Performance of Machine Learning Algorithms to Predict the Rainfall Data of Khulna and Jashore District in Bangladesh

Abstract

Predicting rainfall accurately is vital for agricultural stability, disaster management, and sustainable water use, particularly in climate-vulnerable zones such as southwestern Bangladesh. This study investigates long-term rainfall patterns and develops predictive models for two key districts, Khulna and Jashore, using four supervised machine learning algorithms: Multiple Linear Regression (MLR), Random Forest (RF), Decision Tree (DT), and K-Nearest Neighbors (KNN). Based on long-term meteorological data for the years 1980–2023, i.e., temperature (maximum and minimum), humidity, wind speed, and sunshine hours, the research manifests clear spatial variations in rainfall behavior. The analysis reveals distinct spatial and temporal rainfall characteristics: Khulna exhibits higher and more variable rainfall due to its coastal location, while Jashore displays relatively stable but lower rainfall, increasing its susceptibility to dry spells. Among the models under consideration, Random Forest gave the highest accuracy value, which was the minimum RMSE and MAE and the maximum R² value, particularly in Khulna. In Jashore, although the overall model accuracy was lower, Random Forest was still the most effective because its RMSE (1.568) and MAE (1.23) were lower than those of the other models. The findings suggest that region-specific forecasting models are essential for understanding local precipitation dynamics. Furthermore, ensemble-based techniques like Random Forest prove especially capable of handling the nonlinear and irregular nature of rainfall, offering practical support for policy formulation in agriculture and climate adaptation strategies.

International Journal of Statistical Sciences, Vol. 25(2), November, 2025, pp 115-126