Data-driven modelling for CO₂ emission reduction in bike-sharing systems: Multi-scale estimation and key determinants

Bike-sharing possesses the potential for CO₂ emission reduction through modal shift effects. However, existing research overlooks the influence of other highly-correlated urban system components such as public transport, Points of Interest (POIs), and socio-economic factors on bike-sharing CO₂ emission reductions (BCERs). Utilizing a large-scale dataset comprising over 170 million trip records, 6 million POI data points and 0.4 million check-ins, we achieve a multi-scale estimation of BCER and explore the impact of urban system factors on BCER potential at the grid level, which is a pioneering work of urban-system-level BCER analysis. Firstly, we estimate city-wide BCER using Bayesian prior probability and cross-scale allocation strategies. Secondly, we apply SHapley Additive exPlanation (SHAP) methods to identify the key determinants affecting BCER. Finally, we perform BCER pattern identification based on feature attribution differences and validate the results using various statistical significance tests to enhance interpretability and reliability. The framework's robustness is validated through the aforementioned real-world dataset and the underlying case study in Beijing, China. The key determinants of BCER include work-related POI density, bus passenger flow, and bicycle lane density, contributing 20.5%, 9.5%, and 7.2% to the total attribution, respectively. Additionally, urban functional and public transport factors significantly influence BCER, while urban perceptual factors are not significant. Notably, areas classified as Pattern I (commercial core areas) and Pattern II (within the 5th Ring Road) show substantial emission reduction potential, accounting for 23.8% and 42.8%, respectively. These findings facilitate the formulation of targeted and effective emission reduction strategies, promoting sustainable urban mobility.