都市の保全緑地における植物調査への深層学習法の応用

抄録

Rivers and green spaces in urban areas are used not only for biodiversity conservation but also as a place for environmental education and relaxation for residents. Maintenance after the natural environment in these urban areas has been developed as a nature reserve is indispensable for achieving the intended function. Vegetation maps showing plant species that form the basis of ecosystems and their distribution are important information that forms the basis of maintenance. In this study, we aimed to apply a plant species identification system by deep learning and conducted a plant survey in a conservation green area that requires monitoring in order to acquire data for improving accuracy. A total of 7 plant surveys were conducted to create a data set of plant images, which could be classified with high accuracy by deep learning. However, the accuracy varies depending on the plant species, and the cause of the low correct answer rate was clarified by using SHAP analysis. By improving the imaging method for each plant and selecting the partial images for learning with a large proportion of target plants, it was possible to improve the identification of plants with low accuracy.