Machine learning offers growers a new tool for predicting crop water use

A new study shows that machine-learning models can accurately predict daily crop transpiration using direct plant measurements and environmental data. By training models on seven years of high-resolution lysimeter data, the research demonstrates strong performance across tomatoes, wheat, and barley. The findings point toward future tools that may support both irrigation management and early detection of plant stress.

When it comes to irrigation, the difference between “just enough” and “too much” water can make or break a season. A new study from the Hebrew University of Jerusalem sheds light on a promising direction: a machine-learning method that predicts plant water use each day, using high-resolution data that captures how healthy plants naturally behave.

The research, jointly led by first authors Shani Friedman and Nir Averbuch under the supervision of Prof. Menachem Moshelion, brings together seven years of continuous monitoring from tomato, wheat, and barley plants grown in semi-commercial greenhouses. Using a high-precision load-cell lysimeter system—technology that records subtle changes in plant weight—the team generated highly accurate measurements of daily transpiration, the evaporation of water through leaves that reflects the plant’s water use.

By feeding these measurements into models such as Random Forest and XGBoost, the study showed that machine learning can reliably predict daily transpiration from environmental conditions and plant characteristics. In independent tests, the XGBoost model reached an R² of 0.82, closely matching measured transpiration even under differing climate conditions and in outside facilities. While the models currently rely on lysimeter-based weight data—technology that growers do not typically use in the field—they highlight an important conceptual step toward plant-driven prediction tools.

Two factors stood out as especially important: plant biomass and daily temperature. “These variables consistently shaped how much water plants consumed,” said Friedman. “Understanding how a healthy, well-irrigated plant is expected to behave on a given day also allows us to detect when something is off.”

Because the model predicts what a healthy plant should be doing, unexpected changes in transpiration may serve as early warning signs of stress, whether caused by drought, salinity, disease, root damage, or other environmental pressures. “If a plant behaves differently than the model predicts, that deviation can be an indicator of abnormal or unhealthy plant behavior,” Friedman added.

Averbuch, whose work focuses on precision irrigation, emphasized the long-term potential. “Today, many irrigation decisions still rely on indirect estimates,” he explained. “Although this model is not yet field-ready, the findings show how future systems could incorporate physiological predictions to support more accurate irrigation scheduling.”

The study comes at a time of rising interest in data-driven agriculture, especially as growers face increasing pressure from drought, heat waves, and fluctuating weather patterns. While the approach is not yet a practical farm-deployable solution, it offers a glimpse into how machine learning, environmental sensing, and plant physiology may eventually combine into tools that support both irrigation management and stress diagnostics.

Importantly, the model performed well when tested on plants grown in a different research greenhouse at Tel Aviv University, suggesting the approach could adapt across climates and production setups.

For growers, the message is clear: machine learning is becoming more than a buzzword. In the near future, predictive models based on real plant behavior may help identify stress earlier, support better water-use decisions, and improve crop resilience.