Stepwise Calibration of Age‐Dependent Biomass in the Integrated Biosphere Simulator (IBIS) Model

Rui Ma, Yuan Zhang, Philippe Ciais, Jingfeng Xiao, Yidi Xu, Daniel S. Goll, Shunlin LiangJournal of Advances in Modeling Earth Systems 2024 · DOI

TL;DR

This paper introduces a novel "stepwise calibration" framework for land surface models (LSMs) that accounts for forest age, a crucial factor often overlooked. By integrating satellite data on GPP, LAI, and biomass with a machine learning-assisted approach, they significantly improved model accuracy and provided a more realistic estimate of forest carbon sinks, especially in young forests.

What Problem Does This Paper Address?

Many existing land surface models (LSMs) simplify forest growth by assuming a "steady-state" where forests are always at their maximum carbon storage capacity. This assumption leads to an overestimation of biomass in young forests and can introduce biases in simulating carbon fluxes and stocks. In reality, young forests are dynamic and play a significant role in the global carbon cycle. Accurately representing this age-dependent growth is crucial for improving the reliability of LSMs in predicting ecosystem responses to climate change.

Methods

The authors developed a "stepwise calibration" framework to address the problem. This approach breaks down the calibration process into three sequential steps:

  1. Fast processes (GPP): Calibrating parameters related to light-controlled carbon fluxes (Gross Primary Productivity - GPP) using satellite-derived GPP and LAI data.
  2. Intermediate processes (LAI): Calibrating parameters related to leaf area index (LAI) and leaf carbon allocation/turnover using satellite-derived LAI.
  3. Slow processes (Biomass): Calibrating parameters related to wood and fine root biomass accumulation and turnover using age-dependent biomass curves.

To overcome the computational burden of running complex LSMs repeatedly, they employed a Modified Adaptive Surrogate Modeling optimization (MASM) approach. This involves training a machine learning emulator to approximate the behavior of the IBIS model, allowing for much faster parameter exploration and optimization.

Datasets

The study utilized the following datasets:

  • Climate Forcing Data: CRUJRA v2.2 (temperature, precipitation, etc.)
  • Land Cover Data: MCD12Q1
  • Tree Cover and Lossyear Data: Hansen_2022
  • Remotely Sensed Products:
    • GLASS GPP (2000-2019)
    • GLASS LAI (2000-2020)
    • Biomass (2000-2019)
    • Forest Age (2010)

The age-dependent biomass curves were derived from the GLASS biomass and forest age products.

Key Insights and Ideas

  • Forest Age Matters: The study strongly emphasizes that ignoring forest age leads to significant overestimations of biomass and biases in carbon flux simulations. Young forests are active carbon sinks, and their dynamics are critical for accurate global carbon budget estimations.
  • Stepwise Calibration is Effective: Breaking down the calibration into "fast" and "slow" processes, guided by different data streams (GPP, LAI, biomass), allows for more targeted parameter optimization and avoids the complexities of simultaneous multi-objective calibration.
  • Machine Learning Accelerates Calibration: The use of ML emulators (MASM) drastically reduces the computational cost of calibrating complex LSMs, making it feasible to explore a wider range of parameter spaces and achieve better fits.
  • Biomass Calibration is Key for NBP: While GPP and LAI improvements are valuable, the most significant impact of age-dependent biomass calibration was seen in the Net Biome Productivity (NBP), indicating a more accurate representation of the forest's role as a carbon sink.

Key Results

  • Error Reduction: The calibrated model achieved substantial error reductions:
  • GPP: 28.5%
    • LAI: 35.3%
    • Biomass: 74.6%
  • NBP Improvement: Age-calibrated parameters increased NBP by an average of 50 gC m⁻² yr⁻¹ across all forests. This increase was particularly pronounced in boreal needleleaf evergreen forests, with an average increase of 118 gC m⁻² yr⁻¹.
  • Improved Biomass Simulation: The calibrated model showed a marked improvement in simulating biomass, especially for forests over 50 years old, and better captured the biomass dynamics of older forests where the original model underestimated carbon accumulation.
  • Computational Efficiency: The MASM approach significantly reduced the number of IBIS model runs required for calibration, from millions to a few hundred, without a substantial loss in accuracy.

Limitations and Future Work

  • Data Uncertainty: The accuracy of the biomass-age curves is crucial and is directly influenced by the uncertainties in the satellite-derived biomass and age maps, which are products of machine learning models.
  • Imperfect Fitting: Averaging age pixels and using a "space for time" approach to fit biomass-age curves can lead to a loss of variability and ignore spatial variations in growth rates.
  • Model Structure Limitations: The paper acknowledges that the IBIS model's phenology parameterization is oversimplified, leading to persistent errors in LAI simulation even after calibration. Similarly, the static allocation of carbon among plant tissues is a simplification that can affect simulated carbon pools.
  • Future Work:
    • Develop more refined phenology models to improve LAI simulation.
  • Investigate more dynamic carbon allocation strategies that account for varying climate conditions.
  • Explore regional or global-scale calibration by upscaling site-level parameters, potentially using machine learning.
  • Address the challenge of representing sub-grid heterogeneity (different PFTs and ages within a grid cell).
  • Reconstruct forest age distributions for periods beyond 2010.
  • Develop methods for incorporating temporal consistency of parameters, acknowledging potential structural and biogeophysical differences between young and old forests.

Summary updated

Abstract

Abstract Many land surface models (LSMs) assume a steady‐state assumption (SS) for forest growth, leading to an overestimation of biomass in young forests. Parameters inversion under SS will potentially result in biased carbon fluxes and stocks in a transient simulation. Incorporating age‐dependent biomass into LSMs can simulate real disequilibrium states, enabling the model to simulate forest growth from planting to its current age, and improving the biased post‐calibration parameters. In this study, we developed a stepwise optimization framework that first calibrates “fast” light‐controlled CO 2 fluxes (gross primary productivity, GPP), then leaf area index (LAI), and finally “slow” growth‐controlled biomass using the Global LAnd Surface Satellite (GLASS) GPP and LAI products, and age‐dependent biomass curves for the 25 forests. To reduce the computation time, we used a machine learning‐based model to surrogate the complex integrated biosphere simulator LSM during calibration. Our calibrated model led to an error reduction in GPP, LAI, and biomass by 28.5%, 35.3% and 74.6%, respectively. When compared with net biome productivity (NBP) using no‐age‐calibrated parameters, our age‐calibrated parameters increased NBP by an average of 50 gC m −2 yr −1 across all forests, especially in the boreal needleleaf evergreen forests, the NBP increased by 118 gC m −2 yr −1 on average, increasing the estimate of the carbon sink in young forests. Our work highlights the importance of including forest age in LSMs, and provides a novel framework for better calibrating LSMs using constraints from multiple satellite products at a global scale.
BibTeX
@article{ma2024stepwise,
  author = {Rui Ma and Yuan Zhang and Philippe Ciais and Jingfeng Xiao and Yidi Xu and Daniel S. Goll and Shunlin Liang},
  title = {{Stepwise Calibration of Age‐Dependent Biomass in the Integrated Biosphere Simulator (IBIS) Model}},
  journal = {Journal of Advances in Modeling Earth Systems},
  year = {2024},
  doi = {10.1029/2023ms004048},
}
Figure