PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM-Driven Training-free Location-Attribute Synergic Fusion: A Closed-Loop Paradigm for Dual-source Encrypted POIs and LULC Mapping

Chang Li, Xingtao Peng, Yongjun Zhang, Yinfei He, Cairun Huang

arXiv:2609.25051v1cs.CLcs.AIcs.CV

TL;DR

DSEP fusion must address nonlinear coordinate offsets and inconsistent POI attributes between encrypted coordinate systems, while training-free LULC mapping from encrypted maps remains an open need. The paper alternates attribute matching and location optimization in a closed loop, and reports convergence in essentially two iterations. It also proposes inheriting map classifications for vector-raster LULC mapping; extension beyond two encrypted sources remains future work.

  • Problem

    DSEP fusion must reconcile location offsets and attribute inconsistencies between encrypted-coordinate sources, while training-free LULC mapping from encrypted maps requires further study.

    Source: §1 · pp. 3–4

  • Method

    The paper alternates attribute matching and location optimization in a bidirectional feedback loop, then reassesses attribute confidence using updated geometric residuals.

    Source: §1 · p. 6, §4.7 · p. 33

  • Results

    The closed-loop paradigm is reported to progressively purify correspondences and converge essentially within two iterations.

    Source: §6 · p. 40

  • Takeaways & Limitations

    The proposed LEPF-LULC method inherits land-use classes from encrypted maps through location fusion to produce training-free vector-raster mapping with sparse POIs.

    Source: §1 · p. 7, §2.5 · p. 12

    The framework currently addresses fusion between two encrypted coordinate systems; scalability to three or more sources remains to be investigated.

    Source: §5.4 · p. 40

Abstract

from arXiv · show

Dual-source encrypted points of interest (DSEP), POIs from two encrypted coordinate systems, suffer from intertwined location and attribute uncertainties, including nonlinear systematic misalignment and naming inconsistency, hindering land-use/land-cover (LULC) mapping. To the best of our knowledge, this paper is the first to propose an LLM-driven, training-free location-attribute synergic closed-loop optimization paradigm for DSEP fusion. The paradigm jointly refines location transformation and attribute correspondences through iterative feedback. Attribute-synergic location fusion uses an LLM-driven attribute matching method to establish DSEP correspondences, reducing matching complexity from O(N^2) to O(N), and refines transformation coefficients using an improved particle swarm optimization algorithm within ISODATA-clustered local subregions. Location-synergic attribute fusion then reassesses attribute confidence from updated geometric residuals through an LLM-fuzzy method. The refined correspondences feed back into location optimization, forming a bidirectional closed loop. Sample purification and adaptive radius contraction enable convergence in essentially two iterations. We further propose a training-free LULC mapping method that inherits land-use classes from encrypted maps through location fusion, producing vector-raster integrated LULC maps. A reference-free POI fusion evaluation method is applied across 31 provincial capitals and municipalities in mainland China. Experiments show that our method achieves an average DSEP location fusion residual of 4.58 m and attribute fusion accuracy of 95.12%, improving upon the open-source baseline and state-of-the-art method by 1.77 m and 14.87%, respectively. Overall, the method provides a training-free solution for DSEP fusion and enables georeferencing of encrypted vector data to WGS-84 without field-surveyed ground control points.

1. Introduction

DSEP fusion must reconcile nonlinear coordinate offsets and semantically inconsistent attributes between encrypted maps. The paper proposes a training-free closed loop that alternates matching and location optimization, evaluates fusion without ground truth, and transfers encrypted-map labels to LULC mapping.

1. Introduction

  • The paper frames DSEP as matching the same geographic entities across sources despite differences in both attributes and location, whose interdependence can support fusion.POI datasets from multiple sources can contribute heterogeneous attribute information.Source: §1 · pp. 2–3, Figure 1 · p. 3
  • DSEP fusion must resolve nonlinear, spatially heterogeneous coordinate offsets and ambiguous names or addresses, which can produce mismatches and gross correspondence errors.The study focuses on BD-09 and GCJ-02, whose encrypted coordinates lack an official conversion API to systems such as WGS-84.Source: §1 · pp. 4–5
  • Accurate alignment of encrypted vector data with WGS-84 imagery is presented as a prerequisite for exploiting frequently updated electronic-map parcels in LULC mapping.The paper contrasts this data source with deep-learning approaches that rely on large annotated datasets.Source: §1 · p. 6
  • The proposed paradigm alternates attribute matching, location optimization, and attribute reassessment in a feedback loop that progressively purifies correspondences and converges within two iterations.Its LLM-driven matching reduces complexity from O(N^2) to O(N), while improved PSO optimizes transformation coefficients in ISODATA-partitioned subregions.Source: §1 · pp. 6–7
  • The proposed reference-free evaluation requires neither ground truth for accuracy evaluation nor labeled training data, and reflects relative accuracy at each fusion stage.This is intended to support automated, large-scale DSEP fusion without manual annotation.Source: §1 · pp. 5–7
  • LEPF-LULC directly inherits land-use classifications from encrypted maps through location fusion, enabling training-free vector-raster mapping with sparse POIs.The method is proposed as an alternative to segmentation approaches requiring large annotated datasets.Source: §1 · p. 7

2. Related work

Related work covers language models and separate approaches to POI location and attribute fusion, while identifying a gap in alternating synergic fusion. The paper also extends this joint semantic-geometric approach to training-free vector-raster LULC mapping.

2.1. Large language model

  • The paper uses ChatGPT and DeepSeek-Chat to distinguish ambiguous POI names and addresses, within a broader training-free matching approach that also applies fuzzy mathematics to attribute ambiguity.The cited models are described as general and Chinese-domain-specific LLMs, respectively.Source: §2.1 · p. 8, §2.3 · p. 9

2.2. Location fusion of POIs

  • Prior POI location-fusion work uses spatial and textual features, while this study reduces residual errors through local intelligent optimization after coarse alignment.The proposed approach targets nonlinear spatial offsets in DSEP.Source: §2.2 · p. 8

2.3. Attribute fusion of POIs

  • Prior attribute-matching approaches include trained embedding and neural methods; the paper instead proposes training-free global-to-local matching and fuzzy mathematics for ambiguity and conflicts.The review notes that existing approaches may require large samples and substantial computation, with limited generalization.Source: §2.3 · pp. 8–9

2.4. Location-and-attribute fusion of POIs

  • Unlike open-loop methods that separate matching and registration, the proposed alternating synergic fusion uses matches to constrain location fusion and aligned locations to guide attribute fusion.The framework is model-agnostic: an LLM, BERT, or traditional string-matching model can initialize attribute similarity.Source: §2.4 · pp. 10–11

2.5. LULC mapping

  • The LULC approach transfers joint semantic-geometric optimization to encrypted vector data and WGS-84 imagery, inheriting map classifications without pixel-level inference.It produces vector-raster output with O(N) coordinate transformation and vectorization, no GPU dependency, and avoids raster-boundary discretization errors.Source: §2.5 · pp. 11–12

3.1. Research route

The research route first aligns the encrypted POI datasets, then alternates local attribute matching and location refinement with fuzzy attribute reassessment. The feedback loop produces the final fused POI dataset upon convergence.

3.1. Research route

  • The dashed loop feeds attribute correspondences into location refinement and returns refined geometry to re-condition attribute confidence until convergence.Convergence yields the final fused POI dataset.Source: Figure 2 · p. 12, §3.1 · p. 13
  • The pipeline acquires Baidu and Amap POIs, removes exact duplicates, and applies empirical trigonometric transformation for initial coordinate unification and correspondences.This global fusion supplies initial transformation coefficients and a correspondence set for the iterative framework.Source: §3.1 · p. 13
  • Within each iteration, LLM matching establishes correspondences in ISODATA-partitioned subregions, followed by local transformation optimization and LLM-fuzzy reassessment of attribute confidence.The reassessment conditions confidence on updated geometric residuals.Source: §3.1 · p. 13

3.2. Closed-loop Joint Optimization Formulation

The formulation treats DSEP fusion as joint minimization of geometric and attribute uncertainty. Its iterative procedure alternates correspondence-guided location updates with geometry-conditioned attribute reassessment and contracts the search radius through feedback.

3.2. Closed-loop Joint Optimization Formulation

  • The method jointly estimates a coordinate transformation and one-to-one POI correspondences to minimize geometric displacement and attribute mismatch.The transformation is parameterized by trigonometric coefficients.Source: §3.2 · p. 13
  • Attribute uncertainty is the complement of similarity, while geometric uncertainty is the normalized post-transformation residual relative to the local search radius.The minimum of name and address similarity scores requires both attributes to be consistent.Source: §3.2 · pp. 13–14
  • The objective weights geometric residuals by attribute similarity and penalizes low attribute similarity, with λ balancing the two terms.The correspondence matrix enforces one-to-one matching.Source: §3.2 · p. 14
  • The convergence parameter ε is set to 10^-2, corresponding to centimeter-level precision deemed negligible for map POI location fusion; a maximum iteration count safeguards against non-convergence.The criterion reflects simultaneous convergence of geometry and attributes; the objective decreases monotonically to a residual noise floor.Source: §3.2 · p. 14
  • Each iteration uses LLM-identified candidate pairs to update transformation coefficients with PSO, then removes pairs whose fuzzy attribute credibility falls below threshold.Subregions whose fitness fails to improve are discarded during location optimization.Source: §3.2 · pp. 14–15
  • The loop recomputes residual dispersion on the retained pairs and contracts the next search radius to three times that dispersion.The contracted neighborhood sharpens candidate matching in the following iteration.Source: §3.2 · pp. 14–15
  • The authors describe the alternating updates as block coordinate descent: each fusion step decreases the objective with the other variable fixed, while feedback contracts the feasible region.They state that the bounded, monotonically non-increasing objective converges to a stationary point.Source: §3.2 · p. 16

3.3. Global POIs fusion

Global POI fusion rapidly initializes coordinate alignment and correspondences before local refinement. The section describes polynomial and non-rigid registration formulations, then identifies trigonometric transformation as the global-registration approach used to initialize the closed loop.

3.3. Global POIs fusion

  • Global POI fusion uses strict string matching to select initial pairs, improving registration efficiency while potentially missing matches because of attribute ambiguity.RANSAC removes mismatches and retains reliable Baidu–Amap pairs.Source: §3.3 · pp. 16–17
  • Quadratic polynomials express X- and Y-direction coordinate rectification, while thin-plate-spline registration adds local nonlinear influence and minimizes bending energy for smooth deformation.The TPS formulation combines a quadratic polynomial component with a TPS kernel.Source: §3.3 · p. 16
  • The trigonometric transformation is used for global and local registration; global fusion provides the initial transformation and correspondence set for the closed-loop algorithm.The authors identify it as a comparison method widely applied in spatial data fusion.Source: §3.3 · p. 17

3.4. Attribute-synergic location fusion

Attribute-synergic location fusion divides POIs into local regions, forms spatially constrained candidate matches, and uses LLMs to resolve ambiguous attributes. These matches support location refinement within the iterative loop.

3.4.1. Cluster-based local regions division

  • ISODATA adaptively splits high-variance clusters and merges low-variance ones to determine local fusion regions, with subregions set at approximately 10–15 km.The scale is calibrated to local encryption distortions; the method is described as relatively low in parameter sensitivity.Source: §3.4.1 · p. 17

3.4.2. Local search circle

  • After global fusion, residual dispersion determines a local search circle that defines potential matches; residuals outside [−3σ, 3σ] are treated as outliers under a normality assumption.The standard deviation is used as the measure of uncertainty.Source: §3.4.2 · p. 18
  • The local search circle narrows candidate sets and reduces matching complexity from O(N^2) to O(K · N), lowering LLM token use and improving matching accuracy.Candidates within the circle are evaluated for same-entity matches despite non-identical names and addresses.Source: §3.4.2 · p. 18

3.4.3. LLM-driven attribute matching

  • The method uses ChatGPT and DeepSeek APIs for attribute matching without training, contrasting this with supervised methods such as BERT that incur training cost.DeepSeek-Chat is described as an open-source 671B-parameter sparse MoE model.Source: §3.4.3 · p. 19
  • Predefined prompts, dialogue templates, and batched multi-round LLM conversations produce formatted matching results for POI attributes.The prompt design supplies semantic context, decision rules, and input-output structure.Source: §3.4.3 · p. 20, Table 3 · p. 20

3.5. Global-to-local location fusion

GTLLF uses attribute-weighted geometric residuals to refine coordinate transformations from global parameters within local regions. Its PSO search uses coefficient-specific bounds and can contract them when local accuracy remains insufficient.

3.5. Global-to-local location fusion

  • PSO iteratively updates particle velocities and positions using inertia, individual and swarm-best positions, and random factors.Each particle retains its best position, while the swarm tracks the global best.Source: §3.5 · p. 21
  • For local POI regression, the objective functions operationalize the joint objective’s geometric term by weighting residual contributions with attribute similarity.For TPS, optimization instead seeks a correction vector that minimizes the total energy functional, starting from A0.Source: §3.5 · p. 21
  • The global transformation’s Ctrig coefficient converts longitude/latitude degrees to radians scaled by the encrypted system’s Mercator projection factor.GTLLF guides PSO to progressively optimize the global transformation’s initial parameters as a gradient-free optimization task.Source: §3.5 · pp. 21–22
  • LLM attribute similarities remain fixed during PSO and weight geometric residuals, so higher-confidence correspondences exert greater influence on the transformation coefficients.PSO optimizes the trigonometric transformation coefficients, not the LLM parameters.Source: §3.5 · pp. 21–22
  • GTLLF initializes PSO with global transformation coefficients and searches within coefficient-specific ±3σ bounds, contracting the range if local accuracy remains insufficient.The bounds reflect each coefficient’s distribution; the stated stopping conditions are surpassing global accuracy or reaching the iteration limit.Source: §3.5 · pp. 21–23

3.6. Location-synergic attribute fusion

This stage uses geometric residuals from location fusion to reassess attribute confidence and update correspondences. Its fuzzy method combines name and address similarities, filters candidates by uncertainty thresholds, and assembles fusion datasets.

3.6. Location-synergic attribute fusion

  • Geometric residuals from GTLLF condition the reassessment of attribute confidence and update the correspondence matrix in the second step of the iterative loop.Small residuals retain or strengthen confidence, while large residuals attenuate it.Source: §3.6 · p. 23
  • After defaults such as threshold, clustering parameters, and PSO hyperparameters are set, the training-free process runs without further human input or intermediate-result adjustment.Source: §3.6 · p. 23
  • The fuzzy method combines SoftMax name and address similarities by taking their minimum, yielding a continuous attribute similarity in [0, 1].The resulting similarity is the attribute weight used in the joint objective.Source: §3.6 · pp. 23–24
  • Within a search circle, candidate pairs above threshold T are retained and the pair with the maximum attribute similarity is selected as the correspondence.T is a lower bound based on uncertainty levels; selected similarities lie in (T, 1].Source: §3.6 · p. 24
  • The fusion workflow groups uncertainty levels I–II as A, III–IV as B, and V–VI as C before combining normalized and initial fusion datasets into D.The input includes unmatched POIs from P1 and P2 and matched pairs from P3.Source: §3.6 · p. 24

3.7. Joint LLM- and encrypted POI-driven training-free LULC mapping

The method maps encrypted vector data to WGS-84, overlays labeled polygons on imagery, and recovers land-use labels from electronic-map semantics. It produces vector-raster integrated maps without requiring dense POI coverage.

3.7. Joint LLM- and encrypted POI-driven training-free LULC mapping

  • Corrected vector coordinates are obtained by applying the PSO-optimized GTLLF transformation, mapping encrypted coordinates to WGS-84 without field-surveyed ground control points.The described inputs are sparse POIs and electronic map vectors.Source: §3.7 · p. 25
  • Corrected vector polygons with land-use labels are rasterized onto the WGS-84 imagery grid to produce a vector-raster integrated product.The result is not pure raster output.Source: §3.7 · p. 25
  • Semantic-layer-driven vectorization recovers each vector element’s land-use category from its RGB color using the encrypted map’s category-to-color mapping.Polygons sharing a semantic layer represent the same land-use type.Source: §3.7 · p. 25
  • The process has time complexity O(N) and does not require dense POI distribution because one POI can classify all polygons sharing its semantic layer.The method inherits LULC classification information from electronic maps, avoiding boundary errors inherent in segmentation networks.Source: §3.7 · p. 26

4. Experiment

Experiments assess DSEP fusion across mainland Chinese cities using reference-free residual evaluation, synthetic ground-truth validation, and downstream LULC mapping. GTLLF and LEPF-LULC outperform their reported comparisons, with performance varying by evaluation setting.

4.1. Experimental setup

  • The study covered 31 mainland provincial capitals or municipalities, using 135,518 Baidu POIs and 530,947 Amap POIs collected via official APIs in 2025.Taiwan, Hong Kong, and Macao were excluded because DSEP data were unavailable there.Source: §4.1 · p. 26

4.2. Baseline and evaluation metrics

  • Evaluation used post-fusion residuals between transformed encrypted coordinates, requiring neither labels nor absolute ground truth.Comparisons included the available open-source location method, OLS, TPS, and attribute-matching baselines.Source: §4.2 · p. 27

4.3. The global residual error

  • The open-source API left random residual errors in direction and distance, showing that its coordinate conversion did not fully invert encrypted transformations.This motivated further fine-grained adjustment.Source: §4.3 · p. 27, Figure 6 · p. 28

4.4. The results of location fusion

  • 4.58±2.82 m mean residual: GTLLF improved on ETT (6.35±3.92 m) and TPS (17.65±9.87 m) across cities.GTLLF reduced mean residual by 1.77 m versus ETT; TPS showed high variability and pathological degradation in some cities.Source: §4.4 · p. 31, Table 6 · p. 31

4.5. Sensitivity analysis

  • Fusion residual means varied by no more than ±0.5 m across the tested ISODATA and PSO parameter ranges.The variation was less than 5% of the baseline residual.Source: §4.5 · p. 31, Figure 8 · p. 32

4.6. Ground truth validation and ablation

  • 37.36 m average residual: GTLLF beat ETT (37.55 m) and TPS (293.98 m) for ChatGPT on the synthetic ground-truth dataset.GTLLF also had the lowest average and median residuals for both LLMs; iterative residuals stabilized after the second outer iteration. ETT served as both the synthetic-data generator and a baseline, a caveat mitigated by added Gaussian noise.Source: §4.6 · pp. 32–33, Table 7 · p. 34

4.7. The results of attribute fusion

  • ChatGPT-based attribute matching reached 95.12% accuracy on ambiguous data and 97.67% on non-ambiguous data, exceeding DeepSeek slightly and BERT significantly.BERT produced more correct matches overall, while LLMs required higher confidence to confirm a match.Source: §4.7 · p. 34, Figure 9 · p. 35

4.8. LLM- and encrypted map-driven training-free LULC mapping

  • 75.14% average mIoU and 83.11% average Macro-F1: LEPF-LULC outperformed EULUC-China 2.0 across the reported city averages.EULUC-China 2.0 means used 26 covered cities, versus 31 for LEPF-LULC.Source: §4.8 · p. 36, Table 8 · p. 36

5. Discussion

The discussion explains how iterative feedback purifies POI correspondences and bounds refinement, then examines efficiency, evaluation validity, downstream LULC mapping, and limitations.

5.1. Mechanism of closed-loop synergic optimization

  • The first iteration’s residuals downgrade semantically plausible but geometrically inconsistent pairs, contracting the search radius for the second-round filtering.After the correspondence set stabilizes, the termination criterion is triggered within two to three iterations in practice; residual variance includes an irreducible component from platform conventions.Source: §5.1 · p. 37

5.2. Efficiency, automation, and LLM reliability

  • Matching falls from O(N^2) to between O(N) and O(1.5N), cutting pairwise comparisons by over 93%.The local search circle supports million-scale scenarios; the table compares pairwise computations.Source: Table 9 · p. 38, §5.2 · p. 38

5.3. Validity of evaluation and downstream transfer

  • Residual-based evaluation uses all matched pairs and is supported by synthetic validation, but it is distinct from absolute georeferencing accuracy.Lower residuals were associated with higher matching precision, and methods with lower residuals consistently achieved higher absolute accuracy in synthetic validation.Source: §5.3 · pp. 38–39
  • The framework transfers encrypted vector data to WGS-84 for training-free vector-raster LULC mapping, without GPU inference or annotated data.The stated 1.77 m improvement over the open-source baseline corresponds to approximately two to four pixels of alignment gain in 0.5 to 1 m imagery.Source: §5.3 · p. 39

5.4. Limitations and future work

  • LULC comparison is limited to coarse first-level categories, while encrypted-map label currency may differ from field-survey data.The discussion recommends temporally close imagery phases and periodic change detection to refresh inherited labels.Source: §5.4 · p. 39
  • GTLLF’s improvement over ETT is 1.77 m, with ETT already reducing origin offset by approximately 99.5%.The authors describe the narrow residual range among locally optimized methods as an asymptotic performance ceiling for deterministic transformations.Source: §5.4 · pp. 39–40
  • The residual-quality assumption was validated synthetically but cannot be directly verified on real encrypted data because decryption algorithms are inaccessible.The authors also note that ETT serves as both synthetic data generator and comparison baseline, so circularity cannot be fully excluded.Source: §5.4 · p. 40
  • The framework addresses BD-09 and GCJ-02; scalability of its closed loop to three or more encrypted sources remains uninvestigated.Source: §5.4 · p. 40

6. Conclusion

The conclusion reports POI fusion and LULC mapping results for the proposed closed-loop paradigm. It also identifies multi-source fusion, finer LULC classification, and real-data verification as future work.

6. Conclusion

  • Across 31 mainland China provincial capitals and municipalities, DSEP fusion achieved a 4.58 m average residual and 95.12% attribute accuracy, improving on the baseline and SOTA by 1.77 m and 14.87%, respectively.The LULC mapping method achieved 75.14% average mIoU and 83.11% average Macro-F1.Source: §6 · p. 40
  • The paradigm also supports training-free vector-raster LULC mapping by aligning encrypted electronic-map vectors with WGS-84 imagery.The reference-free evaluation method supports stagewise quality monitoring in large-scale automated fusion pipelines.Source: §6 · p. 40
  • Future work targets multi-source encrypted fusion, finer-grained LULC classification, and direct verification on real encrypted data.Source: §6 · p. 41

CRediT authorship contribution statement

The contribution statement assigns conceptualization, methodology, supervision, project administration, funding, and resources to Chang Li, with data, analysis, software, validation, and writing roles distributed among the coauthors.

CRediT authorship contribution statement

  • Chang Li led conceptualization and methodology; Xingtao Peng handled formal analysis, software, validation, and the original draft, alongside other distributed roles.Yongjun Zhang provided supervision and resources; Yinfei He contributed data curation, software, and validation; Cairun Huang contributed investigation and visualization.Source: CRediT authorship contribution statement · p. 41
Loading 2609.25051v1…