Intelligent On-Site Lithology Identification Based on Deep Learning of Rock Images and Elemental Data
Zhenhao Xu, Heng Shi, Peng Lin, Wen Ma
Abstract
To reduce the disadvantages of factors such as limited expression of two-dimensional images, weathering, or man-made pollution on image-based methods and realize rapid and accurate on-site lithology identification, we propose an intelligent lithology identification method based on the deep learning of rock images and element information. The initial lithology probabilities and locations are first determined from the images using a object detection model, then the initial lithology probabilities and element data are used as input to correct the lithology probabilities using self-constructed fusion model, and finally all lithology detection results are integrated and redundant results are removed. The verification experiment show that the accuracy of the proposed method is 13.73% higher than that of the image-based method. The proposed method improves the performance of lithology identification by improving the classification of lithology and influencing the choice of bounding box. The proposed fusion identification method is a real-time and accurate method for the identification of complex lithology and provides a new idea for the rapid and intelligent identification of on-site lithology.