Comparative analysis of machine learning models to restore gaps in multivariate time series leaf wetness sensor data

• Leaf wetness missing data imputation through machine learning models. • Feature selection, R 2 , MSE, and KGE were evaluated for ML models’ accuracy in LW prediction. • Standalone ML models’ performance in training and validation suggests hybrid model integration. • The hybrid stacked model outperformed in predicting LW with the highest R 2 , lowest MSE, and best KGE. IoT (Internet of Things) has been utilized in agriculture in many ways. When collecting data from IoT-based sensors, errors can occur due to various factors. Exploring the ways to predict missing data is needed to ensure the reliability, accuracy, and completeness of datasets for effective decision-making and analysis. This study focused on predicting missing leaf wetness (LW) sensor data using Machine Learning (ML) models, including Multiple Linear Regression (MLR), Support Vector Regression (SVR), K-Nearest Neighbors (KNN), Gradient Boosting (GB), Long Short-Term Memory (LSTM), Random Forest (RF), Deep Neural Network (DNN) and a hybrid stacked model. Cross-validation and correlation were performed for feature selection to select significant variables and interactions. Several environmental parameters, including temperature (T; ˚C), humidity (H; %), soil moisture (SM; %), rainfall (R; mm), rain event (RE) and their interactions (R x T, R x H, R x SM) were used to predict the LW through ML models. To overcome the limitations of traditional methods in handling the spatial and temporal complexity of environmental data, a hybrid stacked machine learning model, which integrates the strengths of RF (for robust imputation on interaction terms) and DNN (for capturing complex patterns) with dropout for L1 and L2 regularization techniques, was tested. The RF model performed better in training with a low Mean Squared Error (MSE) of 0.032, a high R 2 of 0.967 and a Kling-Gupta Efficiency (KGE) of 0.919 but failed to capture LW peaks. DNN and Fused Regularized DNN (FDNN) showed moderate MSE (AM: 0.46, 0.45; PM: 0.33,0.49) and satisfactory R 2 (AM: 0.53, 0.54; PM: 0.66,0.50), KGE (AM: 0.54, 0.61; PM: 0.7, 0.56), but slightly underestimated LW in training dataset. FDNN performed well compared to RF and DNN in testing data sets, while the hybrid model outperformed other ML models. Through evaluation, it was observed that the hybrid ML model significantly outperformed individual models in both training and validation phases, achieving lower MSE (training = 0.002, validation = 0.113), higher R 2 (training = 0.963, validation = 0.887), and improved KGE (training = 0.968, validation = 0.848). This ensemble modeling approach improves the accuracy of gap-filling in LW data and contributes to precise agricultural decision-making and management practices.