October 2002
The Status of Site-Specific Crop
Response Estimation
Jess Lowenberg-DeBoer, Director
Rodolfo Bongiovanni, PhD
Fine-tuning crop management using site-specific crop
response information was part of the original vision of precision agriculture.
The hope was that producers could use yield monitor, soil test and other
information to identify more accurately how crops respond to inputs on their
farms. This vision has proven difficult to achieve and some even consider it
impossible, but there has been some progress. The goal of this article is to
outline the state of the art of site-specific crop response estimation.
Researchers
in the U.S. and Argentina have estimated site-specific crop response curves
from yield monitor data using various spatial statistics techniques. A study by
Lambert and colleagues reviews those studies and the methods used
(http://www.agriculturadeprecision.org/ingles/indice.htm). The preliminary
conclusions coming out of these studies are:
1)
Designed trials are essential with current technology – data mining
with yield maps for crop response may not be worthwhile.
2)
Crop responses can vary widely from place-to-place within fields and as
a result the profit maximizing input choices also vary.
3)
Ordinary Least Squares (OLS) regression can provide reasonably accurate
estimates of the spatial variation, but the statistical reliability of these
estimates is not good.
4)
The various spatial regression methods seem to result in roughly the
same numerical estimates and similar statistical tests. The choice between
spatial regression methods may be influenced by ease of use, data density, and
software availability.
5)
Site-specific crop response can vary widely from year-to-year depending
on the weather. Making input choices will require risk management.
Data Mining - almost all the
successful examples of reliable crop response information from spatial crop
data comes from designed trials. The term “data mining” in this context means analysis
that extracts useful information from observations on systems without any
planned treatments. One of the key problems with data mining in crop production
is that the range of naturally occurring variables may not be enough for
estimation of robust response curves.
The study by R.J. Florax and fellow researchers in the Netherlands
shows the pitfalls of data mining for crop data. They use hand harvested yield
data from a farmer’s field in the Republic of Niger, West Africa, to estimate
statistically significant responses of millet to soil phosphorous and potassium
test levels, but their estimated coefficients are unreasonably large, probably
because the naturally occurring soil phosphorous and potassium levels cover
only a narrow range of low test levels. For example, most phosphorous tests are
below 6 ppm.
Spatial
Variability of Response – Studies in Argentina and in Minnesota have shown that
crop response can vary dramatically within fields. Both these studies were on
nitrogen application on corn. The study in Argentina used a commercially
available yield monitor and employs a strip trial design that could be used by
any producer with a yield monitor and GPS. The Minnesota work relied on strips
harvested by a plot combine on fifty-foot long segments. This would not be
practical on a commercial scale, but it does provide an important test of the
robustness of results with alternative ways of collecting the harvest data. The
constraint is no longer data collection, but it is the software used to make the
analysis. The Argentine data was analyzed with a statistical analysis tool
called SpaceStat (www.SpaceStat.com). This is a DOS program with a good
interface with ArcView.
OLS
Coefficients – It should be expected that OLS and spatial regression coefficients have
similar values. Statistical theory tells us that OLS coefficients are unbiased
estimators even when there is spatial correlation in the data. The problem with
OLS is that in the presence of spatial correlation the standard deviations are
inflated and consequently many statistical tests are not significant. Spatial
regression tends to show many more statistically significant variables than
OLS. This facilitates decision making because the estimated coefficients are
more reliable
Alternative
Approaches to Spatial Statistics – Spatial statistics have developed by several
different scientific disciplines. For example, spatial econometrics comes out
of geography and regional economics. Geostatistics was developed originally by
geologists and mining engineers. Nearest neighbor an agronomist initially
proposed analysis. It is not surprising that these different approaches make
different assumptions. Geostatistics was originally used primarily for creating
interpolated maps from sample sites, so it is logical that it assumes that
spatial variability changes continuously. Regional economists and geographers
often worked with data from political and administrative units of government
(e.g. counties, cities, neighborhoods) and so they assumed that spatial correlation
was a discrete relationship between these “polygons.” The work by Lambert and
colleagues suggests that all the spatial regression approaches result in the
same decision for the Argentine nitrogen response data.
Risk
Management
– In the Argentine data nitrogen response in a wet year was greatest on the
hilltop. In a dry year the response tended to be greatest in the lowland. In
any given year the nitrogen response differed greatly from one area of the
landscape to another, but ‘expected response’ was very similar everywhere. By
‘expected response” we mean the curve that results when taking a weighted
average of the estimated coefficients for each year. In the Argentine data the
different responses in the wet and dry years effectively balance each other,
resulting in an average response that differs little from one part of the field
to another. While variable rate nitrogen was profitable “after the fact” for
all fields, the expected response seldom favored variable rate application
because there was little spatial difference in expected responses.
Maximizing
the expected net value of crop response is a simple approach to dealing with
crop variability, but is it the best way? Optimistic producers might prefer to
always fertilizer for the wet year, while pessimistic producers might favor a
very conservative approach of always fertilizing for the dry year. The best way
to manage risk in spatial management of nitrogen is a current topic of
research.
Conclusions
Estimation
of site-specific crop responses from yield monitor data is possible with
designed trials. It can be accomplished with data from simple strip trials. The
major constraints to estimation of site-specific response curves is that the
software is not very user friendly and the skills needed to interpret results
are not widespread. The key to robust statistical estimation is that the
spatial correlation of the observations must be explicitly modeled. Spatial correlation cannot be ignored.