開発機械・岩盤工学・資源開発技術・資源経済と社会システム（3）

[3601-05-02] Deep learning as an emerging tool for Igneous rock delineation from hyperspectral images

○Sinaice Brian¹、Kawamura Yohei¹、Shibuya Takeshi²、Sasaki Jo¹、Yoshimoto Hibiki²、Ito Yutaka¹、Ryosuke Tsuruta³、Toru Taho³、Utsuki Shinji³ (1. Akita University、2. Tsukuba University、3. Hazama Ando Corporation)

司会:笹岡 孝司(九州大学)

キーワード：Deep learning、Convolution Neural Network、Hyperspectral imaging

It is without reasonable doubt that industries today are always seeking ways to incorporate modern technology to better their outputs in terms of productivity, reliance, accuracy and time management; hence this study attempts to employ machine learning in the effort to discover the limits of deep learning technology in its abilities to delineate Igneous rocks. The underlying motivation lies in the difficulty to identify, class and name rocks for non-geologists who may need such information before conducting any type of work such as engineering of the rocks, ore, strata or the general categorization of rocks in the mining industry amongst others. Apparatus needed for the fulfilment of this study include a hyperspectral camera capable of capturing images at wavelengths beyond our human capabilities, and a deep learning software, which for this study was the deep learning. To conduct the study, it was vital to use hyperspectral images of the rocks, hence making it possible to gather the needed visual information of the rocks. The following step being training of the deep learning system to class plutonic and volcanic rocks, thereafter using the now trained system to identify and delineate each individual rock through a testing process. From the attained results, is it pretty much safe so say ‘hyperspectral imaging flawlessly assimilates with deep learning as seen from the accuracy ratings of the learning and testing processes, the two make a remarkable amalgamation. To conclude, it is hoped that this study will give incites as to how these two combined technologies can be incorporated to achieve a common goal of easy computerized rock identification with the unlimited option of expanding the database as wide as one deems adequate for their specific purpose.