Crop yield prediction is one of the key factors that need to be considered for sustainable agriculture and food security due to climate change and increasing population. A reliable prediction model is required to determine how much yield will be produced since climate variability exacerbates global food scarcity. Existing methodologies do not account for sequential relationships and complex interactions existing in agricultural data. Moreover, another drawback with traditional methods is that their generalization is low, and hyperparameter optimization is limited, which results in inefficient models. Therefore, this study proposes a hybrid predictive method to overcome the mentioned problems. It includes 1D CNN and BiLSTM branches optimized with the Ant Colony Optimization (ACO) and Whale Optimization Algorithm (WOA), respectively. Both branches operate on the input time series simultaneously. What makes this research unique is the fact that previously there has been no research done where the CNN and BiLSTM branches have been optimized separately using different metaheuristic approaches in the agroclimatic yield prediction scenario. The CNN helps achieve effective feature extraction, whereas the BiLSTM models capture bidirectional temporal relations. The separately optimized features are aggregated to form the final predictions. Benchmarks for statistical accuracy are used to evaluate the performance of the proposed model compared to various baseline models, including the CNN–BiLSTM model with and without attention mechanisms and optimization algorithms. With RMSE = 132.1194, MAE = 104.3635, MAE ± SD of 104.3635 ± 81.0172, and R2 = 0.9678, the proposed attention-based metaheuristic CNN–BiLSTM model has produced satisfactory results. These findings show improved performance over baseline models and competitive results against existing hybrid and attention-based methods.
Pareek et al. (Sat,) studied this question.