JoWUA

Abstract

One prominent area of research focuses on predicting crop production using parameters related to crop characteristics, soil properties, water availability, and environmental conditions. Critical crop characteristics are often gathered for prediction using machine learning algorithms. Although these methods show promise in tackling the yield prediction challenge, they still possess certain limitations and drawbacks. It is unable to map the initial information as well as crop yield figures directly, either non-linearly or linearly; the quantity of the attributes obtained has a significant impact on such algorithms' performance. Deep reinforcement learning encourages new aspects and dimensions for improving the previously listed shortcomings. Deep reinforcement learning creates a comprehensive framework for predicting crop production by fusing the smarts of reinforcement training with deep learning, which enables the mapping of unprocessed information to crop 2forecasting values. In this work, we suggested a Deep Q-Network model for predicting various crop yields. Further, feature selection methods such as PCA, SelectKBest, and Random Forest Methods for feature selection. Based on that, rice, wheat, maize, cotton, and soyabean crops yield is predicted. In our comparative analysis, the Deep Q-Network model demonstrated a significant improvement over traditional methods, achieving an accuracy of 92% in predicting rice yield as opposed to the Random Forest model's 90% accuracy. The model outperformed the RF model in terms of sensitivity and specificity, both of which scored greater than 85%.