Subfield management class delineation using cluster analysis from spatial principal components of soil variables

• We propose the use of a new method named KM-sPC to management class delimitation.
• KM-sPC uses a multivariate algorithm that takes into account spatial correlation.
• The new method relies on fuzzy k-means clustering of spatial principal components.
• Management classes were identified using KM-sPC and unconstrained fuzzy k-means method.
• Differences in yield data between delineated classes were better shown by KM-sPC.

Understanding spatial variation within a field is essential for site-specific crop management, which requires the delineation of management areas. Several soil and terrain variables are used to classify the field points into classes. Fuzzy k-means cluster analysis is a widely used tool to delineate management classes in the multivariate context. However, this clustering method does not consider the presence of spatial correlations in the data. The MULTISPATI-PCA algorithm is an extension of principal component analysis that considers spatial autocorrelation in the original variables to produce synthetic variables. We propose and illustrate the implementation of a new method (KM-sPC) for subfield management class delineation based on the joint use of MULTISPATI-PCA and fuzzy k-means cluster. To assess the performance of KM-sPC, we performed clustering of the original soil variables and of both spatial and classical principal components on three field data sets. KM-sPC algorithm improved the non-spatial clustering in the formation of within-field management classes. Mapping of KM-sPC classification shows a more contiguous zoning. KM-sPC showed the highest yield differences between delineated classes and the smallest within-class yield variance.