PaperlayerView exact revision on arXiv

Source-linked AI summary

Estimating Pedestrian Volumes from GIS-Derived Built-Environment Features: A Machine Learning Framework

Bahareh Golchin, Banafsheh Rekabdar, Sirisha Kothuri, Joseph Broach

arXiv:2609.12173v1cs.LG

TL;DR

Pedestrian counts are costly and sparse, limiting spatial estimates needed for transportation planning. This study uses GIS-derived features and a small-sample machine-learning pipeline to estimate pedestrian volumes at Portland intersections, with the selected model reducing RMSE by 27% over the published Negative Binomial GLM baseline.

  • Problem

    Pedestrian volumes are systematically undermeasured, and expensive manual counts and costly-to-maintain automated counters leave most locations uncounted.

    Source: §1 · p. 1

  • Method

    The pipeline combines GIS-derived features, L1 Lasso feature selection, tuned histogram-based gradient boosting with Poisson loss, and repeated 3×10 cross-validation.

    Source: §2 · p. 2, §3.3 · p. 2, §5 · p. 4

  • Results

    27% lower RMSE than the published Negative Binomial GLM baseline was reported for the selected model.

    Source: §5 · p. 4

  • Takeaways & Limitations

    The selected predictors align with established pedestrian planning theory, supporting the model’s conceptual validity.

    Source: §5 · p. 4

Abstract

from arXiv · show

Transportation agencies need pedestrian volume estimates across entire road networks to prioritize safety investments, yet manual counts are expensive and cover only a small share of intersections. We present a machine learning pipeline that predicts 2-hour PM peak pedestrian volume at 101 urban intersections in Portland, Oregon, from built-environment, land-use, and street-network features drawn from open GIS data. Starting from the Negative Binomial GLM used in practice, we add feature selection, count-aware gradient boosting, and repeated cross-validation, selecting one configuration by a combined rank over RMSE, MAPE, and SMAPE across four cross-validation strategies. The winner, a histogram-based gradient boosting model with Poisson loss and L1 Lasso feature selection, reduces cross-validated RMSE by 12% over the GLM baseline (89.8 to 78.7) and holdout RMSE by 19% (108.0 to 87.9). Code is released on GitHub.

1 Introduction

Pedestrian volumes are poorly measured, while manual counts are costly and sparse. The paper proposes a small-sample machine-learning pipeline to estimate volumes from open GIS features and benchmark it against the Negative Binomial GLM.

1 Introduction

  • Sparse pedestrian counts limit reliable spatial estimates needed for crossing prioritization, land-use impact evaluation, and crash-risk exposure analysis.Manual counting is expensive, and automated counters are costly to maintain; direct demand models using open GIS variables could estimate volumes across the network at marginal cost.Source: §1 · p. 1
  • The Negative Binomial GLM handles overdispersed counts but imposes functional-form assumptions and cannot represent nonlinear interactions among built-environment variables.Machine-learning methods face small-sample constraints, and using all 92 one-hot-encoded predictors for 101 sites produces severe overfitting.Source: §1 · p. 1
  • The proposed pipeline combines feature selection, repeated cross-validation, and tuning, benchmarking seven models, four selectors, four validation strategies, and five feature counts across roughly 70,000 fits.Its combined-rank procedure avoids selecting a configuration based on only one metric.Source: §1 · p. 1

2 Data and Study Area

The study combines manual PM-peak pedestrian counts in Portland with open GIS-derived features. Count sites were not randomly distributed, and features were measured around buffered intersections and across larger surrounding areas.

2 Data and Study Area

  • Strava Metro supplied GPS-traced walking and running activity mapped to the OSM street network, while static network and sociodemographic data came from open sources.Strava’s privacy masking and fine-time-scale sparsity limited its usefulness to annual totals.Source: §2 · p. 1
  • Manual PM-peak counts from 192 City of Portland PDF files covered 101 locations, mainly during spring and summer, with sites concentrated downtown and along project corridors rather than randomly distributed.Figure 1 maps the count locations and scales markers by observed 2-hour PM peak entering volume.Source: §2 · p. 2, Figure 1 · p. 2
  • Each inferred count point was represented by a 100-foot-radius buffer because precise intersection-node identification proved inconsistent across locations.About 10% of locations required manual geocoding placement.Source: §2 · p. 2
  • Strava trips were estimated by summing distance traveled within each intersection buffer and dividing by an assumed 200-foot traversal length, with background activity measured in larger buffers.Built-environment and sociodemographic variables from OSM, ACS, and LEHD were measured locally and within the same buffer distances, using area- or household-weighted aggregation as appropriate.Source: §2 · p. 2

3 Methodology

The methodology predicts 2-hour PM peak counts from GIS-derived site features, comparing count-model and tree-based approaches. It addresses the small-sample, high-dimensional setting through feature selection, repeated cross-validation, and multiple evaluation metrics.

3 Methodology

  • The target is the observed 2-hour PM peak count at each site, predicted from GIS-derived features for generalization to unseen sites.A few high-volume intersections dominate RMSE, while low-volume sites drive percentage errors.Source: §3 · p. 2

3.1 Baseline and candidate models

  • The Negative Binomial GLM is the baseline, while seven candidate models include Poisson-loss histogram boosting, whose log-link and count assumption make it the closest counterpart.Other candidates isolate loss choice or test robustness, variance reduction, and interpretability.Source: §3.1 · p. 2

3.2 Feature selection

  • Four selectors are compared: L1 Lasso, mutual information, Random Forest importance, and F-regression; each is fit on training folds only to prevent holdout leakage.Lasso yields a sparse set of interpretable nonzero coefficients on the log-link scale; mutual information captures nonlinear dependence and Random Forest importance captures interaction-driven contributions.Source: §3.2 · p. 2

3.3 Validation and tuning

  • Repeated 3×10 cross-validation averages 30 folds, tests every site three times, and supplies the headline estimates; preprocessing is fit within training folds.A single 20-site holdout is noisy, with a few high-volume sites shifting RMSE by ±20; the figure distinguishes holdout Phases 1–2 from cross-validated Phases 3–5.Source: §3.3 · p. 2, Figure 2 · p. 3

3.4 Evaluation metrics

  • RMSE emphasizes high-volume sites, MAPE low-volume sites, and bounded SMAPE is stable across the range; training-fold pseudo-R2 is diagnostic, not comparable across model families.These metrics capture different properties of prediction errors.Source: §3.4 · p. 3

4 Experiments and Results

Experiments compare model architectures, feature selectors, feature counts, tuning choices, and extensions. The tuned HistGB_Poisson with L1 Lasso is the selected winner under repeated 3×10 cross-validation, with holdout and cross-validation agreeing on the winner.

Finding 1: architecture matters

  • On the identical R-code feature set, HistGB_Poisson cuts holdout RMSE from 108.0 to 92.7 (14%) and MAPE from 56.8% to 53.4%.The shared feature set isolates the model architecture from feature engineering.Source: §4 · p. 3

Finding 2: feature selection

  • With all 92 encoded inputs and 81 training samples, the GLM’s test predictions average 254× the true volumes, while tree-ensemble RMSEs are 102.6 and 106.5.The GLM also reaches a training pseudo-R² of 1.000; irrelevant features dilute the signal for tree ensembles.Source: §4 · p. 3

Finding 3: selector comparison

  • Lasso reaches RMSE = 79.8 before tuning, a 26% baseline reduction; RF importance is second, while mutual information and F-regression trail.Under the GLM, every selector still yields error of 145–538, so selection alone does not rescue the parametric model.Source: §4 · p. 3

Finding 4: feature-count sweep

  • Among the tested feature counts, Lasso’s lowest RMSE is 78.9 with 20 features; below 15, informative predictors are excluded, and above 30, irrelevant ones re-enter.The study fixes 20 features for interpretability.Source: §4 · p. 3

Finding 5: tuning

  • Across 140 tuning runs, tuning lowers RMSE from ∼79 to 73.7; repeated 3×10 CV is selected by five of seven tuned winners.Low learning rates (0.01), moderate depth (7), and explicit L2 regularization recur among the selected settings.Source: §4 · p. 3, Table 3 · p. 3

Findings 6 and 7: extensions

  • Log-transforming the target raises HistGB_Poisson RMSE from 73.7 to 106.8, while the two ridge-meta stacks score RMSEs of 76.1 and 79.7.The paper attributes the log-transform result to a double log-link and reports that the stacks trail the tuned single model.Source: §4 · p. 3

Final configuration

  • HistGB_Poisson with L1 Lasso (20 features) scored RMSE = 78.72, MAPE = 51.79, and SMAPE = 49.93 under repeated 3×10 CV.Holdout RMSE = 87.85 versus CV RMSE = 78.72; stratifying folds by count quintiles leaves the ranking unchanged and moves winner RMSE to 81.05.Source: §4 · pp. 3–4

What the model uses

  • Permutation importance ranks stv_mi ahead of sig_int_em, although sig_int_em has the largest Lasso coefficient (+45.37).Crowdsourced walking activity, central-business-district proximity, signalization, retail floor area, zero-car households, and sidewalk length are identified as established direct-demand-model correlates.Source: §4 · p. 4, Figure 3 · p. 4

5 Conclusion

The conclusion presents a reproducible machine-learning pipeline designed for small-sample, high-dimensional pedestrian volume estimation. Its selected model reduces RMSE over the published Negative Binomial GLM baseline, while common extensions do not improve results in this setting.

5 Conclusion

  • 27% lower RMSE than the published Negative Binomial GLM baseline came from Histogram-based Gradient Boosting with Poisson loss, L1 Lasso selection of 20 features, and Repeated 3×10 cross-validation.The result followed approximately 70,000 model fits across 526 configurations.Source: §5 · p. 4
  • Log-transformation and stacking did not improve results in this setting.Source: §5 · p. 4
  • The selected predictors align with established pedestrian planning theory, supporting the model's conceptual validity.Source: §5 · p. 4
Loading 2609.12173v1…