Data-Driven Crop Yield Prediction for Indian States: A Bi-Directional Feedforward Q-Network Based Ensemble Model and Hybrid Feature Selection

The process of predicting crop yield in Indian states includes the estimation of the quantity of crops that will be harvested in certain places. Data analytics is employed to predict yields by incorporating many factors such as historical records, weather patterns, soil conditions, and agricultural methods. An inherent limitation of conventional crop production prediction techniques for Indian states is their dependence on basic models that frequently fail to consider the complex nature of agricultural systems. These techniques may have difficulties in capturing the intricate interaction among many elements that impact agricultural yields, resulting in less precise projections. This study proposes the Bi-Directional Feedforward Q-Network (BiFFQ) model as a solution to this limitation. The technique of predicting agricultural production encompasses many sequential processes. Initially, data preprocessing is carryout, including Data cleaning and normalizing using methods like Z-score normalization. Then, the execution of feature extraction encompasses a range of statistical, temporal, geographical, climatic, and crop-specific variables. The selection of these characteristics is based on their pertinence to the forecast of crop output. Next, the implementation of feature selection through the utilization of a hybrid algorithm known as Canid Swarm Optimization (CSO), which is the combination of Golden Jackal Optimization (GJO) and White Shark Optimizer (WSO). Finally, the utilization of the proposed model Bi-Directional Feedforward Q-Network (BiFFQ) model to estimate crop production. The proposed model incorporates many functionalities, including Feedforward Neural Network (FNN), Bidirectional RNN, and Deep Q-Network (DQN), to effectively predict crop yields by utilizing the collected characteristics. The proposed model is executed utilizing the PYTHON and its effectiveness is evaluated by metrics including accuracy, precision, F-score, specificity, sensitivity, MCC, NPV, FPR, and FNR. When compared to existing methods, suggested BiFFQ achieves superior accuracy (97.64%) and precision (96.46%).