Five land-use/land-cover categories were automatically classi-
fied from the remote sensing images, including forest and orchard,
double-cropping land, single-cropping land, no-vegetable land,
and water. The three-date NDVI images within each crop calendar
year were stacked together and then classified by using a
supervised maximum likelihood classifier. Training sites were
first delineated from the natural-color composite of the original
Landsat images and then transferred to the NDVI composite. For
details, refer to Lu et al. (2011).
We randomly selected 144 points for each individual category
and read their land-use/land-cover types from the map in 2000.
Then, the results were compared to ground truth data collected in
field surveys and visual interpretation of China & Brazil Earth
Resource Satellite (CBERS) images and other high-spatial-resolution
images in Google Earth. The overall classification accuracy is
over 85% (Table 1). The area of double-cropping systems in
2010 was compared with that from statistical yearbooks; the
Pearson correlation coefficient was above 9.0 (Local Statistic
Bureau, 2010).