Estimating within-field variation using a nonparametric density algorithm

The application of site-specific techniques and technologies in precision farming requires subdividing a field into a generally small number of contiguous homogeneous zones. The proposed algorithm of clustering is based on nonparametric density estimate, where a cluster is defined as a region surrounding a local maximum of the probability density function. Soil samples were collected in a 2-ha field of the experimental farm of the Agricultural Research Institute, located in Foggia (Southern Italy) and some of the most production-affecting soil properties were interpolated by using the geostatistical techniques of kriging and cokriging. The application of the clustering approach to the (co)kriged surface variables produced the subdivision of the field into five distinct classes. The proposed algorithm proves quite promising in identifying spatially contiguous zones, which are more homogeneous in soil properties than the whole-field. Its great advantage consists in giving an additional description of the residual variation within the class and such a piece of information is very useful in precision farming as a basis for the variable-rate application of agronomic inputs.