Few-shot cotton leaf spots disease classification based on metric learning.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Cotton diseases significantly impact crop yield and quality.
• Accurate identification of cotton leaf diseases is crucial for timely intervention.
• Few-shot learning offers a promising approach for classifying diseases with limited data.

Purpose of the Study:

• To develop and evaluate a few-shot learning framework for cotton leaf disease spot classification.
• To compare segmentation methods for optimal feature extraction.
• To design and implement a convolutional neural network for disease classification.

Main Methods:

• Segmentation of disease spots using Support Vector Machine (SVM) and thresholding.
• Establishment of a disease spot dataset from segmented images.
• Development of a parallel two-way convolutional neural network with weight sharing for feature extraction.
• Metric learning to create a feature space for classification.
• Comparison of convolutional neural network architectures (VGG, DenseNet, ResNet) with a spatial structure optimizer (SSO).

Main Results:

• SVM segmentation retained more informative features (color, shape, texture) compared to thresholding, aiding classification.
• DenseNet architecture demonstrated the highest classification accuracy among the tested models.
• The spatial structure optimizer (SSO) further improved the classification accuracy of DenseNet by an average of 7.7%.

Conclusions:

• The developed few-shot learning framework effectively classifies cotton leaf diseases.
• DenseNet combined with SSO shows significant potential for improving cotton disease diagnosis.
• This research provides a benchmark for future few-shot learning applications in agriculture.