Evaluation of deep learning models using explainable AI with qualitative and quantitative analysis for rice leaf...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning models excel at disease detection but lack transparency, raising trust issues.
• Traditional metrics like accuracy don't assess if models use relevant features.
• Explainable AI (XAI) is crucial for understanding model decisions.

Purpose of the Study:

• To develop and validate a three-stage methodology for evaluating deep learning model accuracy and reliability.
• To combine traditional performance metrics with XAI visualization analysis.
• To assess both classification performance and feature selection relevance.

Main Methods:

• Evaluated eight pre-trained deep learning models (ResNet50, InceptionResNetV2, etc.).
• Employed a three-stage methodology: 1) traditional metrics, 2) LIME for feature visualization and IoU/DSC analysis, 3) a novel overfitting ratio metric.
• Quantitatively assessed feature selection and model reliance on insignificant features.

Main Results:

• ResNet50 achieved the highest accuracy (99.13%) and reliability (IoU: 0.432, overfitting ratio: 0.284).
• InceptionV3 and EfficientNetB0 showed high accuracy but poor feature selection (low IoU, high overfitting ratio), indicating potential reliability issues.
• The proposed methodology provides a comprehensive evaluation beyond simple accuracy.

Conclusions:

• The developed methodology enhances the evaluation of deep learning models for accuracy and reliability.
• It enables the creation of more trustworthy AI systems, particularly for agricultural applications.
• The generic methodology can be extended to other domains requiring transparent AI.