Prediction of Rainfall using Machine Learning Algorithms over the Guinean and Sudanian Zones in a changing Climate in Benin Republic

Abstract

Accurate precipitation forecasting is crucial for efficient climate risk management and resource planning. This study focuses on rainfall prediction using Machine Learning techniques over Guinean and Sudanian ecological zones in Benin Republic. The main objective is to develop a model capable of accurately predicting rainfall in these ecological zones. The dataset used included observed daily data on rainfall, humidity, minimum and maximum temperature, insolation, and wind speed from 1991 to 2021. Three classification algorithms were used for analysis, namely Logistic Regression, Random Forest, and Decision Tree, along with three regression algorithms, including Linear Regression, Random Forest, and Decision Tree. The results highlight the good performance of the classification algorithms in accurately distinguishing between rainy and no-rainy days during the test period in the two ecological zones. In the Guinean zone, the performance of the models is evaluated at 68%, 68% and 62%, but in the Sudanian zone, the performance is even better, with values of 81%, 81% and 75% respectively for Logistic Regression, Random Forest and Decision Tree. With regard to the prediction of daily rainfall, none of the regression models gave satisfactory results, although the Random Forest showed the best performance. In the Guinean zone, the performance of the regression models were evaluated at 9%, 16% and 2%, while in the Sudanian zone, the performances were 20%, 24% and 7% respectively for Linear Regression, the Random Forest and the Decision Tree. In order to assess the models' ability to predict rainfall on a different time scale, monthly forecasts were analysed, and satisfactory performances were observed. In the Guinean zone, the performance of the regression models was 56%, 62% and 33%, while in the Sudanian zone, the performance of the linear regression, random forest and decision tree models was 86%, 87% and 72% respectively. These results contribute to the understanding of rainfall forecasting in the context of the evolution of climate models and provide information on the applicability of machine learning techniques for rainfall prediction.