# Water status diagnosis in greenhouse drip-irrigated tomato and celery using leaf turgor dynamics and machine learning

## TL;DR

This study uses a non-invasive probe and machine learning to monitor water stress in greenhouse tomatoes and celery, enabling more precise irrigation.

## Contribution

A novel non-invasive method combining leaf turgor dynamics and machine learning for real-time crop water status monitoring in drip-irrigated greenhouses.

## Key findings

- Diurnal patterns of Pp in tomato and celery showed two distinct states corresponding to different moisture stress levels.
- Machine learning models integrating Pp states and environmental factors achieved high accuracy in predicting water status.
- Non-full irrigation increased Pp extremes, indicating higher stress levels in both crops.

## Abstract

Accurate crop water status monitoring is crucial for optimized irrigation in controlled environments, but traditional approaches relying on damaging measurements or sporadic sampling frequently restrict real-time evaluation.

This study explored the non-invasive leaf patch clamp pressure (LPCP) probe to evaluate the water status of drip-irrigated tomato and celery. Leaf turgor dynamics analysis enabled the characterization of the LPCP probe’s output parameter (Pp) and its environmental drivers, and the development of predictive machine learning models.

The results indicated that diurnal patterns of Pp in drip-irrigated tomato and celery exhibited two distinct states: State I (unimodal) and State II (troughed), corresponding to moisture conditions with no or mild stress, and severe stress, respectively. The soil water content (SWC) thresholds for State I were set at SWC > 20% (tomato) and SWC > 19% (celery), whereas those for State II were set at SWC < 18% (tomato) and SWC < 16% (celery). For State I, Pp was positively associated with solar radiation but negatively associated with SWC (in tomato) and wind speed (in celery). For State II, the associations between Pp and environmental parameters were less than those in State I. Interestingly, compared to full irrigation, non-full irrigation treatments not only showed a higher proportion of State II but also resulted in an increase in both Pp,max and Pp,min by 15.39%–138.39% in tomato and 3.44%–94.02% in celery. These analytical results yielded four model parameter combinations based on the inclusion of SWC and the management of distinct Pp states. The prediction model that integrated Combination 4 (substate Pp prediction based on meteorological factors and SWC) with the random forest approach exhibited the highest accuracy (R2 = 0.995, MSE = 2.419, RMSE = 1.540, and MAE = 0.531), with SWC identified as its key feature parameter.

These findings provide a scientific foundation for optimizing the precision irrigation of greenhouse vegetables in drip systems.