Agriculture, Natural Resources Conservation Ser- vice. 2022.Land Resource Regions and Major Land Resource Areas of the United States, the Caribbean, and the Pacific Basin. Number 296 in Agriculture Handbook. U.S. Department of Agriculture

    2022

  57. [66]

    Aoran Xiao, Weihao Xuan, Junjue Wang, Jiaxing Huang, Dacheng Tao, Shijian Lu, and Naoto Yokoya. 2024. Foundation models for remote sensing and Earth Observation: A survey.arXiv preprint arXiv:2410.16602(2024)

    2024 arXiv

  58. [67]

    Yonghong Yi, John S Kimball, Richard H Chen, Mahta Moghaddam, Rolf H Reichle, Umakant Mishra, Donatella Zona, and Walter C Oechel. 2018. Charac- terizing permafrost active layer dynamics and sensitivity to landscape spatial heterogeneity in Alaska.The Cryosphere12, 1 (2018), 145–161

    2018

  59. [68]

    Kenji Yoshikawa, Mikhail Kanevskiy, Yuri Shur, Vladimir Romanovsky, Sergei Marchenko, Guido Grosse, Jerry Brown, and Ben Jones. [n. d.]. Permafrost characteristics of Alaska

  60. [69]

    Yu Zhang, Junhua Li, Xiping Wang, Wenjun Chen, Wendy Sladen, Larry Dyke, Lynda Dredge, Jean Poitevin, Donald McLennan, Heather Stewart, et al. 2012. Modelling and mapping permafrost at high spatial resolution in Wapusk National Park, Hudson Bay Lowlands.Canadian Journal of Ear...

    2012

Pith tools

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