Shinshu University Researcher Directory

Yusuke Hayashi, Songqiu Deng, Masato Katoh and Ryosuke Nakamura: Individual tree canopy detection and species classification of conifers by deep learning. Jpn. J. For. Plann. 55: 3~22, 2021　Recently, machine leaning （ML） and deep learning （DL） have been used to grasp tree species at the single tree level. However, the traditional methods need much experience and work time of analyzer, and it was so difficult to re-use the model to new data. Therefore, we applied a one of DL method Mask R−CNN to the UAV−based ortho image and UAV− and ALS−based shape characteristics （canopy height model （H）, slope model （S））, and tried to build a model that can full-automatically delineate individual tree crown and classify species at a new site. We created three data sets （RGB, RGB+H, RGB+S） from the multi-period data of the Shinshu University’s campus forest, and built original models to detect and classify the dominant coniferous tree species: red pine （Pinus densiflora）, larch （Larix kaempferi）, cypress （Chamaecyparis pisifera）. Using these models, Individual tree crowns and tree species were estimated at four sites located in Ina City and Minamiminowa Village, Nagano Prefecture. As a result, it became clear that the RGB+S model had the most generalization of the models with a detection rate of 0.905 and a classification accuracy of 0.955, and that it was highly reusability for new sites. In the future, it is necessary to build a model that does not depend on environmental conditions more, and optimization of DL method, improvement of learning efficiency, accumulation of data, etc. are issues.