15:00 〜 15:15
[SCG55-06] 深層学習を活用した東青ヶ島カルデラにおける海底熱水活動の探査
キーワード：海底鉱物資源、海底熱水鉱床、深層学習、物体検出
Seafloor massive sulfide deposits are expected to be a prominent resource for various metals such as Pb, Zn, Pb, Au, Ag [1, 2]. Previous studies have shown that the signatures of hydrothermal activities can be detected as acoustic scattering anomalies in the water column images obtained by multi-beam echo sounder (MBES images) [3–5]. However, the conventional method required skilled observers carefully observing numerous MBES images to detect hydrothermal emission signals, which have required enormous time and efforts.
Recently, we have shown that object detection, a type of deep learning (DL) technique, is applicable to finding the signals from MBES images, which is fast enough for real-time observations [6]. We have also proposed a method to reduce false positives (misdetections) by time-series analysis [7].
Here, we report the practical application of the DL-based observations. MBES images were captured during research cruises KM22-11C, KM23-01 and KM23-02 at the Higashi-Aogashima Knoll Caldera hydrothermal site, Izu-Bonin Arc. First, we have confirmed the applicability of the DL system by surveying on known sites. Then, a wide-area survey within the caldera was conducted. The DL model successfully identified prominent sites of hydrothermal activities based on a vast number of images. Notably, during the succeeding cruises KM23-08_09C and KM24-09, an existence of the hydrothermal activity was newly confirmed from one of the prominent sites by ROV survey.
References:
[1] Hannington et al. (2011) Geology, 39, 1155–1158.; [2] Iizasa et al. (2019) Mineralium Deposita, 54, 117–132.; [3] Nakamura et al. (2015) Geochem. J., 49, 579–596.; [4] Kasaya et al. (2015) Geochem. J., 49, 597–602.; [5] Kaneko and Kasaya (2022) Geoinformatics, 33, 87–94. [6] Mimura et al. (2023) IEEE-JSTARS, 16, 2703–2710.; [7] Genda et al. (2025) TechRxiv. https://doi.org/10.36227/techrxiv.173895078.87591845/v1.
Recently, we have shown that object detection, a type of deep learning (DL) technique, is applicable to finding the signals from MBES images, which is fast enough for real-time observations [6]. We have also proposed a method to reduce false positives (misdetections) by time-series analysis [7].
Here, we report the practical application of the DL-based observations. MBES images were captured during research cruises KM22-11C, KM23-01 and KM23-02 at the Higashi-Aogashima Knoll Caldera hydrothermal site, Izu-Bonin Arc. First, we have confirmed the applicability of the DL system by surveying on known sites. Then, a wide-area survey within the caldera was conducted. The DL model successfully identified prominent sites of hydrothermal activities based on a vast number of images. Notably, during the succeeding cruises KM23-08_09C and KM24-09, an existence of the hydrothermal activity was newly confirmed from one of the prominent sites by ROV survey.
References:
[1] Hannington et al. (2011) Geology, 39, 1155–1158.; [2] Iizasa et al. (2019) Mineralium Deposita, 54, 117–132.; [3] Nakamura et al. (2015) Geochem. J., 49, 579–596.; [4] Kasaya et al. (2015) Geochem. J., 49, 597–602.; [5] Kaneko and Kasaya (2022) Geoinformatics, 33, 87–94. [6] Mimura et al. (2023) IEEE-JSTARS, 16, 2703–2710.; [7] Genda et al. (2025) TechRxiv. https://doi.org/10.36227/techrxiv.173895078.87591845/v1.