2019 Volume 101 Issue 6 Pages 278-288
This study aims to comprehend the association between visual stem indicators in natural forest management under the selection system by employing Bayesian network analysis. A tree-marking exercise was conducted at a research plot comprising 184 trees with natural forest management at the University of Tokyo Hokkaido forest, located in northern Japan. We constructed the Bayesian network model for seven learning models (model1 - model7), with two retrieval algorithms, such as HC (pattern1 - patten3) and SA (learning iteration＝1,000, 10,000 and 100,000). Performance evaluation of the proposed model was performed with the help of Matthews correlation coefficient (MCC) index. Verification for classification accuracy for unknown data was performed with m-fold cross-validation method. The most classification accuracy without cross-validation model indicated that MCC was 0.87, which was a SA model (learning iterations＝10,000 and 100,000). That of with cross-validation model indicated that MCC was 0.44, which was a HC model (model1, pattern2). As for the association between visual stem indicators, number of parent nodes was three visual indicators for HC model and five visual indicators for SA model. Two visual indicators as parent nodes were common to two retrieval models.
