A lightweight hybrid model for scalable and robust plant leaf disease classification

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Plant diseases cause significant crop yield losses and economic damage, threatening food security.
• Accurate and early plant disease diagnosis is critical for effective management and treatment.
• Existing automated systems face challenges with computational resource limitations in field environments.

Purpose of the Study:

• To develop a lightweight and compact convolutional neural network model for efficient plant leaf disease classification.
• To address the need for rapid disease identification in field conditions with limited computational power.
• To propose the Hybrid Plant Disease Classification Network (HPDC-Net) for enhanced agricultural diagnostics.

Main Methods:

• Designed HPDC-Net with a hybrid block architecture: Depth-wise Separable Convolution Block (DSCB), Dual-Path Adaptive Pooling Block (DAPB), and Channel-Wise Attention Refinement Block (CARB).
• Utilized depth-wise separable convolutions in DSCB to extract robust features efficiently, ensuring a lightweight model.
• Trained and evaluated HPDC-Net on three datasets for classifying potato and tomato leaf diseases.

Main Results:

• Achieved over 99% accuracy on all three datasets for plant leaf disease classification.
• Maintained low computational complexity with 0.06 GFLOPs and 0.52M parameters (10 classes) / 0.17M parameters (3 classes).
• Demonstrated high inference speeds: 19.82 FPS on CPU and 408.25 FPS on GPU.

Conclusions:

• HPDC-Net is an accurate, lightweight, and efficient model for early plant disease detection.
• The model's performance indicates its suitability for real-world agricultural applications with computational constraints.
• The availability of the code on GitHub facilitates further research and implementation.