Application of Binary Classification Method of ANN in Rainfall Prediction of Minna, Niger State, Nigeria: The Multiple Thresholds Approach

Abstract

In a previous work, an improved rainfall prediction model for Minna metropolis using Artificial Neural Network (ANN) was arrived at by the application of the default (single) threshold of 0.5 in the binary classification method of ANN. In this present work, attempt was made to further improve the work using the multiple (variable) thresholds method. The threshold was varied from 0.1 to 1.0 in steps of 0.1. Sensitivity rating was chosen as the performance metric for rain prediction. The best result was obtained using a threshold of 0.4 which has a 72.5% sensitivity rating against that of 63% of the default threshold as previously obtained, implying that with a 0.4 threshold, the network predicts (classifies) rainfall of Minna much better. Other metrics obtained are specificity and accuracy values of 85% and 73% respectively. This improved result has 60 rainfall days predicted out of 32 rainfall days available in the data set which is an indication of more rainfall and probable flooding. The Area Under the Receiver Operating Characteristics (ROC) curve value of 0.57% obtained correspond to the interval 0.5 > 0.59. This places the rainfall classification (prediction) carried out in this work in the “Good” category. More rainfall days predicted serves as an alarm to residents and relevant stakeholders toembark on mitigation in order to minimise the negative impact of heavy rainfall. It was recommended that for more accurate rainfall prediction, robust data encompassing more atmospheric parameters be used for network training.