[H-RE12] New Developments in Engineering Geology

[HRE12-P08] Estimation of Arsenic Contaminated Area in Mudstone and Tuff in Boring Cores Using Hyperspectral Imaging.

*Kaito Takizawa¹, Yoko Ohtomo², Natsuo Okada¹, Shinichiro Nakamura¹, Hirotada Kuroki³, Youhei Kawamura² (1.Graduate School of Engineering, Hokkaido University, 2.Faculty of Engineering, Hokkaido University, 3.Mechatronics Engineering Department, Technical Research Institute, Aoki Asunaro Construction Co.)

Keywords:Spectroscopy, Environmental assessment, Convolutional Neural Network, Tunnel excavation, Arsenic contamination, Naturally occurring heavy metals

Excavated rocks contaminated by hazardous metals has been regarded as a serious problem in tunnel construction. The treatment of the contaminated rocks is costly and time-consuming, which leads to an increase of the unplanned total construction cost and delays. Moreover, existing methods for identifying the contaminated rocks require experts. Here, we proposed a new user-friendly method to distinguish the arsenic contaminated zones in boring core of the tunnel construction that may leach arsenic to water quickly and inexpensively by hyperspectral imaging and deep learning. Tunnel boring cores including zones where arsenic dissolved in water above environmental standard values were subjected to this study. These cores mainly consisted of mudstone, tuff breccia and gravel. Mudstone were partially silicified probably due to the local hydrothermal alteration. arsenic concentrations in the boring cores were determined by handheld X-ray Fluorescence analyzer. Mudstone generally showed low arsenic concentrations, while tuff breccia showed higher arsenic concentrations than those of other rock types. arsenic concentrations in the boring cores indicated a rough positive correlation with amounts of arsenic elution. Microscopic observations and chemical composition analysis using electron probe micro analyzer (EPMA) of thin sections presented that pyrite was the main host mineral for arsenic. Hyperspectral images of the boring cores were classified based on the rock types and arsenic concentrations, input into a Convolutional Neural Network (CNN) to predict the arsenic contaminated areas. The experimental results indicated that the prediction using algorithm based on the individual characteristics of the spectral shape with different arsenic concentrations showed low accuracy of 40~70%, whereas algorithm based on the arsenic concentrations combined with the detail rock classifications indicated high accurate prediction of 80%. These results suggest that hyperspectral imaging combined with CNN is promising to estimate arsenic contaminated areas in the boring cores, which leads to low-cost and rapid detection of hazardous metal-contaminated excavated rocks in tunnel construction.