Intelligent lithology identification via spectral-image fusion
Year: 2025
Type: article
Abstract: • A novel Spectral-Image fusion method for lithology identification was proposed. • Non-destructive, non-contact, fast identification of lithology was achieved via hyperspectral imaging process. • Key features were effectively captured through fusion of spectral-spatial attention mechanism. • Engineering application was implemented via customized transfer learning, non-restricted to limited sample size. Lithology identification is of vital significance for fundamental geological research and engineering applications. Traditional methods rely on rock image features and often cause confusion among visually similar rocks. To enhance identification accuracy, spectral features are integrated to better represent rock composition. Nonetheless, spectral testing inevitably damages samples and is prone to challenges of the occurrence of similar spectra for different materials. This study explores the advantages of hyperspectral imaging technology, enabling the integration of spectral and image data without damage or contact. A novel spectral-image fusion method is proposed for lithology identification. It uses a dual-channel residual neural network that combines spectral and texture feature channels. Additionally, key features are effectively captured by spectral-spatial attention mechanisms. Finally, a customized transfer learning approach is utilized to fine-tune the pre-trained network on a small dataset for lithology identification at the tunnel site, facilitating rapid model adaptation. The research indicates that employing the ResNetX2-50 network for integrating spectral-image information yields optimal identification results, with a fusion accuracy of over 99% on the test set, which is 2 percentage points higher than the accuracy of a single spectral model and about 20 percentage points higher than the accuracy of a single texture model. Research findings provide robust technical support for non-destructive, non-contact, fast lithology identification in field investigations and construction projects.
Source: Underground Space
Authors: Zhenhao Xu, Shan Li, Peng Lin, Qianji Li
Institution: Shandong University
Language: English
Subfield: Media Technology
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
Awards: G4342660735, G5897319517, G7177447721
OpenAlex ID: W4415220269