17:15 〜 18:45
[SRD20-P01] Wide-area ore deposits exploration system using a drone equipped with a multispectral camera
★Invited Papers
キーワード：スマートマイニング、深層学習、ハイパースペクトル、マルチスペクトル、ドローン
Recent severe climate change has accelerated the shift from fossil fuels to renewable energy. Mine development is faced with challenging task of balancing increasing demands for mineral resources associated with green technology and minimizing the environmental pollutions. Under these circumstances, effective and high-spatial resolution exploration of ore deposits are strongly required to establish sustainably designed mine development model. Here, we developed wide-area exploration system designed for use in a drone equipped with a multispectral camera and spectral analysis by Convolutional Neural Networks (CNN). Traditional visible light cameras fall short in accurately identifying specific rocks and minerals, whereas hyperspectral cameras providing comprehensive assessments are hindered by bulkiness and time-consuming data processing. In this study, specialized multispectral bands were determined for ore deposits identification through dimensionality reduction of hyperspectral data to operate a multispectral camera on site.
Case study was performed to develop system for identifying the mineral sand composition on placer deposits in the Kunisaki Peninsula, Oita, Japan. Mineral sand samples were collected from No. 9 Kumano mining area at the east coast and subjected to Mineral Identification Analyzer (MLA) to quantify the accurate mineral composition ratios. Major constituent minerals of the examined samples were augite, anorthite, hornblende, albite, magnetite, quartz, and ilmenite. Commercially available samples of these minerals were also prepared and pulverized to 150–250μm diameter to demonstrate natural mineral sands with various ratios and identical grain sizes. CNN models trained with hyperspectral data of powdered commercial mineral samples could distinguish iron host minerals, magnetite and ilmenite, from other minerals in pure mineral segmentation. The model also could be applied to identify ratios of iron host minerals in natural mineral sands with sufficient accuracy. On the other hand, the leaning models were not accurate to distinguish between magnetite and ilmenite. Our experimental results also indicated that these mineral segmentation models were not applicable to cases where hyperspectral data was dimensionally reduced to 5 bands, whereas regression models using volume ratios of constituent minerals indicated sufficient accuracy in iron host mineral mapping of mixed powder samples. Our results suggest that the regression model with mineral volume ratios is promising for efficient and accurate ore deposit exploration using a multispectral drone within the mining industry.
We also introduce our new project of drone/satellite exploration of lithium ore deposits in Mongolia entitled “Next-generation mine development model for lithium ore deposits in Mongolia”, adopted for Bilateral Programs of Japan Society for the Promotion of Science in 2024.
Case study was performed to develop system for identifying the mineral sand composition on placer deposits in the Kunisaki Peninsula, Oita, Japan. Mineral sand samples were collected from No. 9 Kumano mining area at the east coast and subjected to Mineral Identification Analyzer (MLA) to quantify the accurate mineral composition ratios. Major constituent minerals of the examined samples were augite, anorthite, hornblende, albite, magnetite, quartz, and ilmenite. Commercially available samples of these minerals were also prepared and pulverized to 150–250μm diameter to demonstrate natural mineral sands with various ratios and identical grain sizes. CNN models trained with hyperspectral data of powdered commercial mineral samples could distinguish iron host minerals, magnetite and ilmenite, from other minerals in pure mineral segmentation. The model also could be applied to identify ratios of iron host minerals in natural mineral sands with sufficient accuracy. On the other hand, the leaning models were not accurate to distinguish between magnetite and ilmenite. Our experimental results also indicated that these mineral segmentation models were not applicable to cases where hyperspectral data was dimensionally reduced to 5 bands, whereas regression models using volume ratios of constituent minerals indicated sufficient accuracy in iron host mineral mapping of mixed powder samples. Our results suggest that the regression model with mineral volume ratios is promising for efficient and accurate ore deposit exploration using a multispectral drone within the mining industry.
We also introduce our new project of drone/satellite exploration of lithium ore deposits in Mongolia entitled “Next-generation mine development model for lithium ore deposits in Mongolia”, adopted for Bilateral Programs of Japan Society for the Promotion of Science in 2024.