2.2.3. Past Land-Cover Maps

CDLs, which have been produced annually using time-series Landsat images and field surveys by the USDA NASS [27,28], provide the state-level crop types at a spatial resolution of 30 m. In this study, as for past land-cover maps, CDLs prior to 2015 were used to define rule information for assigning the class labels for the self-learning process. From a preliminary test, meaningful rule information for minority classes such as sorghum and other hay could not be extracted from CDLs prior to 2010. Thus, five years of CDLs, from 2010 to 2014, were considered to extract the rule information for classification

of the 2015 data. In addition, the CDL in 2015, which was not used for classification, was used to extract reference data sets for computing classification accuracy statistics.

As classification was conducted at a 250 m spatial resolution, there was a mismatch of spatial resolution between CDLs and MODIS data sets. The CDLs were upscaled to a 250 m spatial resolution by assigning the most frequently occurring CDL class label within each 250 m pixel to that corresponding pixel. Due to mixed pixel effects, some pixels in the upscaled CDL had higher uncertainty and significantly affected the classification accuracy. To prevent this, only reliable pixels with higher confidence were considered as reference data sets. More specifically, pixels were selected if the fraction of the most prevalent class label within the pixel was greater than 0.8. In the end, a total of 64,000 pixels for six merged classes were used to compute accuracy statistics (Table 2).


