Skills Assessment of Selected Supervised Machine Learning Algorithms in Predicting Seasonal Rainfall over Bauchi in Nigeria

Abstract

An attempt is made to use four selected machine learning algorithms (MLAs) to predict the seasonal and monthly amount of rainfall over a Savana station in Nigeria. The four MLAs are the artificial neural network (ANN), Random Forest model (RFM), K-Nearest Neighbor (KNN), and kernel basis Support Vector Machine (SVM). Monthly mean rainfall and monthly mean air temperature data from June to October over a period of 34 years (1986-2019) were used and seventeen atmospheric variables are used to develop the model during training period. The period is divided into two, the training (1986 - 2013) and testing (2014 - 2019) periods. The results show that SVM and ANN better reproduce both monthly and annual rainfall amount over the study area by accessing their skills during training period and also having lowest RMSE and MAE during testing period. SVM is the most suitable among the four MLAs. Though, some show better results for specific month(s), the SVM and ANN summary yield 84% and 82% respectively of good forecasts for seasonal rainfall amount over Bauchi. The web interface was developed using R (ShinyR Package) programming has a very interactive and good graphical user interface (GUI) for user with little or no computer knowledge. It is recommended that the two MLAs can be used to predict monthly and seasonal rainfall over Savana climatic zone of West Africa using the seventeen input variables and hence other variables can be selected for forecasting other rainfall properties like onset, cessation and length of rainy season over West Africa sub-region. The results also show the importance and weight of each of the seventeen input variables has in reproducing the dependent variable and hence be useful in choosing which input variable can be used in further studying the dynamics of West African rain producing systems.

Keywords: Machine Learning, rainfall amount, training period, error analysis.

DOI: 10.7176/JEES/12-10-06

Publication date:October 31^st 2022