ABSTRACT
Geostatistics were originally developed for the mining industry to estimate the location, abundance, and quality of ore over large areas from soil samples to optimize future mining efforts. These methods have been adapted for many different situations. In this chapter, geostatistics are used to examine the weed distribution inside a single production Želd, variations of distribution over time, skewed data distributions, and correlations with species traits. A geostatistical study starts with selecting a sampling plan that “catches” the spatial relationships among the variables of interest. Exploratory data analysis then examines data distributions and checks whether the prerequisites for a geostatistical analysis are fulŽlled. If necessary, data are transformed and detrended to meet these prerequisites. Then, empirical semivariograms are calculated and used to (1) explain small-scale spatial trends (e.g., weed patch shapes and progress in time as a function of species dispersal and germination traits), (2) determine the variances for unsampled distances to allow prediction of values in unsampled points and maps to be plotted, using kriging, and (3) reduce estimation errors at unsampled points. Crosssemivariograns and cokriging describe covariation of variables in space, and these relationships are used to estimate a sparsely sampled primary variable with the help of an extensively sampled secondary variable. Here, these methods were adapted to predict weed maps with past observations and variograms. Last, error analysis evaluates how close predicted weed maps are to observations and the risk of spraying insuf-Žciently or unnecessarily when basing herbicide spraying in precision agriculture on weed maps predicted with past observations using cokriging.
