PaperlayerView exact revision on arXiv

Source-linked AI summary

Spatio-temporally complementary feature propagation on graphs for longitudinal AADT estimation

Linghang Sun, Qishen Zhou, Michail A. Makridis, Anastasios Kouvelas

arXiv:2609.29906v1cs.LG

TL;DR

Sparse and uneven sensor coverage makes network-wide, multi-year AADT estimation challenging. The paper combines loop-detector measurements and a transportation model through directed-graph feature propagation using Poisson energy minimization and flow ratios. In Zurich, the authors report validation NMAE under 10%.

  • Problem

    Estimating AADT across an urban network is challenging because sensor coverage is costly, sparse, and uneven, despite AADT's importance for transportation planning and infrastructure work.

    Source: §1 · p. 1

  • Method

    The approach fuses loop-detector measurements with transportation-model information through directed-graph Poisson-energy feature propagation using intersection flow ratios.

    Source: §2.1 · p. 5, §2.3 · p. 7, §2.2 · p. 7

  • Results

    Validation in Zurich achieved NMAE under 10%.

    Source: §4 · p. 12

  • Takeaways & Limitations

    The algorithm provides longitudinal AADT statistics that can serve as an accident-exposure index in safety analyses in Zurich.

    Source: §4 · p. 12

    The approach depends on data and model quality, and transportation demand models are resource-intensive and not commonly accessible in every city.

    Source: §4 · pp. 11–12

Abstract

from arXiv · show

The estimation of Annual Average Daily Traffic (AADT) is vital for transportation planning and infrastructure maintenance, yet obtaining accurate values for an entire urban network across multiple years remains challenging due to the high cost and spatial sparsity of physical sensors. This research proposes a novel spatio-temporally complementary feature propagation framework that leverages the strengths of two distinct data sources: spatially sparse but temporally dense loop detector data, and a spatially complete but temporally sparse macroscopic transportation model. The methodology highlights a feature propagation algorithm on directed graphs, formulated as a Poisson energy minimization considering residues. The standard binary adjacency matrix is replaced with flow ratio matrices to capture real-world vehicle turn ratios at intersections. Validated in the city of Zurich, the algorithm demonstrates high computational efficiency, achieving convergence within minutes. Results indicate that the framework effectively reconciles theoretical models with empirical ground truths, yielding a normalized mean absolute error below $10\%$. This scalable approach provides a feasible solution for spatio-temporal network-wide AADT estimation through combining real-world limited sensor coverage and traffic models.

1. Introduction

Urban AADT estimation is constrained by sparse, uneven sensor coverage, while established approaches rely on measurements, road attributes, network topology, or transportation models. The paper proposes longitudinal feature propagation that combines complementary data sources and uses Poisson energy minimization with transport modeling.

1. Introduction

  • Sensor coverage remains sparse and uneven because installation and maintenance costs limit dense monitoring, leaving much of the urban network unobserved.AADT is used in applications including pavement maintenance, environmental assessments, intelligent transportation systems, and safety analysis.Source: §1 · p. 1
  • The proposed longitudinal algorithm combines Poisson energy minimization and transport modeling to propagate AADT from observed to unobserved locations using complementary spatial and temporal information.The paper presents loop-detector measurements and demand modeling as complementary sources, and describes feature values as deduced across both dimensions.Source: §1 · pp. 2–4
  • Established estimation commonly starts with direct AADT measurements and factors them using recommendations from local authorities.Measurements can come from permanent sensors or portable counters used for limited periods.Source: §1 · p. 2
  • Feature-driven methods correlate observed AADT with road attributes to estimate unobserved segments, but GWR and GNNs are described as data-hungry.Attributes may include road type, lanes, land use, speed limit, and population density.Source: §1 · p. 2
  • Topology-driven methods such as kriging use spatial relationships, but ordinary kriging can overestimate minor roads near major roads and DEM can produce a similar neighborhood effect.Regression-kriging incorporates related features to address the ordinary-kriging issue.Source: §1 · pp. 2–3
  • The proposed approach is presented as producing longitudinal AADT, enabling estimation of intersection turn ratios, reflecting flow conservation, and requiring neither road-feature data nor large samples.The paper also notes that high-fidelity demand models are required.Source: §1 · p. 3

2. Methodology

The methodology propagates AADT features over directed road-link graphs by minimizing Poisson energy while combining detector observations with a transportation model. Flow-ratio matrices and residue vectors adapt propagation to traffic turns and network boundaries.

2.1. Feature propagation following Poisson energy

  • The directed graph treats road links as nodes and legal movements between links as edges, with inbound and outbound residues representing network sources and sinks.The residues can represent private drives, parking garages, and links at the network border.Source: §2.1 · pp. 4–5
  • Unknown link features are estimated by minimizing Poisson energy subject to observed sensor values, with Dirichlet smoothing constrained by inbound and outbound residuals.The equilibrium is governed by a Poisson equation, and the gradient-flow formulation is approximated with Jacobi iterations that reset known features.Source: §2.1 · pp. 5–6

2.2. Feature propagation considering flow ratio

  • Replacing binary adjacency with inbound and outbound flow-ratio matrices represents turn proportions, while residue vectors account for traffic entering or leaving network boundaries.The resulting flow Laplacian is asymmetric in Euclidean space and is described as a non-Euclidean Poisson-inspired minimization.Source: §2.2 · pp. 6–7

2.3. Incorporation of the data sources and assumptions

  • Zurich fusion combines temporally complete but spatially sparse loop-detector data with spatially complete but temporally sparse GVM; detectors anchor observed links.GVM supplies the residue vectors and flow-ratio matrices used in propagation.Source: §2.3 · p. 7
  • Fusion assumes GVM approximates empirical mobility without being absolute ground truth, and that flow ratios change more slowly across years than AADT.The model treats detector data and GVM as mutually calibrating inputs; detector data can flag and calibrate later changes in flow ratios.Source: §2.3 · p. 8

3. Results

The Zurich evaluation combines the GVM model with loop-detector data from 2018–2022. Data-fusion NMAE was within 0.10 overall, with 2020 at 0.105; the study also validates the assumptions and tests propagation on GVM data alone.

  • The Zurich network contained 9,947 road segments after mode filtering, and its 1,943 detectors covered 16.0% of the network.The GVM began with 45,100 directed edges; detector groups were manually checked to compute link-level AADT.Source: §3 · p. 8
  • With GVM data alone and varying proportions of masked features, unobserved-node error decreased toward an extremely small value; 5,000 iterations took minutes.The error plot uses a logarithmic y-axis, so a linear trend represents exponential error decay.Source: §3 · p. 8, Figure 5 · p. 9
  • Figure 6 validated both assumptions: about 60% of observable nodes differed by at most 15%, and flow-ratio CV was mostly below 0.05 versus AADT's 0.05–0.10 mode.The GVM-to-detector ratio uses a log2 scale.Source: §3 · p. 10, Figure 6 · p. 10
  • NMAE was within 0.10 overall, while 2020 was an exception at 0.105 in the Zurich data-fusion evaluation.The GVM model was generated for 2019; the authors suggest COVID-19 may have contributed to the 2020 exception.Source: §3 · pp. 10–11, Figure 7 · p. 11

4. Conclusion

The conclusion presents directed-graph Poisson propagation with turn-ratio information as a way to combine loop detectors and a transportation model for cross-year AADT estimation. Zurich validation yields average NMAE below 10%, while estimates depend on input quality and flow-ratio information.

  • NMAE averages below 10% in the Zurich validation, which the authors describe as highly accurate for the task.Source: §4 · pp. 11–12
  • The method fuses loop-detector data with Zurich's GVM model through directed-graph Poisson energy minimization that accounts for node residues.Source: §4 · p. 11
  • Replacing binary adjacency with intersection flow ratios is the transport-specific adaptation; any source providing flow-ratio information can be leveraged.Source: §4 · pp. 11–12
  • Estimates depend on data and model quality, and transportation demand models are resource-intensive and not commonly accessible in every city.Source: §4 · pp. 11–12
  • For planned Zurich safety analyses, longitudinal AADT is an accident exposure index; benchmarking against universal kriging and GWR remains future work.Source: §4 · p. 12
Loading 2609.29906v1…