Multimodal Hybrid Deep Learning Approach to Detect Tomato Leaf Disease Using Attention Based Dilated Convolution...

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Tomato crops are highly susceptible to diseases, impacting yield.
• Manual disease detection is labor-intensive and prone to errors.
• Deep learning shows promise but has limitations in current plant disease detection.

Purpose of the Study:

• To develop a high-performance automated system for tomato leaf disease detection.
• To improve the speed and accuracy of identifying plant diseases.
• To address data imbalance and noise in leaf disease datasets.

Main Methods:

• Utilized an attention-based dilated Convolutional Neural Network (CNN) for feature extraction.
• Employed Bilateral filtering and Otsu segmentation for image preprocessing.
• Generated synthetic data using Conditional Generative Adversarial Network (CGAN) to handle data issues.
• Classified diseases using a logistic regression (LR) classifier after feature normalization.

Main Results:

• Achieved state-of-the-art performance on the Plant Village dataset.
• Demonstrated high accuracy: 100% training, 100% testing, and 96.6% validation accuracy for multiclass detection.
• The proposed multimodal approach offers precise, simple, and quick detection.

Conclusions:

• The attention-based dilated CNN logistic regression (ADCLR) model is effective for precise and rapid tomato leaf disease detection.
• The multimodal approach successfully handles data imbalance and noise.
• Future work includes developing a cloud-based automated classification system for various plants.