Dataset

Japan Aerospace Exploration Agency (JAXA) Earth Observation Research Center (EORC) has released several High-Resolution LULC products for Mainland Vietnam (HRLULC-Vietnam) to provide basic information for a variety of applications for the country such as biodiversity evaluation, natural disaster countermeasures, urban planning, etc. This time, JAXA has released HRLULC-Vietnam "2020VNM_v23.09", which reflects the LULC status over Mainland Vietnam (excluding some remote islands in the east) as of the year 2020 with a spatial resolution of 10m. We aimed to increase classification accuracy by using time series satellite data in combination with the CNN approach developed in the High-Resolution LULC mapping of Japan (2018-2020JPN_v21.11). We succeeded in creating a 12-category LULC for Vietnam with very high accuracy (see the accuracy verification section), which is much higher than the previous 10m LULC maps for Vietnam. The LULC classification system was inherited from the previous studies. The main differences and improvements compared with the previous HRLULC-Vietnam with a 10m spatial resolution (2016VNM_v20.06 and 2007VNM_v21.04) are as follows:

• Higher overall accuracy
• Used time series of satellite data (Sentinel-2, Sentinel-1, PALSAR-2/ScanSAR with 6 seasons)
• Applied CNN approach which spans over time and features dimensions
• A better spatial detail compared to previous versions.
• Added Aquaculture category and remove Evergreen Needleleaf Forest category.

Validation data were collected using the stratified random sampling method, following the good practice of accuracy assessment for LULC change proposed by Olofsson et al. (2014). 600 points were distributed randomly on 12 strata corresponding to 12 LULC categories. An error matrix of sample count (Table 1) was created based on the validation data. After that, the error matrix of area proportion was calculated by multiplying each row of Table 1 with the map’s area proportion of the category in that row. The overall accuracy, user’s accuracy, producer’s accuracy, and their standard deviation were calculated directly from Table 2.