|Sudduth, Kenneth - Ken|
Submitted to: Plant and Soil
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/6/2006
Publication Date: 10/6/2006
Citation: Martin, N., Bollero, G.A., Kitchen, N.R., Kravchenko, A.N., Sudduth, K.A., Wiebold, W.J., Bullock, D.G. 2006. Two classification methods for developing and interpreting productivity zones using site properties. Plant and Soil. 288(1-2):357-371.
Interpretive Summary: Starting in about the mid 1990's, farmers began installing yield monitors on their combines and making yield maps. Many farmers now have a number of years of yield maps for the same field and want to use these maps to identify those areas within fields that yield similarly from year-to-year. Areas with similar yield production have been called "productivity zones" or "yield zones." Farmers are interested in identifying productivity zones because key management decisions made early in the growing season (such as nitrogen fertilizer rate and seeding population) could be based on the expected yield. Further, relating these yield zones to soil and landscape characteristics could help reveal what soil management practices might be used to increase productivity. This study developed productivity zones using two different statistical methods and then evaluated those zones with respect to measured soil and terrain properties. Field data from two fields in Missouri and one in Illinois, all in corn-soybean rotations, were analyzed. The more flexible statistical method, called k-nearest neighbor discriminant analysis, gave better results. In some cases, yield zones corresponded well to areas of specific soil or terrain properties. In general, the strongest correspondence was seen with sensor-obtained measurements, including apparent soil electrical conductivity (ECa) and remotely sensed images. The terrain attributes of elevation and slope also corresponded to yield zones in some cases. This means that farmers may be able to use some combination of ECa, image, and terrain data to create productivity zones for fields where they do not have yield map information. Results of this study will benefit farmers and crop consultants by helping them develop cost effective ways for creating site-specific management plans.
Technical Abstract: Crop performance is often shown as areas of differing grain yield. Many producers utilize simple GIS color ramping techniques to produce visual yield maps with delineated clusters. However, a more quantitative approach such as an unsupervised clustering procedure is generally used by scientists since it is much less arbitrary. Intuitively the yield clusters are due to soil and terrain properties, but there is no clear criterion for the delineation. We compared the effectiveness of two delineation or classification procedures: quadratic discriminant analysis (QDA) and k-nearest neighbor discriminant analysis (k-NN) for the study of how yield temporal patterns relate to site properties. This study used three production fields, one in Monticello, IL, and two in Centralia, MO. Clusters were defined using maize (Zea mays L.) and soybean (Glycean max (L.) Merr.) yield from three seasons. The k-NN had greater and more consistent successful classification rates than did QDA. Classification success rate varied from 0.465 to 0.790 for QDA while the k-NN classification rate varied form 0.794 to 0.874. This shows that areas of certain temporal yield patterns correspond to areas of specific site properties. Although profiles of site properties differ by crop and production field, areas of consistent low maize yield had greater shallow electrical conductivity than those of consistent high maize yield. Furthermore, areas of consistent high soybean yield had a lower soil reflectance than those areas of consistent low yields.