Estimation of Leaf Wetness Duration Using Machine Learning Models

Abstract

The leaf wetness duration (LWD) is one of the most critical parameters related to the infection rate and development of plant diseases, as many pathogens require the presence of free water on plant organs to infect leaf tissue. For this reason, this study evaluated three machine learning models: Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network Multilayer Perceptron (MLP), using hourly surface meteorological inputs to estimate LWP. The models were trained and tested using 20 years of meteorological data for seven agricultural locations situated in the southern region of the state of Minas Gerais and the Triângulo Mineiro region. The models were compared with an empirical model based on observed relative humidity (NHUR>90%). The results indicated that machine learning models are capable of estimating LWD; however, among all methods tested, the MLP model demonstrated the best performance, with high accuracy (mean R² = 0.98) and low errors (mean RMSE = 27.6 minutes and mean MAE = 19.8 minutes). The results also showed that the models are sensitive to the study locations, with models for the southern Minas Gerais locations performing better than those for the Triângulo Mineiro region. This discrepancy is due to the lower LWD in the Triângulo Mineiro locations, which were underestimated by the models, resulting in lower performance.