Abstract Transportation planning requires anticipating how built environments, population characteristics, and aggregate travel indicators evolve over time. Conventional forecasting approaches often treat these dimensions separately, limiting their ability to represent spatial-demographic change. This study introduces the Demographics-Informed Neural Network (DINN), a deep learning framework that forecasts next-year county-level satellite imagery from historical image sequences while enforcing demographic consistency through a frozen demographic predictor. DINN combines an encoder-decoder image predictor, gated skip connections, demographic bottleneck fusion, and demographic consistency regularization. The frozen demographic encoder is then reused as a fixed feature extractor to estimate aggregate county-level commute-mode and vehicle-availability indicators. Experiments using US county data from 2012 to 2023 show that DINN achieves an MSE of 0.00379 and SSIM of 0.8342 for satellite image prediction. Across repeated grouped spatial splits, DINN provides modest but consistent improvements over ConvLSTM, transformer-based, and encoder-decoder baselines, with stronger performance on both image reconstruction and change-detection metrics. The demographic predictor explains approximately 80% of the variance in population characteristics from satellite imagery. Transfer learning results show higher accuracy for absolute travel-indicator counts than for scale-controlled shares, indicating that part of the signal reflects county population scale. Overall, the findings suggest that satellite-derived spatial representations capture built-environment and demographic correlates of aggregate travel indicators; however, the results should be interpreted as county-level predictive associations rather than causal behavioral inference.
Abstract To address the severe safety challenges facing Highway Work Zones (HWZs) in mixed traffic flow environments, particularly the stochastic disturbances introduced by human “Driver Compliance Heterogeneity,” this paper proposes a macroscopic dynamic merging control strategy based on the Proximal Policy Optimization (PPO) algorithm. This study formulates the dynamic merging optimization problem as a Markov Decision Process (MDP), constructing a comprehensive state space that integrates lane-specific traffic states, vehicle composition, driver compliance heterogeneity, and real-time conflict counts. Through coordination with upstream VMS, the PPO agent learns a near-optimal control policy to dynamically adjust the merge point location in response to real-time traffic fluctuations. Within the Simulation of Urban Mobility (SUMO) simulation environment, the proposed strategy is rigorously benchmarked against traditional static merging (Early/Late) strategies, a Poisson regression-based dynamic strategy, and a Deep Q-Network (DQN) baseline. Experimental results demonstrate that the PPO algorithm excels in balancing safety and efficiency, yielding a significant reduction in traffic conflicts compared to other control methods. Further analysis reveals that the strategy effectively suppresses the generation of shockwaves through effective control in high-risk scenarios characterized by low compliance rates, validating its immense potential for proactive risk mitigation in complex work zone scenarios. The analysis confirms that the agent learns to establish an adaptive “Safety Buffer” by proactively shifting merge points upstream in low-compliance conditions. This finding validates the potential of deep reinforcement learning to dynamically trade off spatial resources for system stability in mixed-autonomy bottlenecks.
Abstract Urban road intersections are among the most critical locations for road safety because they involve complex interactions among multiple road users, particularly during turning maneuvers, where conflicts between pedestrians and vehicles frequently occur. This study adopts an observational approach to develop four statistical models for identifying the factors that significantly influence pedestrian–vehicle interactions during both right- and left-turn maneuvers. The study contributes to the existing literature by explicitly incorporating left-turn maneuvers and proposing a practical simplification of social force variables into directly observable indicators. These simplified indicators proved to be statistically significant and effectively captured the dynamics of pedestrian–vehicle interactions, enabling safety analyses in environments where advanced sensing technologies or automated traffic-monitoring systems are unavailable. The proposed simplification of the social force methodology incorporates variables such as pedestrian direction change, pedestrian conflict, crossing behavior, crossing speed, trajectory within the intersection, and the pedestrian paths. Vehicle-related surrogate safety measures, including crossing time and vehicle speed, were also incorporated into the analysis. Specifically, near-crash events were defined using the post-encroachment time (PET) indicator, linking conflict-based safety measures with behavioral and geometric explanatory variables. Predictive performance was further evaluated through receiver operating characteristic (ROC) analysis. The final specification (Model 4) demonstrated excellent predictive accuracy, achieving an area under the curve (AUC) of 0.992. Traffic conditions, geometric characteristics, and the simplified social force variables emerged as the strongest predictors of near-crash occurrence. The most influential predictors included pedestrian direction change, zebra-crossing time, curve length, crossing length, and the absolute difference between design and available sight distance.
Abstract Nationwide travel behavior data are important for transportation planning, but household travel surveys are costly, infrequent, and difficult to implement consistently across diverse regions. This study develops a two-stage hybrid learning framework for estimating county-level travel behavior from satellite imagery. Stage 1 uses a gradient-boost-enhanced encoder–decoder (GBEED) to infer 26 demographic variables from Landsat 8 imagery, while Stage 2 uses a metalearning gradient boost (MLGB) model to estimate 17 travel behavior indicators from the inferred demographic profile. The framework was developed and evaluated using 3,132 US counties, 2022 American Community Survey estimates, and geographically balanced training, validation, and test partitions. GBEED achieved an average R 2 of 0.867 and outperformed encoder–decoder, ResNet-50, vision transformer, transfer-learning CNN, and DenseNet-style baselines. Under end-to-end evaluation, in which Stage 2 receives demographic variables inferred by GBEED rather than observed ACS demographics, MLGB achieved an average R 2 of 0.934 and an average PET-20 of 62.3%. It outperformed the evaluated conventional regression, tree-based, boosting, and transportation-specific baseline models. The geographic transferability index remained above 0.86 for Stage 1 and 0.90 for Stage 2 across the evaluated regional cluster configurations. Zero-shot evaluation in Puerto Rico showed stronger transfer for aggregate indicators than for sparse and equity-sensitive outcomes such as zero-vehicle households, taxicab use, and bicycle use. SHAP analysis indicated that population scale, age structure, racial and ethnic composition, educational attainment, income, and inequality contributed to the travel behavior predictions. The results demonstrate the potential of satellite imagery for scalable and interpretable county-level travel behavior screening, while sparse and equity-sensitive outputs require local validation before project-level planning or investment decisions.
Abstract Work zones and off-ramp queue spillback are two major bottlenecks on expressways in developed urban areas. However, to our best knowledge, few studies have addressed these two types of bottlenecks simultaneously in expressway traffic control. This paper proposes a coordinated ramp metering and variable speed limit strategy to jointly mitigate work zone congestion and off-ramp queue spillover. A model predictive control framework, consisting of a traffic evolution module and an optimization module, is developed to update optimal control actions in a rolling-horizon manner. The traffic evolution module captures fixed lane drops caused by work zones and predicts dynamic lane drops induced by off-ramp queue spillback. The optimization module incorporates on- and off-ramp storage capacity constraints and designs a hierarchical penalty structure to proactively suppress queue spillover risks. A case study based on a congested urban expressway located in the central urban area of Shanghai is conducted to evaluate the proposed strategy. Results show that the proposed approach effectively mitigates work zone congestion and prevents on-ramp and off-ramp queue spillover. Compared with the no-control scenario, a 23.4% reduction in total time spent is achieved, outperforming traditional coordinated control strategies that neglect off-ramp queue dynamics.
