地球・資源分野ショート講演

[2K0101-12-05] VNIRとCNNを用いた花崗岩風化度評価システムの開発(発表者：修士課程)

○金 在元¹、川村 洋平¹、西川 治¹、ブライアン シナイシ　ビノ¹、岡田 夏男¹ (1. 秋田大学)

キーワード：ディープラーニング、CNN、風化度合い判断、ハイパースペクトル画像、VNIR

Experts like geotechnical engineers and geologists, can evaluate how much a rock is weathered from years of experience. Evaluating the degree of rock weathering is a very critical process as weathered rocks pose a threat to the overall structure safety such as in a dam and tunnel construction sites, and, mining sites to mention a few; reason for this being, weathering processes make rocks weaker as compared to their un-weathered counterparts. This study is proposed evaluating system to determine the weathering degree of rocks using a hyperspectral imaging and a deep learning network called Convolutional Neural Network (CNN). Using hyperspectral image data from weathered, fresh, and moderately weathered rocks, CNN learning data was obtained after data extractions had been completed. This study employs a hyperspectral image camera capable of detecting electromagnetic radiation from 400nm-1000nm with a spectral resolution of 204 bands. The CNN algorithm is based on modified a VGG-16 architecture. The total amount of data employed amounts to 19,456 files. 90% of the total data was used as learning data and the remaining 10% as testing data. Preliminary results show that the system has an average precision of over 94%. Furthermore, ‘Grad-Cam’ was used for confirming which part of hyperspectral image has the most effect on the results. The study, therefore, concludes that the conjunction of hyperspectral imaging with CNN, is a practicable process for evaluation of the degree of weathering rock without the need for specialized human expertise or prejudice.