Machine Learning-Based Rainfall Forecasting with Multiple Non-Linear Feature Selection Algorithms

The present research examined the potential of two important feature selection methods, Bayesian Networks (BN) and Recursive Feature Elimination (RFE), in identifying the optimum predictors for forecasting rainfall in the Godavari Basin in India. Initially, a set of ‘probable hydro-climatological variables’ is chosen based on previous studies. Following a correlation analysis between these probable predictors and monthly rainfall, the most correlated contiguous zones were chosen as ‘potential predictors' which were subsequently used as inputs to the two feature selection algorithms, BN and RFE for selecting the ‘optimum predictors’. The optimum predictions were further utilised to develop seven state-of-the-art Machine Learning (ML) models, including Support Vector Regression (SVR), Gaussian Process Regression (GPR), Multivariate Adaptive Regression Splines (MARS), Random Forest (RF), Parallel Multi-Population Genetic Programming (PMPGP (5 demes (Case 1), 9 demes (Case 2), and 17 demes (Case 3)). The result of the correlation analysis revealed that the domain for Relative Humidity, Total Precipitable Water should be considered over the study area, and for U-Wind, V-Wind and Surface Pressure, the zones around Indian Ocean, Arabian Sea and Persian Gulf should be considered respectively. BN, as a feature selection technique for choosing the optimum predictors, was found to be more effective than RFE. In terms of prediction models, GPR and PMPGP models outperformed others, both when used alone and in conjunction with feature selection methods. The R2 values for GPR models vary from 0.82–0.41, whereas the same varies from 0.81–0.31 for PMPGP models.