Precision Agriculture, 1, 39᎐52 1999
ᮊ 1999 Kluwer Academic Publishers. Manufactured in The Netherlands.
Weed Mapping with Co-Kriging Using
TORBEN HEISEL Torben.Heisel@agrsci.dk
Department of Crop Protection, Danish Institute of Agricultural Sciences, Research Center Flakkebjerg, 4200
ANNETTE KJÆR ERSBØLL firstname.lastname@example.org
Department of Mathematical Modelling, Technical Uni
ersity of Denmark, Building 321, 2800 Lyngby,
CHRISTIAN ANDREASEN email@example.com
Department of Agricultural Sciences, The Royal Veterinary and Agricultural Uni
1871 Copenhagen, Denmark
Abstract. Our aim is to build reliable weed maps to control weeds in patches. Weed sampling is time
consuming but there are some shortcuts. If an intensively sampled variable e.g. soil property can be
used to improve estimation of a sparsely sampled variable e.g. weed distribution , one can reduce weed
sampling. The geostatistical estimation method co-kriging uses two or more sampled variables, which
are correlated, to improve the estimation of one of the variables at locations where it was not sampled.
We did an experiment on a 2.1ha winter wheat field to compare co-kriging using soil properties, with
kriging based only on one variable. The results showed that co-kriging Lamium spp. from 96 0.25m
sample plots ha
with silt content improved the prediction variance by 11 % compared to kriging.
With 51 or 18 sample plots ha
the prediction variance was improved by 21 and 15 %.
Keywords: Co-kriging, soil properties, weed densities and distributions, weed mapping
Chemical weed control on a whole-field basis is common. Weeds, however, are not
uniformly distributed in arable fields, but tend to appear in patches. The combina-
tion of a weed map, positioning devices, e.g. GPS Global Positioning System
Tyler, 1993 , and a sprayer to differentiate herbicide application on the basis of
weed infestation Paice et al., 1995 , makes it possible to reduce herbicide use,
costs and environmental pressure. Position specific weed control requires either a
fully automatic real-time system or maps of the weed infestation on fields. At
present real-time systems are not applicable. Christensen et al. 1996 have
obtained a 50 % reduction in herbicide use, when treating fields site specifically
according to weed maps.
Kriging is a local estimation method that uses the spatial dependence of a
particular variable for the estimation procedure, e.g. weed distribution. Previously,
we have found that weed distribution maps are estimated well with kriging on the