Submitted to: International Conference on Precision Agriculture Abstracts & Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 7/17/2002
Publication Date: 12/1/2002
Citation: KITCHEN, N.R., FRIDGEN, J.J., SUDDUTH, K.A., DRUMMOND, S.T., WIEBOLD, W.J., FRAISSE, C.W. PROCEDURES FOR EVALUATING UNSUPERVISED CLASSIFICATION TO DERIVE MANAGEMENT ZONES. PROCEEDINGS 6TH INTERNATIONAL CONFERENCE ON PRECISION AGRICULTURE. 2002. CD-ROM (UNPAGINATED). AMERICAN SOCIETY OF AGRONOMY. MADISON, WI. Interpretive Summary: Once farmers started collecting data with their combines and mapping crop yield information, they realized that yield differences could be extreme within fields -- as much as 100% difference between the low and the high producing areas of a field. This motivated farmers to want to create sub-field management zones so that inputs could be varied according to crop need. In order to do this, they have expressed a need for decision aids that can help identify sub-field management zones. A procedure that has been tested by researchers for creating management zones is called clustering analysis. This procedure uses the computer to group the field into different classes based on two or more measurements. Elevation, slope, and nutrient availability are some of the measurements that have been used. The first purpose of this research was to show that this clustering procedure can generate different outcomes with the same data set. A second step was to show some ways in which the clustering outcomes could be evaluated in order to give the user some confidence that clustering is still a good method of delineating management zones. We illustrated this method with soil and crop yield measurements taken from a Missouri claypan soil field. We found that with only two or three clusters for potential management zones, the outcomes of the clustering analysis were much more different than if four or more clusters were defined. These results will be useful to producers and crop consultants as they seek better ways to develop site-specific management plans.
Technical Abstract: Procedures for creating and then testing management zones using unsupervised classification have not been well established. The objectives of this research were to (1) demonstrate the existence of multiple outcomes from an unsupervised fuzzy clustering procedure for potential management zones and (2) provide test procedures for considering which is the best classification outcome. A fuzzy c-means clustering algorithm was selected for the purpose of partitioning data into groups or clusters for consideration as potential management zones. Soil EC, elevation, and slope were used as clustering variables for a 34 ha claypan soil field. When clustering for 2 to 8 clusters on this data set, we found multiple outcomes were possible for all but 4 clusters. Variance reduction within clusters both for the classification variables and for yield data indicated between 4 and 7 clusters were appropriate. Variance reduction from the different outcomes within a cluster number were similar when the cluster number was 5 or greater. Clustering performance indices (Fuzziness Performance Index, Normalized Classification Entropy, and Separation Index) generally supported the selection of 4 clusters. The field operator was also asked to use his experience to assess maps created from the clustering procedure. His top three choices used 7 clusters. With each test procedure for validation of the clustering outcomes, confidence is gained that an appropriate decision can be made for delineating management zones.