This study proposes a hybrid Seasonal Autoregressive Integrated Moving Average (SARIMA)–Bidirectional Long Short-Term Memory (Bi-LSTM) model for monthly rainfall forecasting in the agroclimatic zones of Chhattisgarh, India. Accurate rainfall prediction is critical for agricultural planning and water resource management, especially under increasing climate variability. The analysis utilizes 120 years (1901–2020) of monthly rainfall data, preprocessed for time series modeling. SARIMA serves as a statistical baseline, effectively capturing linear and seasonal trends, while Bi-LSTM, a deep learning model, is adept at learning long-term and non-linear dependencies. The hybrid SARIMA–Bi-LSTM model leverages the strengths of both approaches to improve forecasting accuracy. Model performance was evaluated using standard metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). Results show that Bi-LSTM outperforms SARIMA, and the hybrid model delivers the best generalization across agroclimatic zones. In the Chhattisgarh Plains, the hybrid model achieved the lowest validation RMSE (41.70 mm), MAE (25.93 mm), and the highest R² (0.906). The study highlights SARIMA’s limitations in capturing non-linearities and Bi-LSTM’s tendency to overfit, both addressed in the hybrid approach. This work demonstrates the effectiveness of hybrid models in enhancing rainfall forecasting and informs climate-resilient agricultural practices.
Naidu et al. (Mon,) studied this question.