Early Detection of Water Stress by Plant Electrophysiology: Machine Learning for Irrigation Management

Problem
This preprint addresses the gap in early detection of water stress in plants, which is crucial for optimizing irrigation management in precision agriculture. Traditional methods often rely on visible symptoms of stress, which can lead to inefficient resource use. The authors propose a machine learning framework that leverages direct physiological sensing through electrophysiological signals to identify stress responses before they manifest visibly.

Method
The authors recorded time-series electrophysiological signals from greenhouse-grown tomato plants subjected to varying levels of water stress. The proposed framework consists of a processing pipeline that includes statistical feature extraction and selection, followed by either automated machine learning (AutoML) or deep learning approaches for classification. The study emphasizes the use of a 30-minute look-back window for input time horizons, which balances rapid decision-making with classification performance. The authors employed sequential backward selection to optimize the feature set while maintaining high accuracy. The final model outputs calibrated probabilities for stress detection, enhancing decision support for irrigation management.

Results
The framework achieved classification accuracies of up to 92% using AutoML, significantly outperforming deep learning methods. The results indicate that the model can effectively detect transitions from healthy to stressed states, even in recordings not included in the training set. This capability is critical for real-time applications in agricultural settings. The authors benchmark their results against standard classification metrics, demonstrating a robust performance across multiple input time horizons.

Limitations
The authors acknowledge that their study is limited to a specific crop (tomato plants) and a controlled greenhouse environment, which may not fully represent field conditions. Additionally, the reliance on electrophysiological signals may not generalize across different plant species or stress types. The study does not address the computational requirements for real-time implementation in large-scale agricultural systems, which could be a barrier to practical deployment.

Why it matters
This work has significant implications for the development of biofeedback-driven irrigation control systems, potentially leading to more efficient resource use in agriculture. By enabling early detection of water stress, the framework can help farmers optimize irrigation schedules, reduce water waste, and maintain crop health. The findings lay the groundwork for future research into automated crop management systems that integrate physiological sensing with machine learning, paving the way for advancements in precision agriculture.

Authors: Eduard Buss, Till Aust, Heiko Hamann
Source: arXiv:2604.28038
URL: https://arxiv.org/abs/2604.28038v1