A Web-Based Machine Learning Approach for Standardized Precipitation Index Prediction

Abstract

Accurate and user-friendly drought forecasting tools are crucial for mitigating the impact of meteorological droughts, particularly in vulnerable areas such as South Sumatra, Indonesia. This study introduces an interactive web-based application built to anticipate drought conditions by forecasting the Standardized Precipitation Index (SPI). The system relies on deep learning techniques trained using three decades of rainfall data collected from the Climatological Station in South Sumatra. In evaluating model performance, both Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) architectures were assessed. While both models delivered comparable short-term predictions, the LSTM experienced a significant decline in accuracy over extended forecasting periods (specifically at SPI-6), primarily due to overfitting. In contrast, the RNN displayed more stable and reliable results, making it the preferable model for this geographical context. Specifically, the RNN achieved a lower Mean Absolute Error (MAE) of 0.4007, a reduced Root Mean Squared Error (RMSE) of 0.4684, and a higher coefficient of determination (R²) of 0.7338. These metrics outperformed those of the LSTM, which recorded a MAE of 0.4115, an RMSE of 0.4840, and an R² of 0.7036. Such results confirm that the RNN offers a more precise and dependable fit for the station’s dataset. The web platform also effectively visualizes the model outputs, providing a dynamic and interactive 24-month forecast view that supports early warning efforts and informed decision-making for regional authorities and stakeholders.