Integrating machine learning methods for computing greenhouse gas emissions baselines in agriculture

Addressing climate change requires significant reductions in global greenhouse gas (GHG) emissions. In Australia, agriculture accounts for approximately 16.8% of total emissions. Despite advancements in emissions accounting, critical gaps remain, particularly concerning the data-driven selection of reference years and the use of aggregated metrics. These issues often overlook sectoral dynamics, highlighting the need for context-specific, disaggregated baselines that more accurately reflect the complexities of emission assessment from agriculture. This study presents a novel analytical methodology that integrates machine learning (ML) techniques, specifically dynamic Generalised Additive Models (GAMs) and Conditional Inference Trees (CITs). Applied to 30 years of CO 2 -e data from Australian inventories, the ML-based approach predicted significant periods of aggregated emissions: from 1990 to 2005 and from 2006 to 2022, with 2005 identified as a critical transition point. The analysis revealed a significant decline in aggregated emissions, with a median of 70.5 Mt CO 2 -e over the past five years, down from a peak of 76.4 Mt CO 2 -e estimated in 2005. By optimising the bias-variance trade-off, the algorithms effectively captured underlying patterns while accounting for variability due to climate-related extreme events, such as the Millennium Drought. Another key contribution of this research is the establishment of accurate and adaptable sector- and activity-level baselines, which enhance decision-making related to resource allocation and performance targets. Although applied to Australian agricultural emissions data, this methodology has global applicability. As climate variability intensifies, its predictive and benchmarking capabilities can improve the utility of historical data, supporting dynamic, data-driven approaches crucial for facilitating the net-zero transition. • Dynamic smoothing splines and recursive decision trees are combined. • Capabilities are demonstrated on a 30-year CO 2 -e time series. • Static and dynamic baselines are provided for sectoral benchmarks. • Multi-factor scenarios and emission factors are exemplified. • Implications for ML-enabled carbon credit mechanisms are outlined.