Pith. sign in

REVIEW 3 major objections 5 minor 69 references

Pairing multimodal LLM place candidates with cross-view image matching pins down ambiguous disaster photos far better than either cue alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Combining MLLM candidate place extraction with DINOv2 cross-view matching of social-media, street-view, and satellite images raises disaster photo geolocation accuracy to 71.6% within 1 km on Hurricane Harvey data.

T0 review reviewed 2026-07-30 challenge →

load-bearing objection Solid two-stage disaster geoloc system with real gains on Harvey; the mechanism claim is a bit ahead of the ablations, but the paper is still worth engaging. the 3 major comments →

arxiv 2607.24856 v1 pith:DWRYMQD2 submitted 2026-07-26 cs.CV cs.AI

DisasterTD: Disaster Toponym Disambiguation Using Multimodal LLMs and Cross-View Geolocalization

Wenping Yin , Ziqi Liu , Naixia Mou , Weijia Li , Danfeng Hong , Hao Li This is my paper
classification cs.CV cs.AI
keywords toponym disambiguationgeolocalizationcross-viewmultimodal LLMdisaster responsesocial media imageryremote sensingDINOv2
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Social media photos from disasters often carry place names that are incomplete, generic, or shared by many real-world sites, so text alone rarely yields a usable coordinate. DisasterTD treats a multimodal large language model as a candidate generator: it reads the photo, extracts toponyms, and queries map services for possible locations inside the disaster region. Those candidates are then checked by matching the social media image against overhead remote-sensing imagery, with street-view imagery as an optional bridge, using a vision-transformer foundation model (DINOv2) to score visual similarity. On a Hurricane Harvey set split by toponym clarity versus ambiguity, the combined pipeline beats both LLM-only geocoding and pure cross-view retrieval, especially when names are vague chain stores or repeated street labels. The practical claim is that semantic shortlisting plus visual verification is what makes fine-grained, street-to-block geolocalization feasible when emergency imagery is noisy and time-critical.

Core claim

On the Hurricane Harvey cross-view benchmark, DisasterTD—MLLM candidate generation followed by DINOv2-based SMI–SVI–RSI verification—consistently outperforms MLLM-only and cross-view-only baselines, reaching geolocalization accuracies of 71.62%, 62.36%, 57.99%, 52.09%, and 47.01% within 1000 m, 500 m, 250 m, 100 m, and 50 m, with mean and median errors of 11.33 km and 0.68 km, and the largest gains on ambiguous toponyms.

What carries the argument

DisasterTD: a two-stage loop in which an MLLM extracts toponyms and builds a region-restricted candidate set, then a DINOv2 encoder ranks those candidates by cross-view similarity among the social media image, optional street view, and remote-sensing patches.

Load-bearing premise

Map services and available overhead or street imagery must usually include the true site among the candidates and remain visually matchable despite missing coverage, damage, or time gaps; if the true place is absent or the reference views are wrong, ranking cannot recover it.

What would settle it

Re-run the same four-way Harvey split with candidate sets that deliberately omit the ground-truth coordinate, or with RSI/SVI collected years after the event, and check whether the reported accuracy lift over MLLM-only and direct SMI–RSI baselines collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Ambiguous chain-store and generic-street photos can be localized to block or POI scale when LLM shortlists are filtered by overhead match scores.
  • Cross-view retrieval alone is a weaker disaster tool than the same matcher fed a small semantic candidate set.
  • Gains hold when the visual backbone is swapped, so the lift is mainly from the generate-then-verify design rather than one encoder.
  • Median error near 0.7 km with a long-tailed mean flags that most cases succeed while a minority of missing or look-alike candidates still fail hard.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same generate-then-verify pattern could transfer to non-disaster VGI where place names are sparse but satellite archives exist.
  • Operational use will hinge less on peak accuracy than on latency and map coverage in poorly mapped regions the paper flags but does not benchmark.
  • Uncertainty scores on candidate completeness could turn the long-tail failures into a human-in-the-loop triage signal for emergency desks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes DisasterTD, a two-stage pipeline for geolocalizing disaster-related social media imagery (SMI): GPT-4o extracts toponyms and generates candidate geolocations via geocoding/place-search restricted to the disaster region, then a DINOv2-based cross-view matcher (with optional street-view imagery as a bridge between SMI and overhead remote-sensing imagery) ranks candidates. Evaluation on a 1,000-sample Hurricane Harvey benchmark, stratified into four categories by toponym number and ambiguity, reports overall Geoloc ACC of 71.62%/62.36%/57.99%/52.09%/47.01% at 1000/500/250/100/50 m, mean/median error 11.33/0.68 km, consistently above MLLM-only, SMI↔RSI, and SMI↔SVI↔RSI baselines, with the largest gains on ambiguous toponyms (e.g., single-ambiguous 29.69%→65.23% at 1000 m). Encoder-swap experiments (ConvNeXt, SAIG-D, TransGeo, Sample4Geo) suggest the gains come from the framework rather than the specific encoder. The evaluation is against held geographic ground truth and is not circular; the empirical protocol is reasonable in design. The main weakness is that the evidence needed to attribute the improvement to visual verification — as opposed to properties of the MLLM-generated candidate pools — is not reported.

Significance. If the attribution claim survives the requested controls, this is a useful and timely contribution to disaster GeoAI: toponym disambiguation for crisis social-media imagery is a real gap, and the paper offers (i) a practical two-stage design combining commodity MLLMs with foundation-model cross-view matching, (ii) a curated 1,000-sample Hurricane Harvey SMI–RSI–SVI benchmark with a sensible four-way difficulty taxonomy, and (iii) multi-threshold, category-stratified evaluation with several baselines including encoder swaps — more evaluation rigor than typical in this subarea. The large, directionally consistent gains on ambiguous toponyms, if verified as ranker-driven, would be a meaningful result for time-critical situational-awareness pipelines. The encoder-swap ablation (Fig. 7c–d) already provides evidence the framework rather than DINOv2 specifically carries the gains, which strengthens generality.

major comments (3)
  1. [§IV.B–C, Tables II–III] Tables II–III and §IV.B: the paper's central attribution — that gains over the MLLM-only baseline come from cross-view visual verification — cannot be established without two numbers the authors already have and never report. (1) Candidate-set recall (oracle accuracy): the fraction of samples whose ground truth lies within d of ANY candidate; §IV.B asserts the true geolocation 'is usually included within the candidate set' without quantifying it. (2) A random-ranker floor: Geoloc ACC when Top-1 is picked uniformly from the same candidate pools. For ambiguous toponyms the region-restricted place search (§III.B, l_j ∈ Ω) can return many candidates densely covering the disaster area, so a random pick can land within 1000 m of the truth. An unknown fraction of the headline single-ambiguous jump (29.69%→65.23% @1000 m) could therefore reflect candidate density rather than visual disambiguatio
  2. [§III.B, §IV.A] The candidate-generation stage is under-specified to a degree that blocks reproducibility and interpretation. The text never states: the definition/extent of the disaster region Ω; the place-search radius; the number of candidates k per toponym (or its distribution); the RSI tile size/footprint fetched around each candidate; or the SVI retrieval radius. §III.B defers stage-one details to prior work [17], but the ranking task's difficulty — and hence the meaning of the reported accuracies — depends directly on these quantities. They should be stated in this manuscript, ideally with the GPT-4o prompts and map-service query parameters in an appendix.
  3. [§IV.A (dataset taxonomy), Fig. 3] The four-way single/multiple × clear/ambiguous taxonomy drives the paper's signature claim ('largest improvements appear in ambiguous toponyms'), yet §IV.A never describes how the labels were produced. Were the 1,000 SMI manually annotated? By how many annotators, with what decision rule for 'ambiguous' (e.g., >1 gazetteer hit within Ω?), and with what agreement? The counts (272/275/302/151) are given, but without a labeling protocol the category-wise tables that carry the main claim rest on an unverifiable partition. A short annotation-protocol subsection and, ideally, released labels would resolve this.
minor comments (5)
  1. [§III.D, Eq. (7)] The denominator is malformed: Σ_j |D_j| appears to intend the total number of evaluated samples (n); as written it sums set cardinalities of ground-truth objects. Please rewrite, and state explicitly whether the metric is Top-1-only (§III.D suggests Top-1 or Top-k is selectable — clarify which is used in Tables II–III).
  2. [§III.C.2, Eq. (3)] The values of the loss weights λ1, λ2, λ3 in Eq. (3), the training data splits for the cross-view encoder (was any Harvey data used in training, or only the 'MultiIan' set?), and the Top-K value used during retrieval are not reported. Also, 'MultiIan dataset' is undefined here — presumably the Hurricane Ian data from [5]; please name and cite it properly.
  3. [Table III, §IV.C.2] Several numerical inconsistencies merit a sentence of explanation: in Table III the multi-ambiguous median error (0.55 km) is lower than multi-clear (0.72 km) and overall-multiple (0.66 km), despite multi-ambiguous being worst on every ACC column; and §IV.C.2 quotes 64.70% as the 250 m multi-clear value where Table III shows 58.27% is Overall-Multiple — please recheck the in-text numbers against the table.
  4. [Tables II–III] Given per-category n of 151–302, please add bootstrap confidence intervals (or at least Wilson intervals) to Tables II–III. Effect sizes are large enough that conclusions are unlikely to change, but the 'consistently outperforms' language in the abstract warrants uncertainty quantification.
  5. [Throughout] Typos and presentation: 'disambiguiation' (Index Terms); 'oveall' (§III.C.2); 'SUBW AY' (§IV.C.2); 'two setting' (§III.D); 'V o' for 'Vo' (ref. [32] and §II.A); duplicated panel label '(b)' in Fig. 5 between (f) and (h). Fig. 7 gives encoder-swap results only as bar charts — a small table of the underlying numbers (partially in text) would help. Please also state dataset/code availability explicitly.

Circularity Check

0 steps flagged

No significant circularity: empirical ACC metrics on held Harvey ground truth are not forced by definition, fit, or self-citation chain.

full rationale

DisasterTD's load-bearing claims are measured Geoloc ACC@d and mean/median geodesic error on the Hurricane Harvey cross-view benchmark against external geographic ground truth, compared to MLLM-only and cross-view-only baselines (Tables II–III, Fig. 7). Accuracy is the standard fraction of predictions within distance d of ground truth (Eq. 7), not a quantity defined from the model's own fitted parameters. MLLM candidate generation and DINOv2/InfoNCE cross-view ranking are methodological components; self-citations to prior MLLM geolocalization [17] and triple-objective cross-view work [5] reuse those components but do not justify or force the reported Harvey numbers. Loss weights, contrastive training, and encoder swaps are ordinary engineering choices and are not equated to the claimed accuracies. Gaps such as unreported candidate-set recall or random-rank floors affect attribution strength, not circularity. The derivation chain is an empirical systems evaluation, not a closed definitional loop.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

The central empirical claim rests on standard contrastive/ViT machinery, proprietary MLLM and map APIs, availability of aligned RSI/SVI, and the assumption that visual similarity among candidates selects the true site. Few numeric knobs are reported as fitted; the main load-bearing premises are domain and tooling assumptions rather than new physical entities.

free parameters (3)
  • Multi-view loss weights λ1, λ2, λ3
    Coefficients on SMI–SVI, SVI–RSI, and SMI–RSI InfoNCE terms in Eq. (3); values not specified in the text but control the bridge role of SVI.
  • Geoloc distance thresholds {50,100,250,500,1000} m = 50/100/250/500/1000 m
    Evaluation cutoffs chosen to map to POI/street/block/neighborhood/district levels (Table I); they define the reported ACC numbers though not trained parameters.
  • Disaster-region search restriction Ω and candidate count k
    Place retrieval is limited to the affected region and returns a finite candidate set; completeness of this set is load-bearing for later ranking.
axioms (5)
  • domain assumption MLLM-extracted toponyms plus commercial geocoding/place search yield a candidate set that frequently contains the true coordinates when restricted to the disaster region.
    Stage-1 design in §III.B; failures when the true site is missing are discussed in §V as residual large errors.
  • domain assumption DINOv2 global (and patch) features make SMI, SVI, and RSI sufficiently comparable that cosine Top-1 among candidates corresponds to correct geography despite viewpoint and disaster-induced appearance change.
    §III.C.2 cross-view matching; domain gap is acknowledged and SVI is introduced as a bridge.
  • standard math InfoNCE contrastive loss with optional KoLeo regularization and iBOT-style patch terms is an appropriate training objective for cross-view alignment.
    Eqs. (3)–(6); standard self-supervised / retrieval practice.
  • domain assumption Reference RSI (NOAA Harvey) and Google-like SVI are spatially registered and temporally close enough to serve as verification evidence for 2017 SMI.
    Data preparation §IV.A; temporal inconsistency listed as a limitation in §V.
  • ad hoc to paper Four-way split by single/multiple and clear/ambiguous toponyms is a meaningful evaluation taxonomy for disambiguation difficulty.
    Category definitions and sample counts in §IV.A / Fig. 3 drive the fine-grained claims about ambiguous cases.
invented entities (2)
  • DisasterTD two-stage framework no independent evidence
    purpose: Name the pipeline that couples MLLM candidate generation with ViT cross-view verification for disaster toponym disambiguation.
    System-level construct; not a new physical entity. Independent evidence is the empirical benchmark comparison, not an external measurement of a postulated object.
  • Hurricane Harvey cross-view SMI–RSI–SVI benchmark (1,000 curated SMI) no independent evidence
    purpose: Provide aligned multi-view evaluation split by toponym clarity/ambiguity.
    Dataset construction from prior Harvey Twitter SMI plus collected RSI/SVI; usefulness depends on public release, which is not stated.

reviewed 2026-07-30 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DisasterTD: Disaster Toponym Disambiguation Using Multimodal LLMs and Cross-View Geolocalization." pith.science (2026). https://pith.science/paper/DWRYMQD2

@misc{pith2026260724856,
  author       = {Pith},
  title        = {Pith review of: DisasterTD: Disaster Toponym Disambiguation Using Multimodal LLMs and Cross-View Geolocalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DWRYMQD2}},
  note         = {Machine review of arXiv:2607.24856}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Social media imagery (SMI) provides timely and fine-grained ground perspectives that are valuable for situational awareness and emergency response. Unlike satellite or aerial imagery, SMI can capture disaster impacts and ground-level conditions in a timely manner. However, geographic references in SMI are often vague or ambiguous, making accurate geolocalization challenging. To address this issue, we propose DisasterTD, a disaster toponym disambiguation framework that integrates multimodal large language model (MLLMs)-based semantic reasoning with cross-view geolocalization. First, MLLMs extract toponyms and generate candidate geolocations from noisy textual inputs. Then, cross-view matching between SMI, remote sensing imagery (RSI), and optionally street-view imagery (SVI) is used to verify and refine these candidate results. We evaluate DisasterTD on the Hurricane Harvey dataset, where SMI is augmented with collected RSI and SVI to construct a cross-view benchmark for disaster geolocalization. The dataset is divided into four categories based on toponym clarity and ambiguity, allowing a fine-grained performance analysis across scenarios. Results show that DisasterTD consistently outperforms MLLM-only and cross-view-only baselines without disambiguation, achieving geolocalization accuracies of 71.62% within 1000 m, 62.36% within 500 m, 57.99% within 250 m, 52.09% within 100 m, and 47.01% within 50 m, while reducing the mean and median errors to 11.33 km and 0.68 km, respectively. The largest improvements appear in ambiguous toponyms, where semantic reasoning with cross-view evidence reduces candidate dispersion and errors. These findings demonstrate the effectiveness of integrating MLLM-based candidate generation with cross-view verification for fine-grained disaster geolocalization.

Figures

Figures reproduced from arXiv: 2607.24856 by Danfeng Hong, Hao Li, Naixia Mou, Weijia Li, Wenping Yin, Ziqi Liu.

Figure 1
Figure 1. Figure 1: A conceptual diagram of DisasterTD for disaster geolocalization, illus [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of DisasterTD, a novel RSI-assisted framework for disaster toponym disambiguation. The pipeline integrates MLLM-based semantic candidate generation with cross-view visual verification, where SMIs are matched with candidate RSIs and SVI serves as an optional bridge. III. METHODS A. Overall Workflow This study proposes DisasterTD, a novel framework for disaster toponym disambiguation that integrates… view at source ↗
Figure 3
Figure 3. Figure 3: Examples of different SMI types, including images with single or [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Data examples. Spatial distribution of NOAA Hurricane Harvey RSI data (a), and examples of cross-view toponym disambiguation samples, including [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Examples of MLLM-based initial geolocalization results and optimized cross-view toponym disambiguation results. SMIs in (a), (c), (e), and (f) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of disaster geolocalization results within a 1000 m error range. Example SMIs in (a) correspond to the geolocalization points shown in [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Geolocalization performance comparison. (a) Performance of DisasterTD and baseline methods. (b) Improvement of DisasterTD over the baselines. (c) Performance of DisasterTD with different cross-view geolocalization models as the visual matching module. (d) Direct SMI-to-RSI geolocalization performance of these baseline models without candidate generation and disambiguation. Across five spatial thresholds (i… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

69 extracted references · 5 linked inside Pith

  1. [1]

    A flood of disaster response chal- lenges,

    M. Dietze and U. Ozturk, “A flood of disaster response chal- lenges,”Science, vol. 373, no. 6561, pp. 1317–1318, 2021

    2021

  2. [2]

    Exploring disaster impacts on adaptation actions in 549 cities worldwide,

    D. Nohrstedt, J. Hileman, M. Mazzoleni, G. Di Baldassarre, and C. F. Parker, “Exploring disaster impacts on adaptation actions in 549 cities worldwide,”Nature communications, vol. 13, no. 1, p. 3360, 2022

    2022

  3. [3]

    Challenges and opportunities in flood mapping and modeling of next-generation geospatial intelligence: a review,

    M. Shirmohammadi, S. Pirasteh, H. Li, M. Akhavan, V . Isazadeh, J. Ji, C. Chen, and Y . Muhammad, “Challenges and opportunities in flood mapping and modeling of next-generation geospatial intelligence: a review,”Geomatics, Natural Hazards and Risk, vol. 17, no. 1, p. 2660866, 2026

    2026

  4. [4]

    Cross-view geolocalization and disaster map- ping with street-view and vhr satellite imagery: A case study of hurricane ian,

    H. Li, F. Deuser, W. Yin, X. Luo, P. Walther, G. Mai, W. Huang, and M. Werner, “Cross-view geolocalization and disaster map- ping with street-view and vhr satellite imagery: A case study of hurricane ian,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 220, pp. 841–854, 2025. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, 2026 13

    2025

  5. [5]

    Triple-objective cross-view geolocalization of disaster- related vgi: the case of hurricane ian,

    W. Yin, F. Deuser, Z. Liu, J. Wei, X. Luo, M. Werner, H. Li, and Y . Xue, “Triple-objective cross-view geolocalization of disaster- related vgi: the case of hurricane ian,”International Journal of Geographical Information Science, pp. 1–23, 2025

    2025

  6. [6]

    Extraction and analysis of natural disaster-related vgi from social media: review, opportu- nities and challenges,

    Y . Feng, X. Huang, and M. Sester, “Extraction and analysis of natural disaster-related vgi from social media: review, opportu- nities and challenges,”International Journal of Geographical Information Science, vol. 36, no. 7, pp. 1275–1316, 2022

    2022

  7. [7]

    Location reference recognition from texts: A survey and comparison,

    X. Hu, Z. Zhou, H. Li, Y . Hu, F. Gu, J. Kersten, H. Fan, and F. Klan, “Location reference recognition from texts: A survey and comparison,”ACM Computing Surveys, vol. 56, no. 5, pp. 1–37, 2023

    2023

  8. [8]

    Geo-knowledge-guided gpt models improve the extraction of location descriptions from disaster-related social media messages,

    Y . Hu, G. Mai, C. Cundy, K. Choi, N. Lao, W. Liu, G. Lakhan- pal, R. Z. Zhou, and K. Joseph, “Geo-knowledge-guided gpt models improve the extraction of location descriptions from disaster-related social media messages,”International Journal of Geographical Information Science, vol. 37, no. 11, pp. 2289– 2318, 2023

    2023

  9. [9]

    A survey of the research progress in image geo-localization,

    G. Huang, Y . Zhou, X. Hu, L. Zhao, and C. Zhang, “A survey of the research progress in image geo-localization,”J. Geo-Inf. Sci, vol. 25, no. 7, pp. 1336–1362, 2023

    2023

  10. [10]

    A faster and more effective cross-view matching method of uav and satellite im- ages for uav geolocalization,

    J. Zhuang, M. Dai, X. Chen, and E. Zheng, “A faster and more effective cross-view matching method of uav and satellite im- ages for uav geolocalization,”Remote Sensing, vol. 13, no. 19, p. 3979, 2021

    2021

  11. [11]

    Sample4geo: Hard neg- ative sampling for cross-view geo-localisation,

    F. Deuser, K. Habel, and N. Oswald, “Sample4geo: Hard neg- ative sampling for cross-view geo-localisation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 16 847–16 856

    2023

  12. [12]

    A satellite-drone image cross-view geolocalization method based on multi-scale information and dual-channel attention mechanism,

    N. Gong, L. Li, J. Sha, X. Sun, and Q. Huang, “A satellite-drone image cross-view geolocalization method based on multi-scale information and dual-channel attention mechanism,”Remote Sensing, vol. 16, no. 6, p. 941, 2024

    2024

  13. [13]

    Hyperspectral imaging,

    D. Hong, C. Li, N. Yokoya, B. Zhang, X. Jia, A. Plaza, P. Gamba, J. A. Benediktsson, and J. Chanussot, “Hyperspectral imaging,”Nature Reviews Methods Primers, vol. 6, no. 1, p. 19, 2026

    2026

  14. [14]

    Scof: Supervised contrastive orthogonal fusion for robust cross-view geolocaliza- tion,

    C. Fang, J. Gao, P. Han, C. Zhao, and B. Gao, “Scof: Supervised contrastive orthogonal fusion for robust cross-view geolocaliza- tion,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–15, 2025

    2025

  15. [15]

    Unsupervised multi-view uav image geo-localization via iterative rendering,

    H. Li, C. Xu, W. Yang, L. Mi, H. Yu, H. Zhang, and G.- S. Xia, “Unsupervised multi-view uav image geo-localization via iterative rendering,”IEEE Transactions on Geoscience and Remote Sensing, 2025

    2025

  16. [16]

    Dstg: Distillation swin transformer for cross-view geo-localization,

    J. Liang, M. Bao, H. Dong, L. Xie, R. W. Liu, and N. Chen, “Dstg: Distillation swin transformer for cross-view geo-localization,”IEEE Transactions on Geoscience and Re- mote Sensing, 2025

    2025

  17. [17]

    Llm-enhanced disaster geolocalization using implicit geoinformation from multimodal data: A case study of hurricane harvey,

    W. Yin, Y . Xue, Z. Liu, H. Li, and M. Werner, “Llm-enhanced disaster geolocalization using implicit geoinformation from multimodal data: A case study of hurricane harvey,”Interna- tional Journal of Applied Earth Observation and Geoinforma- tion, vol. 137, p. 104423, 2025

    2025

  18. [18]

    Foundation models in remote sensing: Evolving from unimodality to multimodality,

    D. Hong, C. Li, X. Li, G. Camps-Valls, and J. Chanus- sot, “Foundation models in remote sensing: Evolving from unimodality to multimodality,”IEEE Geoscience and Remote Sensing Magazine, 2026

    2026

  19. [19]

    How can voting mechanisms improve the robustness and generaliz- ability of toponym disambiguation?

    X. Hu, Y . Sun, J. Kersten, Z. Zhou, F. Klan, and H. Fan, “How can voting mechanisms improve the robustness and generaliz- ability of toponym disambiguation?”International Journal of Applied Earth Observation and Geoinformation, vol. 117, p. 103191, 2023

    2023

  20. [20]

    Towards generative location awareness for disaster response: A probabilistic cross-view geolocalization approach,

    H. Li, F. Deuser, W. Yin, S. Knoblauch, W. Zhao, F. Biljecki, Y . Xue, and W. Huang, “Towards generative location awareness for disaster response: A probabilistic cross-view geolocalization approach,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 237, pp. 130–145, 2026

    2026

  21. [21]

    Toponym usage in social media in emergencies,

    R. Grace, “Toponym usage in social media in emergencies,” International Journal of Disaster Risk Reduction, vol. 52, p. 101923, 2021

    2021

  22. [22]

    Location extraction from social media: Geop- arsing, location disambiguation, and geotagging,

    S. E. Middleton, G. Kordopatis-Zilos, S. Papadopoulos, and Y . Kompatsiaris, “Location extraction from social media: Geop- arsing, location disambiguation, and geotagging,”ACM Trans- actions on Information Systems (TOIS), vol. 36, no. 4, pp. 1–27, 2018

    2018

  23. [23]

    Deep learning for toponym resolution: Geocoding based on pairs of toponyms,

    J. Fize, L. Moncla, and B. Martins, “Deep learning for toponym resolution: Geocoding based on pairs of toponyms,”ISPRS International Journal of Geo-Information, vol. 10, no. 12, p. 818, 2021

    2021

  24. [24]

    Toponym resolution leveraging lightweight and open-source large lan- guage models and geo-knowledge,

    X. Hu, J. Kersten, F. Klan, and S. M. Farzana, “Toponym resolution leveraging lightweight and open-source large lan- guage models and geo-knowledge,”International Journal of Geographical Information Science, pp. 1–28, 2024

    2024

  25. [25]

    Neurotpr: A neuro-net toponym recognition model for extracting locations from social media messages,

    J. Wang, Y . Hu, and K. Joseph, “Neurotpr: A neuro-net toponym recognition model for extracting locations from social media messages,”Transactions in GIS, vol. 24, no. 3, pp. 719–735, 2020

    2020

  26. [26]

    Integrating google earth imagery with landsat data to improve 30-m res- olution land cover mapping,

    W. Li, R. Dong, H. Fu, J. Wang, L. Yu, and P. Gong, “Integrating google earth imagery with landsat data to improve 30-m res- olution land cover mapping,”Remote Sensing of Environment, vol. 237, p. 111563, 2020

    2020

  27. [27]

    Cross-city matters: A multimodal remote sensing benchmark dataset for cross-city semantic segmentation using high-resolution domain adaptation networks,

    D. Hong, B. Zhang, H. Li, Y . Li, J. Yao, C. Li, M. Werner, J. Chanussot, A. Zipf, and X. X. Zhu, “Cross-city matters: A multimodal remote sensing benchmark dataset for cross-city semantic segmentation using high-resolution domain adaptation networks,”Remote Sensing of Environment, vol. 299, p. 113856, 2023

    2023

  28. [28]

    City-scale landmark identification on mobile devices,

    D. M. Chen, G. Baatz, K. K ¨oser, S. S. Tsai, R. Vedantham, T. Pylv ¨an¨ainen, K. Roimela, X. Chen, J. Bach, M. Pollefeys et al., “City-scale landmark identification on mobile devices,” inCVPR 2011. IEEE, 2011, pp. 737–744

    2011

  29. [29]

    Im2gps: estimating geographic information from a single image,

    J. Hays and A. A. Efros, “Im2gps: estimating geographic information from a single image,” in2008 ieee conference on computer vision and pattern recognition. IEEE, 2008, pp. 1–8

    2008

  30. [30]

    Predicting ground-level scene layout from aerial imagery,

    M. Zhai, Z. Bessinger, S. Workman, and N. Jacobs, “Predicting ground-level scene layout from aerial imagery,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition, 2017, pp. 867–875

    2017

  31. [31]

    Wide-area image geolocalization with aerial reference imagery,

    S. Workman, R. Souvenir, and N. Jacobs, “Wide-area image geolocalization with aerial reference imagery,” inProceedings of the IEEE International Conference on Computer Vision, 2015, pp. 3961–3969

    2015

  32. [32]

    Localizing and orienting street views using overhead imagery,

    N. N. V o and J. Hays, “Localizing and orienting street views using overhead imagery,” inEuropean conference on computer vision. Springer, 2016, pp. 494–509

    2016

  33. [33]

    Optimal feature transport for cross-view image geo-localization,

    Y . Shi, X. Yu, L. Liu, T. Zhang, and H. Li, “Optimal feature transport for cross-view image geo-localization,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 07, 2020, pp. 11 990–11 997

    2020

  34. [34]

    University-1652: A multi- view multi-source benchmark for drone-based geo-localization,

    Z. Zheng, Y . Wei, and Y . Yang, “University-1652: A multi- view multi-source benchmark for drone-based geo-localization,” inProceedings of the 28th ACM international conference on Multimedia, 2020, pp. 1395–1403

    2020

  35. [35]

    Im2city: image geo-localization via multi-modal learning,

    M. Wu and Q. Huang, “Im2city: image geo-localization via multi-modal learning,” inProceedings of the 5th ACM SIGSPA- TIAL International Workshop on AI for Geographic Knowledge Discovery, 2022, pp. 50–61

    2022

  36. [36]

    Spatial-aware feature aggre- gation for image based cross-view geo-localization,

    Y . Shi, L. Liu, X. Yu, and H. Li, “Spatial-aware feature aggre- gation for image based cross-view geo-localization,”Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  37. [37]

    Deep learning and stereo vision based detection of post-earthquake fire geolocation for smart cities within the scope of disaster management: ˙Istanbul case,

    T. Kustu and A. Taskin, “Deep learning and stereo vision based detection of post-earthquake fire geolocation for smart cities within the scope of disaster management: ˙Istanbul case,” International journal of disaster risk reduction, vol. 96, p. 103906, 2023

    2023

  38. [38]

    Toponym resolution in text: annotation, eval- uation and applications of spatial grounding,

    J. L. Leidner, “Toponym resolution in text: annotation, eval- uation and applications of spatial grounding,” inACM SIGIR IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, 2026 14 Forum, vol. 41, no. 2. ACM New York, NY , USA, 2007, pp. 124–126

    2026

  39. [39]

    Text-based geolocation prediction of social media users with neural networks,

    I. Lourentzou, A. Morales, and C. Zhai, “Text-based geolocation prediction of social media users with neural networks,” in2017 IEEE International Conference on Big Data (Big Data). IEEE, 2017, pp. 696–705

    2017

  40. [40]

    Web-a-where: geotagging web content,

    E. Amitay, N. Har’El, R. Sivan, and A. Soffer, “Web-a-where: geotagging web content,” inProceedings of the 27th annual international ACM SIGIR conference on Research and develop- ment in information retrieval, 2004, pp. 273–280

    2004

  41. [41]

    Simple supervised document geolo- cation with geodesic grids,

    B. Wing and J. Baldridge, “Simple supervised document geolo- cation with geodesic grids,” inProceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies, 2011, pp. 955–964

    2011

  42. [42]

    A latent variable model for geographic lexical variation,

    J. Eisenstein, B. O’Connor, N. A. Smith, and E. Xing, “A latent variable model for geographic lexical variation,” inProceed- ings of the 2010 conference on empirical methods in natural language processing, 2010, pp. 1277–1287

    2010

  43. [43]

    A coherent unsupervised model for toponym resolution,

    E. Kamalloo and D. Rafiei, “A coherent unsupervised model for toponym resolution,” inProceedings of the 2018 world wide web conference, 2018, pp. 1287–1296

    2018

  44. [44]

    Street- level geolocalization using multimodal large language mod- els and retrieval-augmented generation,

    Y . S. Bicakci, J. Shingleton, and A. Basiri, “Street- level geolocalization using multimodal large language mod- els and retrieval-augmented generation,”arXiv preprint arXiv:2509.01341, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Distinctive image features from scale-invariant keypoints,

    D. G. Lowe, “Distinctive image features from scale-invariant keypoints,”International journal of computer vision, vol. 60, no. 2, pp. 91–110, 2004

    2004

  46. [46]

    Surf: Speeded up robust features,

    H. Bay, T. Tuytelaars, and L. Van Gool, “Surf: Speeded up robust features,” inEuropean conference on computer vision. Springer, 2006, pp. 404–417

    2006

  47. [47]

    Planet-photo ge- olocation with convolutional neural networks,

    T. Weyand, I. Kostrikov, and J. Philbin, “Planet-photo ge- olocation with convolutional neural networks,” inEuropean conference on computer vision. Springer, 2016, pp. 37–55

    2016

  48. [48]

    Geolocsft: Efficient visual geolocation via supervised fine-tuning of multimodal foundation models,

    Q. Yi and L. Shan, “Geolocsft: Efficient visual geolocation via supervised fine-tuning of multimodal foundation models,”arXiv preprint arXiv:2506.01277, 2025

    Pith/arXiv arXiv 2025

  49. [49]

    Com- ing down to earth: Satellite-to-street view synthesis for geo- localization,

    A. Toker, Q. Zhou, M. Maximov, and L. Leal-Taix ´e, “Com- ing down to earth: Satellite-to-street view synthesis for geo- localization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 6488– 6497

    2021

  50. [50]

    Mul- timodal geolocation estimation in news documents,

    G. Tahmasebzadeh, E. M ¨uller-Budack, and R. Ewerth, “Mul- timodal geolocation estimation in news documents,” inEvent Analytics across Languages and Communities. Springer Nature Switzerland Cham, 2024, pp. 17–45

    2024

  51. [51]

    Where am i? cross-view geo-localization with natural language descriptions,

    J. Ye, H. Lin, L. Ou, D. Chen, Z. Wang, Q. Zhu, C. He, and W. Li, “Where am i? cross-view geo-localization with natural language descriptions,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 5890– 5900

    2025

  52. [52]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  53. [53]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,” inInternational conference on machine learning. PMLR, 2023, pp. 19 730–19 742

    2023

  54. [54]

    A conceptual density-based approach for the disambiguation of toponyms,

    D. Buscaldi and P. Rosso, “A conceptual density-based approach for the disambiguation of toponyms,”International Journal of Geographical Information Science, vol. 22, no. 3, pp. 301–313, 2008

    2008

  55. [55]

    Toponym disambiguation using events

    K. Roberts, C. A. Bejan, and S. M. Harabagiu, “Toponym disambiguation using events.” inFLAIRS, 2010

    2010

  56. [56]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” inProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), 2019, pp. 4171–4186

    2019

  57. [57]

    Improving language understanding by generative pre-training,

    A. Radford, K. Narasimhan, T. Salimans, I. Sutskeveret al., “Improving language understanding by generative pre-training,” 2018

    2018

  58. [58]

    Map-based vs. knowledge-based toponym disambiguation,

    D. Buscaldi and P. Rosso, “Map-based vs. knowledge-based toponym disambiguation,” inProceedings of the 5th workshop on geographic information retrieval, 2008, pp. 19–22

    2008

  59. [59]

    Which melbourne? augmenting geocoding with maps

    M. Gritta, M. T. Pilehvar, and N. Collier, “Which melbourne? augmenting geocoding with maps.” Association for Computa- tional Linguistics, 2018

    2018

  60. [60]

    Gazetteer- independent toponym resolution using geographic word pro- files,

    G. DeLozier, J. Baldridge, and L. London, “Gazetteer- independent toponym resolution using geographic word pro- files,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 29, no. 1, 2015

    2015

  61. [61]

    Cross-view object geo- localization in a local region with satellite imagery,

    Y . Sun, Y . Ye, J. Kang, R. Fernandez-Beltran, S. Feng, X. Li, C. Luo, P. Zhang, and A. Plaza, “Cross-view object geo- localization in a local region with satellite imagery,”IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–16, 2023

    2023

  62. [62]

    Spectralgpt: Spectral remote sensing foundation model,

    D. Hong, B. Zhang, X. Li, Y . Li, C. Li, J. Yao, N. Yokoya, H. Li, P. Ghamisi, X. Jiaet al., “Spectralgpt: Spectral remote sensing foundation model,”IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 8, pp. 5227–5244, 2024

    2024

  63. [63]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Trans- formers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  64. [64]

    Dinov2: Learning robust visual features without super- vision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al., “Dinov2: Learning robust visual features without super- vision,”arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  65. [65]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bo- janowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inProceedings of the IEEE/CVF interna- tional conference on computer vision, 2021, pp. 9650–9660

    2021

  66. [66]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, 2022, pp. 11 976–11 986

    2022

  67. [67]

    Simple, effective and general: A new backbone for cross-view image geo- localization,

    Y . Zhu, H. Yang, Y . Lu, and Q. Huang, “Simple, effective and general: A new backbone for cross-view image geo- localization,”arXiv preprint arXiv:2302.01572, 2023

    Pith/arXiv arXiv 2023

  68. [68]

    Transgeo: Transformer is all you need for cross-view image geo-localization,

    S. Zhu, M. Shah, and C. Chen, “Transgeo: Transformer is all you need for cross-view image geo-localization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1162–1171

    2022

  69. [69]

    Hurricane harvey twitter dataset,

    M. E. Phillips, “Hurricane harvey twitter dataset,” 2024, https://digital.library.unt.edu/ark:/67531/metadc993940

    2024

This paper was first reviewed by grok-4.5 on July 30, 2026.