2022 Volume 42 Issue 3 Pages 199-216
Land use and land cover (LULC) maps provide essential data for ecosystem service assessment, agriculture, resource management, disaster management, etc. We have developed a multi-temporal LULC classification algorithm called "SACLASS2" based on a convolutional neural network (CNN) in a two-dimensional space spanned by a temporal axis and a feature axis. This allows for better generalizability and lowers computational costs through a simple treatment of the characteristics of time-series remote sensing data. Moreover, it can keep fine spatial patterns that may be lost when CNN is used in a geographic space. Using this algorithm in combination with a well-qualified training dataset, we took data from Sentinel-2 and ALOS-2/PALSAR-2 together with some ancillary data as input and created a new LULC map of all of Japan for the period from 2018 to 2020. The Japan Aerospace Exploration Agency (JAXA) has released this product free of charge under the title of "JAXA HRLULC version 21.11". It has been greatly improved in terms of both number of categories (12 categories, 88.85 %) and overall accuracy from the earlier version (JAXA HRLULC v18.03; 10 categories, 81.62 %), which used a previous algorithm, SACLASS, based on a kernel density estimator. Compared to other LULC maps of Japan (made by the European Space Agency, Esri, the Ministry of the Environment, and the Ministry of Land, Infrastructure, Transportation and Tourism), HRLULC-Japan v21.11 has the multiple advantages of high-spatial resolution, description of the most recent situation, suitable categories for typical LULC maps of Japan (rice paddy fields, solar panels, bamboo forests, etc.), and overall accuracy.