Optimizing irrigation in peach trees by predicting trunk water potential using machine learning based on FloraPulse microtensiometers

Efficient water management in agriculture is a critical challenge to ensure sustainability, especially in water-constrained regions. Advanced technologies, such as FloraPulse microtensiometers, offer an innovative solution by enabling accurate, real-time monitoring of plant water status by measuring Trunk Water Potential (TWP). This information is essential for assessing water stress and optimising irrigation decisions. This work is the first stage of an overall objective to predict the optimal timing and amount of irrigation. It is important to note that the current models were trained and tested using data from the 2023 season, when all trees were irrigated to 100% of E T c . In this first phase, predictive models are developed that uses advanced Machine Learning (ML) and Deep Learning (DL) techniques to estimate TWP, providing a key tool to anticipate plant water stress as a function of climatic and edaphic variables. According to the initial comparison where ML models (RF, SVR and KNN) and DL model (LSTM) were implemented without using the previous TWP values as input to make the next hour prediction, the LSTM model outperformed the ML models, achieving average R 2 =0.75, MAE=0.19 MPa and CVRMSE=19.22%. Using the lagged TWP values, the LSTM model showed a significant improvement, with average R 2 =0.98, MAE=0.036MPa, and CVRMSE=4%. In addition, LSTM models with 6, 12 and 24 h prediction horizons were developed, all with R 2 greater than 0.89, MAE less than 0.12 MPa and CVRMSE less than 17%. These results allow the precise prediction of future TWP values and thus offer the design of more accurate and efficient irrigation systems. • Trunk Water Potential (TWP) data were collected using the FloraPulse microtensiometer sensor. • Machine Learning (ML) and Deep Learning (DL) models were used to predict the TWP values. • Comparisons were made between ML (RF, SVR, and KNN) and DL (LSTM) models. • LSTM models of 1, 6, 12, and 24 h prediction horizons were implemented. • Results reveal great potential for TWP prediction and irrigation scheduling.