Pith. sign in

REVIEW 5 major objections 6 minor 74 references

DIGMAPPER: A Modular System for Automated Geologic Map Digitization

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DIGMAPPER is a modular system that turns scanned geologic maps into georeferenced vector data in under 25 minutes per map.

desk verdict The speed claim is contradicted by the paper's own Table 7 (polygon extraction alone averages 93 minutes), but the underlying system is a real, deployed engineering contribution worth a serious referee's time. read the letter →

arxiv 2506.16006 v1 pith:CL3F7YD4 submitted 2025-06-19 cs.CV cs.AI

Weiwei Duan , Michael P. Gerlek , Steven N. Minton , Craig A. Knoblock , Fandel Lin , Theresa Chen , Leeje Jang , Sofia Kirsanova
show 3 more authors
Zekun Li Yijun Lin Yao-Yi Chiang
This is my paper · ORCID
classification cs.CVcs.AI
keywords geologicmapdigitizationlayoutanalysisin-contextlearningpolygonextractionlinepointsymboldetectiongeoreferencingsyntheticdata
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

DIGMAPPER is a modular pipeline for turning scanned historical geologic maps into analysis-ready geospatial data, and the paper aims to show that it does so accurately enough to support national-scale critical mineral assessments. The system first segments each map into content and legend regions, then uses the legend to drive extraction of polygon, line, and point features, while a separate path georeferences the map to real-world coordinates. Evaluations on over 100 annotated maps from the DARPA-USGS dataset report high accuracy for feature extraction and reliable georeferencing, with total processing time under 25 minutes per map compared with hours of manual digitization. The work matters because assessments of lithium, cobalt, nickel, and rare-earth resources depend on digitized geologic units and structures that currently sit in undigitized historical maps.

What carries the argument

The load-bearing mechanism is the legend-mediated extraction loop: layout analysis (LayoutLMv3 segmentation plus GPT-4o in-context learning) turns the map legend into structured symbol-description pairs, and the feature-extraction modules consume those pairs as semantic anchors, interpreting polygon colors, markings, and text labels relative to the map's own legend rather than a fixed palette. The other carrying components are TOPAZ (a transformer/U-Net polygon extractor that fuses color, marking, and text cues), LDTR (a line-graph detector whose N-hop connectivity head lets each node aggregate spatial context from nearby and distant points), YOLOv8 with synthetic point-symbol data, and a two-stage georeferencing pipeline that first extracts text-based ground control points and then refines them with LightGlue keypoint matching and RANSAC homography estimation. Dockerized modules coordinated by a Luigi directed-acyclic-graph orchestrator make the chain parallel and replaceable, which is what allows the reported under-25-minute per-map runtime.

What would settle it

Run DIGMAPPER on an uncurated random sample of historical USGS geologic maps, including faded scans, irregular non-rectangular legends, and maps without clean point-symbol annotations, and compare its polygon, line, and point outputs and georeferencing against expert manual digitization using the paper's own IoU/F1 and RMSE metrics. If the median scores on the uncurated sample fall to the levels reported for the fair category while such maps dominate the sample, the claim that the system significantly accelerates the creation of analysis-ready geospatial datasets would not generalize beyond the curated evaluation.

Watch

Extended reading notes

Core claim

The central claim is that a modular, workflow-orchestrated system can automate the full map-digitization chain, from a raw scan to a georeferenced map with attributed polygon, line, and point layers. The layout module fine-tunes LayoutLMv3, a document-layout transformer, to separate map content from the legend and uses GPT-4o, a multimodal vision-language model, with in-context learning to pair legend symbols with their textual descriptions. Those legend pairs anchor the feature extractors: TOPAZ, a transformer/U-Net model, jointly predicts polygon shapes, their colors in the map image, and their association with legend items; LDTR, a line-graph detector built on Relationformer with an N-hop connectivity head, extracts fault lines as vector graphs; and a YOLOv8 object detector trained on synthetic plus human-labeled data finds point symbols such as bedding and foliation markers. Georeferencing combines text-based extraction of corner coordinates and place names with visual keypoint matching against georeferenced USGS topographic maps. The authors state that the system is deployed at USGS and has processed hundreds of maps for critical-mineral assessments.

Load-bearing premise

The headline accuracy figures come from curated subsets of the DARPA-USGS dataset, with maps showing significant annotation errors excluded and point symbols evaluated only on maps with reliable human labels, so the operational claim assumes those curated maps represent the full population of historical USGS scans.

Editorial extensions

If this is right

  • If the claimed accuracy holds, historical USGS map sheets can be converted into georeferenced polygon, line, and point layers at national scale without per-map manual digitization.
  • Critical-mineral assessments that currently wait on expert map reading could run directly on the resulting vector layers, with legend descriptions supplying the rock-unit, fault-type, and bedding-orientation labels that prospectivity models need.
  • Because each module is an independent Docker image, replacing one component, such as a better text spotter or a stronger line refiner, improves the whole pipeline without retraining the other modules.
  • The synthetic-data recipe demonstrated for point symbols suggests an analogous route for underrepresented polygon and line styles, reducing the need for expert annotations in new map collections.
  • The excellent/good/fair quality categories give operators a way to predict which incoming maps will need manual review before spending GPU time.

Reading between the lines

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

  • A natural extension the paper does not develop is to use the quality categories as an automatic triage gate: send excellent maps straight to vector output, route good maps to human curation, and mark fair maps for rescanning or manual digitization, turning the reported throughput into an operational policy.
  • The legend-item and description pairs produced by in-context learning could serve as weak supervision for pretraining TOPAZ and the line and point models on the large unannotated National Geologic Map Database collection, compounding the synthetic-data gains already shown.
  • If the visual georeferencing refinement generalizes beyond the nine maps reported, the text-based and visual methods could be fused by confidence score rather than run sequentially, which would reduce the share of maps that end up in the fair georeferencing category.
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

5 major / 6 minor

Summary. The paper presents DIGMAPPER, a modular, Dockerized pipeline for digitizing historical geologic maps. The system performs layout analysis (using LayoutLMv3 and GPT-4o-based in-context learning), polygon extraction (TOPAZ), line extraction (LDTR), point extraction (YOLO-v8 with synthetic data), and text- and visual-based georeferencing. The authors report evaluations on the DARPA-USGS dataset and claim that DIGMAPPER achieves high accuracy across polygon, line, and point feature extraction, reliable georeferencing, and processes a map in under 25 minutes. The system is asserted to have been deployed at USGS for critical-mineral assessments.

Significance. The paper describes a substantial practical system: a fully containerized, workflow-orchestrated architecture integrating multiple specialist modules, with explicit attention to synthetic-data generation and in-context learning to mitigate scarce annotations. If the claims were fully supported, the system would be a valuable contribution to historical-map digitization and could accelerate the creation of analysis-ready geospatial datasets. The paper also provides quantitative evaluations across all modules and honest reporting of per-module performance in tabular form, which is a strength. However, several load-bearing claims, especially the 25-minute speed guarantee and the unqualified 'high accuracy' statement, are contradicted or undercut by the paper's own tables and evaluation design; these issues must be resolved before the contribution can be assessed as stated.

major comments (5)
  1. [§3, Table 7, Abstract, §5] The claim that 'DIGMAPPER processes a map in under 25 minutes' is contradicted by the paper's own timing data. Table 7 reports Polygon Extraction at 5,596.54 seconds (~93.3 minutes), and the text states that 'modules execute in parallel, and the total runtime is bounded by the slowest module rather than the sum of all.' Since Polygon Extraction is the slowest module, the pipeline's lower bound is approximately 93 minutes per map on the stated AWS g4dn.8xlarge hardware, not under 25 minutes. The abstract and §5 restate the speed claim without qualification. Please report actual end-to-end wall-clock timings, explicitly state whether polygon extraction is optional or requires different hardware, or remove the speed claim.
  2. [Abstract; §2.2.2 Table 2; §2.3.2 Table 3; §2.4.2 Table 4] The abstract's claim of 'high accuracy across polygon, line, and point feature extraction' is unqualified. In the paper's own fair category, polygon extraction achieves IoU 0.17 and F1 0.28 (Table 2), fault-line extraction achieves correctness 0.40 and completeness 0.24 and thrust-fault extraction achieves correctness 0.02 and completeness 0.01 (Table 3), and point extraction achieves F1 0.05 (Table 4). These are not high-accuracy results. The abstract and §5 should explicitly state that high accuracy holds for the excellent and good map categories and that fair-category maps require substantial human curation.
  3. [§2.4.2, Table 4] The point-extraction evaluation assigns maps to the Excellent, Good, or Fair performance category based on the model's own F1 scores ('We categorize the 10 maps into three performance levels: Excellent, Good, or Fair, based on their F1 scores'). This makes the category-level reporting circular: 'Excellent' is defined by high model performance, so the statement that the model performs well on excellent maps is tautological. Please define performance categories a priori from map characteristics (e.g., scan resolution, noise, layout regularity) or report results without categorical labels.
  4. [§2.2.2, Table 2] Table 2 reports an Excellent-category candidate ratio of 8.54, yet the text claims that for excellent and good maps 'instance counts closely match the reference.' The candidate ratio is defined as the number of extracted polygons over ground truth, with values closer to 1 indicating better performance; a value of 8.54 indicates severe over-segmentation (roughly 8.5 extracted polygons per ground-truth polygon). This contradicts the text's claim and is load-bearing for the polygon-extraction accuracy statement. Please address the over-segmentation, for example by reporting the distribution of candidate ratios or by revising the accompanying narrative.
  5. [§2.5.3, Table 6; §2.2.2; §2.3.2; §2.4.2] The abstract's 'reliable georeferencing performance' and 'evaluations on over 100 annotated maps' are not supported by the reported evaluation. Text-based georeferencing has a median RMSE_geo of 0.104 km but a mean of 14.62 km; 24 of 63 maps have RMSE_geo >= 1 km, and only 34 of 63 (54%) achieve pixel-normalized RMSE below 0.1. In addition, feature-extraction evaluations use small, curated subsets: 11 maps for polygons, 10 for lines, and 10 for points, with maps excluded for annotation errors or missing point annotations. The 'over 100 annotated maps' refers to the dataset size, not the evaluated subsets. Please report per-module evaluation sizes and distributional statistics, and qualify 'reliable' to reflect that a substantial fraction of maps have large georeferencing errors.
minor comments (6)
  1. [Table 4] The header contains a typo: 'Precison' should be 'Precision'.
  2. [§2.4.2, Table 5] The text states that inclined bedding has 10,635 human-labeled training samples, but Table 5 reports #Train as 10,625 for inclined bedding; these numbers should be reconciled.
  3. [§2.4.2] The sentence 'Overall, our model achieves an F1 score of 0.82 across 10 maps' is ambiguous because the 0.82 in Table 5 is instance-weighted, whereas Table 4 reports map-level F1 values of 0.89, 0.56, and 0.05 for the three categories. Please clarify that the 0.82 is not a map-level average.
  4. [§5] The conclusion states that DIGMAPPER 'processes a single map under 25 minutes' without the parallel-execution qualification given in §3; the claim should be made consistent with Table 7 or removed.
  5. [References] Reference [51] contains the placeholder 'Accessed: YYYY-MM-DD', and the paper still contains ACM template placeholders such as the conference title, ISBN, and DOI; these should be completed or removed before submission.
  6. [§2.3.2, Figure 8] The line-refinement result shown in Figure 8d is described only as 'preliminary' and is not quantitatively evaluated; either provide metrics or clearly mark this as future work and remove it from the main evaluation narrative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DIGMAPPER's accuracy claims are measured against external DARPA-USGS ground truth, and no prediction reduces to a fitted input or self-citation chain.

full rationale

The paper's central performance claims are evaluated against external DARPA-USGS ground-truth annotations, not against the system's own outputs or fitted parameters. Each module's metrics (IoU, F1, precision/recall, RMSE) are computed on held-out map subsets with fixed evaluation conventions, such as IoU threshold 0.5, a spatial buffer for points, and RMSE thresholds for georeferencing; none of these are fitted to the test set and then renamed as predictions. The point-extraction evaluation does assign 'Excellent, Good, or Fair' labels based on the model's F1 scores, making the category-level contrast partly tautological, but this is a presentational stratification and does not support any load-bearing claim; the overall F1 of 0.82 is an external measurement. The use of mapKurator, GeoLM, and Palette is component-level reuse of prior author tools, not a self-citation chain that forces the accuracy results, and no uniqueness theorem from the authors' prior work is invoked to forbid alternatives. The apparent contradiction between the advertised 'under 25 minutes' runtime and Table 7's slowest-module time of 5,596.54 seconds is a serious internal consistency or correctness concern, but it is not circularity: a false or unsupported speed claim does not make the evaluation equal to its inputs. Overall, the derivation chain is self-contained with respect to circularity, so a score of 0 is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The system relies on several domain-specific assumptions about map structure (four corners, USGS basemap, legend layout) and on empirically tuned thresholds (10-km buffer, 0.5 confidence, 15 ICL examples). The paper does not ship code or data, so these parameters are not independently reproducible.

free parameters (6)
  • Candidate topographic map buffer distance = 10 km
    Used in visual georeferencing to select candidate maps; 'determined empirically' (Section 2.5.2), a fitted threshold.
  • Visual match confidence threshold = 0.5
    Used to filter low-confidence matches and RANSAC inliers (Section 2.5.2); chosen without independent validation.
  • Number of ICL examples = 15
    Found through experiments to yield the most accurate legend extraction (Section 2.1.2).
  • Point symbol matching buffer = 2 x 10^-4 x map diagonal length
    Defines correct detection in point evaluation; approximately symbol size, an evaluation parameter that influences precision and recall.
  • IoU-count weighting for polygon metrics = weighted metrics as reported
    Polygon evaluation weights metrics by IoU count; removing weighting increases IoU and F1 by at least 0.30 for fair maps (Section 2.2.2), so the weighting materially changes reported performance.
  • Corner buffer size = 1000 x 1000 pixels
    Region around detected corners searched for geocoordinate labels (Section 2.5.1); a study-specific choice.
assumptions (5)
  • domain assumption Each map contains four corners with exactly two distinct latitudes and two distinct longitudes
    Assumed in text-based georeferencing heuristics (Section 2.5.1); many maps, especially non-rectangular or partial sheets, violate this.
  • domain assumption Most geologic maps use USGS topographic maps as basemaps, so place names can be matched against a ~250,000-map database
    Stated in Section 2.5.1; if a map uses a non-USGS basemap or a region not covered, retrieval fails.
  • domain assumption Legend items and descriptions follow a left-right structure with a moderate number of entries
    Section 2.1.2 reports degraded performance for top-bottom layouts and legends with over 100 entries; the method assumes the common layout.
  • standard math RANSAC homography estimation with at least four inlier matches is appropriate for aligning map scans
    Used in visual georeferencing (Section 2.5.2); assumes planar map geometry and sufficient textured overlap.
  • domain assumption Colors in legend items are informative for associating polygons even when scans introduce color shifts
    TOPAZ uses color spectra from legends and content; Section 2.2.2 reports that significant color shifts cause fragmentation, so this assumption is load-bearing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DIGMAPPER: A Modular System for Automated Geologic Map Digitization." pith.science (2026). https://pith.science/paper/CL3F7YD4

@misc{pith2026250616006,
  author       = {Pith},
  title        = {Pith review of: DIGMAPPER: A Modular System for Automated Geologic Map Digitization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CL3F7YD4}},
  note         = {Machine review of arXiv:2506.16006}
}
read the original abstract

Historical geologic maps contain rich geospatial information, such as rock units, faults, folds, and bedding planes, that is critical for assessing mineral resources essential to renewable energy, electric vehicles, and national security. However, digitizing maps remains a labor-intensive and time-consuming task. We present DIGMAPPER, a modular, scalable system developed in collaboration with the United States Geological Survey (USGS) to automate the digitization of geologic maps. DIGMAPPER features a fully dockerized, workflow-orchestrated architecture that integrates state-of-the-art deep learning models for map layout analysis, feature extraction, and georeferencing. To overcome challenges such as limited training data and complex visual content, our system employs innovative techniques, including in-context learning with large language models, synthetic data generation, and transformer-based models. Evaluations on over 100 annotated maps from the DARPA-USGS dataset demonstrate high accuracy across polygon, line, and point feature extraction, and reliable georeferencing performance. Deployed at USGS, DIGMAPPER significantly accelerates the creation of analysis-ready geospatial datasets, supporting national-scale critical mineral assessments and broader geoscientific applications.

Figures

Figures reproduced from arXiv: 2506.16006 by the authors.

Figure 1
Figure 1. Examples of complex visual context in geologic [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. DIGMAPPER architecture for automated map digitization. The system takes a scanned map as input (blue) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Workflow of our ICL approach for extracting legend [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: An example of extracting polygon features from [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: LDTR Overview. LDTR inputs images and outputs [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Thrust fault lines follow the style but are distinguished by [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 9
Figure 9. Figure 9: Twelve predefined point symbols To address limited training data, we propose a synthetic data gen￾eration approach consisting of three steps: (1) Given the predefined set of point symbols, our approach uses a GPT-based method to automatically identify table rows in [58…
Figure 7
Figure 7. Figure 7: Examples of fault line extraction results across [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Example of line refinement using a language model. [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 10
Figure 10. Figure 10: DIGMAPPER System Architecture consists of two [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 66 canonical work pages

  1. [1]

    Abdollahi, B

    A. Abdollahi, B. Pradhan, and A. Alamri. Sc-roaddeepnet: A new shape and connectivity-preserving road extraction deep learning-based network from re- mote sensing data. IEEE Transactions on Geoscience and Remote Sensing , 60:1–15, 2022

    work page 2022

  2. [2]

    M. G. Arteaga. Historical map polygon and feature extractor. In Proceedings of the 1st ACM SIGSPATIAL International Workshop on MapInteraction , pages 66–71, 2013

    work page 2013

  3. [3]

    Arzoumanidis, J

    L. Arzoumanidis, J. Knechtel, J.-H. Haunert, and Y. Dehbi. Semantic segmenta- tion of historical maps using self-constructing graph convolutional networks. Cartography and Geographic Information Science , pages 1–11, 2025

    work page 2025

  4. [4]

    Budig and T

    B. Budig and T. C. van Dijk. Active learning for classifying template matches in historical maps. In Discovery Science: 18th International Conference, DS 2015, Banff, AB, Canada, October 4-6, 2015. Proceedings 18 , pages 33–47. Springer, 2015

    work page 2015

  5. [5]

    C.-C. Chen, C. A. Knoblock, and C. Shahabi. Automatically and accurately conflating raster maps with orthoimagery. GeoInformatica, 12:377–410, 2008

    work page 2008

  6. [6]

    Chiang, W

    Y.-Y. Chiang, W. Duan, S. Leyk, J. H. Uhl, and C. A. Knoblock.Using Historical Maps in Scientific Studies: Applications, Challenges, and Best Practices . Springer Publishing Company, Incorporated, 1st edition, 2019

    work page 2019

  7. [7]

    Chiang, C

    Y.-Y. Chiang, C. A. Knoblock, C. Shahabi, and C.-C. Chen. Automatic and accurate extraction of road intersections from raster maps. GeoInformatica, 13:121–157, 2009

    work page 2009

  8. [8]

    Chiang, S

    Y.-Y. Chiang, S. Leyk, and C. A. Knoblock. A survey of digital map processing techniques. ACM Comput. Surv., 47(1), May 2014

    work page 2014

Show all 74 references
  1. [9]

    B. Dadi, M. Ouchchen, F. Faik, S. Boutaleb, D. El Azzab, Y. Mamouch, L. Achk- ouch, A. Bajadi, E. H. Abia, and B. Sadeghi. Hydrothermal fluid pathways and mineralization potential in the high atlas massif (morocco) using fuzzy logic and multifractal modeling. Ore Geology Revie...

    2024

  2. [10]

    L. Dai, G. Zhang, and R. Zhang. Radanet: Road augmented deformable attention network for road extraction from complex high-resolution remote-sensing images. IEEE Trans. Geosci. Remote Sens. , 2023

    2023

  3. [11]

    Q. Dong, L. Li, D. Dai, C. Zheng, J. Ma, R. Li, H. Xia, J. Xu, Z. Wu, T. Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022

    2022 arXiv

  4. [12]

    W. Duan, Y. Chiang, C. A. Knoblock, S. Leyk, and J. Uhl. Automatic generation of precisely delineated geographic features from georeferenced historical maps using deep learning. In Proceedings of the AutoCarto , 2018

    2018

  5. [13]

    Duan, Y.-Y

    W. Duan, Y.-Y. Chiang, C. A. Knoblock, V. Jain, D. Feldman, J. H. Uhl, and S. Leyk. Automatic alignment of geographic features in contemporary vector data and historical maps. In Proceedings of the 1st workshop on artificial intelligence and deep learning for geographic knowle...

    2017

  6. [14]

    Duan, Y.-Y

    W. Duan, Y.-Y. Chiang, S. Leyk, J. Uhl, and C. Knoblock. A label correction algorithm using prior information for automatic and accurate geospatial object recognition. In IEEE Big Data, pages 1604–1610. IEEE, 2021

    2021

  7. [15]

    Duan, Y.-Y

    W. Duan, Y.-Y. Chiang, S. Leyk, J. H. Uhl, and C. A. Knoblock. A label correction algorithm using prior information for automatic and accurate geospatial object recognition. In 2021 IEEE International Conference on Big Data (Big Data) , pages 1604–1610. IEEE, 2021

    2021

  8. [16]

    Dumakor-Dupey and S

    N. Dumakor-Dupey and S. Arya. Machine learning—a review of applications in mineral resource estimation. Energies, 14(14):4079, 2021

    2021

  9. [17]

    Georeferencing a raster automatically

    Esri. Georeferencing a raster automatically. Accessed: 2025-05-07

    2025

  10. [18]

    M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381–395, 1981

    1981

  11. [19]

    Garcia-Molsosa, H

    A. Garcia-Molsosa, H. A. Orengo, D. Lawrence, G. Philip, K. Hopper, and C. A. Petrie. Potential of deep learning segmentation for the extraction of archaeologi- cal features from historical map series. Archaeological Prospection, 28(2):187–199, 2021

    2021

  12. [20]

    Gede and L

    M. Gede and L. Varga. Automatic georeferencing of topographic map sheets using opencv and tesseract. In Proceedings of the ICA , volume 4, pages 1–4. Copernicus GmbH, 2021

    2021

  13. [21]

    Girshick

    R. Girshick. Fast r-cnn, 2015

    2015

  14. [22]

    Gritta, M

    M. Gritta, M. T. Pilehvar, and N. Collier. A pragmatic guide to geoparsing evaluation: Toponyms, named entity recognition and pragmatics. Language resources and evaluation, 54:683–712, 2020

    2020

  15. [23]

    L. Han, L. Hou, X. Zheng, Z. Ding, H. Yang, and K. Zheng. Segmentation is not the end of road extraction: An all-visible denoising autoencoder for connected and smooth road reconstruction. IEEE Transactions on Geoscience and Remote Sensing, 61:1–18, 2023

    2023

  16. [24]

    Y. He, R. Garg, and A. R. Chowdhury. Td-road: Top-down road network extraction with holistic graph construction. In ECCV 2022, pages 562–577. Springer, 2022

    2022

  17. [25]

    Heipke, H

    C. Heipke, H. Mayer, C. Wiedemann, and O. Jamet. Evaluation of automatic road extraction. IAPRS, 32(3 SECT 4W2):151–160, 1997

    1997

  18. [26]

    Heitzler and L

    M. Heitzler and L. Hurni. Cartographic reconstruction of building footprints from historical maps: A study on the swiss siegfried map. Transactions in GIS, 24(2):442–461, 2020

    2020

  19. [27]

    Hetang, H

    C. Hetang, H. Xue, C. Le, T. Yue, W. Wang, and Y. He. Segment anything model for road network graph extraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2556–2566, 2024

    2024

  20. [28]

    N. R. Howe, J. Weinman, J. Gouwar, and A. Shamji. Deformable part models for automatically georeferencing historical map images. In Proceedings of the 27th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, pages 540–543, 2019

    2019

  21. [29]

    Hronsky and O

    J. Hronsky and O. Kreuzer. Applying spatial prospectivity mapping to exploration targeting: Fundamental practical issues and suggested solutions for the future. Ore Geology Reviews, 107:647–653, 2019

    2019

  22. [30]

    Huang, Q

    W. Huang, Q. Sun, A. Yu, W. Guo, Q. Xu, B. Wen, and L. Xu. Leveraging deep con- volutional neural network for point symbol recognition in scanned topographic maps. ISPRS International Journal of Geo-Information , 12(3):128, 2023

    2023

  23. [31]

    Huang, T

    Y. Huang, T. Lv, L. Cui, Y. Lu, and F. Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. InProceedings of the 30th ACM international conference on multimedia, pages 4083–4091, 2022

    2022

  24. [32]

    Opencv: Open source computer vision library

    Itseez. Opencv: Open source computer vision library. In OpenCV Library, 2015

    2015

  25. [33]

    Jocher, A

    G. Jocher, A. Chaurasia, Laughing-Q, J. Fang, and A. V. Yolov8 - ultralytics. https://github.com/ultralytics/ultralytics, 2023. Accessed: 2025-05-11

    2023

  26. [34]

    J. Kim, Z. Li, Y. Lin, M. Namgung, L. Jang, and Y.-Y. Chiang. The mapkurator system: A complete pipeline for extracting and linking text from historical maps. In Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems, SIGSPATIAL ’23,...

    2023

  27. [35]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. White- head, A. C. Berg, W.-Y. Lo, et al. Segment anything.arXiv preprint arXiv:2304.02643, 2023

    2023 arXiv

  28. [36]

    Lewis, Y

    M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V. Stoyanov, and L. Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019

    1910 arXiv

  29. [37]

    T. Li, R. Li, S. Ye, Z. Zhang, Z. Yin, S. Li, and Z. Pan. Crtgan: Controllable road network graphs generation via transformer based gan. In 2024 International Joint Conference on Neural Networks (IJCNN) , pages 1–8. IEEE, 2024

    2024

  30. [38]

    Li, Y.-Y

    Z. Li, Y.-Y. Chiang, S. Tavakkol, B. Shbita, J. H. Uhl, S. Leyk, and C. A. Knoblock. An automatic approach for generating rich, linked geo-metadata from historical map images. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining ,...

    2020

  31. [39]

    Z. Li, J. D. Wegner, and A. Lucchi. Topological map extraction from overhead images. In ICCV, pages 1715–1724, 2019

    2019

  32. [40]

    Z. Li, W. Zhou, Y.-Y. Chiang, and M. Chen. Geolm: Empowering language models for geospatially grounded language understanding. arXiv preprint arXiv:2310.14478, 2023

    2023 arXiv

  33. [41]

    F. Lin, C. A. Knoblock, B. Shbita, B. Vu, Z. Li, and Y.-Y. Chiang. Exploiting polygon metadata to understand raster maps-accurate polygonal feature extraction. In Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems, pages 1–12, 2023

    2023

  34. [42]

    S. Lin, X. Yao, X. Liu, S. Wang, H.-M. Chen, L. Ding, J. Zhang, G. Chen, and Q. Mei. Ms-agan: Road extraction via multi-scale information fusion and asymmetric generative adversarial networks from high-resolution remote sensing images under complex backgrounds. Remote Sensing,...

    2023

  35. [43]

    Lin and Y.-Y

    Y. Lin and Y.-Y. Chiang. Hyper-local deformable transformers for text spotting on historical maps. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 5387–5397, 2024

    2024

  36. [44]

    Lindenberger, P.-E

    P. Lindenberger, P.-E. Sarlin, and M. Pollefeys. Lightglue: Local feature matching at light speed. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023

    2023

  37. [45]

    Luft and J

    J. Luft and J. Schiewe. Automatic content-based georeferencing of historical topographic maps. Transactions in GIS, 25(6):2888–2906, 2021

    2021

  38. [46]

    S. Luo, A. Saxton, A. Bode, P. Mazumdar, and V. Kindratenko. Critical minerals map feature extraction using deep learning. IEEE Geoscience and Remote Sensing Letters, 20:1–5, 2023

    2023

  39. [47]

    Mei, R.-J

    J. Mei, R.-J. Li, W. Gao, and M.-M. Cheng. Coanet: Connectivity attention network for road extraction from satellite imagery. TIP, 30:8540–8552, 2021

    2021

  40. [48]

    Q. Miao, P. Xu, X. Li, J. Song, W. Li, and Y. Yang. The recognition of the point symbols in the scanned topographic maps. IEEE Transactions on Image Processing, 26(6):2751–2766, 2017

    2017

  41. [49]

    Musleh and M

    M. Musleh and M. Mokbel. A demonstration of kamel: A scalable bert-based sys- tem for trajectory imputation. In Companion of the 2023 International Conference on Management of Data , pages 191–194, 2023

    2023

  42. [50]

    Nazary, Y

    F. Nazary, Y. Deldjoo, T. Di Noia, and E. Di Sciascio. Xai4llm. let machine learning models and llms collaborate for enhanced in-context learning in healthcare.arXiv preprint arXiv:2405.06270, 2024

    2024 arXiv

  43. [51]

    Gpt-4o technical report

    OpenAI. Gpt-4o technical report. https://openai.com/research/gpt-4o, 2024. Accessed: YYYY-MM-DD

    2024

  44. [52]

    L. Qiu, D. Yu, C. Zhang, and X. Zhang. A semantics-geometry framework for road extraction from remote sensing images. IEEE Geoscience and Remote Sensing Letters, 20:1–5, 2023

    2023

  45. [53]

    N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Weiwei Duan, et al. arXiv:2408.00714, 2024

    2018 arXiv

  46. [54]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection, 2016

    2016

  47. [55]

    Z. Shen, R. Zhang, M. Dell, B. C. G. Lee, J. Carlson, and W. Li. Layoutparser: A unified toolkit for deep learning based document image analysis. In Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Procee...

    2021

  48. [56]

    S. Shit, R. Koner, B. Wittmann, J. Paetzold, I. Ezhov, H. Li, J. Pan, S. Sharifzadeh, G. Kaissis, V. Tresp, et al. Relationformer: A unified framework for image-to- graph generation. In ECCV, pages 422–439. Springer, 2022

    2022

  49. [57]

    E. S. Smith, C. Fleet, S. King, W. Mackaness, H. Walker, and C. E. Scott. Estimating the density of urban trees in 1890s leeds and edinburgh using object detection on historical maps. Computers, Environment and Urban Systems , 115:102219, 2025

    2025

  50. [58]

    G. D. Subcommittee. Fgdc digital cartographic standard for geologic map sym- bolization. Technical report, Citeseer, 2006

    2006

  51. [59]

    Z. Tang, Z. Yang, G. Wang, Y. Fang, Y. Liu, C. Zhu, M. Zeng, C. Zhang, and M. Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 19254–19264, 2023

    2023

  52. [60]

    C. Tao, J. Qi, Y. Li, H. Wang, and H. Li. Spatial information inference net: Road extraction using road-specific contextual information. IJPRS, 158:155–166, 2019

    2019

  53. [61]

    B. Wang, Q. Liu, Z. Hu, W. Wang, and Y. Wang. Ternformer: Topology-enhanced road network extraction by exploring local connectivity. IEEE Transactions on Geoscience and Remote Sensing , 2023

    2023

  54. [62]

    C. Wang, R. Xu, S. Xu, W. Meng, R. Wang, J. Zhang, and X. Zhang. Toward accurate and efficient road extraction by leveraging the characteristics of road shapes. IEEE Transactions on Geoscience and Remote Sensing , 61:1–16, 2023

    2023

  55. [63]

    Q. Wang, H. Bai, C. He, and J. Cheng. Fe-linknet: Enhanced d-linknet with attention and dense connection for road extraction in high-resolution remote sensing images. In IGARSS, pages 3043–3046. IEEE, 2022

    2022

  56. [64]

    X. Wang, Z. Yu, S. De Mello, J. Kautz, A. Anandkumar, C. Shen, and J. M. Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14176– 14186, 2022

    2022

  57. [65]

    Woodhead and M

    J. Woodhead and M. Landry. Harnessing the power of artificial intelligence and machine learning in mineral exploration—opportunities and cautionary notes. SEG Discovery, (127):19–31, 2021

    2021

  58. [66]

    S. Wu, Y. Chen, K. Schindler, and L. Hurni. Cross-attention spatio-temporal context transformer for semantic segmentation of historical maps. In Proceedings of the 31st ACM International Conference on Advances in Geographic Information Systems, pages 1–9, 2023

    2023

  59. [67]

    S. Wu, M. Heitzler, and L. Hurni. Leveraging uncertainty estimation and spatial pyramid pooling for extracting hydrological features from scanned historical topographic maps. GIScience & Remote Sensing , 59(1):200–214, 2022

    2022

  60. [68]

    X. Xia, D. Zhang, W. Song, W. Huang, and L. Hurni. Mapsam: adapting segment anything model for automated feature detection in historical maps. GIScience & Remote Sensing, 62(1):2494883, 2025

    2025

  61. [69]

    Z. Xu, Y. Liu, L. Gan, Y. Sun, X. Wu, M. Liu, and L. Wang. Rngdet: Road network graph detection by transformer in aerial images. IEEE Trans. Geosci. Remote Sens., 60:1–12, 2022

    2022

  62. [70]

    R. Yao, Y. Wu, C. Wang, J. Xiong, F. Wang, and X. Liu. Elevating legal llm responses: Harnessing trainable logical structures and semantic knowledge with legal reasoning. arXiv preprint arXiv:2502.07912, 2025

    2025 arXiv

  63. [71]

    Yeomans, R

    C. Yeomans, R. Shail, S. Grebby, V. Nykänen, M. Middleton, and P. Lusty. A machine learning approach to tungsten prospectivity modelling using knowledge- driven feature extraction and model confidence. Geoscience Frontiers, 11(6):2067– 2081, 2020

    2020

  64. [72]

    Zhang, X

    J. Zhang, X. Hu, Y. Wei, and L. Zhang. Road topology extraction from satellite imagery by joint learning of nodes and their connectivity. IEEE Trans. Geosci. Remote Sens., 61:1–13, 2023

    2023

  65. [73]

    Zhang, C

    P. Zhang, C. Li, L. Qiao, Z. Cheng, S. Pu, Y. Niu, and F. Wu. Vsr: a unified frame- work for document layout analysis combining vision, semantics and relations. In Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–1...

    2021

  66. [74]

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 , 2020

    2010 arXiv

Pith tools

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