Plant disease detection model for edge computing devices.

Area of Science:

• Agricultural Technology
• Computer Science
• Machine Learning

Background:

• Deep learning models show promise for crop monitoring but face resource constraints on edge devices.
• Accurate plant disease detection is crucial for effective agricultural management and yield optimization.
• Current models often lack the efficiency needed for practical deployment on resource-limited agricultural hardware.

Purpose of the Study:

• To develop and optimize a deep learning model for high-accuracy plant disease classification on edge computing devices.
• To address the resource limitations of edge devices in agricultural applications.
• To enable practical, end-user accessible solutions for image-based crop monitoring.

Main Methods:

• Utilized the MobileNetV3-small architecture for image-based plant disease classification.
• Trained and evaluated the model on the PlantVillage dataset, encompassing 14 crop species and 6 disease groups.
• Applied post-training quantization to reduce model size and parameters while preserving accuracy.
• Converted the optimized model to the ONNX format for cross-platform compatibility.

Main Results:

• Achieved a test accuracy of approximately 99.50% in classifying healthy and diseased plant leaves.
• Reduced model parameters from ~1.5 million to 0.93 million through quantization, maintaining 99.50% accuracy.
• The final model in ONNX format is suitable for deployment on various platforms, including mobile devices.

Conclusions:

• High-accuracy deep learning models for plant disease detection are feasible on resource-constrained edge devices.
• Model optimization techniques like quantization are effective in reducing computational load without sacrificing performance.
• The developed ONNX-formatted model offers a cost-effective and practical solution for precision agriculture.