2026-09-28

A Systematic Evaluation of Infrastructure-Based Radar System for Highway Traffic Monitoring

This study evaluates infrastructure-based radar performance using DRaT, a new dual-modality dataset of vehicle trajectories collected at a highway merging segment in Fort Worth, Texas, validated against drone-derived ground truth. The radar achieved a 78% precision and 57% recall for individual vehicle detection, with performance dropping under congestion and at greater distances. While it tracked trajectories with an IDF1 score of 0.699 and estimated space-mean speed with a mean absolute percentage error under 4%, it underestimated traffic density and volume by roughly 23% due to missed detections.

Why it matters — It establishes the first systematic, multi-level benchmark for roadside radar systems against high-fidelity drone imagery, quantifying the exact margin of error practitioners must expect when using radar for macroscopic traffic flow estimation.

Caveat: The evaluation is based on a single highway merging segment, meaning the performance metrics may not generalize to more complex urban intersections or different geometric layouts.

arXiv · pdf

Electric Vehicle Charging Station Location Selection using Geospatial Artificial Intelligence (GeoAI)

The study develops a geospatial artificial intelligence framework that combines a variational autoencoder and a graph convolutional network to identify optimal electric vehicle charging station locations. Tested on real-world geospatial, land-use, population, and traffic data from Bryan-College Station, Texas, the model achieved an F1-score of 0.87 in identifying existing station locations and proposed 27 new candidate sites based on spatial similarity.

Why it matters — It introduces a method to select charging sites based on the high-dimensional spatial characteristics of successful existing stations, rather than relying solely on traditional cost or distance minimization metrics.

Caveat: The framework's performance and the identified candidate locations are demonstrated using data from only a single mid-sized metropolitan area in Texas.

arXiv · pdf · doi

Critical Data Studies in the Anthropocene

This work introduces the concept of the Anthropocene to critical data studies, examining the environmental impacts and extractive logics of datafication and artificial intelligence. It presents two case studies from Spain and Chile to demonstrate how digital infrastructure intersects with local power dynamics, shaping which ecological and social harms are recognized or ignored.

Why it matters — It expands the scope of critical data studies beyond digital politics to directly address the material and environmental costs of computing, offering a framework that connects science and technology studies with political ecology.

arXiv · pdf · doi

AlphaEarth distinguishes cities but compresses urban variation

This study audited the AlphaEarth satellite foundation model by analyzing the geometry and geography of embeddings across 1,000 urban areas in 162 countries. The analysis revealed that cities occupy a distinct, shifted region on the hypersphere (62.7° from the global mean direction), with continent and climate explaining 24.3% of the variation between urban centers. However, internal urban variation is unevenly captured, with embedding dispersion within cities being 14.1% greater per standard deviation of national development, and annual representations shifting eight times more than pixel changes explain.

Why it matters — It demonstrates that while global satellite foundation models successfully support cross-regional urban comparisons, their representations compress internal urban variation in developing countries and are highly sensitive to sensor changes, making them currently unreliable for tracking urban changes over time.

Caveat: The findings are based on a specific audit of AlphaEarth, though the structural limitations regarding urban variation and temporal instability likely apply to other Earth observation embeddings.

arXiv · pdf

A comparative assessment of global building and settlement datasets across geographic and settlement contexts

The study benchmarks seven global building and settlement datasets, including Overture Maps, Google Open Buildings 2.5D Temporal, and WSF Tracker, against harmonized reference footprints across 135 study areas. Overture achieved the highest median city-level vector F1 score of 0.786, while raster performance varied by resolution, with Google's product leading at 10m (F1 of 0.642) and WSF Tracker leading at 100m (F1 of 0.862). The analysis also revealed that raster accuracy correlates with building density and that small candidate buildings frequently cause false positives in vector datasets.

Why it matters — It establishes a standardized, reproducible benchmark that clarifies how different global spatial products perform across varying geographic contexts, helping researchers select the appropriate dataset based on their specific resolution and settlement density needs.

Caveat: The reported accuracies are conservative in rapidly growing areas due to temporal mismatches, though aligning the WSF Tracker with reference imagery improved its mean F1 score by 0.060.

arXiv · pdf

Glassy dynamics of metropolitan human mobility frozen near the gravitational equilibrium

The researchers developed the Gravity-based Home Swapping Model (GHSM) using Metropolis dynamics to simulate urban commuting as a physical system of interacting particles, where total commuting time represents system energy. The model was calibrated and validated using 30 million mobility records spanning one year across six Japanese cities, demonstrating that simulations initialized from arbitrary states successfully converge to empirical origin-destination matrices.

Why it matters — It establishes that metropolitan residential dynamics behave like glassy physical systems, proving that complex urban commuting patterns can be reconstructed and analyzed using minimal physical mechanisms such as ageing, hysteresis, and freezing near equilibrium.

arXiv · pdf

Monitoring Urban Traffic Dynamics at Fine Spatiotemporal Resolution Using Distributed Acoustic Sensing and Deep Learning

This study develops a deep learning framework that processes raw ground vibration waveforms from underground fiber-optic cables to detect vehicle trajectories and infer traffic volume and speed. Tested on a distributed acoustic sensing (DAS) deployment along a roadway network in College Station, Texas, the system uses a hybrid training strategy combining synthetic and manually annotated images to improve vehicle detection under noisy and congested conditions.

Why it matters — It demonstrates that existing underground fiber-optic infrastructure can be repurposed as a passive, privacy-preserving traffic observatory, achieving meter-level and second-level traffic monitoring without installing traditional cameras or road sensors.

Caveat: The framework's performance and detection accuracy were evaluated on a single roadway network in one city, and its scalability to different urban soil types or deeper cable installations remains unmeasured.

arXiv · pdf

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

The study introduces a spatio-temporally complementary feature propagation framework on directed graphs to estimate Annual Average Daily Traffic (AADT) across urban networks over multiple years. The model combines temporally dense but spatially sparse loop detector data with a spatially complete but temporally sparse macroscopic traffic model, replacing standard binary adjacency matrices with flow ratio matrices to represent intersection turn ratios. Tested in Zurich, the algorithm converged within minutes and achieved a normalized mean absolute error below 10%.

Why it matters — It provides a computationally efficient method to reconcile theoretical traffic models with empirical sensor data, enabling network-wide, multi-year traffic estimation without the prohibitive cost of dense physical sensor deployment.

Caveat: The framework's performance and high accuracy were validated using data from a single city, Zurich.

arXiv · pdf

Advancing the Physical Internet with GraphRAG: A New Way to Review and Integrate Existing Research

This study implements a Graph Retrieval Augmented Generation (GraphRAG) system using the GPT-4o mini language model and a Neo4j graph database to systematically analyze and synthesize the literature on the Physical Internet. The system structures scattered research findings into an interconnected knowledge graph to map emerging trends and identify research gaps in collaborative freight transportation.

Why it matters — It introduces a new methodological capability for logistics researchers, demonstrating how graph-based retrieval-augmented generation can make complex, fragmented academic literature more accessible and structured than traditional review methods.

arXiv · pdf · doi

Learning Local Heterogeneity and Cross-Region Context for Large-Scale Traffic Forecasting

The researchers developed LoReST, a spatial-temporal network model that captures local road heterogeneity and long-range spatial context for large-scale traffic forecasting. The model uses relation-aware local aggregation to handle road- and direction-specific features, alongside an inter-region attention mechanism that pools and broadcasts regional representations to reduce computational costs. Evaluated on four datasets from the LargeST benchmark, the model achieved average relative reductions of 4.78% in MAE, 3.60% in RMSE, and 5.75% in MAPE compared to existing baselines.

Why it matters — It provides a computationally efficient way to scale traffic forecasting to large road networks without sacrificing the fine-grained, direction-specific local details that dictate real-world traffic flow.

arXiv · pdf

Smartphone-Based Method for Automated Speed Enforcement

This study designs and tests an automated pipeline that uses smartphone cameras and computer vision to estimate vehicle speed and identify license plates, colors, and makes/models. Testing on 1,800 images from a Brazilian open-source dataset yielded a 46% accuracy rate for license plate text recognition, while tests on 115 videos from Austin, Texas achieved 60.8% accuracy for color detection, 48.6% for vehicle manufacturer identification, and 16.3% for speed estimation within a 20% error margin.

Why it matters — It establishes the baseline performance and technical feasibility of using consumer-grade mobile devices for automated traffic enforcement, moving beyond specialized, expensive roadside infrastructure.

Caveat: The system's current accuracy rates are relatively low, particularly for precise speed estimation and specific make/model identification, which may limit its immediate utility for formal legal enforcement.

arXiv · pdf

Unused power surge compromises U.S. road vehicles sustainability

This study reconstructs the long-term growth dynamics of U.S. road vehicles, distinguishing between professional and consumer assets. It introduces a 'usage factor' metric to analyze the divergent patterns of installed power, mass, and fuel energy use across vehicle categories, quantifying decoupling mechanisms like engine oversizing and fleet redundancy.

Why it matters — It demonstrates that avoiding oversized vehicle engines during the transition to electric vehicles can significantly reduce overall material demand, providing a new quantitative link between material immobilization and energy consumption for transport policy.

Caveat: The analysis relies on macro-level historical reconstructions of the U.S. vehicle fleet rather than direct, real-time tracking of individual vehicle usage.

arXiv · pdf