A deep learning model for rapid classification of tea coal disease

Area of Science:

• Agricultural Science
• Plant Pathology
• Computer Vision

Background:

• Tea coal disease (Neocapnodium theae Hara) significantly reduces tea yield and quality.
• Traditional disease identification relies on labor-intensive, subjective visual inspection.
• Developing rapid, objective disease detection methods is crucial for tea cultivation.

Purpose of the Study:

• To develop and evaluate deep learning models for rapid tea coal disease classification.
• To compare the efficacy of RGB and hyperspectral imaging for disease detection.
• To establish an accurate, non-destructive monitoring method for tea coal disease.

Main Methods:

• Utilized deep learning architectures including ResNet18, VGG16, AlexNet, and LSTM with both RGB and hyperspectral images.
• Compared classification performance of various models (e.g., WT-ResNet18 for RGB, CARS-LSTM for hyperspectral).
• Evaluated models based on classification accuracy for identifying tea coal disease.

Main Results:

• Hyperspectral imaging achieved higher classification accuracies than RGB imaging.
• The CARS-LSTM model, using hyperspectral data, demonstrated superior performance with 95% accuracy.
• Optimal RGB model (WT-ResNet18) achieved 70% accuracy, significantly lower than hyperspectral models.

Conclusions:

• Hyperspectral imaging combined with deep learning (CARS-LSTM) provides a highly accurate method for tea coal disease classification.
• This approach offers a non-destructive, efficient alternative to traditional visual inspection.
• The developed model supports effective monitoring and management of tea coal disease.