Geological Background Prototype Learning-Enhanced Network for Remote-Sensing-Based Engineering Geological Lithology Interpretation in Highly Vegetated Areas

Abstract

Remote-sensing-based interpretation of engineering geological lithology units provides essential lithological information for engineering planning, design, and construction projects. However, dense vegetation cover can easily interfere with the lithological spectral characteristics of optical remote sensing. Utilizing known geological background prototype-guided knowledge as a guide can mitigate interpretation interference from vegetation. Hence, for the interpretation of engineering geological lithology units, we aimed to develop a novel semantic segmentation network, LSBPnet, which combines a geological background prototype-guided learning model with a U-Net-based model. The geological background prototype-guided learning model employs lithological prototype features learned from the surrounding geological units of the predicted target to identify similarities with the predicted target's features. The model encompasses three steps: inputting known background images and known labels, learning prototype features of the background information, and learning target features based on the background information prototype perception. Using the engineering geological lithology of the Yangtze River's middle reaches as the focus area, which is a region with high vegetation cover, the proposed network achieved a 13.94% increase in the overall accuracy compared with that of commonly prevalent semantic segmentation networks. Incorporating the geological background prototype-guided knowledge helped to strengthen the distinction and relationship between the target and its surrounding lithological features, thereby improving the assessment of prediction outcomes. The diverse ablation experiments further highlighted the importance of incorporating background prototype-guided knowledge.

Keywords

• Deep learning
• engineering geological lithology units
• prototype learning
• remote sensing
• semantic segmentation