JAXA高解像度土地利用土地被覆図の改良に向けた関東域におけるカテゴリ追加および高精度化の試行

抄録

Understanding land use and land cover (LULC) is crucial for various purposes such as sustainable land use, agriculture, and environmental conservation. In Japan, the Japan Aerospace Exploration Agency (JAXA) develops and periodically releases High-Resolution Land Use and Land Cover maps of the entire Japanese territory (HRLULC-Japan). However, JAXA’s existing LULC maps have certain limitations, and some categories with unique characteristics and importance are oversimplified or aggregated into broader classes. This lack of detail hinders the map’s ability to accurately depict diverse LULC patterns. To address this issue, this study tried to add seven new categories (Wetland, Greenhouse, Deciduous orchard, Evergreen orchard, Rice intercropping zone, Lotus field, Tea farm) to HRLULC-Japan version 21.11 that was the latest version available at the start of this research, which has 12 categories. Since accuracy generally decreases as the number of categories increases, this study incorporated ensemble learning to stabilize classification results and reduce misclassification, while enhancing the overall level of detail on the LULC map. Using the Kanto area as a test site, data from Sentinel-2/MSI, Sentinel-1/GRD and ALOS-2/PALSAR-2, as well as from some other ancillary datasets such as Digital Surface Model, were used as input data for the classification algorithm “SACLASS2,” followed by ensemble learning. This method produced an LULC map with an overall accuracy of 93.5 ± 2.0 % across the 19 categories, demonstrating not only that the inclusion of new categories can improve classification detail without sacrificing high accuracy, but also that ensemble learning is effective in achieving these results.