Automated parameterization of stomatal conductance models from thermal imagery by leveraging synthetic images generated from Helios 3D biophysical model simulations

Abstract

Stomatal conductance ( g ) is a critical plant biophysical variable that reflects plant regulation of CO uptake and associated water loss, yet its direct measurement is often prohibitively time-consuming. Estimating the impacts of g indirectly through leaf temperature ( T ) is a common practice, but is complicated by confounding factors such as ambient conditions, measurement aggregation scale, sample size, and measurement time. Using T measurements to instead determine parameters of a model for g that can remove these external factors can provide quasi-traits that are more reliable and heritable. Our objective was to develop an automated pipeline for g model parameterization using thermal data, which could be applied within a 3D biophysical model to predict the impacts of trait variation on canopy-level processes related to water-use efficiency. Field experiments were conducted on common bean, cowpea, and sorghum crops, involving high-resolution thermal measurements obtained from a robotic sensing platform. Subsequently, a deep learning algorithm was trained using synthetic thermography data generated using Helios 3D model simulations encompassing canopy structure, ambient conditions, and T , enabling the prediction of long-wave radiation and incident shortwave radiation for each thermal image pixel. Following this, a leaf-surface energy budget analysis was applied to the collected field thermal data to predict g parameters. Validation of these predictions was performed through comparisons with ground-truth leaf-level gas exchange data. This pipeline offers a promising pathway to predictive simulations of water status and transpiration-related traits, regardless of environmental variation, ultimately enhancing our understanding of plant responses to changing environmental conditions.

Code and data availability

The supplied block contains only the title page and abstract; no public dataset, image, code, model, or supplement availability statement with an authors' URL is present.

No evidence-backed public reproduction asset is currently recorded.