Precision and accuracy assessment of soybean cultivated area classification for United States area estimates

Soybean cultivated area has been estimated for the conterminous United States in 2011 with the use of a sampling strategy to allocate and classify samples of time-series Landsat imagery. To evaluate the Landsat soybean classification's 5m spatial resolution, RapidEye images were characterized for soybean extent for a subset of the sample population. The National Agricultural Statistical Service's (NASS) Cropland Data Layer (CDL) was also used as a basis for comparison, to investigate discrepancies at 5m spatial resolution. The Landsat and RapidEye sample block results featured nearly balanced commission (9.5%) and omission (7.5%) errors and a regression slope of 1.004. By comparison, the CDL exhibited greater commission error (15.2%) than omission error (4.5%) and a regression slope of 1.142. Locational error was also examined with respect to errors 1) within and 2) along the edges of fields. RapidEye was more sensitive to no soybean areas within fields, where commission errors found in CDL were much greater than those found in the Landsat products. Along edges, Landsat had nearly balanced omission and commission errors while the CDL layer had commission errors two times the rate of omission errors when compared to RapidEye. Additional investigation of the RapidEye data and field measurements at the field scale exposed valuable information about the accuracy and precision of the soybean area estimates for the United States and are used to calibrate and validate the national and sub-national area estimates for soybean cultivated area. RapidEye data proved to be a viable reference for assessing medium spatial resolution crop type classification. Results indicate the utility of per pixel Landsat mapping for estimation of soybean cultivated area within the United States.