Deep learning black box and pattern recognition analysis using Guided Grad-CAM for phytolith identification.

Area of Science:

• Computational archaeology
• Deep learning applications in archaeobotany
• Image analysis for plant fossil identification

Background:

• Deep learning models, such as VGG19, act as 'black boxes' in identifying multi-cell phytoliths.
• Visual explainers are crucial for understanding the decision-making processes of these AI models.
• Traditional archaeobotanical methods rely on expert visual identification of phytolith characteristics.

Purpose of the Study:

• To apply visual explainers to a VGG19 model for identifying phytoliths from Avena, Hordeum, and Triticum genera.
• To demonstrate the model's learning by highlighting key phytolith features used in classification.
• To compare the AI model's identification methods with those of human archaeobotanists.

Main Methods:

• Utilized Grad-CAM, Guided Backpropagation, and Guided Grad-CAM visual explainer techniques.
• Employed Guided Grad-CAM to highlight relevant regions and emphasize details in microscope images.
• Applied these methods to a trained VGG19 model for phytolith identification.

Main Results:

• The wave pattern was identified as a key decision-maker in 91% of cases.
• Papillae were significant in 86% of Avena images and 94% of images containing papillae.
• Dendritic long-cell shape was a distinctive feature in 38% of Triticum images.

Conclusions:

• Guided Grad-CAM validated established phytolith identification patterns, like the wave pattern's importance.
• Different phytolith characteristics are prominent for various genera.
• Dendritic long-cell shape was identified as a distinct classification feature.
• This research contributes to computer vision best practices in computational archaeology.