Performance Evaluation of Machine Learning and Deep Learning for Rainfall Forecasting

Abstract

Climate change is a significant challenge for both humans and the environment, with its impacts increasingly felt across various regions of the world. The most evident consequence is the alteration of extreme weather patterns, which often lead to destructive and life-threatening natural disasters. Among these, extreme rainfall was the most damaging factor, frequently triggering floods. However, the increasing occurrence of related events outlined the urgent need for developing more accurate rainfall forecasting systems as a strategic measure for disaster risk reduction. This research adopted daily rainfall data from Samarinda City, collected between 2004 and 2012, to conduct prediction using both machine and deep learning methods. The implementation of machine learning methods, such as Support Vector Regression (SVR), enabled the model to learn from historical data and uncover complex patterns, resulting in accurate forecasts and improved adaptability to climate variability. Meanwhile, deep learning models, including Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM), enhanced prediction performance by capturing more intricate and abstract data relationships. Performance evaluations conducted using Mean Absolute Error (MAE) and Mean Squared Error (MSE) showed that deep learning outperformed machine learning in accuracy. The LSTM model achieved the best performance, with loss values of 0.0482 and 0.0527 for MSE and MAE, respectively. The advantage of deep learning lies in its ability to build more complex models for handling non-linear problems and to learn data representations at various levels of abstraction, which has led to more accurate results. Furthermore, LSTM surpassed RNN by effectively overcoming the vanishing gradient issue, allowing for more stable and efficient training that led to superior predictive performance.