Crop Yield Prediction Using Combined Long Short-Term Memory (LSTM) And Deep Belief Networks (DBN) Models: A Hybrid Deep Learning Approach

Abstract

Predicting crop yields accurately is essential for both statistical analysis and economic assessments at the farm level, helping guide investment decisions and manage agricultural imports and exports. However, crop yield prediction using artificial intelligence (AI) faces several complexities, particularly when handling heterogeneous data sources. This paper proposes a novel crop yield prediction method using a hybrid approach based on enhanced feature ranking fusion processes. The method begins with data normalization to cleanse unnecessary information, followed by the application of an improved SMOTE algorithm to enhance data for feature extraction. Feature extraction includes correlation-based features, statistical features, entropy-based features, and original information analysis to capture detailed crop growth patterns. Optimal feature selection is achieved through an enhanced feature ranking fusion technique, which incorporates chi-square, relief, and RFE methods. The prediction model integrates Long Short-Term Memory (LSTM) and Deep Belief Networks (DBN) to capture both temporal and spatial dependencies within the data. The hybrid model is validated using key performance metrics such as accuracy, precision, specificity, and sensitivity, demonstrating superior performance compared to traditional classifiers like LSTM, DBN, Convolutional Neural Networks (CNN), Bi-GRU, and Support Vector Machines (SVM). The results show that the proposed hybrid approach effectively predicts crop yield with improved accuracy and efficiency, offering valuable insights for agricultural decision-making.