Cooperation from the U.S Department of Agriculture Farm Service Agency (FSA) was crucial to the success of this project. FSA county-level offices collect and archive the most extensive farm field data available. FSA data provides information on crop type, acreage, and field location. For this project selected samples of the field data were used to both "train" computer software to differentiate between crop types as well as to assess the overall accuracy of the final LULC map product. No information regarding crop yield, owner/operator characteristics, or USDA program participation was generated or sought. Nine of the 27 study area counties were chosen to be sampled. The nine counties were chosen based on 4 broad factors: geographic distribution (an even spread north to south and east to west across the study area), crop diversity/specialization (some counties must be taken into account for their specialized crop production: Arkansas County for rice, Mississippi County for cotton), total harvested acres of principal crops, and driving distances between county offices (3-county proximity groups for travel logistics purposes). The acquisition of all field data required three weeks. Visits to individual county offices ranged from one to one-and-one-half days, and did not exceed two 8 hour days. The counties sampled were Arkansas, Chicot, Clay, Craighead, Lee, Lincoln, Lonoke, Mississippi, and Monroe. The project goal was to sample not less than 2.5% of the harvested acreage for each crop in each sampled county. The actual sampling was approximately 5% for each crop in every county. Approximately 2% of all cropland in the MAV was sampled. In addition to the data collected at the county offices, extensive field notes were produced from field observations and from discussions with farmers and FSA personnel. These notes provided valuable insight on farming practices not obtainable from other sources.

FSA filing method: Each county office archives a complete coverage of the most recent USGS NAPP aerial photographs for their county. A standard Arkansas Highway Department county series map overlain with an alpha-numeric grid system (B-13, J-4, etc.) corresponding to individually archived NAPP photographs. To find a particular farm field one must first go to the index map, find the approximate location of the farm on the map, pull the corresponding photograph, and find the farm/field on the photo. Each farm and field is given a unique identifier (farm 54 field C) That identifier is prominently labeled (indexed) on the photograph. To obtain crop information for a particular farm/field one must pull the file folder which corresponds to that index number. (Usually several fields are found in each farm folder).

("Heads Up" digitizing of farm/field data)

A random letter-number pair generator was developed and used to select photographs from the photo index mentioned above. If the selected photo had a farm-field containing a targeted crop one or more of the fields on that photo was sampled. If no fields were located on the photo it was be passed over for the next randomly selected photo. A spreadsheet was used to keep a running total of all crop categories in order to obtain crop target numbers and to maintain known proportion between crop categories (i.e. soybeans 47% of sampled area, rice 28%, cotton 16%, sorghum 8%, corn 1%). Both crop target numbers and the known proportion came from the 1992 USDA Census of Agriculture. Crop target numbers were simply the total number of sampled acres desired for each crop type. The target number for each crop was calculated by, first, determining the total number of acres for each target crop using 1992 Census of Agriculture figures, and then by multiplying the result by 2.5%. A laptop computer was used to collect the training data in the field offices. Field acreages were recorded using Microsoft Excel. PCI ImageWorks ran concurrently with the spreadsheet and was used to display a satellite image of the county with roads and Public Land Survey System overlain. First, farm fields on the NAPP photos were located on the satellite image. Next, employing heads-up digitizing methods (Edit Graphic in PCI's ImageWorks), field/crop information was reproduced as a collection of binary maps (one for each sampled crop, see "Landuse Landcover Themes" below). Before use, the binary maps were "cleaned" to remove noise. All pixels outside of a three standard deviation ellipse on a scatterplot of Summer (Spring for Wheat) TM bands 3 and 4 were removed from the samples. Thirty percent of the sampled area selected would be used for training data during feature extraction (supervised classification). Seventy percent were reserved for accuracy assessment purposes.