Authors: David Johnson*, USDA/National Agricultural Statistics Service
Topics: Agricultural Geography, Land Use and Land Cover Change, Remote Sensing
Keywords: Landsat, crop, classification, historical, change, remote sensing
Session Type: Paper
Start / End Time: 3:05 PM / 4:45 PM
Room: Balcony B, Marriott, Mezzanine Level
Presentation File: No File Uploaded
This research showcases the ability to generate crop maps for the United States (US) annually from 1984 through 2007. If successful, this would mimic the Cropland Data Layer (CDL) classifications that already exist from 2008 to present and provide a basis for future in-season mapping. A methodology is now feasible due to a convergence of factors including freely available and calibrated Landsat imagery that is analyzable with web-based distributed computing systems like Google Earth Engine. These retrospective crop cover classifications were created through steps summarized as: 1) Landsat surface reflectance data cloud-screened and median-composited into four seasonal epochs, 2) the individual composites for years 2008 - 2011 stacked and intersected with respective CDLs and random samples drawn, 3) the four years of training samples combined to provide a generalized set of predictors, 5) the samples trained using a classifier deriving ubiquitous decision tree rules, and 6) the rules applied against each of the 1984 - 2007 composite stacks creating unique classifications for each year. Analysis was constrained to the county-level in agriculturally intensive areas. Crop areas were tabulated and compared to county-level historical planted area statistics from the US Department of Agriculture for validation. Classified area by crop showed a nearly universal positive correlation over the 24-year period. However, the correlations were weak averaging an R-squared of only 0.19 for corn. Soybeans and wheat performed similarly with average correlations of 0.17 and 0.16, respectively. Thematic utility of the information was still found promising, particular after removing obvious outlier years.