ABSTRACT

Remotely sensed images of a Mississippi cotton field were collected with high spatial and temporal resolution during the growing season. The image data were overlaid in a GIS (geographic information system) with historical spatial data from the field: topography, soil texture, and historical cotton yield. The combination of all these data was studied to determine relationships with yield data collected at the end of the season. The motivation was to develop the ability to predict yield, well in advance of harvest, on a spatially variable basis. Such an ability would be an excellent new farm-management tool, allowing producers to better understand, in a spatially variable context, the monetary risks and returns involved in applying costly inputs such as pesticides, fertilizers, etc. Statistical analyses were conducted at grid-cell sizes from 10 m square (100 m2) to 100 m square (10,000 m2) in 10-m increments. The relationships at each cell size were calculated with data available at the beginning of the season, at the first image date, at the second image date, and so on until the last image date. Stepwise linear regression was used as the procedure for selecting variables at each date that would make up the most appropriate model to predict yield. Results indicated that the accuracy of the models at most dates was highest at the 100-m cell size. Remotely sensed data apparently added a great amount of information to the models, with most of that information being provided in the first 2 months after harvest. The highest R2 value produced by any model was 0.92, and the prediction error associated with that model was on the order of 150 lbs/ac, that in a field with yields varying from about 1500 to 2600 lbs/ac.