Hybrid Vision and Sequence Learning for Crop Disease Detection: A Multi-Stage Deep Learning Approach

Abstract

Rapid and precise diagnosis and identification of plant diseases are crucial to enhance agricultural production and ensure food security. There are many Machine Learning (ML) models that cannot effectively generalize in different environmental conditions, plant varieties and disease symptoms. To overcome these drawbacks, this research proposes a deep transfer learning system which is adaptive for detection and classification of cotton and rice diseases. The proposed system includes 5 major stages, which are robust feature extraction, feature fusion, Bidirectional Long Short-Term Memory (BiLSTM) network, self-attention mechanism, classification, and the rule-based fuzzy logic system. The hybrid feature extraction technique, which integrates DenseNet, EfficientNet and ResNet, is one of the major accomplishments of this research, as it extracts a variety of complementary and useful spatial features from the leaf images of disease. The extracted features are then further enhanced through the application of attention mechanisms such as Squeeze-and-Excitation (SE) Networks and Convolutional Block Attention Module (CBAM) blocks, which enable the model to better focus on the most relevant channel-wise and spatial information. Final decision-making process is enhanced by the robust and accurate fuzzy logic system, and the temporal dependencies are learned effectively with the BiLSTM network. The proposed framework is validated in the effectiveness using the publicly available datasets from Kaggle which are related to cotton leaf disease and rice leaf disease. The classification accuracy obtained from the experimental results is 99.50% on average. The comparison shows that the proposed deep learning framework outperforms the existing method in terms of accurate detection and classification of plant diseases.