*2.2. Data*

#### 2.2.1. In-Situ Reference Data

Ground samples of six land cover types (cropland, forest, urban area, desert, waterbody, and wetlands) were visually identified from Google Earth high-resolution images in terms of polygons. These sample polygons were resampled to generate randomly placed points for each cover type (Table 2). Individual fields of five crop species were collected in a field campaign in July 2018. Crop sample points were randomly extracted from each field under the condition that the minimum distance between any two points must be no less than 20 m. This was done in ArcGIS Data Management toolbox. Stratified K fold was applied in the units of crop fields; that is, sample points from the same field can only be used for training or testing. Thus, training samples are not spatially correlated to the samples used in validation. In the crop type classification, five folds were used to verify the accuracy, and the details are shown in Table 3.




**Table 3.** Ground samples collected for major crop types in the study area.
