Hiroko Yamaguchi and Katsuhisa Kohroki: Development and issues of forest management in the Doshi water resource forest owned by Yokohama City: Focusing on the change in management policy since the 1990s. Jpn. J. For. Plann. 56: 1～11, 2022　Recently, the public functions of forests have become increasingly important. The water resource forest in Doshi village, managed by the Yokohama City Waterworks Bureau, has been under forest management since 1919. In 1991, the management policy was changed from a policy that mainly focused on wood production with a focus on water source recharge functions to a policy that made wood production a sub-objective and prioritized conservation of head waters. This study aimed to clarify how the change of management policy regarding the Doshi water resource forest has changed the resource composition, forest operations, forestry workers, exchanges between city and mountain village, etc. We conducted a literature search and oral surveys. Analysis of the resulting data revealed that zoning has changed, moving more toward “stable natural forests”, that pruning and road network construction have been discontinued, and that timber production has become a subordinate objective in terms of operations. Since the policy change, the forest has been shifting toward becoming a fully natural forest by thinning artificial forest areas. However, it is unclear to what extent the transformation toward “stable natural forest” is a result of operations based on the management plan. Thus, it is considered necessary to evaluate the project based on vegetation surveys.

Yasushi Minowa and Shun Nakatsukasa: Identification of coniferous trees using deep learning. Jpn. J. For. Plann. 56: 13～23, 2022　We used deep learning to identify coniferous tree species based on photographs taken outdoors. We took 300 photographs of each of the eight coniferous tree species at the Kyoto Prefectural University Campus and Kyoto Botanical Garden and used these to produce 28,800 256×256-pixel images. We used Caffe as the deep learning framework, and AlexNet and GoogLeNet as the deep learning algorithms. The learning models were as follows: LM-1 had a 9:1 ratio of training data to test data without data augmentation, LM-2 had a 9:1 ratio of training data to test data with data augmentation, LM-3 had an 8:2 ratio of training data to test data without data augmentation, and LM-4 had an 8:2 ratio of training data to test data with data augmentation. To assess the performance of the learning models, 10 simulations were conducted for each learning model, and the average correct ratio for the test data was calculated. The model with the highest average correct ratio was 86.5%（LM-1, epochs = 200, GoogLeNet)without data augmentation and 88.5%（LM-3, epochs = 200, GoogLeNet)with data augmentation. The accuracy of tree species identification was slightly improved by data augmentation. As for patterns of species misclassification, Tsuga sieboldii was less likely to be misidentified, while Abies firma was more likely to be misidentified.