Prediction of three vital rainfall characteristics: Advanced hybrid tree- or lazy-based learner?

• Rainfall volume, maximum daily rainfall, and number of rainy days are predicted. • Hybrid machine learning models of tree- and lazy-based models developed and compared. • Different input combinations are investigated. • Rotation forest-Kstar hybrid model was found to be the most effective model in all cases. • ML-based models are efficient tool in quick and accurate prediction of rainfall characteristics. Rainfall is considered the most important factor affecting floods, soil erosion, water quality, and groundwater resources. Therefore, accurate rainfall prediction is imperative for the efficient management of land and water resources. This study was aimed at investigating the capabilities of ensemble machine learning algorithms in predicting rainfall characteristics, including the monthly rainfall volume, maximum daily rainfall, and number of rainy days. Specifically, tree-based (random forest and dual perturb and combine tree), and lazy learner (Kstar and instance-based K-nearest neighbors) models, along with their hybridized with rotation forest model (ROF) were used for the rainfall predictions. To meet the aim, nine rainfall-related weather variables, including minimum, maximum, and mean relative humidity; solar radiation; minimum, maximum, and mean temperature; evaporation; and wind speed were considered as inputs in Fars province, Iran. All models were evaluated using qualitative and standard statistical metrics. Minimum relative humidity was found to be the most significant/sensitive input variable in all cases (correlation coefficient between 0.77 and 0.85), while wind speed was found to be the least effective variable (correlation coefficient between 0.15 and 0.20). Minimum relative humidity data alone were sufficient to build efficient input scenarios for all models for the given region and timeframe. Based on the performance metrics, the ROF-Kstar hybrid model was found to be the most effective and robust for the prediction of all the response variables with Nash Sutcliff Efficiency ranges between 0.65 and 0.70. Overall, lazy-based learner models outperformed the tree-based models. This study highlights the potential of proposed novel techniques to predict rainfall characteristics, which can be extended to other domains.