Linear geostatistics can provide, by kriging from soil and environmental data, local estimates that are unbiased and have minimum and known variance, and are in this sense optimal. The procedure assumes a model, usually that the variables are realizations of random processes and intrinsic sensu Matheron. It can also be used for global estimation. Variograms, or covariance functions, are vital and must themselves be estimated from data. That demands fairly heavy sampling. Classical sampling also provides global estimates. It needs no assumptions about the nature of the variation, but to avoid bias and to estimate the errors sampling must be randomized.