Japan Geoscience Union Meeting 2023

Presentation information

[J] Oral

H (Human Geosciences ) » H-TT Technology & Techniques

[H-TT16] Geographic Information System and Cartography

Wed. May 24, 2023 3:30 PM - 4:45 PM 201A (International Conference Hall, Makuhari Messe)

convener:Mamoru Koarai(Earth Science course, College of Science, Ibaraki University), Kazunari Tanaka(Department of Civil Engineering and Urban Design, Faculty of Engineering, Osaka Institute of Technology), Kazuhiko W. Nakamura(The University of Tokyo), Chairperson:Mamoru Koarai(Earth Science course, College of Science, Ibaraki University), Kazunari Tanaka(Department of Civil Engineering and Urban Design, Faculty of Engineering, Osaka Institute of Technology), Kazuhiko W. Nakamura(The University of Tokyo)

3:45 PM - 4:00 PM

[HTT16-02] Land cover mapping using Earth observation satellite at cloud-prone regions

*YUHWAN SEO1, TAKAFUMI YOSHIDA1 (1.Northwest Pacific Region Environmental Cooperation Center)


Keywords:Image classification, Parcel-based classification, Binary classification, Medium spatial-temporal

Most land cover (LC) mapping is created by the multispectral imagery from earth observation satellite that has spatial and temporal advantages. Meanwhile, creating a high-accuracy LC map by previous methods for cloud-prone regions is considered difficult because it is difficult to obtain clear (cloud-free) multispectral imagery. In this study, targeting Toyama Prefecture in Japan, we developed a new LC mapping method that applies open-source data to multispectral imagery and created a LC map dataset with medium spatial-temporal resolution (MST) from 1985 to 2020. The MST in this study means a LC map dataset with biennial mapping and 30 m spatial resolution. We collected 181 Landsat images with a cloud coverage ratio of less than 50% from 1985 to 2020 and applied supervised image classification by random forest twice to make an initial MST LC map. For the first image classification, we used to classify land cover into 6 classes (water, urban, agricultural land, bareland, grassland, and forest) using the collected 181 images. For the second image classification, we stacked the 181 images biennially to classify again into 6 classes while using changes of land use among the images to improve the accuracy of the MST LC map. Next, we reclassified the obtained initial MST LC map by parcel-based classification using road and railway data from OpenStreetMap and then, binary classification using an area survey value from the Ministry of Agriculture, Forest and Fisheries in Japan. Parcel-based classification is a method used to unify the land cover within a parcel, and we performed by the most frequent land cover class in all of the parcels in the target area. If any agricultural land is mixed with other land covers within a parcel, and the agricultural land is 10 % or more of the parcel, we decided it as agricultural land. The reason for this is to avoid misclassification of orchards, vegetable fields and others, which are agricultural land, as bareland, grassland, or others and to correct this mistake, and to apply the binary classification in the next step. Binary classification is for reclassifying the overestimated agricultural land so that it matches the area survey data, and we performed it only on agricultural land. In our binary classification, we calculated the spatial index that the maximum value of the sum of the maximum NDVI and maximum NDWI was 1, for each pixel biennially. When the value of the spatial index exceeded that of the area survey data, that was set a threshold value: the value higher than the threshold was classified as agricultural land and lower one was classified as urban. Finally, we confirmed the validity of the LC change and created the final MST LC map by performing temporal correction. As a result, we could obtain a high-quality LC map dataset with an accuracy of 92 % or higher from 1985 to 2020 for Toyama Prefecture. It is the reproduced LC map dataset with high accuracy and high resolution over the past 30 years.