Enhanced rock recognition via EVSS-integrated YOLO11: A deep learning approach for precise geological classification

Area of Science:

• Geology
• Computer Science
• Artificial Intelligence

Background:

• Traditional rock identification methods are inefficient and struggle with fine-grained features.
• Accurate rock classification is crucial for resource exploration, engineering, and hazard assessment.

Purpose of the Study:

• To develop an advanced deep learning model for automated rock image identification.
• To improve the efficiency and accuracy of classifying igneous, sedimentary, and metamorphic rocks.

Main Methods:

• An improved You Only Look Once version 11 (YOLO11) model was developed, integrating the Efficient Visual State Space (EVSS) module.
• The EVSS module enhances feature extraction by modeling long-range spatial dependencies, overcoming convolutional network limitations.
• The model was evaluated against Vision Transformer (ViT), ResNet, and standard YOLO11 for classification and object detection.

Main Results:

• The EVSS-enhanced YOLO11 achieved a 92% classification accuracy, surpassing ViT (85%), ResNet (74%), and standard YOLO11 (87%).
• For object detection, EVSS-YOLO11 reached a 91.8% mAP50, outperforming the original YOLO11 (87.7%).

Conclusions:

• The EVSS-YOLO11 framework demonstrates effectiveness and robustness for rock image identification.
• This intelligent approach provides strong technical support for geological analysis and resource characterization.