A Hybrid RF-Attentive BiLSTM Framework for the Agroclimatic Mapping of Arecanut Yield in Central Karnataka

Climate variability, geographical disparity, and the lack of agricultural datasets make it difficult to estimate Arecanut crop production in Central Karnataka. Existing machine learning and deep learning models often fail to capture nonlinear climatic interactions as well as temporal climate dynamics. To address this, this research proposes a novel hybrid Random Forest (RF)–attentive Bidirectional Long Short-Term Memory (BiLSTM) framework for the agroclimatic mapping of the Arecanut crop. To handle the class imbalance problem and to enhance feature representation, the Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors (SMOTE-ENN) resampling technique is employed in the present study. Here, the RF module extracts nonlinear climatic interactions, whereas the attentive BiLSTM module captures temporal climate patterns that strongly influence the Arecanut growth cycle. The use of a self-attention model helps improve the dynamic weighting of historical observations, whereas the BiLSTM model learns relevant features, improving overall performance. The effectiveness of the developed model is evaluated through comparative analysis with other state-of-the-art solutions. The hybrid model outperforms all baseline models, achieving the highest coefficient of determination (R² = 0.914), the lowest Root Mean Square Error (RMSE = 0.39), the lowest Mean Absolute Error (MAE = 0.30), and the lowest Mean Absolute Percentage Error (MAPE = 8.1%). The high correlation coefficient (r = 0.956) further demonstrates the strong predictive capability of the model in capturing crop–climate interactions and confirms the superiority of the proposed hybrid model in agroclimatic mapping.