A systematic review of urban road traffic CO₂ emission models

Data requirements

Microscopic emission models, which are suitable for estimating emissions at the individual vehicle level, compute emissions based on vehicle trajectories or operating conditions at a second-by-second resolution, resulting in substantial data requirements. These models incorporate dynamic parameters that continuously change during operation, such as acceleration and speed. In contrast, macroscopic and mesoscopic emission models rely on traffic aggregated variables.

Local governments typically have access to aggregated road traffic data, summarized at the traffic flow level rather than classified by individual vehicle^[9]. Urban traffic infrastructure, such as surveillance cameras and inductive loop detectors, is widely deployed, making it relatively easy to obtain parameters such as flow, speed, and density. Therefore, the data requirements of macroscopic and mesoscopic emission models are generally easier to satisfy.

For instantaneous vehicle emission models, trajectory data are a critical input for evaluating urban traffic emissions. Currently, there are three primary sources of such data: (1) GPS data or on-board diagnostics (OBD) data collected from commercial fleets, including ride-hailing services^[45], taxis^[46], buses^[47,48], trucks^[49], and increasingly, electric vehicles^[50]; (2) mobile phone signaling data and user location information obtained from navigation applications; (3) roadside sensors, surveillance cameras, and electronic toll collection (ETC) systems, particularly at signalized intersections and key roadway segments^[51].

Sparse trajectory data, which are often collected at intervals longer than one second, are common; however, most vehicle carbon emission estimation models require inputs at a one-second resolution. As a result, researchers have increasingly focused on the problem of trajectory reconstruction for emission calculations^[52,53]. Ma et al.^[54] proposed a trajectory reconstruction method based on interpolation of acceleration distributions, demonstrating that the reconstructed trajectory closely approximates the real trajectory, achieving 2% to 17% higher accuracy compared to other methods. Shang et al.^[55] developed a traffic energy consumption model that integrates macroscopic and microscopic data, and reconstructed trajectories using a nonparametric kernel smoothing algorithm combined with variational theory. For data with varying sampling frequencies, this approach significantly improved the accuracy of emission estimates based on reconstructed trajectories.

Accuracy

Macroscopic and mesoscopic emission models rely on aggregated data (such as regional average speed and fleet age distribution), which obscure individual driving heterogeneity. This simplification can lead to the smoothing of emissions in localized hotspots, such as intersections, inherently limiting the accuracy of these models. Smit et al.^[56] found that based on data from portable emission measurement system (PEMS) for five SUVs, the emission factors under actual urban hot operating conditions were seven times higher for nitrogen oxides and four times higher for nitrogen dioxide than those estimated by the COPERT model in Australia.

Microscopic emission models, by contrast, capture individual driving behavior and thus offer the potential for higher accuracy than macroscopic and mesoscopic models, providing a more detailed representation of real-world vehicle emissions. However, the accuracy of these models is largely dependent on the precision of input parameters. The actual results may be constrained by sensor noise and the high costs of data cleaning, often resulting in effective accuracy that falls below laboratory calibration values^[57].

Moreover, emission models are frequently integrated with traffic simulation models. Existing research reveals inconsistencies in whether microscopic traffic simulation models can reliably reproduce real-world pollutant emission characteristics. One notable issue is the discrepancy between the VSP distributions generated by traffic simulators and those observed in reality^[58]. Lejri et al.^[59]found that COPERT may overestimate NOx emissions during smooth traffic flow and underestimate them during congestion. Although PHEM performs well in terms of accuracy, its lack of detailed information on the actual fleet composition may introduce certain biases. The study noted that when using the PHEM model, the absolute global relative errors for fuel consumption and NOx emissions could reach 5.0% and 9.2%, respectively. Gräbe et al.^[60] coupled MATSim with HBEFA to compare simulated emissions with PEM-based measurements. Their results showed that the greenhouse gas CO₂ emissions from light and heavy vehicles over a 61.7-kilometer route were 77% and 57% of the measured values, respectively, which may be related to errors in road classification and the HBEFA model’s failure to account for road gradient effects.

Complexity

Microscopic emission models capture the state and behavior of individual vehicles, requiring highly detailed input data and substantial computational processing. As a result, these models can be computationally intensive and time-consuming when applied to large-scale traffic networks. For instance, coupling the microscopic traffic simulator VISSIM with the PHEM emission model involves extensive calibration and detailed speed profiles^[61]. In contrast, mission models based on average speed or aggregated traffic conditions enable much faster processing, making them more suitable for large-scale or real-time applications.

Transparency

Transparency refers to the extent to which a model’s internal mechanisms, including its structure, parameters, and training process, and whether these elements can be accessible and comprehensible to humans. Zhou et al.^[8] categorize the emission models reviewed in this study as black-box models, which are data-driven and expressed through mathematical relationships. According to this framework, traditional emission models can be classified into physical models (e.g., CMEM, PHEM), hybrid models (e.g., MOVES, COPERT, HBEFA), and data-driven models (VT-Micro).

Transferability

Vehicle emission characteristics vary significantly across regions, heavily influenced by local traffic conditions, driving behaviors, vehicle emission levels, and fuel quality. Emission models are typically developed based on the local driving cycles, such as the U.S. Federal Test Procedure (FTP), and rely on extensive experimental or empirical datasets. However, many countries and regions have yet to establish comprehensive and standardized motor vehicle emission factor databases. This situation results in inconsistent methodologies and fragmented data sources. Although some researchers have attempted localized calibration of these emission factor models to enhance their applicability and accuracy^[62], challenges persist regarding their precision and practical usability in real-world applications.

Overall evaluation

Figure 4 presents a radar chart that offers a qualitative overview of five widely used traditional emission models (COPERT, HBEFA, MOVES, CMEM, and VT-Micro) evaluated across five key criteria: Data Requirements, Accuracy, Complexity, Transparency, and Transferability. The scoring values, ranging from 1 to 5, are primarily derived from accumulated domain knowledge and general experience reported in the literature, rather than from a standardized quantitative evaluation framework.

Figure 4. Radar chart evaluating traditional emission models across five criteria. Higher scores (1-5) indicate more favorable performance for each respective attribute.

The radar chart clearly illustrates the inherent trade-offs among these models. Microscopic models such as CMEM and VT-Micro tend to achieve higher scores in Accuracy and Transparency, reflecting their detailed representation of vehicle behavior and model structure. However, these models require more granular and extensive input data, as evidenced by their higher Data Requirements scores. In contrast, macroscopic models like COPERT and HBEFA demonstrate strengths in lower data demands and reduced Complexity, making them more suitable for applications with limited data availability or computational resources. Nevertheless, this often comes at the expense of reduced accuracy and diminished capacity to capture localized emission variability. The model MOVES offers a balanced performance, with moderate scores across the evaluated criteria.

Given the diverse application scenarios and varying data complexities involved in deploying each model, there is currently no comprehensive scientific study that rigorously compares these models under uniform conditions. Differences in modeling assumptions, input data quality, geographic context, and computational methods pose significant challenges to direct benchmarking. To advance the field, future research efforts should prioritize the development of standardized evaluation frameworks and the establishment of shared benchmark datasets.

Statistical regression models

Various regression and classification algorithms have been employed in emission prediction, including Simple Linear Regression (SLR), Support Vector Regression (SVR), Decision Trees, and Random Forests. For instance, Chandrashekar et al^[65]. developed a speed-based emission model using SLR, SVR, and Piecewise Linear Regression (PLR), finding that PLR outperformed the others in predicting CO, CO₂, HC, and NOx emissions. Similarly, Jia et al.^[66] used real-world vehicle emission data and applied a Random Forest model to estimate emissions of CO₂, CO, NOx, and THC, achieving strong predictive accuracy (R² > 0.85 for all pollutants).

Artificial Neural Networks (ANNs) have also been widely used to model nonlinear relationships in vehicle emissions^[67,68]. For example, Madziel^[69] proposed a two-dimensional emission model for hybrid vehicles using artificial neural networks, achieving an R² coefficient of 0.73 and a Mean Squared Error (MSE) of 0.91.

Time-series models

Instantaneous vehicle emissions are inherently temporal, with key variables, such as speed, acceleration, and engine conditions, exhibiting strong time dependencies. As a result, time-series models have proven effective in modeling dynamic emission patterns based on high-frequency vehicle trajectory data. Popular methods include LSTM, Gated Recurrent Unit (GRU), and recurrent neural networks (RNNs)^[70].

Among them, LSTM is particularly effective in capturing both long-term and short-term dependencies in sequential data through its unique gating mechanisms, including the input gate, forget gate, and output gate, to regulate the flow and storage of information. This architecture enables the LSTM network to dynamically determine which information to retain, which to update, and which to output, thereby capturing complex dynamic patterns in time series data.

However, the effectiveness of these models is often constrained by the availability and quality of input data. In real-world applications, the collection of fine-grained vehicle operation data can be challenging due to limited sensor coverage, privacy concerns, or high data acquisition costs. To address this limitation, Sun et al.^[71] constructed a supervised system based on parallel theory to improve the estimation accuracy of vehicle CO₂ emissions. This system combines physical models with data-driven models, leveraging the robustness of the former when data are scarce and the high precision of the latter when data is abundant. Through two real-world case studies, it was verified that this system can effectively enhance the accuracy of emission estimates, and the physical model can maintain its robustness even when some parameters are unknown, serving as a complement to the data-driven model.

Table 4 summarizes representative studies that have employed time-series modeling approaches (e.g., LSTM, RNN, GRU) for vehicle-level emission estimation, highlighting their model structures, data sources, and input variables across three major dimensions: dynamic driving behavior, vehicle attributes, and environmental factors.

Statistical regression models

At the road segment level, Grote et al.^[19] employed loop detector data to develop traffic emission prediction models. They analyzed the correlation between emissions derived from GPS data and traffic parameters captured by loop detectors. Their study compared the predictive performance of Multiple Linear Regression (MLR) and Multilayer Perceptron (MLP), achieving emission predictions for 24 types of vehicles.

At the intersection level, Song et al.^[77] proposed an emission model based on delay correction. This model associates emissions with traffic performance metrics such as average delay time and number of stops. By establishing baseline emission factors for different intersection types and adjusting them based on congestion severity, they developed a delay correction model capable of dynamically adjusting emissions. Comparative evaluations demonstrated the model’s robustness and practical applicability.

Spatiotemporal models

Due to the spatial non-stationarity and temporal variability of vehicle emissions, spatiotemporal models have emerged to account for the topological structure of road networks and dynamic environmental factors. These models leverage techniques such as Geographically Weighted Regression (GWR), Graph Neural Networks (GNN), and GCN to model the complex spatial-temporal relationships in traffic emissions.

Liu et al.^[78] utilized large-scale vehicle trajectory datasets to estimate road-level CO₂ emissions and developed a Geographical Convolutional Neural Network Weighted Regression (GCNNWR) model. This hybrid model integrates convolutional neural networks to capture nonlinear spatial interactions and regression components to model emission intensity. Empirical evidence from Beijing demonstrated that the GCNNWR model significantly outperformed traditional spatial regression models in capturing spatial heterogeneity, providing valuable insights for urban emission mitigation strategies.

Table 5 summarizes several representative spatiotemporal emission models at the traffic level, highlighting their diverse data sources and input variables used to capture the complex spatiotemporal dynamics of traffic emissions.

Data requirements

The data requirements for data-driven emission models are broadly comparable to those of classical emission models. However, unlike classical models that typically rely on predefined emission factor databases, data-driven approaches, particularly those employing machine learning and deep learning, require independent model training. This entails substantially higher demands in terms of both data volume and quality^[82]. Spatiotemporal models represent an emerging paradigm in traffic emission modeling, yet they have not been widely adopted in traffic management practice. One major barrier is their exceptionally high data requirements. For instance, models based on GNNs or GNNs require detailed Geographic Information System (GIS) data of the road network, high-resolution dynamic traffic flow information, and meteorological parameters for pollutant dispersion. These datasets are typically maintained by different government agencies or private entities, and they often vary significantly in format, coordinate systems, and temporal resolution. Beyond challenges in data acquisition, low data quality and the complexity of integrating heterogeneous data sources present substantial obstacles to the effective deployment of spatiotemporal emission models in real-world road networks.

Accuracy

The accuracy of data-driven emission models is influenced by multiple factors, including data quality, preprocessing methods, feature engineering, model architecture, algorithm adaptability, and training strategies such as hyperparameter optimization, making straightforward evaluation challenging^[83].

Complexity

In data-driven machine learning approaches, traditional models such as linear regression and decision trees offer high computational efficiency. Ensemble models (e.g., XGBoost and random forests), while requiring careful feature selection and overfitting control, still exhibit significantly lower computational demands than deep learning models. When applied to road network-level emission estimation, the strong spatiotemporal dependencies inherent in vehicular emissions substantially increase model complexity. External factors such as weather conditions, traffic patterns, and points of interest further compound this challenge. Due to the involvement of graph convolution operations and spatiotemporal attention mechanisms, these models have many parameters and require handling high-dimensional tensor operations during training, demanding significant GPU memory and computational resources. This is particularly challenging in large-scale road networks, where issues related to the storage and propagation efficiency of adjacency matrices may arise. For instance, spatiotemporal GNNs typically scale quadratically with sequence length and the number of links in the graph, hindering their application in large graphs and long time series^[84].

Transparency

When a linear relationship exists between model inputs and emission outputs, the model tends to exhibit higher transparency and interpretability^[77,85]. However, machine learning methods such as neural networks and deep learning are considered black-box models. Due to their complex internal structures and lack of explicit analytical formulations, these models offer limited insight into the decision-making process. When input data are noisy or uncertain, the opacity of such models may undermine confidence in the attribution of emissions, posing challenges for traffic managers and policymakers seeking transparent and explainable results.

Transferability

Regarding transferability, statistical regression-based emission models generally demonstrate strong transferability. When the feature distribution of the target dataset closely resembles that of the training dataset, these models can be directly transferred while maintaining high accuracy. Even in the presence of differences, quick adaptation can be achieved through local feature recalibration. However, for neural networks, the transferability of time series models is limited by vehicle type and driving behavior, while spatiotemporal models face restrictions related to road network topology, necessitating the re-encoding of node attributes.

Overall evaluation

To facilitate a structured comparison, five representative data-driven emission models were evaluated across five dimensions: Data Requirements, Accuracy, Complexity, Transparency, and Transferability. The scores (1 = lowest, 5 = highest) are based on a synthesis of performance reports in the literature and prevailing expert understanding of model behavior in emission estimation contexts. While these values are not derived from uniform benchmarking experiments, they offer a qualitative overview of the trade-offs and design considerations associated with each modeling approach.

Figure 5 presents a radar chart summarizing the evaluation results of these data-driven models: Random Forest, LSTM, GCN, XGBoost, and a Hybrid Physics-Deep Learning approach. Among them, Random Forest and XGBoost demonstrate relatively strong performance in terms of Data Requirements, Complexity, and Transferability, highlighting their practical utility in diverse scenarios where data availability and model generalization are critical. LSTM and the Hybrid Physics-DL models achieve the highest scores in Accuracy, reflecting their ability to capture complex temporal dependencies and physical relationships in emission dynamics. However, these models tend to be more complex and less transparent, which may pose challenges in interpretability and computational demand. GCN, while offering moderate accuracy and transferability, scores lower on data requirements and complexity, possibly due to the need for structured graph data inputs and sophisticated model architectures. This qualitative assessment underscores inherent trade-offs in selecting data-driven emission models: simpler models with lower data demands and higher transparency may sacrifice some accuracy, while more complex models can better fit intricate emission patterns but at the cost of interpretability and data needs.

Figure 5. Radar chart evaluating data-driven emission models across five criteria. Higher scores (1-5) indicate more favorable performance for each respective attribute.