Rainfall Prediction using Decision Tree: A Case Study of CST, Phuntsholing

Abstract

In this study we perform hourly rainfall prediction. Climatic data is chaotic in nature and performing regression analysis for short time periods, using limited data recorded by the weather station does not yield good results. Hence, in this study we consider rainfall prediction as a binary classification problem and classify rainfall events into two classes: rainy (positive class) or non-rainy (negative class). Using the independent climatic parameters of the current hour the rainfall status of the next hour is predicted. The dataset used was collected from CST weather station and contains records of 8 weather parameters recorded hourly. We want to study the usability of this data collected by CST weather station for predictive tasks. Since, there is no baseline prediction result on this dataset, we used logistic regression as the baseline model. The accuracy score of logistic regression was 73%. Decision tree which is the focus of this study to perform binary rainfall classification is a popular supervised machine learning algorithm, which forms a flowchart like structure where each internal node represents a feature. The optimization of parameters was conducted through grid search and we used k-fold validation with k value of five and we achieved an accuracy score of 79 percentage.