Non-coding deep learning models for tomato biotic and abiotic stress classification using microscopic images.

Area of Science:

• Plant Pathology
• Artificial Intelligence
• Computer Vision

Background:

• Accurate plant disease classification traditionally requires expert knowledge and laboratory analysis.
• Microscopic examination of plant samples is crucial for diagnosing diseases.

Purpose of the Study:

• To evaluate the efficacy of non-coding deep learning (NCDL) platforms for classifying tomato plant diseases using microscopic images.
• To assess the performance of various NCDL platforms in disease symptom classification.

Main Methods:

• Utilized microscopic images (×30) of diagnostically validated tomato plant samples.
• Employed multiple NCDL platforms including Amazon Rekognition Custom Label, Clarifai, Teachable Machine, Google AutoML Vision, Microsoft Azure Custom Vision, and Apple CreateML.
• Performed external validation to assess model robustness.

Main Results:

• NCDL platforms demonstrated high performance, with mean F1 scores ranging from 91.6% to 98.5%.
• Accuracy varied across platforms, with Amazon Rekognition Custom Label reaching 99.8% and Apple CreateML at 87.3%.
• External validation showed a minimal drop in accuracy (≤7%) for most tested NCDL platforms.

Conclusions:

• NCDL platforms offer a viable, accurate, and efficient method for classifying tomato plant diseases from microscopic images.
• These models can support the development of mobile/web applications for disease diagnosis and management.
• NCDL can improve the speed and efficiency of diagnostic laboratory sample processing through early triage.