Decadal rainfall forecasting using CNN–BiLSTM: a case study in Indramayu, Indonesia

Abstract

Failures in rice cultivation often result from extreme rainfall variability—ranging from droughts due to low precipitation to crop damage caused by excessive rainfall. This makes rice production highly sensitive to climate fluctuations, emphasizing the need for accurate rainfall forecasting to optimize planting and harvesting schedules. In Indramayu Regency, unpredictable rainfall patterns pose significant forecasting challenges. With the Indonesian Meteorological Agency’s (BMKG) average accuracy reaching only 63%, alternative approaches are urgently needed. This study employs a hybrid deep learning model—Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM)—to address these limitations, as previous methods such as RNN and LSTM have struggled to capture seasonal patterns and the complexity of decadal rainfall data. The model was trained on rainfall data from January 2000 to July 2022 and tested on data from August 2022 to January 2025. The model architecture includes two convolutional layers (filters 16 and 32), max pooling, and three Bi-LSTM layers (64, 50, and 32 neurons), trained using the Adam optimizer (learning rate = 0.0001), a batch size of 64, and Mean Square Error (MSE) as the loss function. Evaluation results indicate a Mean Absolute Percentage Error (MAPE) of 17.49%, classified as “good” forecasting accuracy. This translates to an overall accuracy of 82.51% (MAPE-based classification). These findings demonstrate that the CNN-BiLSTM model effectively predicts decadal rainfall in Indramayu and has the potential to reduce crop loss by optimizing agricultural strategies such as drainage, fertilization, and harvesting aligned with rainfall projections for February to April 2025.