Thomas Goddard, Len Kryzanowski, Karen Cannon, Cesar Izaurralde, Tim Martin
Precision farming aims to optimize the use of soil resources and external inputs(fertilizers and herbicides) on a site specific basis. Precision farming takes advantage of rapidly evolving GPS technology together with electronic sensors and controllers to monitor crop response under variable inputs and landscape position. Objectives of this study were: (i) to discover the soil-landscape-input relationships governing crop yields in characteristic Alberta landscapes, (ii) to test the performance of an agricultural simulation model using site-specific data, and (iii) to develop a method to analyze high-resolution data by linking the model to a GIS. Crop yields were monitored during 1994 and 1995 at four sites in Alberta using a high-precision 3-D DGPS. The Erosion-Productivity Impact Calculator (EPIC) model was run on a sub-field, site-specific basis using soil pit information and two levels of climatic data. The crop growth routines were compared against two years of yield maps obtained from 40 - 100 ha fields. A program was designed and built to couple the EPIC model with the GRASS-GIS. Our results show the potential to depict and analyze the variance in crop yield, leaching potential, erosion risk and economics on a farm field scale. The application of this approach in precision farming would allow for optimizing the use of resources on a site-specific basis thereby contributing to minimize detrimental environmental impacts such as nitrate leaching or erosion.
Farming systems are continuing to change in response to economic, technological and social
trends. Farming practises are becoming questioned not only by farmers but the public at
large. Concerns are profitability and environmental impact. Farmers look to adopt new
technology. In Canada, the trend over the past two decades has been to reduced tillage
systems (direct seeding) and increase use of fertilizers and herbicides. The cost savings of
reduced tillage does not outweigh the added cost of crop chemicals so profit margins have
remained narrow. Our labour costs are high and our natural resources finite so we must
utilize technology to maintain competitiveness in a world market. Developments in sensor-
controller technology, computers and positioning systems now bring new opportunities for
farm management (Goddard et al. 1995).
The development of the publicly available global positioning system (GPS) has opened new
doors in opportunities for spatial data. This is a passive positioning system from a
constellation of 24 orbiting radio-navigation satellites. They provide continuous position
data provided the receivers have a line of site access (or nearly so) to the satellites.
Positioning can be two or three dimensional in real-world coordinates.
Differential GPS (DGPS) uses a stationary monitor receiver to calculate the difference
between the true position and the determination of a position from the satellites for
a point in time. This allows the differential correction of a roving (remote) receiver so that
the errors, induced as part of the military declassification, can be removed. This allows the
positions of a roving receiver to have a three dimensional accuracy of 20 cm or better
(Lachapelle et al. 1994). The use of a radio modem allows the
transfer of differential corrections
and corrected position determinations in real-time. The combination of accuracy and real-
time determinations presents possibilites for guidance of farm equipment and the
development of accurate digital elevation models (DEMs) as a by-product of other GPS aided
farm
operations (e.g. yield mapping). Experience in Alberta has shown that harvesting operations
on a 80 hectare field can also yield 100,000 elevation measurements. Subsequent computer
processing and terrain analyses can provide useful information to augment yield map
interpretation.
Detailed yield map interpretation combined with terrain analysis from high quality DEMs
and site specific soil sampling will provide new opportunities for the use of integrated crop
models. Modeling landscapes and crops may negate the desire of expensive grid sampling
which is the current recommendation to those entering the practise of precision
farming.
One model that holds some potential for this application is the Erosion Productivity Impact
Calculator (EPIC). Using either simulated weather (based upon monthly parameters) or daily
records it will predict
crop growth under a variety of conditions. It contains a soil model, tillage models, water
erosion models (variations of USLE), a wind erosion model, hydrologic processes and
pesticide movement, all dynamic on a daily time step. Economic parameters are included
in the model as static parameters. EPIC has been shown to be effective for a variety of
purposes and recent interest has been expressed on the Canadian prairies to evaluate EPIC
with long-term research plots (Moulin and
Beckie 1993; Toure et al 1995).
A research project was initiated in 1993 in Alberta to look at the application of GPS and
related technologies at a farm scale. The objective was to use DGPS to allow yield mapping
of
fields, site specific sampling and variable rate fertilizer application. The goal was to develop
the ability to apply the optimal rate of fertilizer for each part of a field for maximum
economic yield and minimum environmental impact. The project presented the opportunity
to examine the use of integrated crop models for application in site specific management
systems.
A second project utilizing the same sites and equipment was initiated in 1994. The
objective of this project was to examine an alternative approach to grid sampling of
landscapes in order to reduce costs to farmers. The project would take advantage of
detailed terrain data provided by high precision DGPS and integrated crop models to study
the development of optimum agronomic management on a site specific basis. The study
would be further aided by the development of software to integrate the EPIC model with
the GIS, GRASS in order to provide a prototype to provide agronomic recommendations
and risk assessment on a site specific basis across a field.
Four farmer cooperators were selected in a north-south line through Alberta in order to
encompass a range of soil and climate conditions (approximately 49-54 degrees N and 112
degrees E). One test field was selected on each farm. One site
was irrigated (57 hectares) and the other three were dryland fields ranging from 32 to 81
hectares in size. Field characteristics were typical for the area (Table 1).
Table 1. Field characterization for three dryland sites in Alberta, Canada.
| HUSSAR | STETTLER | MUNDARE | |
| Frost Free Days | 115 | 115 | 100 |
| Annual Total Degree Days (above 5C) |
1550 | 1450 | 1400 |
| Mean Annual Precip. (mm) | 350 | 400 | 450 | Topography | Strongly rolling | Strongly hummocky | hummocky |
| Elevation range (m) | 44 | 12 | 5 |
| Parent Material | till | fluvial/till | till/fluvial |
| Soils | Dark Brown Chernozemic, Regosolic, Gleysolic |
Thin Black Chernozemic, Gleysolic, Regosolic |
Black Chernozemic, Solonetzic, Gleysolic |
Commercial continuous yield monitors were installed on the four different types of farm
combines. The yield monitors were interfaced with the portable GPS receivers when each
of the project fields were harvested. Variable rate fertilization was done with a pneumatic
banding applicator with two tanks for individual control of the rate of nitrogen (urea) and
phosphate (mono ammonium phosphate) fertilizer. Research in the Canadian prairies has
shown this to be the most efficient and environmentally benign method of fertilizer
application. The delivery rate from the two tanks of fertilizer could be adjusted instantly,
according to a prescription map, as the tractor-applicator moved across the
field.
The GPS receivers used were 10 channel C/A code narrow correlator spacing NovAtel 951
GPSCard connected to NovAtel Model 501 antennas (Fenton et
al. 1991, Van
Dierendonck et al. 1992). Differential corrections were done in post processing of harvest
operations. Three dimensional accuracy at the sub-decimeter level were achieved. For
variable rate fertilizer application the base station set up in the field corner and the mobile
receiver on the tractor were linked with two short range wireless radio modems in the low
900 MHz RF range. This provided DGPS positions in real-time. The integrated GPS system
and it's performance is described by Lachapelle et al. (1994).
For the parent project 18 to 26 hectares at each location was soil sampled using a 68 by 68
m grid in the fall of each year. At each grid-line intersection, composite samples of 12 to
15 cores were taken at four increments to a depth of 90 cm. The soil test results were used
in conjunction with yield maps, aerial photographs and topographic maps to construct
fertilizer application maps. Map units with similar levels of N and P were defined. The
optimum application rates for N and P were estimated based on soil test values, yield and
landscape features. The Bow Island and Mundare sites were mapped for salinity using an
EM38 salinity meter and GPS as described by Cannon et al. (1994).
Fertilizer rate experiments were conducted in strips 13 m wide across the full length of the
fields. Four constant rates of N with one rate of P were used in one block. The other block
had three constant rates of P with one rate of N. Each rate was replicated twice. On an
adjacent area, variable rate applications (N and P) were compared to constant rate
applications in alternating strips.
The public domain, raster based GIS, GRASS (Geographic Resource Analysis Support
System) was used for mapping, overlays and data analyses (U.S. Army, 1993). The landform
regimes were described using the system of Pennock et al. (1987).
In 1994 and 1995 site specific sampling points were characterized at all sites for the
purposes of providing input for the EPIC model. Holes were dug at the shoulder, backslope
and footslope position of four hills at each site. Soil profiles were described and sampled
for physical and chemical analyses. The position of each hole was established with
DGPS.
A GRASS statistical routine was developed (s.rstats) to search the raw yield data around
the soil description hole for yield data from the combine harvestor. The routine will for
a desired number of yield data points (e.g. n=15) around a point providing the mean and
standard deviation of yield as well as the closest value to the point and its distance along
with the mean distance of all points. This allows the GIS operator to determine and control
the areal extent of the yield data to conform to landscape position, fertilizer treatment or
other field features. The resultant yields and their variances are used to assess the
performance of the EPIC crop model for site specific management.
The EPIC model version 3090 was used on a DOS platform. Detailed profile data were used to generate EPIC input files. Although some climate data was recorded at each field site, the simulations reported here are using the nearest weather station data (less than 30 km distance). Daily data from the nearest stations was obtained and assembled into the appropriate format for EPIC. The crop at two of the sites in 1994 (Hussar, Stettler) was spring wheat and canola at the other (Mundare). The plant growth parameters for the Canadian prairies from Kiniry et al. (1995) were used instead of the defaults.
Only results from the three dryland farms will be presented and discussed. Delays in
obtaining the yield data for 1995 have prevented presentation of that data at this
time.
The performance of the DGPS receivers and the software for solution of position allowed
for vertical accuracies in the range of 70 mm and are futher reported elsewhere (Lachapelle
et al., 1994). The amount of elevation data collected in each field was dependent upon the
duration of harvest operations but was usually not less than 50,000 observations for an 80
hectare field. The consequence of using high precision DGPS
each year, or several times per year (e.g. planting, harvesting, spraying) is that very large
data sets can be accumulated in a very short period of time as a by-product of other
precision farming operations. The data was not geometrically even as the GPS receiver was
mounted
on the combine harvester which cut the crop at six to nine meter widths depending upon
the crop and combine design. The data collection rate was constant (1 Hz) and since
combines have hydrostatic drives it allows a continuously variable forward speed which
usually had data spaced from one to 2.5 m within the lines. This presented some issues of
either de-densifying data or using robust interpolation techniques.
Satisfactory topographic contours were developed in GRASS using the thin plate spline
method with
tension and smoothing ("s.surf.tps"). Linearity of the raw data was reflected in the result of
using an inverse distance weighted routine ("r.idw"). If several data sets were to be
combined as mentioned above, the data density would become so great that any
interpolation method would likely be adequate.
The characteristics of the soils varied greatly with landscape position (Table 2). Often the
difference in classification at each position was at the soil order level of the Canadian System
of Soil
Classification.
Yields measured by the combine harvesters were found to vary by slope position at all sites.
An example of one site and year is presented in Table 3 of data from the whole field when
segmented according to landscape element (after Pennock et al. 1987).
The measured yields by landscape position at the four transects in each field also displayed
the same trend however, a difference with respect to the lower slope position is evident
between the site with conventional tillage (Stettler) and the site which has been under a
direct seeding system for 10 years (Hussar) (Figure 1). The effect of the thick straw mulch
on temperature and moisture as well as the finer soil textures have all contributed to less
of a difference between slope position. The lower slope position at Hussar was very moist
and crop lodging did occur to further decrease yield. The wheat grown at Hussar in 1994
was a newly licensed variety of a new wheat type (Canada Prairie Spring wheat, AC Taber
variety). EPIC crop parameters may not be appropriate for that type of wheat.
EPIC yields were consistently lower for the three sites in 1994 ( 1.4 Mg/ha and 2.0 Mg/ha
respectively). Likewise, the coefficient of variation for EPIC predicted yields was nearly
half the CV of the measured yields (CV for predicted was 27% and for actual, 49%).
Linear regression analysis indicated a significant, but weak, relationship between the
predicted and measured yields (Figure 2). Inspection of the canola crop and wheat crops
independently showed no better agreement of one over the other. Recent calibration and
validation work with EPIC in France showed a coefficient of determination of half the value
obtained here (Cabelguenne et al. 1990). They used nearly same number of data points and
the crop was only wheat. Wheat was found to have the lowest coefficients of determinations
compared to other crops such as soybeans and sorghum (Cabelguenne et al. 1990).
Improved relationships are expected with daily weather data from an on site station as well as
mass transfer of runoff from one landscape element to the other.
In order for integrated crop models to be accepted, the ability to fairly represent yield
potential on a site specific basis is needed. The benefit of using an integrated model is that
other features of the model can then be used with little additional effort. This latter
feature may be very appealing for wide spread use in the agriculture industry. Models for
pesticide leaching, erosion and tillage are more likely to be used if they come as part of an
agronomic model that can be used for prescription mapping and risk assessment.
Table 2. Soil characteristics by landscape position for three dryland sites.
| Landscape Position | Footslope | Backslope | Shoulder | ||||
|---|---|---|---|---|---|---|---|
| Site | Mean | STD | Mean | STD | Mean | STD | |
| Hussar | OM% | 3.2 | 0.3 | 3.9 | 2.8 | 3.3 | 2.3 |
| Depth to B horizon (cm) |
11 | 1 | 20 | 9 | 10 | 1 | |
| pH | 6.4 | 0.3 | 6.7 | 0.9 | 6.6 | 0.9 | |
| Sand% | 37 | 8 | 40 | 6 | 38 | 6 | |
| Clay% | 29 | 6 | 26 | 4 | 28 | 5 | |
| Nitrate (ppm) | 10 | 1 | 16 | 15 | 17 | 19 | |
| Phosphate (ppm) |
13 | 4 | 12 | 8 | 9 | 5 | |
| Mundare | OM% | 3.6 | 1.2 | 3.6 | 1.7 | 3.3 | 1.5 |
| Depth to B horizon (cm) |
44 | 26 | 30 | 3 | 16 | 7 | |
| pH | 5.8 | 0.2 | 7.0 | 1.1 | 6.1 | 0.7 | |
| Sand% | 43 | 10 | 50 | 8 | 49 | 7 | |
| Clay% | 20 | 3 | 20 | 3 | 21 | 2 | |
| Nitrate (ppm) | 20 | 2 | 16 | 8 | 12 | 7 | |
| Phosphate (ppm) |
22 | 11 | 16 | 8 | 14 | 8 | |
| Stettler | OM% | 4.5 | 0.1 | 2.5 | 0.3 | 2.5 | 0.8 |
| Depth to B horizon (cm) |
58 | 23 | 25 | 13 | 13 | 3 | |
| pH | 6.4 | 0.7 | 6.5 | 0.2 | 7.0 | 0.5 | |
| Sand% | 61 | 1 | 70 | 6 | 71 | 9 | |
| Clay% | 15 | 0 | 13 | 3 | 14 | 4 | |
| Nitrate (ppm) | 25 | 1 | 12 | 7 | 11 | 3 | |
| Phosphate (ppm) |
10 | 5 | 13 | 8 | 13 | 9 | |
Table 3. Yield of fertilized and unfertilized spring wheat by landscape position at the Hussar site, 1994.
| Spring wheat yield (Mg/ha) | |||
|---|---|---|---|
| Landscape position | |||
| Fertilizer Treatment | Shoulder | Backslope | Footslope |
| Fertilized (F) | 3.59 | 3.74 | 3.67 |
| Unfertilized (U) | 2.54 | 2.64 | 3.32 |
| F - U | 1.05 | 1.11 | 0.35 |
Regression ANOVA Table for Figure 2. Source Sum of sq. Deg of Free. Mean Sq. F-Ratio Prob>F Model 1.13 1 1.13 8.31 0.01 Error 3.59 28 0.14 Total 4.72 Coefficient of determination 0.24 Variable Coefficient S.E.E. T-Statistic Prob>t Constant 1.16 0.17 6.87 0.00 Yield 0.22 0.08 2.81 0.01
Funding for this project was provided by the Canada-Alberta Environmentally Sustainable Agriculture (CAESA) agreement. The project benefited from the field and computer skills of Germar Lohstraeter and Sheilah Nolan of Alberta Agriculture, Food and Rural Development. GPS support was provided by the University of Calgary, Department of Geomatics.