|Sudduth, Kenneth - Ken|
Submitted to: World Congress on Neural Networks
Publication Type: Proceedings
Publication Acceptance Date: 5/4/1998
Publication Date: N/A
Interpretive Summary: Producers and scientists are currently investigating and adopting precision farming techniques and strategies. Many of these strategies are based upon conventional input recommendation procedures. These recommendations necessarily assume that all factors that limit yield are included in the recommendation process. For areas that are affected by other limiting factors, the recommendations may be invalid. A better understanding of th complex relationships between crop yield and yield limiting factors is required. In this study, we attempted to functionally relate crop yield to possible limiting factors, using various advanced computing techniques known as feedforward neural networks. Several of these methods accurately estimated the nonlinear relationship between soil parameters and crop yield on the test data set, while retaining good generalization characteristics. Yield maps estimated by the various techniques generally agreed well with actual yield maps. These neural techniques may benefit scientists by providing them with a new tool for investigating crop response to limiting factors. Producers and agribusiness may also benefit through the improved recommendations and management strategies developed with these 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, pesticides, etc. A deeper understanding of the functional relationship between yield, soil and site properties is of critical importance to precision farming. A number of feedforward neural network methods were investigated in an attempt to identify techniques able to functionally relate soil properties and crop yields on a point by point basis. The dataset used included a representative year of soybean yield, site and soil property data on a single field and consisted of 344 points. Both training accuracy and generalization ability were evaluated for several previously reported neural techniques, through the use of cross validation. Training accuracies were found to be quite good, with a standard error of calibration (SEC) less than 176 kg/ha for many of the methods investigated. Generalization results were also quite reasonable, with the standard error of prediction (SEP) less than 213 kg/ha for those same methods. Resilient backpropagation (rprop) showed the best results, both in terms of accuracy (SEC=174 kg/ha) and in generalization ability (SEP=200 kg/ha).