Within-field zoning using a region growing algorithm guided by geostatistical analysis

Region growing methods are of potential interest to define within-field zones and resulting site-specific management. These methods are unsupervised and based on regions which grow from the initial seeds according to homogeneity criteria. However, the determination of seed number and seed locations has strong repercussions on the zoning output. This paper proposes an approach to allow knowledge inherited from geostatistical analysis to guide the seed initialization (seed number and location) of a region growing-based segmentation method. In this study, the segmentation method is a general division/merging procedure which, in this study, is used for merging and region growing. An original point is the possibility to use it either for irregularly located data or data arranged on a regular grid. The initialization of the segmentation method is guided by a prior analysis where a few parameters of the semi-variogram model are used to set: (1) the number of seeds required to initialize the region growing procedure; (2) to decide their relative locations (i.e. minimal distance between seeds); and (3) to identify potential outliers as seeds that may flaw the growing procedure. Both methods were tested on two data sets: (1) three hypothetical fields of known distribution and known spatial organization; (2) a real field where yield monitor data were obtained. A qualitative analysis of the results is presented, as well as the evolution of the variance explained by the model.