Machine learning for gap-filling in greenhouse gas emissions databases

Area of Science:

• Environmental Science
• Data Science
• Climate Change Research

Background:

• Greenhouse gas (GHG) emissions datasets frequently contain incomplete information due to inconsistent reporting and lack of transparency.
• Accurate GHG data is crucial for effectively targeting strategies to accelerate emissions reductions.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) methods for automating the completion of incomplete GHG emissions datasets.
• To provide guidance on selecting appropriate gap-filling methods based on dataset characteristics and gap complexity.

Main Methods:

• Utilized three datasets with varying complexity and 18 distinct gap-filling techniques.
• Compared the performance of simple interpolation, extrapolation, and various ML models.
• Analyzed feature importance from ML models to identify data collection priorities.

Main Results:

• Simple interpolation is most accurate for minor gaps (e.g., missing time steps) with limited features.
• Machine learning methods outperform extrapolation for complex gaps involving non-reporting emitters when more features are present.
• Graph-based methods demonstrate scalability and ease of updating predictions with new data and multimodal sources.

Conclusions:

• Machine learning offers a powerful approach to automate GHG dataset completion, enhancing accuracy and transparency.
• Feature importance analysis from ML models can guide efforts to improve data collection efficiency.
• The study provides a framework and practical guidance for developing integrated systems for automated GHG emissions estimations.