Seasonal auto-regressive integrated moving average with bidirectional long short-term memory for coconut yield prediction

Crop yield prediction helps farmers make informed decisions regarding the optimal timing for crop cultivation, taking into account environmental factors to enhance predictive accuracy and maximize yields. The existing methods require a massive amount of data, which is complex to acquire. To overcome this issue, this paper proposed a seasonal auto-regressive integrated moving average-bidirectional long short-term memory (SARIMA-BiLSTM) for coconut yield prediction. The collected dataset is preprocessed through a label encoder and min-max normalization is employed to change non-numeric features into numerical features and enhance model performance. The preprocessed features are selected through an adaptive strategy-based whale optimization algorithm (AS-WOA) to avoid local optima issues. Then, the selected features are given to the SARIMA-BiLSTM to predict the coconut yields. The proposed SARIMA-BiLSTM is adaptable to handling a widespread of various seasonal patterns and captures spatial features. The SARIMA-BiLSTM performance is estimated through the coefficient of determination (R2), mean absolute error (MAE), mean squared error (MSE), and root mean square error (RMSE). SARIMA-BiLSTM attains 0.84 of R2, 0.056 of MAE, 0.081 of MSE, and 0.907 of RMSE which is better when compared to existing techniques like multilayer stacked ensemble, convolutional neural network and deep neural network (CNN-DNN) and autoregressive moving average (ARIMA).