Optimised MobileNet for very lightweight and accurate plant leaf disease detection

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Accurate plant disease classification is crucial for food security and climate change adaptation.
• Existing models often lack the efficiency required for real-time applications in precision agriculture.

Purpose of the Study:

• To develop a lightweight, accurate, and efficient multi-class plant leaf disease classification model.
• To design a model suitable for diverse crop types and real-time detection.

Main Methods:

• Developed PlantNet, a novel lightweight model based on a modified MobileNet architecture.
• Utilized depthwise separable convolutions, Batch Normalization (BN), and Rectified Linear Unit (ReLU) activation.
• Incorporated a multi-stage design for enhanced feature extraction.

Main Results:

• PlantNet achieved high training (99%), validation (97%), and test (98%) accuracies.
• Demonstrated excellent precision, recall, and F1-scores (0.97-1.0) across most classes.
• Outperformed larger models like ResNet-50 and Inception V3 in computational efficiency (1.46 MB, 389,286 parameters, 0.676s inference).

Conclusions:

• PlantNet offers a computationally efficient and accurate solution for plant disease detection.
• The model's compact size and performance make it ideal for real-time precision agriculture applications.
• PlantNet provides a scalable approach to managing crop health amidst global challenges.