The Yu-no-oku gold mining ruin in the Mt. Kenashi area of Yamanashi prefecture includes the Nakayama, Uchiyama and Kayagoya gold mining ruins, which were operated during the Sengoku and Edo periods in the 15th - 17th centuries of Japanese history. The Nakayama gold ruin is registered as a Japanese national heritage site as the first mining area at the geological outcrops in Japanese history. Before gold mining at the outcrops, it involved collecting gold from rivers and/or river terraces. In this study, we created a red relief image map with a horizontal accuracy of 25 cm at the site of the Yu-no-oku gold mining ruins. The map reveals the creation of the artificial slope, which is likely to have been made in the Sengoku period. The geographical size of the slope is approximately 2 km from east to west, 1.3 km from north to south, with an altitude ranging from 900 m to 1,850 m, and an area of approximately 150 hectares. Based on the description in the ancient documents, the engineers who might create the slopes were referred to as “Kinzan-shu” and contributed to destroy the outer shell of the Fukasawa castle during the attack on the Hojo clan by Shingen Takeda in 1570. Key words

We propose a noise removal method for point cloud data acquired by multibeam echo sounders (MBES) using sparse modeling based on a continuous dictionary. For each local region (patch) of the point cloud, a local coordinate system is constructed via Singular Value Decomposition (SVD), and a coefficient matrix is formed using Discrete Cosine Transform (DCT) bases derived from the x and y coordinates. This matrix is then combined with a pre-trained dictionary, and sparse coefficients are estimated using the Orthogonal Matching Pursuit (OMP) algorithm to reconstruct the z-values and remove noise. Validation experiments using both synthetic and real MBES data showed favorable performance: the F1 score reached approximately 0.95 for synthetic data and around 0.70 for real data, reflecting a good balance between detection accuracy and false negatives. Although the recall was relatively low for real data due to missed noise points, the ability to train the dictionary using only synthetic terrain data without requiring manually labeled real data demonstrates a practical advantage. Future work includes improving reconstruction accuracy through refined threshold settings, better local region configurations, and adaptive dictionary learning using real data that resemble the target seafloor topography.