Fan et al. (2026) Bias-variance trade-off in radiative transfer model inversion drives uncertainty in leaf area index estimation

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-09-21
• Authors: Dasheng Fan, Xihan Mu, Tim R. McVicar, Yongkang Lai, Donghui Xie, Guangjian Yan
• DOI: 10.1016/j.rse.2026.115662

Research Groups

• State Key Laboratory of Remote Sensing and Digital Earth, Faculty of Geographical Science, Beijing Normal University, Beijing, China
• Beijing Engineering Research Center for Global Land Remote Sensing Products, Faculty of Geographical Science, Beijing Normal University, Beijing, China
• CSIRO Environment, Canberra, ACT, Australia

Short Summary

This study characterizes four common radiative transfer model inversion methods within a unified Bayesian framework, revealing that differences in their statistical principles lead to distinct bias-variance trade-offs, causing significant variations (30%–35%) in leaf area index (LAI) retrieval accuracy.

Objective

• To characterize how and why leaf area index (LAI) retrievals vary with the choice of radiative transfer model inversion method, within a unified Bayesian framework.

Study Configuration

• Spatial Scale: Large-scale, global/continental for satellite-based LAI retrieval.
• Temporal Scale: Not explicitly defined, but implied for continuous monitoring of LAI.

Methodology and Data

• Models used: Radiative Transfer Models (RTMs) are inverted using four representative methods characterized as statistical point estimators: numerical optimization (OPT), look-up table (LUT), LUT with solution averaging (LUT-mean), and neural network (NN)-based. Alternative machine learning methods like random forest (RF) and Gaussian process regression (GPR) were also considered.
• Data sources: Satellite observations, simulated datasets, and in situ datasets.

Main Results

• Differences in statistical principles among inversion methods result in distinct bias-variance trade-offs, causing LAI retrieval Root Mean Square Errors (RMSEs) to vary by 30%–35% across simulated and in situ datasets.
• The numerical optimization (OPT) and look-up table (LUT) methods approximate the maximum a posteriori estimator, exhibiting low bias but high variance, and produce funnel-shaped scatter plots between retrieved and reference LAIs.
• The LUT with solution averaging (LUT-mean) and neural network (NN)-based methods reduce variance by introducing bias, leading to overestimation in the mid-LAI range and underestimation in the high-LAI range, thereby producing S-shaped scatter plots.
• The NN-based method approximates the posterior mean estimator and achieves the lowest RMSE among the methods studied.
• Random forest (RF) and Gaussian process regression (GPR) methods share the same statistical interpretation as the NN-based method and exhibit similar retrieval performance.

Contributions

• Identifies the bias-variance trade-off of inversion methods as a significant, previously underappreciated component of uncertainty in LAI monitoring.
• Provides a unified Bayesian framework for understanding and comparing different RTM inversion methods.
• Highlights the critical need to consider the bias-variance trade-off in future LAI product development and evaluation.

Funding

• Not specified in the provided text.

Citation

@article{Fan2026Biasvariance,
  author = {Fan, Dasheng and Mu, Xihan and McVicar, Tim R. and Lai, Yongkang and Xie, Donghui and Yan, Guangjian},
  title = {Bias-variance trade-off in radiative transfer model inversion drives uncertainty in leaf area index estimation},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115662},
  url = {https://doi.org/10.1016/j.rse.2026.115662}
}