Bayesian optimized multimodal deep hybrid learning approach for tomato leaf disease classification.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Manual identification of tomato leaf diseases is labor-intensive and prone to inaccuracies.
• An automated system is crucial for early disease detection and timely intervention to maintain crop yield and quality.

Purpose of the Study:

• To develop and evaluate robust Bayesian optimized deep hybrid learning models for automated tomato leaf disease classification.
• To compare the performance of seven hybrid models, including Convolutional Neural Network (CNN) and various machine learning classifiers, for disease identification.

Main Methods:

• Proposed seven deep hybrid learning models combining CNN for feature extraction with classifiers like Random Forest, XGBoost, SVM, and others.
• Incorporated a Boruta feature filtering layer for statistically significant feature selection.
• Utilized the PlantVillage dataset for training and testing, employing various statistical classification metrics.

Main Results:

• The CNN-Stacking model demonstrated the highest classification performance among the seven hybrid models.
• Achieved average precision, recall, f1-score, MCC, and accuracy exceeding 98% on an unseen dataset.
• The models exhibited high time efficiency, with testing times as low as 0.174 seconds, and demonstrated generalizability across challenging image conditions.

Conclusions:

• The developed hybrid deep learning approach offers a computationally inexpensive and superior alternative to existing methods for tomato leaf disease diagnosis.
• The high accuracy and efficiency of the CNN-Stacking model support its potential for integration into a real-time smartphone application for farmers.
• This technology can empower farmers with timely disease diagnosis and management strategies, thereby enhancing tomato production.