Page 104 - Read Online

Yu et al. Carbon Footprints 2025, 4, 17 https://dx.doi.org/10.20517/cf.2025.12 Page 19 of 23

dynamic, open, and complex system composed of various elements including people, vehicles, roads, and
the environment, leading to multi-dimensional and multi-scale data characteristics. Therefore, advanced
deep learning techniques, including spatiotemporal graph convolutional networks or adaptive graph
recurrent units, hold promise for accurately modeling such nonlinear, time-varying relationships.
Incorporating cross-attention mechanisms to dynamically fuse data sources of varying temporal granularity
and quality could further enhance prediction robustness in real-world deployments.

Given the geographical differences in emission characteristics, localizing and personalizing models become
crucial to meet the unique requirements of different regions and vehicle types. Hence, balancing
personalized modeling (customized for specific cities or scenarios) with generalization (the ability to be
applied across regions and scenarios) is key to enhancing the practical value of these models. Future
research should aim to explore flexible architectures-such as modular backbones with city-specific adapters
or domain adaptation techniques-that balance personalization with generalization. Defining quantifiable
metrics to evaluate this trade-off, and developing continuous learning mechanisms for adaptive
deployment, will be essential for enhancing the long-term utility and policy relevance of emission modeling
systems.

CONCLUSION
In urban road traffic operations, key factors contributing to emission heterogeneity include traffic activity
intensity, vehicle and fuel types, real-world operating conditions, and environmental influences. Both
traditional emission factor models and emerging data-driven models explicitly or implicitly account for
these factors in their formulations.

This review reveals the coexistence of classic emission models and modern data-driven approaches, each
with distinct strengths and limitations. Traditional models typically estimate emissions via regression fitting
or binning of vehicle operation parameters. While these models offer a transparent structure and strong
interpretability, they often rely on manually defined parameters and lack adaptability to the variability of
complex urban traffic conditions, which constrains their predictive performance.

Despite increasing research progress, urban traffic managers still face considerable challenges in leveraging
emission models for actionable CO mitigation. While commercial emission tools are available, trade-offs
2
between model precision, generalizability, and computational cost remain unresolved. Moreover, the
superiority of micro-scale emission models over macro-level approaches continues to be a topic of ongoing
debate in both academia and practice.

DECLARATIONS
Authors’ contributions
Responsible for overall research design, including structure and framework creation: Yu, C.
Provided research direction and specific guidance: Yang, X.
Conducted comprehensive literature collection and organization: Mu, J.
Performed proofreading and formal analysis: Liu, S.

Availability of data and materials
Not applicable.