Articles
Preliminary assessment of machine learning models to estimate stomatal conductance using capacity-based leaf sensors
Article number
1468_47
Pages
357 – 364
Language
English
Abstract
The study evaluated a low-cost capacitive leaf transpiration sensor, which detects the condensation of water vapor resulting from the leaf transpiration process.
This measure of water vapor condensation enables an estimate of plant water status by comparing the daily patterns of leaf transpiration and solar radiation.
The study was conducted on one-year-old potted lemon trees grown in a greenhouse.
Measurements included sensor capacitance, stomatal conductance (gs) with a gas exchange analyzer, and climate data.
Pearson correlation analysis was used to select relevant parameters, followed by a machine learning approach to develop a predictive model for gs. The model incorporated capacitance data with leaf temperature and vapor pressure deficit (VPD) as parameters.
Model performance was evaluated using both a standard data-splitting approach (80% training, 20% for test) and Leave-One-Out Cross-Validation (LOOCV). Support Vector Regression (SVR) and Random Forest (RF) were found to be the most effective models.
Under a data splitting approach, SVR achieved R2=0.63 and mean absolute error (MAE)=14.6 mmol m‑2 s‑1, while RF showed R2=0.58 and MAE=15.7 mmol m‑2 s‑1. LOOCV confirmed the reliability of both models, with SVR reaching R2=0.52 and MAE=15.2 mmol m‑2 s‑1, and RF achieving R2=0.45 and MAE=16.8 mmol m‑2 s‑1. These findings confirm that capacitance alone is insufficient for accurate gₛ prediction, while combining it with VPD substantially improves performance.
The capacitive leaf transpiration sensor, when integrated with additional parameters, may thus serve as a reliable, low-cost integration tool for real-time assessment of leaf gs and continuous estimates of plant water status.
This measure of water vapor condensation enables an estimate of plant water status by comparing the daily patterns of leaf transpiration and solar radiation.
The study was conducted on one-year-old potted lemon trees grown in a greenhouse.
Measurements included sensor capacitance, stomatal conductance (gs) with a gas exchange analyzer, and climate data.
Pearson correlation analysis was used to select relevant parameters, followed by a machine learning approach to develop a predictive model for gs. The model incorporated capacitance data with leaf temperature and vapor pressure deficit (VPD) as parameters.
Model performance was evaluated using both a standard data-splitting approach (80% training, 20% for test) and Leave-One-Out Cross-Validation (LOOCV). Support Vector Regression (SVR) and Random Forest (RF) were found to be the most effective models.
Under a data splitting approach, SVR achieved R2=0.63 and mean absolute error (MAE)=14.6 mmol m‑2 s‑1, while RF showed R2=0.58 and MAE=15.7 mmol m‑2 s‑1. LOOCV confirmed the reliability of both models, with SVR reaching R2=0.52 and MAE=15.2 mmol m‑2 s‑1, and RF achieving R2=0.45 and MAE=16.8 mmol m‑2 s‑1. These findings confirm that capacitance alone is insufficient for accurate gₛ prediction, while combining it with VPD substantially improves performance.
The capacitive leaf transpiration sensor, when integrated with additional parameters, may thus serve as a reliable, low-cost integration tool for real-time assessment of leaf gs and continuous estimates of plant water status.
Authors
R. Massenti, P.T. Bulacio Fischer, A. Carella, G. Milazzo, M. Thalheimer, R. Lo Bianco
Keywords
AI, Citrus × limon, capacitive leaf transpiration sensor, plant water status, random forest, transpiration, vapor pressure deficit
Groups involved
Online Articles (70)