A Novel Method of Multitype Hybrid Rock Lithology Classification Based on Convolutional Neural Networks

Area of Science:

• Geology and Earth Sciences
• Computer Science and Artificial Intelligence

Background:

• Accurate rock lithology recognition is crucial for geological surveys, mineral exploration, and mining engineering.
• Current methods face challenges with objectivity, rock variability, and complex hybrid lithologies, limiting accuracy and efficiency.
• Multitype hybrid rock lithology identification remains an under-researched area.

Purpose of the Study:

• To propose a novel method for multitype hybrid rock lithology detection using convolutional neural networks (CNNs).
• To enhance model inference efficiency through neural network model compression techniques.
• To evaluate the proposed method's performance against existing algorithms.

Main Methods:

• Collected datasets of four fundamental rock types: sandstone, shale, monzogranite, and tuff.
• Generated multitype hybrid rock lithology datasets using data augmentation.
• Trained and evaluated a CNN model on the hybrid datasets, comparing it with three other algorithms.
• Employed neural network model compression for efficient inference.

Main Results:

• The proposed CNN method demonstrated superior performance in precision, recall, and overall efficiency compared to the other three evaluated algorithms.
• The model achieved an inference time twice as fast as the compared methods.
• Single image detection was accomplished in just 11 milliseconds.

Conclusions:

• The developed CNN-based method offers a highly accurate and efficient solution for multitype hybrid rock lithology detection.
• The model's speed and efficiency make it suitable for industrial applications, including deployment on embedded hardware or Android platforms.