Analysis of Factors Influencing Driving Safety at Typical Curve Sections of Tibet Plateau Mountainous Areas Based on Explainability-Oriented Dynamic Ensemble Learning Strategy

Abstract

The complex topography of China’s Tibetan Plateau mountainous roads, characterized by diverse curve types and frequent traffic accidents, significantly impacts the safety and sustainability of the transportation system. To enhance driving safety on these mountain roads and promote low-carbon, resilient transportation development, this study investigates the mechanisms through which different curve types affect driving safety and proposes optimization strategies based on interpretable machine learning methods. Focusing on three typical curve types in plateau regions, drone high-altitude photography was employed to capture footage of three specific curves along China’s National Highway G318. Oblique photography was utilized to acquire road environment information, from which 11 data indicators were extracted. Subsequently, 8 indicators, including cornering preference and vehicle type, were designated as explanatory variables, the curve type indicator was set as the dependent variable, and the remaining indicators were established as safety assessment indicators. Linear models (logistic regression, ridge regression) and non-linear models (Random Forest, LightGBM, XGBoost) were used to conduct model comparison and factor analysis. Ultimately, three non-linear models were selected, employing an explainability-oriented dynamic ensemble learning strategy (X-DEL) to evaluate the three curve types. The results indicate that non-linear models outperform linear models in terms of accuracy and scene adaptability. The explainability-oriented dynamic ensemble learning strategy (X-DEL) is beneficial for the construction of driving safety models and factor analysis on Tibetan Plateau mountainous roads. Furthermore, the contribution of indicators to driving safety varies across different curve types. This research not only deepens the scientific understanding of safety issues on plateau mountainous roads but, more importantly, its proposed solutions directly contribute to building safer, more efficient, and environmentally friendly transportation systems, thereby providing crucial impetus for sustainable transportation and high-quality regional development in the Tibetan Plateau.

1. Introduction

Tibet, a pivotal region in Western China, presents formidable challenges to transport safety due to its unique geographical characteristics—an average altitude exceeding 4000 m and a topography predominantly characterized by high mountains and deep valleys. The intricate road conditions, encompassing diverse pavement distresses such as cracks, potholes, and ruts induced by intense ultraviolet radiation, permafrost, and frequent landslides, combined with volatile and harsh climatic conditions like dense fog and sudden temperature drops leading to ice/snow cover that reduces pavement friction, severely test drivers’ safe operating capabilities. The cumulative effect of these hazardous factors directly imperils the fundamental safety of the transport system, thus posing a significant constraint on regional sustainable development. For instance, the nearly one hundred accidents in Nagqu City in 2019, attributed to adverse weather [1], not only resulted in substantial casualties but also inflicted considerable and undeniable damage to the local economy. In high-altitude regions, the perilous nature of roads renders post-accident rescue operations exceptionally difficult and time-consuming. This significantly exacerbates loss of life and economic burdens, directly contravening the sustainable transport objective of “reducing losses and enhancing efficiency.” Of particular concern is a major traffic accident in Nyemo County, documented in [2], which was triggered by driver misconduct, leading to 44 fatalities, 11 injuries, and economic losses exceeding 39 million RMB. This incident was not merely an isolated safety event but a profound blow to the region’s development resilience, underscoring the urgent need to enhance driver behavioral compliance and the overall reliability of road operations—key tenets in the construction of a sustainable transport system. Furthermore, the hypoxia prevalent in Tibet’s high-altitude areas can induce altitude sickness in some drivers, manifesting as dizziness, sluggish reactions, and diminished judgment and response times, with severe cases even posing a risk of syncope. Research indicates that drivers at high altitudes exhibit more risky driving behaviors, endure greater mental and psychological loads, and possess lower perceptual capabilities [3,4]. These physiological and behavioral challenges not only directly compromise driving safety but also potentially decrease transport operational efficiency, threatening the lives of travelers. This directly contradicts the sustainable transport principle of ensuring safe and equitable passage rights for all road users.

National Highway G318 serves as the primary east–west transportation artery in China, spanning approximately 5476 km across 7 provincial-level regions including Tibet, Sichuan, and Chongqing. The Sichuan-Tibet Highway (as the core section of G318) features multiple curve types such as hairpin turns, successive curves, steep gradient curves, and long downhill curves. These curves are characterized by poor sight distances, small turning radii, and difficult passing maneuvers, demanding advanced driving skills. As the vital access route to Tibet, this highway functions as the primary corridor for freight transportation. With rising living standards, an increasing number of tourists now opt for self-driving tours to Tibet, causing a surge in road transport demand. However, the current road network density is low and the road grade is not high. Many drivers lack experience when dealing with curves and steep slopes. Insufficient familiarity with road conditions will increase the probability of traffic accidents [5,6,7,8]. With the increase in altitude, the reaction time becomes longer. Inexperienced drivers will have weaker perception ability in high-altitude environments [9], and are very likely to have traffic accidents. In recent years, the transportation department has carried out quality improvement and renovation on the G318 National Road. During the peak tourist season, more traffic police are dispatched to patrol some tourist attractions. However, traffic accidents still occur. Current interventions, while valuable, still possess considerable room for enhancement in addressing the complex and dynamic realities of road conditions and driver behaviors, particularly regarding the critical bottleneck areas—curves. Therefore, an in-depth investigation into driving safety on curved sections of highways in the Tibetan Plateau mountainous region, alongside the identification of influencing factors and the quantification of their weights, is paramount for constructing a more resilient and sustainable transportation system.

1.1. Literature Review

Currently, safety research on curves in mountainous areas can be broadly categorized into two types. One type directly assesses curve safety from the perspective of road design, while the other indirectly determines safety influencing factors based on driving behavior.

Regarding research on curve safety from a road design perspective, Kar P et al. [10] proposed a non-stationary model based on Surrogate Safety Measures (SSM) and the Peaks-Over-Threshold (POT) approach of Extreme Value Theory (EVT). This model investigated the impact of horizontal curve geometric characteristics in mountainous terrain on the rollover (ROR) crash risk for Heavy Commercial Vehicles (HCVs). It addressed the challenge of proactive safety assessment amidst the scarcity of microscopic driving data in mountainous areas and proposed targeted risk mitigation measures. Yin Q [11] extracted the trajectories of Hazardous Material Vehicles (HMVs) on curved and graded sections (literally: “mountainous curve-slope sections”) in mountainous areas to conduct safety research. It was determined that both road alignment and vehicle type/structure influence safe speed. Machine learning was then used to obtain road parameters and classify vehicle lateral stability, finding that lateral instability states of HMVs on mountainous highways are concentrated at small-radius curves and continuous curves (or successive curves), and that vehicle operating speed is significantly influenced by curve radius. Kou Y [12] studied the rollover/sideslip speed thresholds for vehicles traveling on small-radius curved and graded sections, identifying the influencing factors of vehicle instability. They selected five small-radius curved and graded sections in Guizhou for experiments to obtain trajectory data and developed a model for the relationship between trajectories and road geometric alignment indicators. Zhang W [13] collected vehicle trajectories on six reverse curve sections composed of combined small and large radii. Using software such as Tracker and MATLABR2023b, they extracted variables like speed, curve radius, and grade to study hazardous zones/areas. The results showed that the length of the tangent segment between circular curves significantly affects vehicle operating speed. Yue L et al. [14] investigated the relationship between the minimum radius of circular curves on bends, superelevation, and other parameters under different design speeds, and validated their findings using simulation software. The study found that curve design is independent of vehicle type; the minimum curve radius is inversely proportional to superelevation and the lateral friction coefficient (or side friction coefficient), and directly proportional to vehicle speed. Yue L et al. [15], based on an analysis of vehicle operating stability, developed an overall safety model for combined sections of steep grades and sharp curves. The results indicated that the minimum radius of the subsequent sharp curve is positively correlated with design speed and negatively correlated with the grade and superelevation of the initial segment. Ehsan Ramezani-Khansari et al. [16] studied the impact of five geometric design features—radius, superelevation, grade, lane width, and shoulder width—on the average speed on horizontal curves of two-lane rural roads. Standardized regression coefficients showed that the most significant factor affecting speed is radius, followed by grade. He J et al. [17] used regression analysis methods to investigate the intrinsic relationships and patterns among variables such as grade, horizontal curve radius, speed, and heart rate on curved and sloped sections (or segments with combined curves and grades). They established a relationship model linking longitudinal grade/slope, horizontal curve radius, and heart rate. They determined the minimum horizontal curve radius values for curves with longitudinal grades at different vehicle speeds. Wen H et al. [18] employed the Delphi method and cluster analysis to identify the main factors of driving risk on mountainous highways and their various combinations. They established safety evaluation indicators for the driving environment on mountainous highways, which include three levels and five sub-factors across two aspects: road environment and climatic environment. Xie S et al. [19], to study the distribution patterns of traffic accidents, used statistical analysis and K-means clustering to analyze the relationship between accident rates and alignment parameters (or geometric parameters of alignment) on two-lane freeways/expressways. The results showed that accident rates on curved sections are higher than on tangent sections (or general sections), collisions are the primary accident type, and road conditions have a significant impact on the occurrence of traffic accidents.

Regarding studies on curve safety from the perspective of driver behavior, Sun Y [20] utilized SILAB to construct a high-fidelity driving simulation scenario for sharp mountain curves, exploring optimal intervention strategies for such curves. The results indicated that drivers exhibited diverse driving patterns when no intervention was applied, and the combined audio-visual intervention yielded the best results. Liu Y et al. [21] collected one year of vehicle speed data from over 40 provincial-level curves to establish a speed difference estimation model. They quantitatively analyzed safe driving speeds, drivers’ expected speeds, and the optimal placement of speed limit warning signs, concluding that safety warning signs are required when the speed difference exceeds 15 km/h. To prevent accident blackspots on mountain curves, Qiao J et al. [22] collected driver behaviors such as vehicle speed and heart rate on mountain two-lane highways, and established a curve operational safety model. Hu H [23] investigated the physiological and behavioral coupling characteristics of drivers when encountering oncoming vehicles on mountain curves through real-vehicle experiments. The study found that driving workload was highest during curve negotiation and lowest upon curve entry. Experienced drivers’ speed changes during curve entry and negotiation were not significantly influenced by oncoming vehicles, whereas for inexperienced drivers, the impact was more pronounced, though speeds decreased for both groups. Zhao H et al. [24] investigated 1067 traffic accidents on small-radius mountain curves, selecting 15 characteristic variables, and utilized a Logit model to analyze the influencing factors of accident severity for rear-end and head-on/side-impact collisions. Yao Y [25] constructed one control group and eight different visual guidance facility groups, employing an AHP-DEA comprehensive evaluation model to assess the safety of driving behavior on mountain curves. The results indicated that after the installation of guidance facilities, drivers’ driving states and reactions improved, and significant differences were observed in the effects of various facility combinations. Song J [26] segmented curves into distinct study areas and, integrating six traffic flow parameters such as vehicle type and travel direction, developed a rear-end conflict identification algorithm for curved road sections. The study investigated the types and occurrence mechanisms of conflicts in each area to determine relative driving safety risk levels. Chen Y [27] conducted real-vehicle experiments on mountain hairpin bends to investigate trajectory characteristics and the potential risks associated with conflicts among various trajectory patterns. The study concluded that the curvature for right turns was greater than the road curvature, and was related to factors such as slope direction, curve angle, and speed. Ji X et al. [28] utilized vehicle trajectory and traffic flow data, combined with hazardous driving behaviors on mountain curve sections, to analyze the relationship between characteristic variables (features) and traffic accident risk using machine learning algorithms, thereby performing traffic accident risk prediction. To identify high-risk driving behaviors and accident-prone road sections, Xu J et al. [29] conducted real-vehicle experiments to obtain vehicle trajectories and road alignment. They found that drivers employing higher speeds when entering curves would more frequently encroach upon the opposing lane or shoulder. The smaller the curve radius, the higher the frequency of trajectories encroaching upon the opposing lane. Yu Z et al. [30] selected 11 hairpin curves for on-site driving tests, analyzing trajectory morphology and patterns, associated risks, and road collision mechanisms. They identified six trajectory patterns for left turns and four for right turns. Over 70% and 60% of drivers, when entering hairpin curves for right and left turns, respectively, encroached upon the opposing lane, indicating a potential collision risk. Wu F et al. [31] conducted driving simulation experiments with 30 participants to investigate the impact of curve radius on their visual search and operational patterns. The results showed that as the curve radius increased, their horizontal gaze shifted to the right; in terms of operational patterns, steering wheel angle exhibited a negative correlation with curve radius. Men et al. [32] investigated vehicle deceleration behavior before entering curves on mountain roads and established a vehicle speed control model. The results demonstrated the model’s feasibility and effectiveness.

In conclusion, both road design and driving behavior can have an impact on the safety of driving on curves. Scholars have revealed the intrinsic connection between vehicle stability, driver operation and accident risk through trajectory analysis, simulation modeling, real vehicle tests and accident data mining, emphasizing the reduction in risks such as rollovers and collisions through design optimization and active intervention (warning facilities, speed control). However, most of the above-mentioned studies were conducted in plain and mountainous areas, and the research indicators mainly included height, slope, curve radius, vehicle speed, trajectory, heart rate, safety factor, accident rate, etc. Mountain curves usually have more return curves and steep sections, so special attention should be paid when driving. Due to the extreme climate conditions, harsh terrain environment and high altitude with low oxygen in the mountainous areas of the Xizang Plateau, it will have a psychophysiological impact on drivers, making drivers face greater driving risks when driving on curves. However, at present, there are relatively few studies related to safe driving on curves in the mountainous areas of the Xizang Plateau. Due to the differences in environmental factors such as altitude, the existing models in the plain mountainous areas may not be applicable to the mountainous areas of the Xizang Plateau. Based on this, it is very necessary to conduct safety research on different types of curves of highways in the mountainous areas of the Xizang Plateau and analyze their influencing factors.

1.2. Research Contributions

The contributions of this paper are twofold:

(1)

Developing an indicator system of influencing factors for various curve types on Tibetan Plateau mountainous highways. This system identifies dominant factors affecting traffic safety under different curve configurations, providing actionable recommendations for road administrators and drivers.

(2)

Proposing an interpretability-oriented dynamic ensemble learning strategy for curve safety analysis. At the micro-level, the model contribution weights are dynamically adjusted based on sample prediction uncertainty. At the macro-level, the base model weights are determined through feature interpretation consistency evaluation. A hierarchical fusion is achieved via linear combination, balancing local interpretability stability and global prediction accuracy.

2. Data and Indicators

2.1. Data Collection

The Tibetan Plateau highway contains numerous curves. Based on historical accident data and curve typology, three representative curves were selected: a single curve (near Dongda Mountain, elevation 5008 m), a reverse curve (K3671+900, elevation 3366 m), and a long downhill curve (K3682+500, elevation 2854 m). All three curves exhibit significant deflection angles approaching 180-degree turns, characteristics consistent with most curves in the Tibetan Plateau mountainous areas. The geometric configurations are shown in Figure 1, Figure 2 and Figure 3.

The curve data collection was conducted using a DJI Air 3S UAV(DJI Technology Co., Ltd., Shenzhen, China) and a Pegasus UAV D2000(Shenzhen Feima Robotics Technology Co., Ltd., Shenzhen, China). Windless weather conditions were prioritized during data acquisition to prevent UAV instability from causing data anomalies, which could reduce the accuracy of subsequent models. As the length and geometry of each curve varied, the hovering height of the UAV was adjusted accordingly to ensure complete coverage of the curve. To avoid interference from data collectors on driver behavior, personnel maintained a safe distance from the target curves during operations, ensuring that the collected data solely reflected natural driving behavior. The effective vehicle passage duration for each curve was collected for 50 min, respectively, resulting in a total dataset of 150 min.

2.2. Indicator Selection

In the safety research concerning highway curves in the mountainous regions of the Tibetan Plateau, the selection of indicators forms the core foundation for constructing risk assessment models, directly influencing the reliability of conclusions and their engineering guidance value. When conducting safety research, indicator selection typically requires comprehensive coverage of all “Human-Vehicle-Road-Environment” elements to prevent single-dimensional bias. Furthermore, the chosen indicators must be precisely quantifiable, eliminating subjective and ambiguous descriptions.

(1)

Driver Behavior Indicators: In traffic safety research, “human factors” primarily refer to direct road users, including drivers, pedestrians, and passengers. Given the remote location of high-altitude mountainous highways, pedestrian activity is minimal, and passengers typically do not impact the traffic system if compliant with traffic rules; thus, both can be disregarded in this study. Driver behavior indicators, as the core element of “human factors”, directly reflect drivers’ decision-making and operational capabilities under complex road conditions. Key metrics include acceleration and cornering preference. Acceleration is represented numerically, while cornering preference categorizes drivers’ tendencies when navigating curves into four types: taking an outer turn, taking an inner turn, approaching the center line, and occupying the lane for driving. Among them, “1” represents taking an outer turn, “2” represents approaching the center line, “3” represents taking an inner turn, and “4” represents occupying the lane for driving.

(2)

Vehicle Indicators: In safety research, vehicle performance receives significant attention. However, due to variations across vehicle types, performance characteristics differ substantially. Therefore, vehicle classification is adopted as the metric in curve safety studies. Due to their negligible proportion in mountainous plateau curves, non-motorized vehicles such as motorcycles and tricycles, along with electric vehicles—which exhibit inferior climbing performance compared to gasoline-powered vehicles and face a scarcity of charging infrastructure in high-altitude environments—are also rarely observed on plateau mountain roads; thus, all are considered negligible for this study. In this paper, vehicle types are categorized into four classes: heavy trucks, light trucks, SUVs, and standard vehicles. Standard vehicles are represented by 1, SUVs by 2, light trucks by 3, and heavy trucks by 4.

(3)

Road Indicators: References [33,34] indicate that road curve direction, curvature, and gradient types can significantly influence driving behavior and increase potential safety risks. Reference [35] demonstrates that the combination of gradients and curves may lead to driver misjudgment of road conditions, thereby increasing accident likelihood. The curvature change rate (CCR) characterizes the speed at which the degree of road curvature changes and is the derivative of curvature with respect to the arc length. Consequently, curvature change rate and gradient are selected for road factor analysis.

(4)

Traffic Flow and Environmental Indicators: Reference [36] investigated driver acceleration/deceleration behaviors during car-following scenarios and developed a safety proximity behavior model. Reference [37] examined driver safety perceptions across 12 oncoming vehicle meeting scenarios, considering variations in vehicle types, quantities, and positions. Results indicate that both car-following and oncoming vehicle meeting entail safety risks. In high-altitude mountainous highways, poor sight distance and complex road conditions necessitate heightened caution during these maneuvers. Altitude, as a characteristic of the plateau, is also incorporated into environmental factors. Therefore, oncoming vehicle meeting, car following and altitude are selected as road environmental indicators. Among them, 1 indicates the occurrence of meeting or following vehicles, while 0 indicates no occurrence. Considering the aforementioned analysis of the transportation system, the metrics chosen for this research are detailed in

Table 1. Research Indicators for Curve Safety on Mountainous Highways in the Tibetan Plateau.

Category	Indicators
Driving Behavior	Acceleration (m/s2)
	Cornering preference
Vehicle	Vehicle type
Road	Longitudinal grade
	Curvature change rate (rad/m2)
Traffic Flow and Environment	Vehicle meeting
	Car following
	Altitude (m)

.

3. Data and Methods

3.1. Indicator Data Extraction

Indicator data are derived from three sources. Driving behavior, vehicle dynamics, and traffic flow/environmental parameters can be directly obtained from trajectory data. Road data: road gradients are captured via oblique photogrammetry, and rate of curvature change is calculated post-image segmentation.

Vehicle trajectory extraction is performed using Tracker 6.2 software. Developed by Cornell University (USA), this open-source video analysis and modeling tool is specifically designed for physics education and experimentation. Its capability to analyze object motion trajectories frame-by-frame through videos and extract physical quantities (position, velocity, acceleration) has led to widespread adoption in trajectory extraction research. After importing the video, parameters such as road width and coordinate axes are configured, with trajectory points sampled at 5-frame intervals. The left front wheel is selected as the trajectory reference point to facilitate cornering preference analysis. Following invalid data removal, a total of 500 valid vehicle trajectories are obtained (120, 206, and 174 trajectories in respective datasets). Speed, acceleration and other parameters are all extracted by the software, the data sample consists of all trajectory points. The preferences for meeting oncoming vehicles, following oncoming vehicles and cornering are objectively judged based on the relative position relationship between the vehicle and the centerline of the line and the curb of the highway. An example of the extracted vehicle trajectory is shown in Figure 4

Following flight path planning with a Pegasus UAV (Unmanned Aerial Vehicle), automated road environment data acquisition for the study area is conducted. The acquired imagery and POS (Position and Orientation System) data are then individually exported to a computer. The POS data is processed to retain latitude, longitude, and elevation. Subsequently, these data are imported into Context Capture Center Master 4.4.16 software for loading and processing, resulting in the output of 3D imagery products. Figure 5 shows an example of the generated 3D terrain. From the generated 3D model, the gradient between any two trajectory points can be derived.

Curvature Change Rate reflects the degree of abruptness of road curves. The three selected types of curves undergo image segmentation using the Segformer model. Figure 6 shows three curved road segments from the image segmentation results of the SegFormer model. Subsequently, the segmented data is imported into CAD 2021 software for the calculation of the Curvature Change Rate.

Speed data were imported into MATLAB, and the execution of the code yielded the speed distributions for each curve type and 85th percentile speeds, as depicted in Figure 7. Statistical analysis reveals normally distributed speed patterns across all three curve types. The single curve demonstrates concentrated speed values predominantly within 10–45 km/h, while reverse curves exhibit vehicular velocities clustered between 15 and 70 km/h. Notably, long downhill curves present a polymodal distribution spanning 25–125 km/h; nearly half of them have a speed exceeding 50 km/h, accounting for the majority. This velocity differentiation correlates strongly with both geometric configurations and environmental conditions captured during data acquisition. Field imagery from single curves indicates snow accumulation and reduced pavement friction coefficients, factors inducing driver caution and consequent speed moderation. Conversely, extended gravitational effects on long downhill amplify vehicular momentum, compounded by favorable road surface conditions that reduce driver inclination for active speed regulation. The observed speed phenomena underscore the critical interdependence between roadway geometry, environmental constraints, and driver behavior patterns in velocity modulation. These findings align with established principles of vehicle dynamics on vertical grades and horizontal curves [38].

3.2. Model Selection and Development

In the study of curve safety in plateau mountainous regions, the relationships among indicators (driving behavior, vehicle, road, traffic flow, and environment) may exhibit complex linear or non-linear characteristics [39]. References [40,41,42] discuss the advantages, limitations, and application scenarios of conventional statistical models versus emerging machine learning methods in road safety research. To evaluate model suitability, this study adopts both linear models and non-linear tree-based models for comparative validation, thereby establishing a foundation for subsequent model selection. Selected existing linear and non-linear models are summarized in

Table 2. Comparison of Various Models.

Model	Core Principle	Advantages and Disadvantages	Applicable Scenarios
Logistic Regression	Maps a linear combination to probabilities using the Sigmoid function, thereby modeling the likelihood of a binary event occurring.	Advantages: Simple model structure, strong interpretability, where the output variable weights (coefficients) directly reflect the influence of features on event/risk probability.	Applicable to binary classification problems, capable of handling mixed-type data (e.g., continuous traffic flow data and categorical road type indicators). Specifically in transportation, suitable for predicting mode choice (e.g., public transit vs. private vehicle), classifying traffic incident presence, or determining road segment congestion status.
Ridge Regression	Ridge Regression is an extension of Ordinary Least Squares (OLS) regression that incorporates an L2-norm regularization term. This term is added to the loss function to penalize the magnitude of the regression coefficients.	Advantages: Computational Efficiency, Variance Reduction, Enhanced Generalization Disadvantages: No Intrinsic Feature Selection, Outlier Sensitivity.	This method is particularly well-suited for scenarios where significant multicollinearity exists among the independent variables, and there is a strategic requirement to retain all features within the model rather than selecting a subset.
Lasso Regression	Adds an L1 norm penalty term to linear regression, forcing some coefficients towards zero to achieve feature selection.	Advantages: Prevents overfitting, performs automatic feature selection. Disadvantages: May mistakenly eliminate important features.	Suitable for high-dimensional data (where the number of features exceeds the number of samples) or scenarios requiring model simplification or feature screening. Relevant in transportation for identifying key influencing factors from a large pool of heterogeneous data (e.g., numerous road network attributes, sensor readings, and external variables).
Support Vector Machine (SVM)	Seeks a hyperplane that maximizes the classification margin, and handles non-linear problems through kernel functions.	Advantages: Performs well with small sample sizes and is effective in high-dimensional spaces. Disadvantages: High computational complexity, challenging to tune parameters.	Applicable for small sample sizes, high-dimensional data (e.g., text classification), and scenarios requiring clear classification boundaries (e.g., image recognition). In intelligent transportation systems, useful for anomaly detection (e.g., identifying unusual traffic patterns with limited training examples), classifying vehicle types from sensor data, or detecting specific road conditions.
Random Forest (RF)	Generates multiple decision trees via Bagging. Each tree is trained by sampling with replacement from the original dataset, and randomly selecting a subset of features at each node split. The final prediction integrates results from individual trees through majority voting (for classification) or averaging (for regression).	Advantages: Strong resistance to overfitting, suitable for high-dimensional data. Provides measures of feature importance. Disadvantages: High model complexity, interpretability is weaker than a single decision tree.	Tasks requiring high-accuracy prediction and handling complex non-linear relationships. Frequently employed in transportation for sophisticated traffic flow prediction, predicting travel delays influenced by numerous interacting factors, or modeling accident severity based on diverse environmental and driver behaviors.
LightGBM (LGB)	A histogram-based gradient boosting framework that employs a Leaf-wise growth strategy.	Advantages: Low memory consumption, extremely fast training speed. Directly supports categorical features (eliminating the need for one-hot encoding). Disadvantages: Potentially more prone to overfitting, requires careful parameter tuning.	Large-scale datasets with extremely high demands for training speed and memory efficiency, particularly suitable for high-dimensional and sparse datasets. Ideal for real-time traffic prediction, large-scale transportation network analysis, or demand forecasting leveraging vast amounts of ITS data where speed and resource efficiency are critical.
Gradient Boosting Tree (GBT/GBDT)	Iteratively trains multiple weak learners, where each new tree aims to fit the residuals or gradients from the predictions of the previous model iteration.	Advantages: High prediction accuracy, suitable for complex non-linear relationships. Disadvantages: Slower training speed, sensitive to parameters.	Complex regression and classification tasks requiring high prediction accuracy. Commonly applied in transportation for highly accurate travel time prediction, complex traffic incident classification, or precise demand forecasting where nuanced relationships are crucial.
XGBoost (XGB)	An optimized version of GBDT, incorporating regularization, second-order derivative optimization, among other enhancements.	Advantages: Fast speed, high accuracy, supports parallel computation. Disadvantages: Categorical features require manual encoding.	Efficient modeling of large-scale datasets. Widely used in transportation for comprehensive traffic flow prediction, intelligent signal control optimization, or large-scale accident risk assessment due to its superior performance on big data.

.

3.2.1. Experimental Setup

The data were divided into three groups based on curve types. Continuous variables (acceleration, CCR, Longitudinal grade) were standardized using Z-score normalization to eliminate dimensional discrepancies, while categorical variables (cornering preference, vehicle type, curve type) underwent one-hot encoding. Input features comprised four categories with seven indicators: driving behavior (acceleration, cornering preference), vehicle (vehicle type), roadway (curvature change rate, longitudinal gradient), and traffic flow/environment (car-following status, oncoming traffic presence, altitude). As stated in the Transportation Research Board (TRB) Special Report on Managed Speed [43], the 85th percentile speed is a critical descriptive statistic for evaluating road safety and the most commonly used metric in speed limit design [44]. References [45,46,47,48] all indicate that lateral acceleration can affect driving safety. Exceeding the critical value can cause rollovers and other issues, thereby leading to traffic accidents. Therefore, in the research of this paper, the safety status is measured jointly by vehicle speed and lateral acceleration. Since this paper studies the sharp curve section in the plateau mountainous area, the critical threshold of lateral acceleration is selected as the upper limit of the low-risk 1 m/s². Taking 85% of the cumulative speed of all vehicles on each curve as the dividing point, if the speed is greater than 85% or the lateral acceleration is greater than 1 m/s², the output is considered dangerous; otherwise, it is considered safe. The output target is a binary safety state (0/1), where 0 represents safety and 1 represents danger.

During data preprocessing, a rigorous quality control procedure was implemented:

(1)

The 5th–95th percentile truncation was applied to continuous variables to eliminate extreme values;

(2)

Logarithmic transformation was performed on the skewed-distributed CCR indicator for normality;

(3)

One-Hot Encoding strategy with preserved original column namespace was used for categorical variables to avoid feature naming conflicts.

(4)

Addressing class imbalance using the Synthetic Minority Over-sampling Technique (SMOTE).

To enhance model expressiveness, three types of derived features were constructed during the feature engineering phase. The dataset was split into training and test sets in a 7:3 ratio via stratified sampling. Four evaluation metrics were selected:

ROC-AUC (Area Under the ROC Curve): Evaluates ranking capability across all classification thresholds, mitigating class imbalance bias;

AP (Average Precision): Represents area under the precision-recall curve, focusing on positive-class prediction quality;

F1 Score: Harmonic mean of precision and recall, balancing false positives and false negatives;

Accuracy: Overall prediction correctness ratio, validating model practicality.

3.2.2. Model Selection and Evaluation

References [49,50] indicate that tree-based models are commonly employed non-linear models for identifying driving behaviors and road characteristics influencing traffic safety, given their efficiency in processing complex data and interaction features. Considering the specific characteristics of driving safety prediction for curves on mountainous plateau highways, this paper selected five representative machine learning models to constitute a comparative experimental framework:

(1)

Logistic Regression, serving as the baseline linear model;

(2)

Multiple Linear Regression, used for validating the effectiveness of feature selection;

(3)

Random Forest (RF), representing traditional ensemble methods;

(4)

LightGBM (LGB) and XGBoost (XGB), as typical implementations of modern gradient boosting trees.

These five models collectively cover the algorithmic spectrum from linear to non-linear, and from traditional to state-of-the-art.

The final output results are the data of five model evaluation indicators at each curve, as shown in Figure 8.

In the illustrated results, a value closer to 1 indicates better model performance, while a value closer to 0 suggests poorer model outcomes. Considering the three figures collectively, the non-linear models demonstrate significantly superior performance across all metrics compared to the linear models. For instance, in single curves, the ROC-AUC values for RF, LGB, and XGB reached 0.859, 0.856, and 0.851, respectively, representing an improvement of approximately 3% relative to the linear models. In terms of Average Precision (AP), the non-linear models all exceeded 0.7, indicating stronger predictive capability for the samples compared to the linear models. Furthermore, for the F1-score and accuracy metrics, the non-linear models also showed significantly higher values than the linear models. Consequently, the probability of false positives and false negatives is considerably lower when employing non-linear models. This trend is similarly observed in adverse curve and long downhill curve scenarios. In summary, these findings unequivocally demonstrate that non-linear models exhibit superior predictive power and enhanced adaptability to diverse scenarios within the context of plateau curve safety research. Therefore, it is strongly recommended to incorporate ensemble learning methodologies in future investigations to further optimize model performance.

3.2.3. Model Construction

Based on the aforementioned validation, to address the critical requirement of feature importance interpretation in predictive tasks, we innovatively developed an Explainability-Oriented Dynamic Ensemble Learning Strategy with dynamically weighted allocation mechanisms based on interpretability optimization across three ensemble models: RF, XGB and LGB, as illustrated in Figure 9. The dynamic weight allocator comprehensively considers model uncertainty and feature stability to assign weights to each model, then performs weighted averaging of Shapley Additive explanation (SHAP) values across models. Model uncertainty typically refers to the confidence level of a model in specific predictions. In classification tasks, if the predicted probability for a sample approaches 0.5, it indicates high uncertainty in the model’s classification, whereas probabilities approaching 0 or 1 denote higher confidence. Feature stability may refer to the consistency of feature importance rankings across different models, or the coherence of SHAP values among various models. Features demonstrating consistently high SHAP values across multiple models suggest reliable impact on predictions, whereas significant SHAP value discrepancies indicate lower stability. In dynamic weight allocation, reducing weights for high-uncertainty models and increasing weights for stable features can mitigate interference from high-variance models, while screening cross-model consistent important features to maintain robustness.

For model uncertainty quantification, the predictive confidence of each model at the sample level is analyzed by integrating analysis of variance (ANOVA), information entropy measurement, and confidence interval estimation through predictive probability, as shown in Figure 9. In feature stability computation, the stability coefficient of cross-model feature interpretation is calculated via Kendall’s Tau concordance test based on feature importance ranking, thereby establishing model-level credibility weights, as illustrated in Figure 9.

Based on running the Python(3.12) code under the aforementioned framework, the model’s performance, uncertainty, feature importance, and stability across different curve types were obtained, as shown in Figure 10 and Figure 11.

Figure 10a illustrates that, based on the probability variance metric, the RF model exhibits the highest stability, demonstrating minimal predictive volatility across various curve types and robust model performance. Conversely, Figure 10b reveals that the information entropy of both XGB and LGB is significantly lower than that of RF, thereby indicating a higher degree of uncertainty in the RF model’s predictions. The information entropy of the long downhill curve is the highest among the three models. The entropy values of the single curve and the reverse curve are not much different among the three models, indicating that the uncertainty in the long downhill scene is relatively high, while the single curve and the reverse curve are relatively stable. The elevated uncertainty associated with long downhill curves can be attributed to their complex interplay of vertical (gravity), longitudinal (acceleration/deceleration), and lateral (turning) dynamics. This intricate environment often results in more diverse or less predictable driver behavior, consequently challenging the models to render highly confident predictions in such scenarios. Confidence interval width serves as an indirect indicator of uncertainty. Figure 10c further illustrates that XGB and LGB consistently yield values greater than 0.9 across all three curve types, whereas RF exhibits comparatively narrower confidence intervals. Synthesizing the insights from these three figures, it becomes evident that the RF model tends to make predictions centered around intermediate probability values, with its outputs consequently concentrated within a relatively narrow range. Conversely, XGB and LGB models are predisposed to generating more definitive predictions, albeit with a broader distribution spread. As depicted in Figure 10d, the predictive probability distribution shapes of all three models exhibit striking similarity, each characterized by a prominent peak in the vicinity of 1.0 and a comparatively minor distribution around 0.0. This congruence underscores a high degree of consistency among the three models in their macro-level predictive probability distributions, suggesting that they have assimilated similar underlying data patterns, with predictive probabilities predominantly concentrated around 1.0.

In Figure 11a, the model correlation values of XGB and LGB reach 0.98, indicating a strong positive correlation between XGB and LGB in terms of feature importance. This suggests that these two models largely concur on which features are significant. The correlations of RF with the other two models are 0.75 and 0.65, respectively, indicating a degree of commonality in their feature importance assessments but also notable discrepancies. These divergences stem from the distinct methodologies employed in constructing Random Forests compared to gradient boosting trees, which can lead to different emphases in feature selection and importance evaluation. Such variations facilitate the capture of highly influential features during ensemble modeling. Figure 11b shows that the stability scores for each feature range between 0.83 and 1.0, all exceeding 0.8, thereby providing a robust foundation for model training and generalization.

4. Results

Given the superior deterministic performance of XGB and LGB in model prediction, these models are consequently accorded greater weighting than the RF model when determining the overall model ensemble weights, as illustrated in Figure 12.

The weighted average is first derived by multiplying the raw importance of each feature by its corresponding model uncertainty weight. This weighted average is then further multiplied by each feature’s global stability score. Subsequently, the final dynamic allocation weights for each curve are obtained by normalizing the integrated SHAP values across all features, as depicted in Figure 13.

As illustrated in Figure 13, the weighting coefficients of safety impact factors exhibit significant variations across different curve configurations. The prioritization hierarchy for single curves manifests as follows in descending order of significance: CCR > Acceleration > Longitudinal grade > Vehicle type > Vehicle meeting > Car-following > Cornering preference > Altitude; In the case of reverse curves, the priority sequence demonstrates: Acceleration > CCR > Vehicle type > Longitudinal grade > Car-following > Cornering preference > Vehicle meeting > Altitude; For long downhill curves, the hierarchy of influencing factors is typically observed as follows: Acceleration > CCR > Vehicle type > Longitudinal grade > Cornering preference > Vehicle meeting > Car-following > Altitude.

In the context of single curves, reverse curves, and long downhill curves, the two paramount influencing factors are consistently the CCR and acceleration, with their combined contribution significantly outweighing other elements. In the domain of road traffic safety research encompassing both plain regions and mountainous terrains, curve geometry (e.g., radius) and driver behavior (e.g., speed) have consistently constituted focal points in accident analyses. The prominence of curvature change rate and acceleration metrics in this study demonstrates substantial alignment with this established consensus. Across diverse topographies, precise perception of road geometric configurations and effective control of vehicular dynamics remain fundamental to ensuring driving safety. Globally, numerous nations grapple with mountainous road safety challenges, exemplified by the Alpine regions in Europe [51], the Rocky Mountains in North America [52], and the mountainous areas of southwestern China [53]. Such investigations frequently emphasize gradient, narrow sections, visibility, and analogous factors. The indicator importance in this study aligns broadly with these previous studies, but its emphasis on “CCR” and “acceleration” may be more in-depth than some studies that solely focus on physical parameters such as road friction coefficient and road width. By systematically evaluating safety-influencing factors across distinct curve types (single curves, reverse curves, and long downhill curves) on plateau mountain highways, this research not only enhances comprehension of traffic safety complexities in rugged terrains but also furnishes critical insights for future road design optimization, targeted safety interventions, and driver training protocols.

4.1. Verification of Model Results

To verify the reliability of the obtained weights and ensure more accurate suggestions for subsequent safe driving on high altitudes, sensitivity verification and permutation feature importance (PFI) were, respectively, conducted for each curve index.

4.1.1. Sensitivity Verification

For continuous variables, such as acceleration and curvature rate of change, their maximum, minimum, and several representative intermediate values were sampled. Conversely, for discrete variables, such as vehicle type and turning preference, each unique category was utilized as a distinct test input. In this verification process, for each individual feature, a new dataset was systematically constructed by perturbing only that specific feature, while holding all other feature data constant. Subsequently, the trained predictive model was employed to independently predict outcomes for each newly generated dataset. The average predicted probability for the positive class within each dataset was then computed. The sensitivity profiles for the eight parameters are shown in Figure 14.

A comprehensive analysis of the eight sets of illustrated data reveals that the weights assigned to each curve and the results from the sensitivity validation exhibit a substantial degree of congruence. This observation underscores the efficacy of the dynamic ensemble learning strategy in yielding reliable weight estimations. Among all the considered metrics, the CCR and acceleration demonstrate the most pronounced fluctuations and steepest trends, intuitively indicating their heightened influence as key determinants of prediction outcomes. Variations in categories such as cornering preference, car-following, and vehicle meeting generally result in more moderate volatility of predicted safety probabilities compared to other factors, suggesting a lower degree of sensitivity. Consequently, these factors consistently rank among the bottom three across all three types of curves. For single curves, a discernible decline in the predicted probability around a gradient of zero indicates pronounced model sensitivity in this specific region. In the context of reverse curves and long downhill curves, the magnitude of changes in vehicle type and gradient leads to their respective rankings in third and fourth position. Since the altitude weight result is 0, the sensitivity verification image cannot be drawn.

4.1.2. Permutation Feature Importance Verification

Permutation feature importance evaluates the contribution of individual features by measuring the change in a model’s baseline performance metrics—such as accuracy or AUC—on a designated test set. The process involves randomly shuffling the values of a specific feature within the dataset, thereby disrupting its original association with the target variable. The model’s performance is then re-assessed using this altered data. The extent of performance deterioration serves as an indicator of the feature’s importance: a substantial decline signifies a strong dependency of the model on that feature, whereas negligible change suggests limited relevance. Previous analyses have elucidated how the model leverages these features in decision-making. Conversely, permutation feature importance examines the feature’s significance from an opposing perspective. It assesses the impact on predictive performance when the feature’s information is selectively obscured. When evidence from two fundamentally different methodologies converges to the same conclusion, the reliability and robustness of the inference are markedly strengthened.

The permutation feature importance of these 8 indicators was verified, respectively, and the permutation feature importance arranged by curve type as shown in Figure 15 was obtained. It can be seen that CCR demonstrates the highest PFI significance in the single curve, reaching an astonishing 0.6434. This is highly consistent with the significance of CCR ranking first in the single curve in the weighting results. For the indicators with relatively small weights such as following vehicles, meeting oncoming vehicles, cornering preferences, and altitude, the importance of PFI is also the lowest. Acceleration has the top weight in the reverse curve and the long downhill curve, and the PFI here is as high as 0.3930 and 0.2746, which once again aligns with the weight results and sensitivity analysis results. The ranking of the importance of several other indicators is also generally consistent with the results of the weight analysis and the sensitivity analysis, further verifying the reliability of the weight results.

5. Conclusions

This study addresses the complex terrain and high accident rates of highways in the Tibetan Plateau mountainous regions, focusing on the factors affecting driving safety across different types of bends. The objective is to provide a scientific basis for enhancing road safety in the area. Environmental data were collected from three typical high-altitude curves along National Road G318, and multi-source software was employed to extract and categorize key safety indicators. Comparative validation results demonstrated that non-linear models significantly outperform linear models in predictive accuracy.

Based on this, the present study employs an explainability-oriented dynamic ensemble learning strategy that integrates model stability and feature importance to compute feature weights. This comprehensive approach enables an in-depth analysis of factors affecting driving safety across various types of curves in the Tibetan highland mountainous region. Specifically, the relative importance distribution of eight key safety indicators—including CCR, acceleration, longitudinal gradient, car following, cornering preference, vehicle type, vehicle meeting and altitude—is ranked under different curve categories. The detailed results are as follows:

Single Curve: CCR > Acceleration > Longitudinal grade > Vehicle type > Vehicle meeting > Car-following > Cornering preference > Altitude;

Reverse Curve: Acceleration > CCR> Vehicle type > Longitudinal grade > Car-following > Cornering preference > Vehicle meeting > Altitude;

Long Downhill Curve: Acceleration > CCR > Vehicle type > Longitudinal grade > Cornering preference > Vehicle meeting > Car-following > Altitude.

Notably, for all three curve types, the aggregate contribution of CCR and acceleration exceeded 49%. In the context of single curve scenarios, this combined weight even reached a substantial 0.7434. Relevant management personnel and drivers should pay particular attention to these two aspects when deploying control measures and navigating curves. Road traffic authorities ought to integrate sustainable development principles into road planning and design. Where conditions permit, prioritizing an increase in curve radius and smoothing out curvature changes should be considered. This not only significantly mitigates the risk of accidents caused by centrifugal force but also, while ensuring safety, can enhance the average vehicle speed, optimize overall traffic flow efficiency, and reduce unnecessary traffic delays and energy consumption. A higher acceleration weight implies the necessity of careful speed control when traversing curves. In high-accident or potentially hazardous curve zones, prominent dynamic speed limit signs can be installed ahead of the curves. These indicators can dynamically adjust speed recommendations based on real-time traffic volume, vehicle type, etc., guiding drivers to actively reduce speed and enabling more refined risk management. If feasible, installing “electronic police” or similar systems in such high-accident curves can be beneficial. This not only effectively constrains drivers’ speeding behavior, creating a deterrent effect, but also allows road traffic authorities to accumulate valuable speeding data. This data can then support subsequent traffic optimization and driver training initiatives, forming a sustainable “monitoring-constraint-optimization” closed loop. Additionally, regular training for drivers is advisable. The training content should encompass driving techniques for complex road conditions in plateau mountainous regions, risk identification capabilities, and emergency response measures. The aim is to enhance drivers’ sense of responsibility and proactive risk avoidance abilities. Given that vehicle types rank high in weight for the three curve categories, drivers should also pay attention to speed control when operating different vehicle types. Due to the varying centers of gravity of these vehicles, heavy trucks, for instance, must reduce cornering speeds significantly more than standard vehicles. Drivers are also required to conduct regular vehicle maintenance to prevent brake failure, tire wear, and other issues that could lead to braking system malfunction and severe consequences.

6. Limitations

The core findings of this study reveal the scenario-dependent nature of safety influencing factors. Although progress has been made in elucidating the safety impact factors of curves on high-altitude plateau roads, certain limitations remain.

Constraints related to data collection facilities and weather conditions resulted in the acquisition of only three primary types of curve data; however, in the Tibetan Plateau mountainous region, there are other representative categories of hazardous curves that were not captured. The spatiotemporal coverage of the data was limited, failing to encompass driving behaviors under extreme weather conditions such as snow and ice or dense fog. Moreover, the sample size and indicators are relatively insufficient compared to previous studies, and data on pavement conditions and driver physiological and psychological states were not collected.

Regarding data processing, the Tracker software—used for physical motion analysis—may lack sufficient precision, especially when extracting data related to complex turning maneuvers. Some indicators could not be effectively extracted and required manual verification, which is time-consuming. Future research will incorporate alternative software capable of more accurately capturing vehicle trajectories for automatic data extraction. Additionally, vehicle type classification was relatively coarse, limited to four categories, thereby neglecting potential differences in vehicle dynamics and control characteristics among various vehicle models. This simplification may adversely affect model prediction accuracy and feature importance analysis.

In terms of model deployment, reliance on machine learning models trained on the current datasets limits predictions to offline scenarios, precluding real-time online feedback. Furthermore, due to certain indicators being unmeasurable—such as vehicle speed and factors like pavement friction—effective modeling of relationships between these factors remains challenging.