Explainable machine learning for predicting stomatal conductance across multiple plant functional types

Abstract

Stomatal conductance (gₛ) is a key leaf-level function controlling water, carbon, and energy exchange between vegetation and the surrounding environment. Conventionally, semi-empirical models have been used to model gₛ, but these models require re-parameterization as ecosystems undergo phenological changes over the growing season. In contrast, machine learning (ML) models offer a potential path to overcome this problem but are less interpretable than process-based models. This study explores ML as an approach to develop flexible and robust models of gₛ for a range of plant functional types (PFTs), including C3 crops, C3 grasses, shrubs, and tree species across different continents. An explainable machine-learning approach (eXML) was used here to provide novel interpretations and insights into the ML model formulations and relative predictor importance. We contrast the performance of three ML architectures: extreme gradient boosting, random forests, and neural networks. Models were developed and examined using many combinations of environmental and physiological predictors. The results demonstrated that ML models significantly outperform conventional semi-empirical models in predicting gₛ responses to the environment, while not requiring re-parameterization as is required in the semi-empirical paradigm. Particular focus is placed on models formulated around predictor sets that are: (a) relevant to gₛ estimation in modern terrestrial biophysical simulation models, and (b) composed of variables describing environmental and physiological drivers that can be remotely sensed non-invasively. “Generalized” models developed using data from all four PFTs demonstrated strong predictive performance using only three predictor variables, capturing 63–80 % of the variability in stomatal conductance across all ML architectures. Four predictor variables resulted in models capturing 79–83 % of gₛ variability, and models developed using all five predictor variables examined here were able to capture as much as 87 % of gₛ variability across all PFTs. Uncertainty in gₛ predictions was quantified using quantile regression. Shapley additive explanations was applied to unravel instance-based positive and negative contributions of environmental and physiological predictors to gₛ modeling, while illustrating that the models are consistent with the underlying ecophysiology. This work demonstrates the power of ML to introduce a new paradigm in the simulation of highly dynamic ecophysiological processes critical to environmental prediction.

Plant phenotyping relevance

機械学習モデルを開発・比較し、植物の生理形質である気孔コンダクタンスを推定・検証しているため、表現型取得・推定手法が研究の中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.