Infill-sampling design and the Cost of classification errors