LEVERAGING MACHINE LEARNING TO IMPROVE EVAPOTRANSPIRATION ESTIMATION AND IRRIGATION SCHEDULING IN GRAPEVINES

Shafik Kiraga

doi:10.7273/000007402

Effective irrigation scheduling depends on accurately estimating reference evapotranspiration (ETr), which is influenced by weather conditions and crop coefficients. Traditional models, such as the standardized Penman–Monteith equation (ASCE-PM), often yield unreliable results in hot, highly advective environments due to the empirical and mechanistic terms in its formulation. Additionally, the daily variability of water stress coefficients in grapevines remains underexplored. This study evaluates the potential of machine learning (ML) models as alternatives to ASCE-PM for estimating ETr across different temporal and spatial scales, assessing their transferability between regions, and exploring their application in quantifying actual evapotranspiration (ETa) in grapevines with improved stress detection. Meteorological data and lysimeter-measured ETr from well-watered alfalfa in Bushland, Texas (1996–1998), were used to develop and validate several ML models. These included support vector regression (GA-SVR), random forest (GA-RF), artificial neural networks (GA-ANN), and extreme learning machines (GA-ELM), each optimized using genetic algorithms. The models were trained to predict ETr at multiple timescales (daily, hourly, and quarter-hourly), and their performance was evaluated using RMSE, MAE, MBE, and R² metrics. Among these, the GA-ELM model achieved the highest accuracy, significantly outperforming the ASCE-PM standard. To improve spatial ETr estimation, the GA-ELM model was integrated into the METRIC (Mapping Evapotranspiration at High Resolution with Internalized Calibration) framework. The GA-ELM-calibrated METRIC model achieved an R² of 0.84 against lysimeter ET, compared to 0.74 for ASCE-PM, reducing estimation errors by up to 20%. To test model transferability, data from Bushland and Prosser, Washington, were analyzed using kernel density estimation and covariance analysis to examine the influence of meteorological variables on ETr. A semi-supervised learning approach was used to retrain the model using data from both sites. Results showed that temperature, relative humidity, and wind speed were key drivers of ETr at both locations. When tested on Bushland lysimeter data, the model achieved an MAE of 1.42 mm/day and an RMSE of 2.02 mm/day, demonstrating strong cross-regional performance. To estimate daily water stress coefficients (Ks) and their relationship to vine water status indicators, field experiments were conducted in a Cabernet Sauvignon vineyard in Prosser over three growing seasons (2022–2024), under three irrigation regimes: Full Irrigation (FI), Regulated Deficit Irrigation (RDI), and Drought Irrigation (DI). Periodic measurements included predawn (ψpd) and midday (ψmd) leaf water potential and drone imagery, while high-resolution data included soil water content and potential, sap flow, stem water potential (ψstem), and maximum daily shrinkage (MDS). ETa was estimated using sap flow (SapT), soil water balance (SWB), drone-based METRIC, the FAO method, and Ks-derived estimates. Ks was computed from normalized ψstem values in relation to relative soil water content (RSWC). Results indicated that ψpd was a more reliable indicator of vine water status than ψmd, with RSWC thresholds of 31.9% and 52.4% corresponding to ψpd values of –0.24 MPa and –0.16 MPa, respectively. Ks ranged from 0.37 to 1 across the seasons and responded to irrigation events when it dropped below 0.8. A strong relationship was observed between Ks and MDS. Although ET estimates varied significantly across methods, ETa was not significantly different but lower than ET estimates based on the FAO method. In contrast, it was significantly different from all the other methods. These findings highlight the value of integrating Ks with other soil and vine water status metrics to enhance irrigation scheduling–– an approach that provides a robust framework for improving vineyard water management through more precise ETa estimation and stress detection.

LEVERAGING MACHINE LEARNING TO IMPROVE EVAPOTRANSPIRATION ESTIMATION AND IRRIGATION SCHEDULING IN GRAPEVINES

Files and links (1)

Abstract

Metrics

Details