Field-Scale Experiments for Site-Speciﬁc Crop
Management. Part II: A Geostatistical Analysis
M. J. PRINGLE*
A. B. McBRATNEY
Australian Centre for Precision Agriculture, MacMillan Building A05, University of Sydney,
NSW 2006, Australia
S. E. COOK,
CIAT, A.A. 6713, Cali, Colombia
Abstract. Part II analyses approach C experiments. Field-scale experiments were applied to four wheat
ﬁelds in the Western Australian wheat belt. Diﬀerent experimental designs were used: two two-dimen-
sional sine-waves, a chequerboard, and a two-factor strip arrangement. In each experiment, the yield
associated with a particular treatment was predicted by kriging to where the other treatments were located.
Diﬀerent forms of kriging were investigated. Co-located cokriging, using the previous-season yield map as
a covariate, was the most promising. The kriged data were then modelled with polynomial yield response
functions. The outcome was a map for each ﬁeld that described the optimum application of experimental
input. The requirements varied continuously across the ﬁeld, and could justify future site-speciﬁc crop
management. The two-factor strip experiment was the most successful of those presented; the ﬁeld on
which it was used showed relatively strong responses to the applied inputs. The other sites were aﬀected by
lack of rain and/or design ﬂaws. The underlying philosophy is sound, but the method proposed is time-
consuming and ineﬃcient. We hope that this paper can stimulate further research on the subject.
Keywords: site-speciﬁc crop management, ﬁeld-scale experiments, kriging, polynomial response
When a controllable input can be applied as a continuous variable to a ﬁeld (e.g.
fertiliser rates, seeding depths), continuous management is feasible. In some cases,
the required input can be estimated by direct calibration with ancillary properties
(e.g. remotely sensed data or soil tests). An alternative, however, is to conduct an
experiment with the input, whereby the design covers the area of interest (as opposed
to just a portion of it), and treatments are allocated systematically, and inputs are
varied about the mean application. Localised yield values can be used, in conjunc-
tion with response functions, to estimate an optimal continuous input application.
This paper is essentially a methodological advance on the analyses presented in
earlier papers (e.g., Cook et al., 1999; Pringle et al., 1999). The method presented
may be summarised as:
*Current address: Rothamsted Research, Harpenden, Hertfordshire, AL5 2JQ, UK.
Precision Agriculture, 5, 625–645, 2004
Ó 2005 Kluwer Academic Publishers. Manufactured in The Netherlands.