Machine learning-based land-use regression models for predicting carbon dioxide concentrations in San Francisco Bay area

Carbon dioxide (CO₂) is a key driver of anthropogenic climate change and cities have been identified as major sources of emissions. Urbanization and land use change are associated with rising urban CO₂ emissions, highlighting the need to study spatiotemporal trends in intraurban CO₂ to inform sustainable city planning. This study investigates the use of land use regression (LUR) to predict intraurban CO₂ concentrations, using data from the BEACO₂N monitoring network in the San Francisco Bay Area. Additionally, LUR is compared to machine learning (ML) algorithms capable of capturing non-linear relationships, representing a two-fold novel contribution. Model performance is evaluated using reserved data from training sensors as well as unseen sensor locations. For training sensors, extreme gradient boosting (XGBoost) and a convolutional neural network (CNN) achieved the highest predictive accuracy (R²=0.58), outperforming traditional LUR (R²=0.34). XGBoost and CNN also outperformed traditional LUR for unseen sensor locations, accounting for up to 42% of the variability in observed CO₂ concentrations. These models offer insight into urban land use and carbon dynamics, supporting more informed approaches to urban planning and decarbonization.