[M-AG33] Satellite Land Physical Processes Monitoring at Medium/High/Very High Resolution

[MAG33-P04] Creation of a 10-m Land Use Land Cover Map for the mainland Vietnam Using a Special Convolutional Neural Network

*Thinh Van Truong^1,2、Sota Hirayama³、Kenlo Nishida Nasahara⁴ (1.Degree Programs in Life and Earth Sciences, Graduate School of Science and Technology, University of Tsukuba, Tennoudai 1-1-1, Tsukuba, Ibaraki 305-8572, Japan、2.Center for VNU Development, Vietnam National University-Hanoi, Vietnam、3.Earth Observation Research Center (EORC), Japan Aerospace Exploration Agency (JAXA), Sengen 2-1-1, Tsukuba, Ibaraki 305-8505, Japan、4. Faculty of Life and Environmental Sciences, University of Tsukuba, Tennoudai 1-1-1, Tsukuba, Ibaraki 305-8572, Japan)

キーワード：Vietnam, LULC, High accuracy, CNN, Time-series

Creating a high spatial resolution and high accuracy land use land cover (LULC) map is necessary to understand human activities' impacts on the land surface. Several global LULC products have been released, such as the Dynamic World, the World Cover, and the Esri Land Cover. These LULC products attracted attention of global users due to their high spatial resolution of 10 m or their capability in providing time series information on a global scale. Even though recent global LULC products achieved significant improvements, their applicability to national scales is facing difficulties due to the lack of a detailed LULC category system, or the accuracy assessment to a national scale needs to be comprehensively conducted. Therefore, creating a high-resolution and high-reliability LULC map for the national scale remains crucial when a more detailed LULC category system or reliable accuracy assessment is needed. This study aims to create a high-accuracy and high-reliability LULC map for Vietnam by fusing satellite images from sensors such as Sentinel-2, Sentinel-1, ALOS-2/PALSAR-2, and Landsat-8. We especially employed probability maps of the Dynamic World product as input features for producing a better LULC map of Vietnam. To make a high-accuracy map, we used a software package, namely SACLASS2 created by Hirayama et al., 2022 (https://doi.org/10.11440/rssj). The algorithm of SACLASS2 is a special Convolutional Neural Network (CNN) spanning over a time-spectrum domain instead of convolution on a geospatial domain as normal CNNs. SACLASS2 combines satellite images from different sensors so that each pixel of the combined image includes information about features (bands/indices) and seasons. By conducting the convolution process over a time-spectrum domain, SACLASS2 can reflect seasonal changes while maintaining the details of spatial patterns. Therefore, it can overcome the limitation of a typical CNN, which is losing spatial detail because of the geospatial convolution process. In this study, we created a LULC map with 12 categories (Figure 1-a) by training the CNN model with more than 120,000 training data. To create a dense distribution of training data over Vietnam, we used the visual interpretation method on Sentinel-2 and Google Earth images while inheriting existing reference data produced by our colleagues at the University of Tsukuba, Japan. By doing a preliminary experiment for accuracy assessment with more than 49,000 reference data taken by visual interpretation, our LULC map achieved 93.7 % of Overall accuracy. This result is much higher than the existing 10-m LULC maps of Vietnam. In addition, we achieved high average user and producer accuracy with about 91 % and 90 %, respectively. The result also showed that SACLASS2 could produce a map with great spatial detail but resulted in many noises in the output map. By adding probability maps of the Dynamic World product, we can remove some noises and increase the accuracy of the output. This study is one of few studies which applied the result of global LULC products to create a better LULC map on a national scale. It, therefore, emphasized the necessity of releasing a probability LULC map in addition to a categorizing LULC map as the standard output of LULC making process in the future.