Appropriate On-Farm Trial Designs for Precision Farming

1
Appropriate On-Farm Trial Designs for Precision
Farming
J. Lowenberg-DeBoer1, D. Lambert1, R.
Bongiovanni2 1Purdue University, Site-Specific
Management Center, Department of Agricultural
Economics, Purdue University, West Lafayette,
Indiana 2Precision Agriculture Project, National
Institute for Agricultural Technology (INTA),
Manfredi, Córdoba, Argentina lowenbej_at_purdue.edu
2
Motivation

• Objective of on-farm trials is different from
  research trials
• Farmers want to make the best economic decisions
  for their operation
• Most farmers do not care about underlying
  mechanisms or whether results are generalizable
• For on-farm trials we need to shift focus away
  from research to farm management decision making

Photo Farmphotos.com
3
Feedback from US Farmers

• For hybrid and variety trials, filling planters
  with small quantities of seed and cleaning boxes
  for the next hybrid or variety takes too much
  time.
• In larger operations, seed is often purchased in
  bulk. This makes it difficult to fill the planter
  with small quantities. Hybrid and variety strip
  trials work better with seed in bags.
• Split planter trials are convenient only if your
  combine head is exactly half the width of the
  planter. That is not always the case.
• For narrow row soybeans, many producers prefer to
  harvest at a diagonal to the rows. This makes it
  impossible to detect narrow strips on the yield
  maps.

4
Farmers prefer large block designs
5
Soil Density Trials, LeRoy, IL, USA, are an
example
Photo Russ Munn
6
Field were split into large blocks (gt10 ha) and
yield data averaged by soil type polygon
7
Tracked Equipment Advantage Occurred Mainly with
Corn on Lowland Fields with Clay Soils
8
On-farm trials provide experience with different
designs, but do not tell us which is best.
9
Why use a Monte Carlo Simulation in developing
alternative trial designs?

• It is cheaper to narrow the range of alternative
  designs with simulation before doing expensive
  field testing
• With spatial heterogeneity field testing cannot
  entirely answer the question since one can only
  do one trial in one place in a given year
• Simulation allows us to test different designs on
  the same set of spatial characteristics with the
  same weather years

10
Pilot Test of Monte Carlo Approach

• 8 scenarios total
• Two experimental designs (3 treatments, no
  blocks 3 treatments, 5 blocks)
• Two estimation methods (OLS and SAR)
• Two levels of spatial autocorrelation (rho 0.5
  and 0.9)
• 100 Monte Carlo trials for each scenario

11
Monte Carlo experimental design detail

• 2 15 x 15 grids
• N treatments 0, 75, 150 kg ha-1
• Topography zones from the Las Rosas (Argentina)
  trials.
• OLS slope coefficients from the Argentina trial
  were used to simulate yields in each grid cell

12
Two Experimental Designs Simulated
13
Monte Carlo experimental design detail

• Treatments were randomly assigned in blocks
  Blocks were randomly assigned over the grid
• Quadratic model to generate yields (with Las
  Rosas OLS coefficients)
• y ßo ß1N ß2N2 di
  interaction terms u
• Spatial model to induce correlation y Xß
  (I ?W)-1u
• u a randomly drawn i.i.d. innovationN(0, s2)
  s2 is from the Las Rosas trial.
• The same vector of disturbances was used for
  each scenario.
• W is an n x n matrix defining a neighborhoods of
  observations.
• Two levels of ? were used to induce correlation
  between grid cells 0.5 and 0.9.

14
Partial Budgeting for the experiment
Profit is maximized when the value of the
increased yield from added N equals the cost of
applying an additional unit or when the marginal
value product equals the marginal factor cost.
15
Spatial error model (SAR)
y Xß e with e ?W e u

1. Obtain OLS residuals, e
2. Given e, estimate ? that maximizes the SAR Log
   likelihood function
3. Given ?, find the GLS estimates
4. Compute a new set of residuals until convergence
5. Given ? and e, compute variance for inferential
   statistics

Anselin, 1988.
16
Results of Pilot Simulations
Neither the single plot or the repetitions were
very successful in correctly identifying spatial
variability
High Spatial Correlation (rho0.9)
17
Spatial analysis and repetitions increase
reliability
Results of Pilot Simulation Study
Single plot data and non-spatial analysis are
least reliable.
Single plot data with spatial analysis is as
reliable as OLS with three repetitions.
Repetitions and spatial analysis most reliable
18
Results from Pilot Simulations

• Neither experimental design is particularly
  successful in identifying spatially variable
  response to nitrogen
• Single plot design was often as successful at
  identifying spatial variability of response as
  the traditional randomised block design
• Traditional design usually results in a more
  reliable decision than the single plot design, in
  the sense that the range and standard deviation
  of returns is smaller

19
Summary

• With rapid technology change farmers need more
  on-farm information to make good decisions
• Farmers often shy away from rigorous on-farm
  comparisons because of logistical problems
• Precision Ag technology facilitates data
  gathering, but classic on-farm trial designs are
  still often too time consuming
• Simulation provides a practical way to narrow the
  range of alternative designs before on-farm
  testing

20
Further research

• Alternative statistical models (e.g. Nearest
  Neighbor, Cressies REML-geostatistic approach)
• Continuous spatial process assumption vs.
  discrete approach
• More Monte Carlo trials the unexpectedly small
  success rate (large Type II error rate) in
  correctly identifying spatial variation of N
  response may in part be due to too few simulation
  runs
• Different designs this preliminary run looked at
  only a single plot and five blocks

21
Questions or Comments?