Machine learning models for prediction of rainfall over Nigeria
Olusola Samuel Ojo, Samuel Ogunjo
Abstract
Investigating climatology and predicting rainfall amounts are crucial for planning and mitigating the risks caused by variable rainfall. This study utilized two multivariate polynomial regressions (MPR) and twelve machine learning algorithms, namely three artificial neural networks (ANN), four adaptive neuro-fuzzy inference system (ANFIS) and five support vector machine (SVM) algorithms, to estimate monthly and annual rainfalls in a tropical location. The ground measured rainfall data were collected from the Nigerian Meteorological Agency (NIMET), Lagos spanning 31 years (1983–2013) spatially distributed across Nigeria. The proposed models employed geoclimatic coordinates such as longitude, latitude, and altitude as input variables. Analyses based on general performance index (c) showed that the adaptive neuro-fuzzy inference system (ANFIS) model’s algorithms outscored the MPR, ANN and SVM models in the ten months of the year. Its the generalized bell-shaped algorithm (ANFIS-GBELL) performed best for January, April, May, July, October and annual rainfalls, the Gaussian algorithm (ANFIS-GAUSS) for November and December, the subtractive clustered algorithms (ANFIS-SC) for August and September rainfalls, and fuzzy c-means algorithms (ANFIS-FCM) for June rainfall. Also, the multivariate polynomial regression of second order (MPR-2) model performed best for February and March rainfalls. These models’ algorithms have general performance index ranging from 0.906 to 0.996 and they are thereby proposed for the estimation of rainfall amounts over Nigeria.