3:30 PM - 5:00 PM
[PPS06-P14] Automatic geological Mapping by Semantic Segmentation of KAGUYA Multiband Images Using Deeplabv3+
Keywords:The Moon, Machine Learning, Segmentation
Automatic geological Mapping by Semantic Segmentation of KAGUYA Multiband Images Using DeepLabv3+
This preliminary research substitutes traditional mapping by hand with automatic mapping by machine learning. This trial is the first step of automatic segmentation and conversion from photo-geologic maps to vector data on Web-GIS.
This research selects DeepLabv3+ [1]. This is one of the most highest performance models in semantic segmentation. Training data is composed of Kaguya MI map (SELENE MOON MI 5 MAP V3.0, resolution: 14.8 m/pixel) [2] and labeled data based on a geological map of Kramer+ (2013) [3]. This labeled data has been prepared by comparison and correspondence between the MI map and the photo-geologic map. The training dataset has 381 pairs of images and labels, which includes 267 pairs as training data and 114 pairs as test data.
This trained model has given semantic segmentations at both polar regions on the Moon. Performance of the segmentation has been verified by visual inspections. We will report this trial in the poster presentation.
References:
[1] Chen et al. (2018) arXiv:1802.02611 ”Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation”
[2] KAGUYA(SELENE) Data Archive https://darts.isas.jaxa.jp/planet/pdap/selene/index.html.ja
[3] Kramer et al. (2013) j.icarus.2012.11.008 “Spectral and photogeologic mapping of Schrödinger Basin and implications for post-South Pole-Aitken impact deep subsurface stratigraphy”
This preliminary research substitutes traditional mapping by hand with automatic mapping by machine learning. This trial is the first step of automatic segmentation and conversion from photo-geologic maps to vector data on Web-GIS.
This research selects DeepLabv3+ [1]. This is one of the most highest performance models in semantic segmentation. Training data is composed of Kaguya MI map (SELENE MOON MI 5 MAP V3.0, resolution: 14.8 m/pixel) [2] and labeled data based on a geological map of Kramer+ (2013) [3]. This labeled data has been prepared by comparison and correspondence between the MI map and the photo-geologic map. The training dataset has 381 pairs of images and labels, which includes 267 pairs as training data and 114 pairs as test data.
This trained model has given semantic segmentations at both polar regions on the Moon. Performance of the segmentation has been verified by visual inspections. We will report this trial in the poster presentation.
References:
[1] Chen et al. (2018) arXiv:1802.02611 ”Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation”
[2] KAGUYA(SELENE) Data Archive https://darts.isas.jaxa.jp/planet/pdap/selene/index.html.ja
[3] Kramer et al. (2013) j.icarus.2012.11.008 “Spectral and photogeologic mapping of Schrödinger Basin and implications for post-South Pole-Aitken impact deep subsurface stratigraphy”