Citrus disease detection using convolution neural network generated features and Softmax classifier on hyperspectral...

Area of Science:

• Agricultural Science
• Computer Science
• Image Processing

Background:

• Citrus fruit quality is impacted by diseases and peel blemishes, necessitating effective identification methods.
• Existing machine vision systems have limited capabilities in distinguishing various citrus peel conditions.

Purpose of the Study:

• To develop an AI-based classification algorithm for detecting eight different citrus peel conditions using hyperspectral imaging.
• To identify the most discriminating spectral bands for classification using Principal Component Analysis (PCA).
• To compare the performance of a Convolutional Neural Network (CNN) model using PCA-selected bands versus randomly selected bands.

Main Methods:

• Acquired hyperspectral images (450-930 nm, 92 bands) of grapefruits with various peel conditions.
• Utilized PCA to select the five most discriminating spectral bands.
• Developed a novel CNN model based on VGG-16 architecture for feature extraction and SoftMax for classification.
• Compared classification performance using PCA-selected bands against randomly selected bands.

Main Results:

• PCA identified five key spectral bands (666.15, 697.54, 702.77, 849.24, 917.25 nm) for classification.
• The CNN model achieved high performance with PCA-selected bands: 99.84% accuracy, 99.84% sensitivity, and 99.98% specificity.
• Randomly selected bands yielded slightly lower performance (98.87% accuracy, 98.43% sensitivity, 99.88% specificity).

Conclusions:

• AI-based algorithms can successfully classify multiple citrus peel conditions using hyperspectral imaging.
• PCA-selected bands enhance classification accuracy, demonstrating their importance in feature selection.
• The findings support the development of real-time machine vision systems for citrus processing and quality control.