PaperlayerView exact revision on arXiv

Source-linked AI summary

ArchiveLens: A Generative Spatial Knowledge Framework for Semantic Reconstruction of Historical Maps

Yanan Wu, Yalin Yang, Chengbin Deng

arXiv:2610.35606v1cs.AIcs.CYcs.LGcs.SI

TL;DR

Historical-map digitization has often emphasized geometry without formalizing cartographic semantics, limiting machine-readable access to building attributes. ArchiveLens reconstructs map features and attributes and integrates them in a provenance-grounded spatial retrieval system. In a controlled benchmark, it achieved high sheet-level and exact-building recall across 1,500 questions.

  • Problem

    Historical-map digitization has focused mainly on geometry, while semantic attributes and links to original cartographic evidence remain insufficiently formalized.

    Source: §1 · p. 2, §2.1 · p. 3

  • Method

    ArchiveLens structures map records for retrieval using dense semantic indexing and parsed spatio-temporal constraints.

    Source: §5.3.1 · p. 13, §5.3.2 · p. 14

  • Results

    Across 1,500 questions, gold-sheet recall was 96.8% at rank 1, 98.8% at rank 3, and 99.1% at rank 5; exact-building recall was 95.0%, 97.3%, and 98.1%, respectively.

    Source: §6.4.1 · p. 24

  • Takeaways & Limitations

    ArchiveLens organizes archival records as entity-centric geospatial knowledge that can support verifiable, large-scale analyses while retaining links to historical sources.

    Source: §9 · p. 33

    Chicago map coverage differs across publication years, so aggregate differences represent cross-sectional records rather than direct longitudinal change.

    Source: §8 · p. 32

Abstract

from arXiv · show

Historical maps contain dense morphological and semantic information about past urban environments, but many digitization workflows focus primarily on extracting geometry, and do not capture cartographic semantics such as building use and construction materials. This study introduces ArchiveLens, a generative spatial knowledge framework that operationalizes historical maps as entity-centric, machine-queryable geospatial databases. The pipeline decouples morphological and semantic reconstruction through three stages: tiled YOLOv11-seg instance segmentation for building footprint extraction; deterministic material classification via hue, saturation, and value (HSV)-based color thresholds; and localized attribute parsing using a quantized vision-language model constrained by domain-specific prompts. Extracted entities are integrated into a hybrid Spatial-Temporal Retrieval-Augmented Generation (ST-RAG) architecture that combines dense semantic retrieval with symbolic spatio-temporal filtering and mandatory provenance grounding. Applied to 445 georeferenced Sanborn map sheets of Chicago (1918-1950), the framework achieves high geometric and semantic accuracy. Beyond performance, ArchiveLens advances GIScience by formalizing historical cartography as structured spatial knowledge representation, allowing researchers to query building attributes and compare spatial patterns across map sheets and years.

1. Introduction

Historical maps encode geometry alongside social, economic, and infrastructural meaning, but digitization has largely formalized geometry rather than cartographic semantics. ArchiveLens converts those embedded classifications into machine-readable records and a provenance-grounded system for querying historical map corpora.

1. Introduction

  • Historical maps encode social organization, economic activity, risk assessment, and infrastructure as well as spatial geometry, supporting analysis of urban change and neighborhood patterns.Their standardized symbols, colors, abbreviations, and typologies embed structured semantic systems in visual form.Source: §1 · p. 1
  • Digitization workflows have emphasized geometric extraction while leaving embedded cartographic classifications largely unstructured, limiting computational use of historical map semantics.Historical archives are also fragmented across localized, rolling volumes that resist continuous spatial or longitudinal analysis.Source: §1 · p. 2
  • ArchiveLens reconstructs historical map knowledge through feature extraction, VLM-based semantic parsing, and integration into a provenance-grounded ST-RAG architecture.The framework was grounded and evaluated using Sanborn maps, whose regularized colors, abbreviations, and building-level annotations provide a test case.Source: §1 · p. 2
  • ArchiveLens advances feature-level semantic GIS, automated corpus-scale enrichment, and provenance-aware integration of generative AI into archival spatial infrastructures.The stated aim is to make maps active, formally modeled knowledge resources for GIScience.Source: §1 · p. 2

2. Literature Review

Historical GIS faces challenges in scaling and synthesizing fragmented archives, while extraction methods, vision-language models, and retrieval-augmented generation offer complementary approaches to structuring, interpreting, and grounding map information.

2.1 The value of Historical Cartography and Sanborn Maps

  • Existing footprint extraction often leaves cartographic attributes and building-level links unformalized, while flat databases can sever extracted records from their source evidence.Historical GIS also relies on localized studies and manually curated layers, leaving archive knowledge siloed and difficult to synthesize computationally.Source: §2.1 · p. 3

2.2 Automated Extraction from Historical Maps

  • Deep learning supports instance-level building extraction, and YOLO's balance of detection accuracy and inference speed motivates its use in large-scale historical-map workflows.Prior Sanborn work used footprint and structure extraction primarily for 3D reconstruction rather than large-scale map parsing.Source: §2.2 · p. 4

2.3 Vision-Language Models for Semantic Information Understanding

  • VLMs can jointly interpret visual and textual signals, while domain-aware prompts constrain their role as semantic reconstruction modules linking vectorized geometry to feature attributes.Map interpretation requires integrating typography and spatial reasoning because meaning is distributed across text, colors, symbols, and adjacency.Source: §2.3 · pp. 4–5

2.4 Retrieval-Augmented Generation for Spatial Knowledge

  • RAG retrieves relevant archival evidence and constrains generated answers to retrieved, citable records, supporting reproducible and provenance-aware interpretation.The paper's agent uses dense semantic retrieval to find records and generates responses only from retrieved evidence.Source: §2.4 · p. 5

3. Data Acquisition: The Old Insurance Maps Platform

The study assembled its Sanborn map imagery from the Old Insurance Maps platform, a public humanities repository that hosts digitized archival sheets.

3. Data Acquisition: The Old Insurance Maps Platform

  • The study used Old Insurance Maps, a digital public humanities initiative hosting high-fidelity digital versions of historical Sanborn documents.The original analog sheets are preserved and digitized by federal institutions such as the Library of Congress, which typically distributes them as isolated raster images.Source: §3 · pp. 5–6
  • The platform's streamlined access supports large-scale data acquisition and systematic corpus construction, while its map images provide visual clarity and geometric reliability for model training.Source: §3 · p. 6

4. Study Area

ArchiveLens was evaluated on 445 georeferenced Chicago Sanborn sheets spanning four publication years and a cityscape with dense, heterogeneous urban forms. The maps encode building footprints alongside parcel-level attributes, addresses, and spatial relationships.

4. Study Area

  • The Chicago study corpus comprises 445 fully georeferenced Sanborn sheets from 1918, 1920, 1923, and 1950.The study presents Chicago's dense, heterogeneous map content as a stress test for resolving overlapping geometries, irregular footprints, and dense annotations.Source: §4 · p. 6
  • Sanborn sheets link discrete building footprints aligned to parcels with labels for use and construction, such as “D” for dwelling and “1B” or “2B” for brick stories.Street-front addresses and adjacent rowhouse layouts also encode topological relationships and urban density.Source: §4 · p. 6, Figure 1 · p. 7

5. Methodology

ArchiveLens combines footprint extraction, material and text attribution, and hybrid spatial-temporal retrieval to turn Sanborn sheets into queryable records. Its methodology evaluates both extraction components and the retrieval agent while preserving source-sheet provenance.

5. Methodology

  • Tiled YOLOv11s-seg produces building masks that are reconciled across overlapping tiles, refined, and transformed back into map coordinates.The methodology evaluates bounding-box and polygon-mask localization separately using precision, recall, mAP50, and mAP50–95.Source: §5 · pp. 9–10

5.1.2 Deterministic Material Inference

  • Material classes are assigned deterministically from median HSV values of in-mask pixels, matched against predefined lookup thresholds.Restricting analysis to the predicted mask avoids spectral contamination from nearby line work and structures.Source: §5.1.2 · p. 10

5.2.2 Building-Level Attribute Extraction

  • Building-level attributes are extracted from buffered crops and parsed into semantic key-value pairs linked to geometric records by shared coordinates.The crop buffer scales with bounding-box dimensions and is clamped; prompts restrict outputs to standard Sanborn abbreviations.Source: §5.2.2 · pp. 12–13

5.3.1 Spatio-Temporal Knowledge Indexing

  • Sheet-level textual records prepend city, state, and year before building uses, street names, and typographic notes for embedding and retrieval.The records are embedded, L2-normalized, and indexed with FAISS for inner-product search.Source: §5.3.1 · p. 13

5.3.2 Intent-Aware Hybrid Retrieval

  • Hybrid retrieval first recalls semantically relevant records, then filters them against parsed geographic, temporal, and intent constraints.For queries spanning multiple years or cities, the system scales the baseline K threshold to balance the final context window.Source: §5.3.2 · p. 14

5.3.3 Provenance-Grounded Generative Synthesis

  • The synthesis model must attach a source-sheet identifier to every factual claim, building reference, or structural description.This requirement makes generated information traceable to its source map and vectorized geometries.Source: §5.3.3 · p. 14

5.4.1 Agent Retrieval Benchmark

  • The agent benchmark tests retrieval, attribute recovery, source citation, and abstention on unsupported requests across 1,500 questions.Questions include 1,200 answerable cases and 300 requests for attributes unavailable in the maps.Source: §5.4.1 · p. 15

5.4.2 Retrieval and Response Evaluation Metrics

  • Evaluation covers spatial reasoning, topology, semantic reconstruction, and hallucination mitigation, alongside a 50-question comparison of dense-only, symbolic-only, and hybrid retrieval.The 1918–1920 utilization comparison is treated as cross-sectional because available sheets do not necessarily cover identical areas; retrieval comparison uses top-5 reference-candidate rate and separately checks metadata compliance.Source: §5.4.2 · p. 15, §5.4.3 · p. 16

6. Results

ArchiveLens combines semantic relevance with explicit metadata filtering to improve retrieval, while its evaluations show strong record retrieval, grounded responses, and cross-year aggregation. Results also reveal limits in recognizing map abbreviations and in interpreting differences across map sheets as historical change.

6.1 Performance of Geometric Context Extraction

  • Bounding-box mAP50 was 0.9361 and segmentation mAP50 was 0.9328; HSV material accuracy was 0.905 per building and 0.924 area-weighted.The material evaluation matched predicted footprints to ground-truth polygons at IoU 0.30; smaller outbuildings were more prone to misclassification. Figure 4 reports the detection and segmentation metrics separately.Source: §6.1 · p. 17, Figure 4 · p. 17

6.2 Performance of Semantic Attribute Extraction

  • Utilization extraction achieved 85.33% exact-match accuracy, micro-F1 0.853, and macro-F1 0.739, with class recall ranging from 40.0% for Post Office to 93.0% for Dwelling.The multilabel task used 627 manually annotated labels; 37,870 utilization codes were extracted across processed sheets, with multiple codes possible per building.Source: §6.2 · pp. 19, 22

6.3 ST-RAG Architecture

  • ArchiveLens linked building records to sheet-level street metadata, situating the Federal Steel Fixture Co. within its street network and attaching a traceable source-sheet identifier.The example synthesis included manufacturing use and internal structural details; the authors note that citation presence alone does not guarantee factual correctness.Source: §6.3 · pp. 23–24

6.4.1 Agent Retrieval Benchmark Results

  • Across 1,500 questions, gold-sheet recall was 96.8% at rank 1 and 99.1% at rank 5, while exact-building recall reached 98.1% at rank 5.Gold-sheet recall accepts any top-k record from the designated sheet; exact-building recall requires the specific target building record.Source: §6.4.1 · p. 24

6.4.2 Retrieval and Response Evaluation Results

  • All 200 answerable responses cited the correct sheet, mean answer correctness was 95.6%, and all 50 unavailable-attribute questions were correctly abstained from.The audit also found abbreviation errors, including “Auto St’ge” read as “Auto Stage” and “Priv. Garage” as “Pit Garage.”Source: §6.4.2 · pp. 24–26

6.4.3 Cross-Year Query Evaluation Results

  • The 1918 and 1920 query produced separate utilization counts and percentages, but the authors characterize the output as cross-sectional rather than evidence of urban change.The map subsets differ in sheet counts and may cover different or non-overlapping areas.Source: §6.4.3 · p. 26

6.4.4 Retrieval Strategy Comparison Results

  • 94% of benchmark questions had a reference candidate retrieved by ST-RAG, versus 84% for keyword/symbolic filtering and 70% for dense-only retrieval.For the metal-refining query, ST-RAG returned a 1918 record supporting the requested function, material, and underground-tank detail; the alternatives missed one or more constraints.Source: §6.4.4 · pp. 26, 27, 29, 30
  • The retrieval comparison shows why combining semantic relevance with temporal and metadata filtering mattered: dense-only retrieval could return a wrong-year match, while filter-only retrieval could return a semantically mismatched record.For the metal-refining query, ST-RAG retrieved a 1918 record matching the requested function, material, and tank note.Source: §6.4.4 · pp. 28–30

7. Discussion

ArchiveLens turns historical maps into entity-centric geospatial records and combines semantic retrieval with archival constraints and source-sheet provenance. In the controlled comparison, ST-RAG retrieved reference candidates more often than either single retrieval strategy and returned no metadata-inconsistent records.

7.1 Transforming Historical Cartography into Entity-Centric Geospatial Records

  • ArchiveLens links individual building footprints with attributes such as construction materials and uses, producing searchable records for building-level and large-scale historical spatial analysis.The records can also provide a foundation for future historical knowledge graphs and analyses of urban form and change over time.Source: §7.1 · pp. 30–31

7.2 Hybrid ST-RAG retrieval

  • 94% reference-candidate retrieval for ST-RAG, versus 84% for filter-only and 70% for dense-only, with no metadata-inconsistent records.The hybrid strategy combines semantic similarity with spatial, temporal, and keyword-based constraints.Source: §7.2 · p. 31

7.3 Archival Rigor and the Mitigation of Generative Hallucination

  • A mandatory-citation mechanism requires source-sheet identifiers for factual claims, keeping generated descriptions and spatial relationships traceable to the original maps.The system is constrained to synthesize retrieved context rather than make unsupported claims.Source: §7.3 · p. 31

8. Limitations and Future Directions

ArchiveLens’s semantic extraction is constrained by source-image quality, the color conventions and acquisition conditions of the Chicago corpus, and uneven map coverage across years. The authors also identify cross-year entity matching and a provenance-aware knowledge graph as future work.

8. Limitations and Future Directions

  • Dense residential outbuildings are more prone to semantic misclassification when abbreviations are truncated or ink overlaps building boundaries.VLMs transcribe typography effectively within larger architectural footprints, but performance is subject to source-material visual constraints.Source: §8 · p. 31
  • HSV material thresholds are validated only for the standardized Sanborn color conventions and digitized Chicago corpus, not as unchanged thresholds across other collections.A new collection may require color-fidelity assessment and recalibration; this study did not test cross-city or cross-provider sensitivity.Source: §8 · pp. 31–32
  • Differences in building counts, materials, or utilization across publication years are cross-sectional because map sheets often cover different Chicago areas.The corpus does not reconstruct missing coverage; overlapping editions and cross-year building matching are proposed for stronger longitudinal analysis.Source: §8 · p. 32
  • A provenance-aware spatio-temporal knowledge graph with cross-year entity-resolution edges is proposed as a future extension, not a developed capability.Developing and validating cross-year matching and temporal relations remain future work.Source: §8 · p. 32

9. Conclusion

ArchiveLens combines instance segmentation, localized semantic parsing, and hybrid ST-RAG to turn historical map data into an interactive spatial query engine. The authors present it as a blueprint for structured, traceable archival analysis beyond Sanborn maps.

9. Conclusion

  • ArchiveLens couples instance segmentation and localized semantic parsing, then integrates the attributes into an intent-aware hybrid ST-RAG spatial query engine.The framework delineates urban boundaries while extracting contextual typography, producing an interactive, provenance-grounded system.Source: §9 · pp. 32–33
  • The framework organizes archival data as an entity-centric geospatial database accessible through natural language while retaining links to historical sources.The authors describe this as supporting verifiable, large-scale analysis of urban morphology.Source: §9 · p. 33
  • The authors propose the framework as a blueprint for structuring fragmented historical records beyond Sanborn maps for digital humanities and GIScience.Its stated scope is processing unstructured records into structured, traceable formats.Source: §9 · p. 33

Appendix A.1 Prompt templates and inference configuration

The full-sheet prompt combines filename clues with visual evidence and requests structured map metadata, road names, and a concise sheet description.

Appendix A.1 Prompt templates and inference configuration

  • The full-sheet prompt asks the model to use filename and image clues and return a JSON object.The filename may indicate city, state, year, volume, and sheet number.Source: Appendix A.1 · pp. 33–34
  • Requested metadata include city, state, year, volume, sheet number, and an approximate north-arrow angle when visible.The prompt permits inferring the city from the filename if the map is blurry.Source: Appendix A.1 · p. 33
  • The prompt requests visible road names with confidence and orientation, capped at 40, plus a concise description of the map-sheet context..Source: Appendix A.1 · p. 33

} Appendix A.2 Building-level attribute extraction

The building-level prompt requests structured extraction of utilization, storeys, names, and other useful text from a single Sanborn building crop.

Appendix A.2 Building-level attribute extraction

  • The prompt directs the model to extract information from a crop showing one building using standard Sanborn abbreviations.The response is constrained to a JSON object with utilization, notes, name, storeys, and extra-text fields.Source: } Appendix A.2 Building-level… · p. 34
  • Utilization is recorded as a list of abbreviations, with examples including dwelling, flat, store, garage, warehouse, and manufacturing.The prompt supplies abbreviation mappings such as D/Dwg, F, S, A/Auto, W, and Mfg.Source: } Appendix A.2 Building-level… · p. 34
  • Storeys should be transcribed as written, using 0 if missing, while names and other useful interior text are captured separately.Examples of additional text include “Tin Shop,” “Hardw.,” and “Flats abv.”Source: } Appendix A.2 Building-level… · p. 34

Appendix A.3 Evidence-grounded synthesis with mandatory citation

Appendix A.3 constrains synthesis to exact, explicitly recorded building evidence and requires sheet-level citations; Appendix A.4 specifies the local computing environment and one-pass deterministic run protocol.

Appendix A.3 Evidence-grounded synthesis with mandatory citation

  • Synthesis must match the exact target, use only fields explicitly recorded, and report when the target or requested field is absent.Answers must remain concise and directly supported by the matched record.Source: Appendix A.3 · pp. 34–35
  • Every factual statement must cite its supporting sheet, and a retrieved source counts as evidence only when its text supports the statement.Source: Appendix A.3 · p. 35

Appendix A.4 Computational environment and run protocol

  • VLM extraction and synthesis ran locally on an RTX 4090; each sheet, crop, and benchmark question was processed once without repeated trials or generation selection.The environment used PyTorch, Transformers, SentenceTransformers, and FAISS.Source: Appendix A.4 · p. 35
Loading 2610.35606v1…