How AI is Reshaping Urban Greenhouse Gas Prediction and Climate Policy Integration

A comprehensive review by researchers from Hiroshima University and the University of Queensland highlights how machine learning is revolutionizing urban greenhouse gas prediction, offering unprecedented accuracy and adaptability. The study underscores the need for wider geographic application and integration of ML insights into real-world urban climate policies.

A groundbreaking review by Yukai Jin and Ayyoob Sharifi, affiliated with Hiroshima University’s Urban Environmental Science Lab and The University of Queensland’s School of Architecture, Design and Planning, maps the rapidly evolving role of machine learning in urban greenhouse gas (GHG) prediction. Their study, published in Renewable and Sustainable Energy Reviews, analyzed 75 scientific papers published between 2003 and 2023. The findings reveal that cities, which generate over 70% of global CO₂ emissions, are increasingly relying on machine learning (ML) to forecast emissions, improve climate modeling, and guide more effective urban policies. The review highlights the power of ML to process vast, complex datasets, ranging from meteorological trends and satellite imagery to transportation metrics, without relying on the rigid assumptions embedded in traditional models. In doing so, ML is helping cities anticipate climate risks and design mitigation strategies with unprecedented precision.

From Trial-and-Error to Optimization Powerhouses

The study identifies a clear trajectory in the field: an initial reliance on artificial neural networks (ANNs) in the early 2000s has since evolved into a dynamic toolkit that now includes hybrid models, deep learning, and optimization algorithms. Models such as Long Short-Term Memory (LSTM), Support Vector Machines (SVM), XGBoost, and hybrid combinations like POA-ELM (Pelican Optimization Algorithm–Extreme Learning Machine) and ICSO-SVM (Improved Chicken Swarm Optimization–SVM) are now leading the field. These models demonstrate superior performance compared to traditional methods, with R² scores reaching up to 0.9989 and Mean Absolute Percentage Error (MAPE) values as low as 0.3%. Hybrid models, in particular, consistently outperform single-model approaches by enhancing model training, reducing overfitting, and optimizing prediction accuracy. Time-series hybrid models often combine metaheuristic optimization techniques, such as genetic algorithms, with predictive models to fine-tune neural networks and other structures, yielding significant improvements in both speed and reliability.

Urban Focus: Big Cities, Big Data, Big Gaps

A surge in publications after 2015, coinciding with the Paris Agreement, marked a turning point for ML-based GHG prediction. The number of studies increased year-over-year, peaking in 2023. However, the research remains geographically concentrated. China alone accounts for nearly half of the reviewed studies, followed by the United States. Cities such as Beijing, Shanghai, and Jakarta were among the most frequently studied, while medium and small cities especially in Africa and South America remain severely underrepresented. The review attributes this imbalance largely to data limitations in smaller or less developed urban areas, making it difficult to train effective models. Yet, the authors argue that expanding research into a more diverse range of cities is essential. Different city types exhibit varying industrial patterns, energy profiles, and infrastructure, and excluding them risks producing models that are too narrow in application.

Getting the Data Right: Preprocessing and Key Variables

Data quality is at the core of successful ML predictions. The reviewed studies place strong emphasis on preprocessing techniques to improve data consistency and reliability. This includes denoising to remove anomalies, imputing missing values through interpolation, balancing skewed datasets using tools like SMOTE (Synthetic Minority Over-sampling Technique), and normalizing data to ensure comparability across variables. Several innovative techniques emerged: one study downscaled national data to city-level emissions across 13,000 cities globally, while others used hyperspherical transformations or multivariate imputation for cleaner input data.

Selecting the right variables is equally crucial. The review finds that environmental factors such as temperature and wind speed, social indicators like population size and transportation activity, and economic metrics such as GDP and energy consumption are the most frequently used predictors. Methods like Principal Component Analysis (PCA) and Random Forests are used to identify the most influential variables, improving both model efficiency and predictive power. However, the study notes a lack of standardization in how these factors are selected and weighted, calling for more consistent criteria in future research.

Beyond Prediction: Shaping the Cities of Tomorrow

The study goes beyond technical analysis to examine the real-world implications of machine learning in urban climate governance. It identifies energy consumption especially from electricity, heating, and transport, as the primary driver of GHG emissions in cities. In Shanghai, for instance, electricity alone accounts for over 60% of residential CO₂ emissions. Transportation is another major contributor; in Osaka, a study linked reductions in CO₂ to more affordable and efficient public transit. Land-use changes, such as the expansion of urban construction into green zones, also contribute significantly to carbon output.

To address these challenges, the authors recommend a multifaceted strategy: restructuring economic models to reduce reliance on high-emission industries, investing in renewable energy, promoting compact urban planning, and incentivizing green consumer behavior. Crucially, they emphasize that machine learning models must be integrated into planning and policy frameworks, not just used as analytical tools. For this to happen, data scientists, policymakers, and urban planners must collaborate closely to translate model insights into action.

In essence, the review positions machine learning not just as a tool for understanding the past or present but as a compass for navigating a more sustainable urban future. As cities continue to grow and the urgency of climate action intensifies, ML could play a decisive role in turning data into decisions and decisions into meaningful change.