|Sudduth, Kenneth - Ken|
Submitted to: International Conference on Precision Agriculture Abstracts & Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 12/1/1996
Publication Date: N/A
Citation: Interpretive Summary: Site-specific management, or precision farming, is a strategy in which cropping inputs such as fertilizers are applied at varying rates across a field in response to variations in crop needs. The effect of these factors on yield is complex and may change from point to point within a field. We evaluated a number of mathematical and statistical methods in order to model the relationship of yield to these other factors. Of the statistica techniques we evaluated, only more complex ones were able to create satisfactory models. We also found that an artificial neural network, a type of computer program that mimics the function of the human brain in a limited sense, was able to model yield. These results are important to scientists and ag professionals seeking to develop improved methods and recommendations for precision farming, since by using these types of analysis techniques they may be able to develop better fertilizer management plans.
Technical Abstract: The full benefit of site specific crop management will not be realized until the critical factors that cause variations in yield can be related to a measured parameter or identified as being caused by unknown parameters. Analysis and correlation methods that would assist in the detection of these critical factors on a point by point basis were evaluated. Methods investigated included standard linear correlation, stepwise multiple linea regression, projection pursuit regression, and back propagation neural network analysis. The data set used for analysis consisted of four field years of yield data, along with soil fertility, topsoil depth, and field topography. Standard statistical techniques such as linear correlation and stepwise multiple linear regression were unable to adequately model the spatial variability. Complex nonlinear, non-parametric methods such as projection pursuit regression and back propagation neural network analysis provided much more accurate yield predictions (up to r*2=0.77 and r*2=0.55 respectively). Nonlinear response curves were obtained from the nonparametric models for yield as a function of each of the input variables. These empirical response curves generally agreed with observed yield limitations on these fields and should prove useful for studying the interactions between multiple critical factors and crop yields.