[A-CG37] Biogeochemical Cycles in Land Ecosystem

[ACG37-06] Development of a low-cost, high-efficiency vegetation classification method using public data: comparison with existing methods

*Takeshi Ise¹, Noriko Kurata² (1.FSERC, Kyoto University, 2.Faculty of Intercultural Studies, Yamaguchi Prefectural University)

Keywords:deep learning, vegetation classification, aerial photographs, chopped picture method

Identification of vegetation type is essential for understanding the biogeochemical cycling over large areas and for land use planning. Since terrestrial ecosystems, especially forests, have a significant impact on the carbon cycle, it is important to understand how they work in order to mitigate climate change, but it has been extremely difficult to identify forest vegetation on a scale large enough to influence the global carbon cycle. In Japan, large-scale data has been maintained by the Ministry of the Environment, but these data are not sufficient in terms of accuracy, cost, and frequency of updating. Therefore, this study aims to achieve both high identification accuracy and cost performance by applying artificial intelligence to aerial photographs released as public data by the Geospatial Information Authority of Japan. In this study, we decided to apply the "chopped picture method," which enables classification of amorphous objects by identifying their texture. National forests in Tsuwano Town, Shimane Prefecture and Toyo Town, Kochi Prefecture were selected as target sites, and aerial photographs were acquired. From the aerial photographs, three types of images were acquired and trained to construct a vegetation identification model: cedar plantation forest, cypress plantation forest, and others. The results showed that the vegetation identification model was able to accurately classify vegetation in the target area. Its performance often exceeded that of the Ministry of the Environment's vegetation survey and GIS data from national forests, indicating that artificial intelligence may be able to replace vegetation classification in large areas, which previously relied on human labor. Because this study utilized free public data, the cost is much lower than in the past. The artificial intelligence model is capable of stable identification regardless of the characteristics, mood, and fatigue of the operator, and is also expected to increase efficiency through automation of the job. The modeling in this study can be done using YOLOv8, a relatively easy-to-implement framework, which is expected to be useful for the widespread use of this technology.