Krish Singh: Using Machine learning to calibrate complex process-based models for carbon stock projection in NSW forests

Atmosphere/ML Workshop 2026 Atmosphere/ML Workshop 2026 Lightning Talks
1

Krish Singh: Using Machine learning to calibrate complex process-based models for carbon stock projection in NSW forests

Author Information:

Krish Singh (1*), Belinda Medlyn (1), Benjamin Smith (1,2), Bin Wang (1,3,4), Fabiano Ximenes (3)

1 Hawkesbury Institute for the Environment, Western Sydney University, Penrith, NSW 2751, Australia
2 University of Lund, Department of Physical Geography and Ecosystem Science, 223 62, Lund, Sweden
3 NSW Department of Primary Industries, Wagga Wagga Agricultural Institute, Wagga Wagga, NSW 2650, Australia
4 Gulbali Institute for Agriculture, Water and Environment, Charles Sturt University, Wagga Wagga, NSW 2678, Australia

  1. Introduction and scientific context
    Enhanced carbon sequestration by terrestrial ecosystems is a key feedback constraining the rate of global warming predicted by Earth System models. Projection of long-term terrestrial carbon sequestration within these models can be achieved using Vegetation Demography Models (VDMs). However, the demographic processes represented in VDMs are uncertain due to limited confrontation with data. Data Assimilation (DA) is a statistical framework that integrates model parameterisation and projections with data.

  2. Methodology/project summary
    Our ultimate aim is to build a DA framework to improve the predictive capabilities of the state-of-the-art VDM, LPJ-GUESS, over the forests of NSW, Australia. Machine learning (ML) is supporting this aim in two steps. We initially conducted an ML-based sensitivity analysis to identify sensitive parameters and their functional relationships to modelled Gross Primary Productivity (GPP) and Aboveground Biomass (AGB). These sensitivities will inform the design of an ML model trained to emulate LPJ-GUESS to reduce computational constraints and perform DA.

  3. Results
    Parameters relating to leaf physiology and water availability were among the most influential parameters for both model outputs. However, parameters influencing mortality and leaf turnover were relatively more important for AGB than GPP. ML enabled us to identify parameters with non-monotonic influences on model outputs.

  4. Audience
    The methodology presented here highlights a simple workflow for how one can examine complex process-based models with ML-based techniques. This methodology is valuable for model developers and users who want to identify the key processes and parameters driving the outputs of their models.

  5. Keywords
    Explainable AI, Vegetation Demography Model, Emulator, Data Assimilation, Sensitivity Analysis