Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

RAG for Geoscience: What We Expect, Gaps and Opportunities

T0 review · 2 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper argues that text-only retrieve-then-generate RAG cannot meet geoscience's evidence-hungry tasks, and proposes Geo-RAG, a four-stage retrieve–reason–generate–verify loop that grounds outputs in multimodal Earth data, physical laws

desk verdict A well-structured vision paper for a physics-aware, multimodal RAG loop; the flagship gap-filling use case lacks a verification target, but as a research agenda it deserves a serious referee. read the letter →

arxiv 2508.11246 v1 pith:VDPFYJB2 submitted 2025-08-15 cs.ET cs.IR

Runlong Yu , Shiyuan Luo , Rahul Ghosh , Lingyao Li , Yiqun Xie , Xiaowei Jia This is my paper
classification cs.ETcs.IR
keywords retrieval-augmentedgenerationgeosciencemultimodalretrievalphysics-awarereasoningscience-gradeartifactsverificationearthobservationgeo-RAG
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

Geoscientific questions—filling gaps in satellite observations, assembling equations and parameters for models, geolocating field photos, and replaying historical cases for policy—demand evidence that is spatial, temporal, and physical, not just textual. The paper argues that mainstream retrieval-augmented generation (RAG), which retrieves text passages and feeds them into a language model, is structurally unable to meet this demand because it is text-only, physics-blind, format-agnostic, and verification-free. It proposes Geo-RAG, a modular retrieve → reason → generate → verify loop that retrieves multimodal Earth data, enforces dimensional and conservation constraints during reasoning, emits machine-readable science artifacts such as CF-compliant rasters and simulator configuration files, and verifies outputs against numerical models, sensors, and experts. If the blueprint holds, geoscience AI systems could become physically grounded, uncertainty-aware, and traceable enough for operational forecasting and policy decisions.

What carries the argument

The load-bearing mechanism is the four-stage retrieve → reason → generate → verify loop itself, which converts RAG from a one-shot text pipeline into an iterative scientific instrument. It is carried by three interacting components: a hybrid indexing architecture (dense CLIP-style embeddings for language and imagery, geohash or Hilbert curves for spatial locality, balanced B-trees for time, and AST hashing for symbolic code and equations) that makes multimodal Earth data retrievable; a neuro-symbolic reasoning stack (dimensional sentry, differentiable constraint solvers, CF/WMO semantic filters) that prunes hypotheses by physical admissibility; and a verification scaffold (model replay, sens

What would settle it

Run the proposed loop on one concrete scenario—say, taking cloud-covered satellite imagery and producing gap-filled rasters with uncertainty—and compare it against text-only RAG and a physics-blind generator on withheld sensor data. The claim fails if the four-stage loop does not beat both on accuracy and physical plausibility, or if a hybrid index over a daily-updating petabyte-scale Earth dataset cannot hold recall without memory expanding by tens of gigabytes per day.

Watch

Extended reading notes

Core claim

The central claim is that a simple retrieve-then-generate pipeline is insufficient for geoscience, and that RAG must be reimagined as a closed, self-refining scientific workflow rather than a text summarizer. Geo-RAG treats Earth itself as the primary knowledge base, retrieving co-registered artifacts—satellite swaths, weather-station time series, DEM tiles, reanalysis cubes, symbolic equations, legacy code—through a hybrid index of dense CLIP-style embeddings, geohash/Hilbert spatial curves, temporal B-trees, and AST hashing for equations. Reasoning happens in metric space rather than token space: a dimensional sentry, differentiable constraint solvers, and CF/WMO semantic filters reject ph

Load-bearing premise

The blueprint depends on the assumption that petabyte-scale, continuously updated Earth data can be indexed for multimodal retrieval cheaply enough, and that physics-constrained decoding and verification can run fast enough for real workflows; the paper itself flags these as unresolved tensions in its implementation section.

Editorial extensions

If this is right

  • If Geo-RAG works, geoscience AI systems can retrieve and fuse non-text data directly—imagery, terrain, reanalysis, equations—without flattening them into lossy text.
  • Generated outputs shift from paragraphs to executable science artifacts: gap-filled rasters with uncertainty bands, simulator configuration files, and geolocation maps that plug directly into GIS and modeling workflows.
  • Physics-aware reasoning and dimensional checks would prevent implausible outputs such as negative rainfall or impossible temperatures before they reach operational forecasts.
  • The verification loop gives every output a testable status: artifacts are replayed in simulators, cross-checked against independent sensors, and reviewed by experts, with failures triggering re-retrieval.
  • Traceability improves: provenance metadata and physics audit trails make outputs auditable for regulatory and policy use.

Reading between the lines

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

  • The components Geo-RAG assembles already exist in isolation; the paper's real wager is that closing the loop—verification feeding back into retrieval—creates a qualitative gain. A head-to-head prototype on one scenario (say, cloud-gap filling) would settle whether the loop outperforms a pipeline that merely adds physics constraints without verification.
  • The retrieve–reason–generate–verify pattern is not geoscience-specific; any evidence-hungry, simulation-backed field (climate adaptation, epidemiology, hydrology) could adopt the same loop, though the paper leaves that generalization implicit.
  • The paper's own implementation section suggests the binding constraint is engineering cost: mutable petabyte-scale indexing, latency of constraint-aware decoding, and the expense of verification. One direct test would be measuring recall and memory growth under a daily-updating satellite feed.
  • A user-study implication: if physics audit trails and provenance tagging measurably increase expert trust and willingness to adopt AI outputs, that would justify the extra machinery; the paper does not run that experiment.
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

2 major / 6 minor

Summary. This position paper argues that mainstream retrieval-augmented generation (RAG) is text-centric and therefore insufficient for geoscience tasks, which the authors characterize as evidence-hungry. They propose Geo-RAG, a modular retrieve → reason → generate → verify loop that combines multimodal Earth-data retrieval, physics- and domain-constrained reasoning, generation of science-grade artifacts (e.g., NetCDF rasters, model configuration files), and verification against models, ground truth, and experts. The paper motivates the proposal with four scenarios (data-gap filling, scientific knowledge retrieval, spatial reasoning from visual inputs, and case-based decision support), identifies four limitations of mainstream RAG, sketches the Geo-RAG architecture, lists five implementation challenges, and briefly discusses governance. No system is implemented or evaluated; the contribution is a problem framing, a blueprint, and a research agenda.

Significance. If realized, Geo-RAG would be a meaningful step toward making generative AI usable in high-stakes geoscience: the four limitations identified (text-only retrieval, physics-blind reasoning, format-insensitive generation, verification-free outputs) are real and the proposed loop is a coherent organizing principle. The paper is commendably honest about open challenges, especially in Section 6, and it names concrete technologies (CLIP-style embeddings, geohash/Hilbert curves, CF conventions, PROV-O) that make the vision falsifiable and implementable. However, as a vision paper it contains no experiments or prototype, and some of its empirical premises are asserted rather than demonstrated. Its value lies primarily in problem reframing and agenda setting.

major comments (2)
  1. [§4.4 and §2 (Scenario 1)] The verification stage is load-bearing for the paper's trustworthiness claim, but it is epistemically mismatched with the flagship motivating scenario. In Scenario 1, the task is filling gaps in Earth observations where ground measurements and independent sensors are, by definition, unavailable or sparse; otherwise there would be no gap. Yet §4.4 defines verification as checking against 'numerical models, ground measurements, and expert assessments'. In the gap-filling setting, sensor cross-checks and withheld observations cannot provide a failure signal, and model replay can test physical consistency but not fidelity to unobserved reality. The self-refining 'if verification fails, retrieve new evidence' mechanism therefore has no clear trigger in the paper's central use case. Please address this mismatch explicitly: state what verification means in data-sparse regimes, how validation is
  2. [§3, Limitation 2] The central claim that mainstream RAG is insufficient for geoscience rests on asserted failure modes such as LLMs producing 'negative rainfall' or 'ocean temperatures exceeding 1000°C'. These examples are illustrative but are not cited or documented. If the paper's reframing is to be credible, it should provide at least one concrete, referenced instance of such a failure in a geoscience RAG or LLM application, or present a small motivating study. Without this, the reader cannot distinguish real gaps from hypothetical ones, and Table 1's 'Open Gaps' column remains entirely qualitative.
minor comments (6)
  1. [§3, Limitation 2] Typo: 'LLMs lack an inherent encode of physical laws' should read 'an inherent encoding'.
  2. [Headers] The running header contains the placeholder 'Conference acronym ’XX, June 03–05, 2018, Woodstock, NY' and must be removed or replaced with the actual venue.
  3. [§5] 'Geo-RAG in Action' is a hypothetical walkthrough, not an evaluation. Consider retitling it 'Illustrative Walkthrough' or adding a sentence clarifying that no implementation is reported.
  4. [§1.2 vs §4.4] The contributions say Geo-RAG integrates 'automated verification', but §4.4 includes a human expert-in-the-loop stage. Please reconcile this wording, e.g., by saying 'semi-automated verification'.
  5. [References] Reference [26] (Co2sum: contrastive learning for factual-consistent abstractive summarization) is cited to support constraint-aware decoding in §6, but summarization is not the same as constrained decoding. Please verify the citation or replace it with a more directly relevant work.
  6. [§7] The governance section states retrieval should 'mitigate rich-data bias through equitable evidence selection' but gives no mechanism or example. A one-sentence illustration or pointer to future work would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Geo-RAG is a vision/blueprint paper with no derivation, fitted parameters, or prediction that reduces to its inputs.

full rationale

This paper is a position/vision piece, not a derivation or empirical study. It contains no equations, no fitted parameters, and no experimental 'prediction' that could be equivalent to a training signal or input by construction. The central claim—that a simple retrieve-then-generate pipeline is insufficient for geoscience and that a retrieve→reason→generate→verify loop would help—is a proposal grounded in motivating scenarios, not a result derived from a premise that already contains the conclusion. Statements such as 'Geo-RAG supports four core capabilities' are definitional descriptions of the proposed architecture, not independent predictions. The self-citations ([38]–[42]) are background references to the authors' prior work on physics-guided learning and geo-foundation models; they are used to suggest that differentiable constraints and physics-aware learning are feasible, but they are not invoked as a uniqueness theorem, nor do they carry the burden of proving that Geo-RAG works. The reviewer-noted issue that verification requires ground truth in gap-filling scenarios is a real feasibility/epistemic limitation, but it is not a circularity: it does not make the paper's conclusion equivalent to its assumptions. No step in the paper's argument chain reduces to its own input. Accordingly, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

The paper introduces no free parameters or physical entities. Its central claim rests on four domain assumptions about the nature of geoscience tasks and the feasibility of the proposed architecture, none of which are validated. The architecture itself (Geo-RAG) is an invented system with no independent evidence.

assumptions (4)
  • domain assumption Geoscientific tasks are evidence-hungry and cannot be served by text-only RAG.
    Section 1.2 asserts this without empirical evidence; it is the motivating premise of the entire proposal.
  • domain assumption Physical laws can be encoded as differentiable soft constraints in the reasoning stage.
    Section 4.2 relies on this to claim physics-aware reasoning is operational; it cites prior work [40,41] but no Geo-RAG-specific implementation exists.
  • ad hoc to paper The proposed hybrid indexing (CLIP embeddings, geohash/Hilbert curves, B-trees, AST hashing) can handle petabyte-scale multimodal Earth data.
    Section 4.1 asserts this architecture without design details or performance evaluation; Section 6 then states streaming index mutation is a significant challenge.
  • ad hoc to paper Verification via model replay, sensor cross-checks, and expert review can be automated affordably.
    Section 4.4 proposes a containerized validation system, but Section 6 notes full physics replay is computationally prohibitive and adaptive validators depend on priors that can become outdated.
invented entities (1)
  • Geo-RAG system
    purpose: Proposed four-stage loop (retrieve-reason-generate-verify) for multimodal, physics-aware, verifiable RAG in geoscience.
    Geo-RAG is described only as a vision in Sections 4 and 5; no implementation, benchmark, or falsifiable prediction is given, so there is no independent evidence of its effectiveness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RAG for Geoscience: What We Expect, Gaps and Opportunities." pith.science (2026). https://pith.science/paper/VDPFYJB2

@misc{pith2026250811246,
  author       = {Pith},
  title        = {Pith review of: RAG for Geoscience: What We Expect, Gaps and Opportunities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VDPFYJB2}},
  note         = {Machine review of arXiv:2508.11246}
}
abstract

Retrieval-Augmented Generation (RAG) enhances language models by combining retrieval with generation. However, its current workflow remains largely text-centric, limiting its applicability in geoscience. Many geoscientific tasks are inherently evidence-hungry. Typical examples involve imputing missing observations using analog scenes, retrieving equations and parameters to calibrate models, geolocating field photos based on visual cues, or surfacing historical case studies to support policy analyses. A simple ``retrieve-then-generate'' pipeline is insufficient for these needs. We envision Geo-RAG, a next-generation paradigm that reimagines RAG as a modular retrieve $\rightarrow$ reason $\rightarrow$ generate $\rightarrow$ verify loop. Geo-RAG supports four core capabilities: (i) retrieval of multi-modal Earth data; (ii) reasoning under physical and domain constraints; (iii) generation of science-grade artifacts; and (iv) verification of generated hypotheses against numerical models, ground measurements, and expert assessments. This shift opens new opportunities for more trustworthy and transparent geoscience workflows.

Figures

Figures reproduced from arXiv: 2508.11246 by the authors.

Figure 1
Figure 1. The retrieve → reason → generate → verify loop. 4 Vision of Geo-RAG Geo-RAG reimagines mainstream RAG as a four-stage, self-refining loop ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. PIER: Physics-Informed Environmental Retrieval for Time-Series Modeling

    cs.LG 2026-07 conditional novelty 5.0 of 10

    PIER augments embedding-based retrieval for lake modeling with a physics-aware stream scored by local verifiers, improving water temperature and dissolved oxygen prediction across 356 lakes.

Reference graph

Works this paper leans on

45 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    PROV-O: The PROV Ontology

    2013. PROV-O: The PROV Ontology. W3C Recommendation REC-prov-o-20130430. World Wide Web Consortium (W3C), Cambridge, MA. https://www.w3.org/TR/ 2013/REC-prov-o-20130430/ Accessed 14 Jun 2025

    work page 2013

  2. [2]

    Miltiadis Allamanis, Earl T Barr, Premkumar Devanbu, and Charles Sutton. 2018. A survey of machine learning for big code and naturalness. ACM Computing Surveys (CSUR) 51, 4 (2018), 1–37

    work page 2018

  3. [3]

    Grigory Isaakovich Barenblatt. 1996. Scaling, self-similarity, and intermediate asymptotics: dimensional analysis and intermediate asymptotics . Number 14. Cambridge University Press

    work page 1996

  4. [4]

    Benjamin Bischke, Patrick Helber, Joachim Folz, Damian Borth, and Andreas Dengel. 2019. Multi-task learning for segmentation of building footprints with deep neural networks. In 2019 IEEE International Conference on Image Processing (ICIP). IEEE, 1480–1484

    work page 2019

  5. [5]

    Pierre Bourhis, Juan L Reutter, Fernando Suárez, and Domagoj Vrgoč. 2017. JSON: data model, query languages and schema specification. In Proceedings of the 36th ACM SIGMOD-SIGACT-SIGAI symposium on principles of database systems . 123–135

    work page 2017

  6. [6]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374 (2021)

    arXiv 2021

  7. [7]

    Jay DeYoung, Iz Beltagy, Madeleine van Zuylen, Bailey Kuehl, and Lucy Lu Wang

    work page

  8. [8]

    Matthias Drusch, Umberto Del Bello, Sébastien Carlier, Olivier Colin, Veron- ica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. 2012. Sentinel-2: ESA’s optical high-resolution mission for GMES operational services. Remote sensing of Environment 120 (2012), 25–36

    work page 2012

Show all 45 references
  1. [9]

    Brian Eaton, Jonathan Gregory, Bob Drach, Karl Taylor, Steve Hankin, John Caron, Rich Signell, Phil Bentley, Greg Rappa, Heinke Höck, et al. 2003. NetCDF Climate and Forecast (CF) metadata conventions. URL: http://cfconventions. org/Data/cf- conventions/cf-conventions-1.8/cf-c...

    2003

  2. [10]

    Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning. PMLR, 1050–1059

    2016

  3. [11]

    Yolanda Gil, Ewa Deelman, Mark Ellisman, Thomas Fahringer, Geoffrey Fox, Dennis Gannon, Carole Goble, Miron Livny, Luc Moreau, and Jim Myers. 2007. Examining the challenges of scientific workflows. Computer 40, 12 (2007), 24–32

    2007

  4. [12]

    Paul Griffioen et al. 2019. A unit-aware matrix language and its application in control and auditing. Ph. D. Dissertation. Universiteit van Amsterdam

    2019

  5. [13]

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. In International conference on machine learning. PMLR, 3929–3938

    2020

  6. [14]

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquín Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. 2020. The ERA5 global reanalysis. Quarterly journal of the royal meteorolog- ical society 146, 730 (2020), 1999–2049

    2020

  7. [15]

    Gautier Izacard and Edouard Grave. 2020. Distilling knowledge from reader to retriever for question answering. arXiv preprint arXiv:2012.04584 (2020)

    2020 arXiv

  8. [16]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. Few-shot learning with retrieval augmented language models.arXiv preprint arXiv:2208.03299 1, 2 (2022), 4

    2022 arXiv

  9. [17]

    Greg Janée. 2009. Preserving Geospatial Data: The National Geospatial Digital Archive’s Approach. In Archiving Conference, Vol. 6. Society of Imaging Science and Technology, 25–29

    2009

  10. [18]

    Herve Jegou, Matthijs Douze, and Cordelia Schmid. 2010. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence 33, 1 (2010), 117–128

    2010

  11. [19]

    John R Jensen. 2009. Remote sensing of the environment: An earth resource per- spective 2/e. Pearson Education India

    2009

  12. [20]

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. 2021. Physics-informed machine learning. Nature Reviews Physics 3, 6 (2021), 422–440

    2021

  13. [21]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open-Domain Question Answering.. In EMNLP (1). 6769–6781

    2020

  14. [22]

    Omar Khattab, Keshav Santhanam, Xiang Lisa Li, David Hall, Percy Liang, Christo- pher Potts, and Matei Zaharia. 2022. Demonstrate-search-predict: Compos- ing retrieval and language models for knowledge-intensive nlp. arXiv preprint arXiv:2212.14024 (2022)

    2022 arXiv

  15. [23]

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017)

    2017

  16. [24]

    Patrick Lewis et al . 2020. Retrieval-augmented generation for knowledge- intensive NLP tasks. arXiv preprint arXiv:2005.11401 (2020)

    2020 arXiv

  17. [25]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2023. Lost in the middle: How language models use long contexts. arXiv preprint arXiv:2307.03172 (2023)

    2023 arXiv

  18. [26]

    Wei Liu, Huanqin Wu, Wenjing Mu, Zhen Li, Tao Chen, and Dan Nie. 2021. Co2sum: contrastive learning for factual-consistent abstractive summarization. arXiv preprint arXiv:2112.01147 (2021)

    2021 arXiv

  19. [27]

    Lei Ma, Yu Liu, Xueliang Zhang, Yuanxin Ye, Gaofei Yin, and Brian Alan Johnson

  20. [28]

    Yu A Malkov and Dmitry A Yashunin. 2018. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE transactions on pattern analysis and machine intelligence 42, 4 (2018), 824–836

    2018

  21. [29]

    Felipe Pezoa, Juan L Reutter, Fernando Suarez, Martín Ugarte, and Domagoj Vrgoč. 2016. Foundations of JSON schema. In Proceedings of the 25th international conference on World Wide Web. 263–273

    2016

  22. [30]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.000...

    2021 arXiv

  23. [31]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084 (2019)

    2019 arXiv

  24. [32]

    Hans Sagan. 2012. Space-filling curves. Springer Science & Business Media

    2012

  25. [33]

    Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2023. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652 (2023)

    2023 arXiv

  26. [34]

    William C Skamarock, Joseph B Klemp, Jimy Dudhia, David O Gill, Dale M Barker, Michael G Duda, Xiang-Yu Huang, Wei Wang, Jordan G Powers, et al. 2008. A description of the advanced research WRF version 3. NCAR technical note 475, 125 (2008), 10–5065

    2008

  27. [35]

    World Meteorological Organization. 2011. Manual on Codes: International Codes Volume I.1. WMO. WMO-No. 306

    2011

  28. [36]

    Shangyu Wu, Ying Xiong, Yufei Cui, Haolun Wu, Can Chen, Ye Yuan, Lianming Huang, Xue Liu, Tei-Wei Kuo, Nan Guan, et al. 2024. Retrieval-augmented gener- ation for natural language processing: A survey. arXiv preprint arXiv:2407.13193 (2024)

    2024 arXiv

  29. [37]

    Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. 2018. DOTA: A large-scale dataset for object detection in aerial images. In Proceedings of the IEEE conference on computer vision and pattern recognition...

    2018

  30. [38]

    Yiqun Xie, Zhaonan Wang, Gengchen Mai, Yanhua Li, Xiaowei Jia, Song Gao, and Shaowen Wang. 2023. Geo-foundation models: Reality, gaps and opportunities. In Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems. 1–4

    2023

  31. [39]

    Runlong Yu, Shengyu Chen, Yiqun Xie, Huaxiu Yao, Jared Willard, and Xiaowei Jia. 2025. Foundation Models for Environmental Science: A Survey of Emerging Frontiers. arXiv preprint arXiv:2504.04280 (2025)

    2025 arXiv

  32. [40]

    Runlong Yu, Chonghao Qiu, Robert Ladwig, Paul Hanson, Yiqun Xie, and Xi- aowei Jia. 2025. Physics-Guided Foundation Model for Scientific Discovery: An Application to Aquatic Science. Proceedings of the AAAI Conference on Artificial Intelligence 39, 27 (Apr. 2025), 28548–28556

    2025

  33. [41]

    Hanson, Yiqun Xie, Yanhua Li, and Xiaowei Jia

    Runlong Yu, Chonghao Qiu, Robert Ladwig, Paul C. Hanson, Yiqun Xie, Yanhua Li, and Xiaowei Jia. 2024. Adaptive Process-Guided Learning: An Application in Predicting Lake DO Concentrations. In 2024 IEEE International Conference on Data Mining (ICDM). IEEE, 580–589

    2024

  34. [42]

    Runlong Yu, Yiqun Xie, and Xiaowei Jia. 2025. Environmental Computing as a Branch of Science. Commun. ACM (June 2025), 5 pages

    2025

  35. [43]

    Liangpei Zhang, Lefei Zhang, and Bo Du. 2016. Deep learning for remote sensing data: A technical tutorial on the state of the art. IEEE Geoscience and remote sensing magazine 4, 2 (2016), 22–40

    2016

  36. [2019]

    ISPRS journal of photogrammetry and remote sensing 152 (2019), 166–177

    Deep learning in remote sensing applications: A meta-analysis and review. ISPRS journal of photogrammetry and remote sensing 152 (2019), 166–177

    2019

  37. [2021]

    arXiv preprint arXiv:2104.06486 (2021)

    Ms2: Multi-document summarization of medical studies. arXiv preprint arXiv:2104.06486 (2021)

    2021 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.