JAMSTEC

Top > Extraction of image feature quantity using machine learning

Extraction of image feature quantity using machine learning

Sparse modeling is a mathematical and information-scientific methodology that assumes that observed high-dimensional data can be represented by a small number of essential variables. As this methodology enables maximum information extraction even from a small amount of data, it is utilized for various applications such as high-speed imaging of medical MRI, and direct observation of black holes. In this research program, we believe that sparse modeling would be effective for the analysis of seafloor geomorphology, as it is similar to the problems of imaging of MRI (Magnetic Resonance Imaging) and the observation of black holes, and by effectively using the limited available data to its fullest extent, the original seafloor geomorphology can be obtained accurately.

We are attempting to improve existing methodologies that enable the super-resolution of photographs, to apply them to seafloor topographic maps. It has become clear that not only the resolution can be improved but also patterns such as microtopography, which are characteristics of the target area, can be captured automatically. This could be considered a unique advantage of sparse modeling, which is based on a simple and clear mathematical methodology. Presently, we are attempting to link the obtained characteristic geomorphologic patterns with the knowledge of the fields of natural science, such as geology and biology, to contribute to several important scientific issues, such as disaster prevention, resource exploration, and environmental problems.

Reconstruction of topography by combination of several image patterns