MULTI STAGE NEURAL NETWORK BASED ON ENSEMBLE LEARNING APPROACH FOR WHEAT LEAF DISEASE CLASSIFICATION

Abstract

Wheat leaf diseases significantly affect crop productivity and quality, making early and accurate detection crucial for sustainable agriculture. This study proposes a multistage neural network-based ensemble learning approach for automated wheat leaf disease classification. The framework integrates deep learning models, specifically EfficientNetB0, ResNet50, and a custom CNN with an Attention Mechanism, leveraging transfer learning for effective feature extraction. A feature fusion (concatenation) algorithm is employed to combine outputs from these three models, followed by a fully connected layer (256 units) with a 0.5 dropout rate for regularization. The system was trained on a Kaggle wheat leaf disease dataset through the Adam optimizer and categorical cross-entropy loss. The proposed architecture achieved a validation accuracy of 95.60% and a low validation loss of 0.1185, demonstrating balanced performance and resilience against overfitting. Furthermore, a real-time prediction module was implemented for practical field deployment. These results confirm the model is robust and scalable for precision agriculture, enabling timely diagnosis and improved crop management.