Submitted to: Geospatial Information in Agriculture and Forestry International Conference
Publication Type: Proceedings
Publication Acceptance Date: January 12, 2000
Publication Date: N/A
Interpretive Summary: Producers are currently adopting precision farming techniques and strategies throughout the U.S. These innovative producers are asking agricultural researchers to develop new and improved recommendations for fertilizers and other inputs, to better use the additional field information available from precision farming. To provide these site- specific management recommendations, a better understanding of the complex relationships between crop yield and site and soil characteristics is required. Our goal was to evaluate the predictive abilities of several neural network methods for relating crop yields to site and soil characteristics, both within ten individual site-years and across multiple site-years of data, including climatological variables. Neural networks are computer software systems that mimic the basic functions and connections of the neurons within the human brain. Neural methods were consistently more accurate on the individual site-year analyses, particularly when compared to linear statistical techniques. Multiple site-year data sets indicated that severe overfitting to climatological variables was occurring. Data sets that contain many climatologically unique site years will be required for accurate analyses of this type. This information will benefit scientists by providing additional tools for the investigation of crop response to limiting factors such as soil fertility or water holding capacity. Producers and agribusiness will also benefit through the improved recommendations and crop management strategies developed with such techniques.
Technical Abstract: Precision farming is a relatively new field of study whose goal is to improve cropping efficiency by variable application of crop treatments such as fertilizers and pesticides. A deeper understanding of the functional relationship between yield and soil and site properties is of critical importance to precision farming. Several supervised feedforward neural network methods were investigated to identify techniques able to functionally relate soil properties and crop yields on a site-specific basis, both within and across multiple site-years of data. Compared to representative linear and nonlinear, non-parametric techniques, neural methods were able to produce minimal generalization errors in every site- year, without exception. In some cases, the standard error of prediction was reduced by more than 20% compared to linear methods. In particular, resilient backpropagation produced the overall minimal generalization errors in 6 out of the 10 site-years studied. Multiple site-year analyses showed that neural techniques had enough flexibility to accurately approximate yield across site-years. However, they also indicated severe overfitting to the few unique climatological observations available and that significantly larger numbers of climatologically unique site-years of data would be required to obtain accurate results with this approach.