Deep Learning-Based Intelligent Apple Variety Classification System and Model Interpretability Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Accurate fruit classification is crucial for agricultural management and quality control.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), show promise for image-based agricultural applications.
• Interpretability of CNN decisions remains a challenge in critical applications.

Purpose of the Study:

• To evaluate and compare the performance of different CNN architectures (series and DAG networks) for apple classification.
• To investigate the impact of dataset configuration on model accuracy and generalization.
• To enhance the interpretability and credibility of CNN models in agricultural contexts.

Main Methods:

• Employed transfer learning with series networks (AlexNet, VGG-19) and directed acyclic graph (DAG) networks (ResNet variants).
• Utilized two distinct training datasets with varying training-to-testing ratios to assess model robustness.
• Applied feature visualization, strongest activations, and Local Interpretable Model-agnostic Explanations (LIME) for model interpretability.

Main Results:

• All models achieved high accuracy (>96.1%) on Dataset A (2.4:1.0 ratio), significantly outperforming Dataset B (1.0:3.7 ratio) (89.4-93.9%).
• VGG-19 demonstrated superior performance, reaching 100% accuracy on Dataset A and 93.9% on Dataset B.
• Deeper networks within the same framework generally showed increased model size, accuracy, and processing time.
• Visualization techniques provided insights into model decision-making processes for apple image classification.

Conclusions:

• Dataset configuration is a critical factor influencing CNN performance in agricultural image classification.
• VGG-19 is a highly effective model for this specific apple classification task.
• Model depth influences performance trade-offs, and interpretability methods enhance trust in AI-driven agricultural solutions.