Comparative Analysis of Traditional and Machine Learning Models for Rainfall Forecasting in Barishal District of Bangladesh

Rainfall prediction is crucial for agricultural planning and water resource management, as Bangladesh’s agriculture heavily depends on rainfed irrigation. Existing forecasting models are complex and costly, both budgetarily and computationally. As a result, our study evaluates the comparative performance of forecasting models, comprising two traditional time series models (Exponential Smoothing (ES) and Seasonal Autoregressive Integrated Moving Average (SARIMA)), and one machine learning model (Long Short-Term Memory (LSTM)). The monthly rainfall data for Barishal, Bangladesh, spanning the period from 1970 to 2022, were obtained from the Bangladesh Meteorological Department. The models' performance was assessed using root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (R), Nash-Sutcliffe efficiency coefficient (NSEC), and Kling-Gupta Efficiency (KGE). The ES and SARIMA models perform closely. With RMSE, MAE, R, NSEC, and KGE values of 109.35, 73.60, 0.79, 0.62, and 0.74, respectively, the ES model performs better than the SARIMA model. On the other hand, the machine learning model LSTM struggled with the test data, resulting in a higher RMSE (150.34), MAE (100.95), and lower R (0.60), NSEC (0.27), and KGE (0.60) values. This indicates that for the small dataset, the LSTM machine learning model is less effective. Therefore, our suggestion is to use a statistical model, especially the ES model, to forecast monthly rainfall in the Barishal division, as it is effective and computationally efficient. These findings are beneficial for policy development, the pesticide industry, tourism, event management, water conservation, and predicting floods and droughts.