CNNAttLSTM: an attention-enhanced CNN-LSTM architecture for high-precision jackfruit leaf disease classification

Area of Science:

• Agricultural Science
• Computer Science
• Deep Learning

Background:

• Jackfruit leaf diseases significantly reduce crop yield and farmer income.
• Manual inspection for disease diagnosis is labor-intensive and lacks scalability.
• Existing deep learning models face challenges in generalization and computational efficiency for real-time field applications.

Purpose of the Study:

• To develop an automated and efficient system for multi-class jackfruit leaf disease detection.
• To propose a novel hybrid deep-learning architecture, CNNAttLSTM, for enhanced classification accuracy.
• To enable real-time, on-field disease diagnosis and support precision agriculture.

Main Methods:

• A hybrid deep-learning model (CNNAttLSTM) integrating Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) units, and an attention mechanism was developed.
• Images were processed into spatial patches, treated as pseudo-temporal sequences for LSTM analysis.
• Spatial and temporal features were extracted and combined with an attention mechanism for disease classification.

Main Results:

• The CNNAttLSTM model achieved 99% classification accuracy, surpassing baseline CNN (86%) and CNN-LSTM (98%) models.
• The model demonstrated high computational efficiency with 3.7 million parameters and 22ms inference time per image.
• The patch-based pseudo-temporal approach effectively captured spatial-temporal dependencies for distinguishing similar disease classes.

Conclusions:

• Combining spatial feature extraction with temporal modeling and attention significantly improves plant disease detection robustness.
• The lightweight CNNAttLSTM design facilitates real-time and edge-device deployment for agricultural applications.
• The study highlights the potential of CNNAttLSTM for scalable, accurate agricultural disease monitoring and precision agriculture.