Evaluation of USDA Large Area Crop Estimation Techniques

The USDA's EDITOR system registers and digitizes the ground truth and raw Landsat data, clusters, classifies, and develops area estimates by regressing the ground truth hectarage onto the number of pixels classified per segment (a sampling unit of one square mile). A research program was conducted to evaluate the performance of EDITOR and make selected improvements to components of EDITOR. It was found that the use of multitemporal data over unitemporal significantly improved crop hectarage estimates. Performance measures on an independent test set and a jackknifed test set decreased, indicating that the current procedure of using a single data set for training the classifier, developing the regressions and evaluating the results leads to overoptimistic performance estimates. An alternative clustering algorithm, CLASSY, when substituted for the current EDITOR clustering method, produced improved estimates. Use of a simpler classifier, namely Mean Square Error classifier, did not produce significantly better hectarage estimates but showed more extendibility of the regression lines to an independent test set. The calibration approach to regression pointed out a fundamental problem in the current regression model and suggested an alternative estimation approach which has several theoretical advantages.