Conterminous United States crop field size quantification from multi-temporal Landsat data

Abstract

Abstract Agricultural field size is indicative of the degree of agricultural capital investment, mechanization and labor intensity, and it is ecologically important. A recently published automated computational methodology to extract agricultural crop fields from weekly 30 m Web Enabled Landsat data (WELD) time series was refined and applied to a year of Landsat 5 Thematic Mapper (TM) and Landsat 7 Enhance Thematic Mapper Plus (ETM +) acquisitions for all of the conterminous United States (CONUS). For the first time, spatially explicit CONUS field size maps and derived information are presented. A total of 4,182,777 fields were extracted with mean and median field sizes of 0.193 km 2 and 0.278 km 2 , respectively. The CONUS field size histogram was skewed; 50% of the extracted fields had sizes greater than or smaller than 0.361 km 2 , and there were four distinct peaks that corresponded closely to sizes equivalent to fields with 0.25 × 0.25 mile, 0.25 × 0.5 mile, 0.5 × 0.5 mile, and 0.5 × 1 mile side dimensions. There were discernible patterns between field size and the majority crop type as defined by the United States Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) cropland data layer (CDL) classification. In general, larger field sizes occurred where a greater proportion of the land was dedicated to agriculture, predominantly in the U.S. Wheat Belt and Corn Belt, and in regions of irrigated agriculture. The results were validated by comparison with field boundaries manually digitized from Landsat 5 and Google-Earth high resolution imagery. The validation was undertaken at 48 approximately 7.5 × 7.5 km sites selected across a gradient of field sizes in each of the top 16 harvested cropland areas in U.S. states that together cover 76% of harvested U.S. cropland. Conventional per-pixel confusion matrix based measures that assess pixel level thematic mapping accuracy, and object extraction accuracy measures, were derived. The overall per-pixel crop field classification accuracy was 92.7% and the overall crop field producer's and user's accuracies were 93.7% and 94.9%. Comparing all the reference and extracted field objects, 81.4% were correctly matched and the extracted field sizes were on average underestimated by 1.2% relative to the reference field objects.