Classification of volcanic ash particles using a convolutional neural network and probability.

Area of Science:

• Geology
• Volcanology
• Computer Science
• Artificial Intelligence

Background:

• Volcanic ash analysis traditionally relies on subjective visual classification or objective but parameter-dependent quantitative methods.
• Classifying complex volcanic ash shapes is challenging, leading to subjectivity in qualitative analysis and the need for specific parameters in quantitative approaches.

Purpose of the Study:

• To develop and apply a convolutional neural network (CNN) for objective and accurate classification of volcanic ash particle shapes.
• To overcome the limitations of traditional methods by enabling quantitative classification of complex ash morphologies without predefined shape parameters.

Main Methods:

• Defined four fundamental volcanic ash particle shapes: blocky, vesicular, elongated, and rounded.
• Trained a CNN model using images of particles composed of single basal shapes, achieving over 90% accuracy in recognition.
• Applied the trained CNN to classify ash particles with multiple basal shapes, interpreting the output as a mixing ratio.

Main Results:

• The CNN accurately recognized the four defined basal shapes with high precision.
• The network successfully classified complex ash particles by determining their mixing ratios of basal shapes.
• Clustering analysis based on CNN output probabilities correlated with different volcanic eruption types.

Conclusions:

• The developed CNN provides an objective and quantitative method for classifying complex volcanic ash shapes.
• This approach eliminates the need for subjective categorization or selection of specific shape parameters.
• The CNN-derived mixing ratios offer a novel way to characterize volcanic ash, potentially leading to a new classification system and improved understanding of eruption dynamics.