ニューラルネットワークを用いた地質データの空間分布推定と分布に影響を及ぼす因子の抽出

概要

論文の詳細を見る
Geostatistics have been widely used for regionalized geologic data having spatial correlation structures. In some cases, these structures cannot be clarified by a semivariogram due to limited data and/or complexity of the data distribution. We applied a feedforward neural network to such kind of problem by choosing chemical content data in a kuroko-deposits area in the Hokuroku district, Akita prefecture in northern Japan as a case study. The content data of multiple chemical components were collected from 1917 sample points at 143 drillhole sites. The network structure consists of three layers: input, middle, and output. Input data of the network are coordinates of a sample point, Bouguer anomaly (gravity anomaly), and rock code. Contents of principal metals of kuroko, Cu, Pb, and Zn, and value of alteration index are estimated through the network. Numbers of neurons in the middle layer and training data vary with the location to improve the estimation accuracy and avoid overtraining of the network. Suitable numbers of neurons and training data were determined by repeating jackknifing at each point. Three-dimensional distributions of Cu, Pb, and Zn were successfully estimated by the proposed method, as shown by anomaly zones localized near the known kuroko deposits and the margins of depressions or uplifts in Pre-Tertiary basement rocks. Using the weights connecting adjacent neurons, sensitivity analysis of the neural network was carried out. Influence factors were obtained in this analysis to clarify geologic factors that affect the distributions of Cu, Pb, and Zn contents. Depth, the dacite in the middle-upper Nishikurosawa geologic stage, and Bouguer anomaly were found to be influenceable factors. Sensitivity vectors were also calculated through the analysis. Large vectors point toward anomaly zones near the kuroko deposits and express the anisotropy in the local distribution of the sample data. It is proved by semivariogram that the data have spatial correlation along the dominant direction of vectors. The sensitivity vector map contributes to infer flow paths of hydrothermal fluids that contained the sources of kuroko ores.
社団法人資源・素材学会の論文
2003-07-25

ニューラルネットワークを用いた地質データの空間分布推定と分布に影響を及ぼす因子の抽出

スポンサーリンク

概要

著者

関連論文

スポンサーリンク