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

Area of Science:

• Environmental Science
• Urban Planning
• Climate Change Research

Background:

• Cities are major sources of anthropogenic carbon dioxide (CO2) emissions, driving climate change.
• Urbanization and land use changes are linked to increasing CO2 emissions within cities.
• Understanding intraurban CO2 spatiotemporal trends is crucial for sustainable city planning.

Purpose of the Study:

• To investigate the efficacy of land use regression (LUR) for predicting intraurban CO2 concentrations.
• To compare LUR with machine learning (ML) algorithms for CO2 prediction.
• To evaluate model performance using both training and unseen sensor locations.

Main Methods:

• Utilized data from the BEACO2N monitoring network in the San Francisco Bay Area.
• Applied traditional land use regression (LUR) models.
• Employed machine learning algorithms, including extreme gradient boosting (XGBoost) and convolutional neural networks (CNN).
• Evaluated predictive accuracy using reserved training data and data from unseen sensor locations.

Main Results:

• For training sensors, XGBoost and CNN achieved the highest accuracy (R²=0.58), surpassing traditional LUR (R²=0.34).
• XGBoost and CNN also outperformed LUR for unseen sensor locations, explaining up to 42% of CO2 concentration variability.
• Machine learning models demonstrated superior ability in capturing non-linear relationships in urban CO2 data.

Conclusions:

• Advanced ML models like XGBoost and CNN are effective tools for predicting intraurban CO2 concentrations.
• These models provide valuable insights into urban land use and carbon dynamics.
• The findings support the development of informed urban planning and decarbonization strategies.