Global wetland samples (based on Sentinel-1/2, Landsat, MODIS images)

To solve the high-frequency sample needs of time series wetland classification, we developed a method for automatically producing global wetland samples based on 13 global and regional wetland-related datasets and millions of images from Landsat 8 OLI, MODIS, Sentinel-1 SAR GRD, and Sentinel-2 MSI sensors. Considering the consistency of types and the separability of spectra, we summarized all classification systems into three types: wetland, water body, and non-wetland.Samples are randomly selected based on the equal-area stratified sampling scheme based on the existence probability of wetlands. In order to ensure sufficient samples, we proposed global sample size of 500,000. According to the global potential wetland distribution data set, the sample size of each grid was allocated, and samples were randomly selected. Based on 13 auxiliary data sets, we first determined the sample type according to the order of water body and wetland and assigned the "non-wetland" attribute to the type of neither water body nor wetland. The 13 auxiliary data sets include GlobeLand30 (Chen et al., 2014), FROM-GLC (Yu et al., 2013), GlobCover (Arino et al., 2010), GLC_FCS30_2020 (Liu et al., 2020), Joint Research Centre Global Surface Water Survey and Mapping map (Pekel et al., 2016), Global Reservoir and Dam Database (GRanD) (Lehner et al., 2011), Global Mangrove Watch (GMW) (Bunting et al., 2018), Global Lakes and Wetlands Database (GLWD) (Lehner et al., 2004), Murray Global Intertidal Change (MGIC) (Murray et al., 2019), CAS_Wetlands (Mao et al., 2020), CA_wetlands (Wulder et al., 2018), National Land Cover Database (NLCD) (Yang et al., 2018), Global Potential Wetland Distribution Dataset (GPWD) (Hu et al., 2017).We also included 139027 Landsat 8 OLI images, 21160 MOD09A1 images, 296479 Sentinel-1 SAR images, and 4553453 Sentinel-2 MSI images globally from January 1 to December 31, 2020. We extracted minimum, maximum, mean, and median information for each band and NDVI, NDWI, MNDWI, and LSWI indexes in four sensors of global wetland samples. In order to remove this part of the noise, this study kept the water, wetland, and non-wetland samples within one standard deviation of the annual mean of each spectral band as the sample's secondary screening conditions to ensure the accuracy of samples.The number of wetland samples determined by each sensor is different. Landsat 8 has a total of 202,111 samples, including 13,176 water bodies, 54,229 wetland samples, and 134,706 non-wetland samples; MODIS has a total of 190,898 samples, including 13,436 water body samples, 50,400 wetland samples, and 127,062 non-wetland samples ; Sentinel- has a total of 185,943 samples, including 10,885 water samples, 54,224 wetland samples, and 120,834 non-wetland samples; Sentinel-2 has a total of 185,484 samples, including 11,225 water samples, 52,142 wetland samples, and 122,117 non-wetland samples.They are stored separately in four shapefiles.