Full text of DOI:10.1002/etc.5620210328

Environmental Toxicology and Chemistry, Vol. 21, No. 3, pp. 672–681, 2002
Printed in the USA
0730-7268/02 $9.00
ϩ .00
EVALUATION OF THE AgDISP AERIAL SPRAY ALGORITHMS IN THE
AgDRIFT MODEL
S
ANDRA L. BIRD,*† STEVEN G. PERRY,‡ SCOTT L. RAY,§ and MILTON E. TESKE࿣
†Ecosystems Research Division, National Exposure Research Laboratory, U.S. Environmental Protection Agency, 960 College Station Road,
Athens, Georgia 30605
‡Atmospheric Sciences Modeling Division, Air Resources Laboratory (MD-81), National Oceanic and Atmospheric Administration,
Research Triangle Park, North Carolina 27711, USA
§Dow AgroSciences, 9330 Zionsville Road, Indianapolis, Indiana 46268-1053, USA
࿣
Continuum Dynamics, 34 Lexington Avenue, Ewing, New Jersey 08618-2302, USA
(
Received 26 September 2000; Accepted 13 August 2001)
Abstract—A systematic evaluation of the AgDISP algorithms, which simulate off-site drift and deposition of aerially applied
pesticides, contained in the AgDRIFT model was performed by comparing model simulations to field-trial data collected by the
௡
Spray Drift Task Force. Field-trial data used for model evaluation included 161 separate trials of typical agriculture aerial applications
under a wide range of application and meteorological conditions. Input for model simulations included information on the aircraft
and spray equipment, spray material, meteorology, and site geometry. The model input datasets were generated independently of
the field deposition results, i.e., model inputs were in no way altered or selected to improve the fit of model output to field results.
AgDRIFT shows a response similar to that of the field observations for many application variables (e.g., droplet size, application
height, wind speed). However, AgDRIFT is sensitive to evaporative effects, and modeled deposition in the far-field responds to
wet bulb depression whereas the field observations did not. The model tended to overpredict deposition rates relative to the field
data for far-field distances, particularly under evaporative conditions. AgDRIFT was in good agreement with field results for
estimating near-field buffer zones needed to manage human, crop, livestock, and ecological exposure.
Keywords—Spray drift
Pesticides
Modeling
BACKGROUND AND OBJECTIVES
to fill this need, is described in detail in a companion article
[9]. Since substantial productivity costs to applicators and
growers and potential damage to the environment hinge on the
result of regulatory decision-making, models used in this pro-
cess must have demonstrated scientific credibility (i.e., model
algorithms must be based on sound science with demonstrated
model performance). AgDRIFT contains empirical curves to
estimate drift from ground and orchard airblast applications
and AGricultural DISPersal (AGDISP) [10] mechanistic al-
gorithms to estimate drift from aerial applications. The primary
purpose of the current article is to provide a systematic eval-
uation of the aerial drift algorithms incorporated into Ag-
DRIFT to support its use in a regulatory setting.
An extensive body of literature has developed with respect
to establishing the scientific soundness of models for use in
environmental assessments. Beck et al. [11] identified two fea-
tures that are intuitively used to judge a model: the composition
of the model and the performance of the model. Compositional
validity is based on the intrinsic structure of the model and
some consensus judgment about the constituent hypotheses
represented in the model. Judgment of composition reflects the
generic properties of the model irrespective of the current task.
Model performance is evaluated based on the suitability of a
model for undertaking a specific task. An important part of
measuring performance is determining the magnitude of the
risk of making a wrong decision stemming from applying the
model to the task at hand. The exposure estimates generated
by AgDRIFT are used to make decisions regarding allowable
pesticide use. If these decisions do not sufficiently restrict
pesticide use, serious environmental consequences occur,
whereas overly restrictive decisions can have significant eco-
nomic repercussions. The magnitude of the decision-making
Drift of air-borne pesticides beyond the target site during
aerial applications is a source of environmental concern due
to the potential for human health impacts, downwind contam-
ination and damage to crops and livestock, and endangerment
of ecological resources. The Spray Drift Task Force (SDTF),
a consortium of pesticide registrants, has gathered field and
laboratory data [1] based on the assumption that pesticide drift
is primarily a function of application techniques (e.g., droplet
size and release height), environmental conditions, and phys-
ical properties of the spray solution and not of the active in-
gredient per se. The sensitivity of drift to numerous factors,
including atmospheric conditions [2–6] and application equip-
ment [6–8] and the inherent variability of field-trial results [8],
makes field testing the full range of possible meteorological
and application scenarios difficult. Modeling provides a co-
herent framework for evaluating the potential risks of spray
operations and the potential effectiveness of mitigation op-
tions. Both the SDTF and the U.S. Environmental Protection
Agency, Office of Pesticide Programs (OPP), support the use
of a spray drift modeling tool incorporating input databases
and postprocessing utilities to improve the efficiency, cost ef-
fectiveness, and reliability of the product evaluation and reg-
istration process. The AgDRIFT௡ model, which was developed
* To whom correspondence may be addressed (bird.sandra@epa.gov).
Paper NERL-ATH-ERD-00-043, U.S. Environmental Protection
Agency. This article has been reviewed in accordance with the U.S.
Environmental Protection Agency’s peer and administrative review
policies and has been approved for publication. Mention of trade
names or commercial products does not constitute endorsement or
recommendation for use by the U.S. Environmental Protection Agen-
cy, the authors, their employers, or the Spray Drift Task Force.
672

Evaluation of the AgDRIFT
௡
aerial spray drift model
Environ. Toxicol. Chem. 21, 2002
673
risk, however, is dependent on the specific product and product
use. Thus, even for a relatively well-defined model application,
the absolute performance requirements are dependent on the
uncertainty level acceptable in the individual case.
ticide registration use will more frequently be for low-flight
row crop application. The primary proposed use for AgDRIFT
in regulation and safety studies is to evaluate exposure from
deposition of pesticides onto aquatic and terrestrial systems
near the application area. Primarily, the model use will be
Agricultural dispersion, a well-documented Lagrangian-
type model [10] still under active development, serves as the
computational engine for AgDRIFT. The selection of AGDISP
as the computational engine for AgDRIFT was based on its
widespread scientific acceptance (i.e., compositional validity)
and the appropriateness of the structure of the model for the
problem of estimating aerial spray drift. The AGDISP has an
extensive history of use and development by several federal
agencies, including the National Aeronautics and Space Ad-
ministration (NASA), U.S. Department of Agriculture Forest
Service, and the U.S. Army. In addition to the mechanistic
aerial application algorithms, AgDRIFT contains empirical
curves to estimate drift from ground and orchard airblast ap-
limited to within 300 m (ϳ1,000 ft) of the edge of the field.
The deposition values generated by AgDRIFT will be used in
the regulatory context to determine the size of buffer zones
in combination with other label restrictions needed to reach a
safe level of exposure. The evaluation efforts thus focus on
model performance in predicting deposition within 300 m of
the application zone for relatively low-flight applications.
The information on the model performance will be pre-
sented in the form of graphical analyses and summary statis-
tics. There are three general questions that these analyses are
intended to address: How well does the model predict the field-
trial results on average? To what extent do the model predic-
tions vary from the measured results? Can the quality of the
fit between the model and the field data be related to specific
application variables or environmental factors? The specific
goal of the buffer zone analysis is to evaluate the potential
differences in buffer zone size developed from model simu-
lations versus those from field-trial observations and to infer
the general magnitude of risk associated with these important
regulatory estimates.
plication of pesticides. It consists of a Visual Basic Windows
௡
(Microsoft, Redmond, WA, USA) interface, access to input
data libraries, a toolbox for analysis of output, and plotting
and export utilities. AGDISP, a FORTRAN model, is compiled
as a digitally linked library incorporated into the visual shell.
Teske et al. [9] describe in detail the structure of AgDRIFT,
the incorporation of AGDISP into this framework, the scien-
tific basis for the algorithms, and recent improvements in the
computational algorithms.
A number of model developers have proposed a variety of
metrics and statistical test methodologies for model evaluation
[16–19]. Reckhow et al. [16] recommend that evaluations in-
clude a combination of graphical comparisons and statistical
tests appropriate to the context of the proposed use. The ap-
proach in this article is to present a variety of measures of
model performance to allow the user to evaluate acceptability
of model performance for various perspectives; i.e., we are
attempting to supply the reader and model user with sufficient
information to perform their own hypothesis testing and to
understand the behavior and limitations of the model. Analysis
is presented both graphically (quantile-quantile [Q-Q] plots
and box-and-whisker plots) and as summary statistics (e.g.,
mean and standard deviation of model-observation differences
at various downwind distances). Our analysis evaluates per-
formance at individual downwind distance locations separately
since the rapid decline in deposition with increasing downwind
distance can mask the response to other variables. In addition
to performance categorized by downwind distance, model per-
formance is also examined as a function of major application
variables and meteorological conditions.
A number of studies have examined the performance of
AGDISP (or AGDISP as the near-wake component in Forest
Service Cramer–Barry–Grim [FSCBG] [12]) with field obser-
vations. A sensitivity analysis and evaluation study of AGDISP
6.1 and two other spray drift models were performed for En-
vironment Canada [13]. For a series of 18 experimental treat-
ments, observed results differed from the three model estimates
by a factor of two or less. The outputs against which the models
were compared included maximum deposition, location of
maximum deposit, integrated deposit, and the median inte-
grated deposit location. Several field studies were performed
by the Forest Service over forested areas, and the results were
compared with FSCBG [12], which also utilizes AGDISP as
the near-field model. These studies involved relatively high
flight (10–60 m) releases. The model to field-data comparisons
were primarily graphical, showing good comparisons with cal-
culated values for R2 between predictions and observations
generally
Ͼ0.5. The New Zealand Forest Research Institute,
Rotorua, New Zealand [14], compared the FSCBG/AGDISP
combination with field results of 12 separate applications over
grasslands with 3 separate nozzle types and a 10.3-m release
height. These researchers concluded that the modeling system
was useful in assessment of the effectiveness of spray equip-
ment and for defining operational and meteorological variables
that will minimize drift. Anderson et al. [15] compared the
FSCBG/AGDISP combination to results from 17 releases over
fully leafed 16-m mixed oak forest. These researchers found
the model acceptable in predicting average (over several trials)
deposition but noted that the model results as expected did
not reflect the high run-to-run variability observed in the field
data.
MATERIALS AND METHODS
Field-trial data
A total of 180 aerial trials were conducted by the SDTF
during a series of three distinct sets of field studies. Details
of these studies are provided in Hewitt et al. [1]. The studies
were conducted in pairs with each pair consisting of a standard
treatment and a variable treatment. The application variables
were held constant for all standard treatments and systemati-
cally altered during the study for the variable treatments. Four
flight lines were applied for most of the trials, with each having
a 13.7-m swath width. Deposition samples were taken on hor-
izontal alpha-cellulose cards along the four parallel flight paths
and along three parallel lines perpendicular to the flight path.
In many cases, the three line replicate cards were composited
The major focus of this study is to provide a performance
evaluation of AgDRIFT, i.e., an evaluation of the model spe-
cifically in the context of agricultural pesticide product reg-
istration. Previous evaluation efforts on the computational en-
gine of AgDRIFT (AGDISP) have focused on the use of the
technology in high-flight (Ͼ10 m) forestry applications. Pes-

674
Environ. Toxicol. Chem. 21, 2002
S.L. Bird et al.
Table 1. Summary of field-trial conditions
Meteorological conditions
Mean
Minimum
Maximum
Wind speed (m/s) at 2-m height
Relative humidity (%)
4.4
57
25
1.3
7
0
7.7
93
35
Temperature (
Њ
C)
Atmospheric stability
Neutral to unstable
Application parameters
Standard
Variable
Drop size (VMD) (
Flight height (m)
Carrier
␮
m)
238
2.5
Water
160–811
1.6–9.3
Water, oil
Ϯ
0.5
Tracer
Diazinon
Malathion, carbaryl, orthene
Aircraft
Target vegetation
AgHusky
Uniform short grass
AgHusky, AirTractor AT-502, WASP
Uniform short grass, cotton canopy
prior to analysis to reduce analysis costs. Downwind deposi-
tion values used in the analysis were either the composite value
or the average value of the three line replicates at a given
downwind distance when analyzed separately. All data were
collected using the U.S. Environmental Protection Agency
mandated Good Laboratory Practice Standards (GLPS). The
greatest uncertainty in the data is at low-deposition, far-field
These inputs describe the aircraft and its power plant (pro-
peller); the nozzle number, location, and the droplet size dis-
tribution they create; the spray material physical properties;
ambient meteorology; and field and spray geometry.
Aircraft and engine characteristics required as model inputs
include the planform area, fixed-wing semispan or helicopter
rotor radius, flight spraying speed, aircraft weight, propeller
blade radius, propeller or helicopter rotor RPM, spatial loca-
tion of the blade hub relative to the tip of the trailing edge of
the wing, engine efficiency, and aircraft drag coefficient. Spe-
cific characteristics of the three SDTF field-test aircraft—Cess-
na AgHusky 188, Air Tractor AT-502, and Aerodyne Wasp—
were obtained from aircraft compendia [20,21] and from tele-
phone conversations with engineers at the three manufacturing
facilities. Actual spraying speeds for each of the field trials
were recorded in the field as described in [1]. Bilanin et al.
[10] suggest that the aircraft drag coefficient be set to 0.1 and
the engine efficiency be taken as 0.8, values that were retained
in these simulations.
Nozzle placement information required as input includes
the spatial location of the spray boom relative to the centerline
of the aircraft, the vertical distance to the tip of the trailing
edge of the wing (or the rotor plane of the helicopter), and the
number and spacing of nozzles on the spray boom. Specific
details of the boom setup for the three SDTF field-test aircraft
were recorded during the field trials.
The droplet size distributions for the nozzles and test sub-
stances sprayed in the SDTF field trials were obtained from
wind tunnel studies conducted at SpraySearch (Werribee, Aus-
tralia) and New Mexico State University (Las Cruces, NM,
USA) as described in [1]. Typically, three replicates were mea-
sured for each nozzle configuration and averaged for use in
model simulations. In a few cases, where wind tunnel data
were not available for a specific formulation, a substitute test
substance (assumed to behave similarly to the field-trial tank
mix) was used as a surrogate. The wind tunnel drop size dis-
tributions were each collected into 32 size classes. Model de-
position as a function of distance is much smoother when none
of the drop size classes contains more than 2% of the cumu-
lative spray volume. Therefore, wind tunnel atomization data
were linearly interpolated on a volume basis into additional
size classes so that no size class held more than 2%.
collection stations (Ն305 m beyond the downwind edge of the
field), where collection efficiency and sample recovery become
potential data-quality issues [1]. Data beyond the 305-m col-
lector station were not used in the analyses presented here.
The range of meteorological conditions and application vari-
ables during the field trials is summarized in Table 1.
For this analysis, 19 of the 180 treatments were removed
from the dataset for the following reasons. Two of the variable
treatments used a spray tank mix containing two tracers (mal-
athion and carbaryl). Since these two tracers did not stem from
independent treatments, only the malathion results were used
in the analysis. The carbaryl and malathion measured results
agreed well in both of the dual tracer treatments (with an
average difference of 15%). Two treatments were eliminated
due to their nonstandard downwind distance coordinate sys-
tem. These treatments were applied deep inside a cotton field
to examine canopy effects. Four single-swath and two full-
field (20-swath) treatments were removed to make the re-
maining treatments (all with 4 swaths) easier to compare. The
number of swaths of spray applied is a significant factor in
the total mass deposited; keeping these treatments in the da-
taset would have increased the complexity of the analysis and
masked more subtle effects. Thirteen treatments were elimi-
nated when the wind angle deviated substantially (i.e., mean
wind angle plus a standard deviation was greater than 45Њ)
from the card line during the course of the experiment and an
analysis of the field geometry indicated that a substantial
amount of material missed the downwind collectors. The re-
sults presented here are based on the remaining 161 treatments.
Model simulations
AgDRIFT inputs drive the various elements of the model
used to approximate the physics within the wake behind the
spray aircraft and subsequent drift through the ambient at-
mosphere [9,10]. These inputs were either measured during
the field trials, measured in the laboratory or wind tunnel, or
estimated from published literature. No input values were ob-
tained by calibrating to the field deposition results. The com-
plete list of AgDRIFT input requirements is detailed in [9].
Information on sprayed material required by the model in-
cludes the volumetric application rate, specific gravity, non-
volatile application rate, active ingredient application, and the
evaporation rate of the tank mix. Volumetric application rates

Evaluation of the AgDRIFT
௡
aerial spray drift model
Environ. Toxicol. Chem. 21, 2002
675
were calculated from field-measured application rates. Specific
gravity was estimated as 1.0 for all water-based test substances
and 0.92 for oil-based carriers. Nonvolatile and active appli-
cation rates were calculated from the tank mix recipes and
measured volumetric rates. Evaporation rates were obtained
from laboratory data [22], with further correction of the data
for low relative speed between the droplet and the local air
stream [23].
Meteorological and environmental characteristics required
for AgDRIFT include wind speed, wind direction, temperature,
relative humidity, and surface roughness. The meteorological
data were obtained at the SDTF field-study sites with instru-
mented 10-m meteorological towers [1]. Data were captured
at intervals of 1 s for wind speed and wind direction and 10
s for temperature and relative humidity. Field logs time-
stamped the beginning of each trial. Meteorological data were
extracted for 10 min from that time to recover representative
average data. Wind speed and wind direction were averaged
following the unit vector approach [24]; temperature and rel-
ative humidity were obtained as simple averages of the raw
data.
AgDRIFT assumes neutral stability for the air layer in
which the spray material is dispersing and depositing. There-
fore, the model determines the wind speed profile from the
best estimate of the wind speed near release height above the
ground (typically taken as 2 m) and a representative aerody-
namic surface roughness. The wind speed and direction were
sampled in the field at four tower heights (0.33, 1.83, 3.05,
and 9.15 m). The 2-m average wind speeds were estimated
from a least-squares curve fit of the average wind speed at the
four levels. The aerodynamic surface roughness was inferred
from extrapolation of the measured wind profiles that occurred
during near-neutral atmospheric conditions. An average sur-
face roughness was computed for each of the three sets of field
tests. Wind direction is taken as that near the release height,
i.e., 1.83 m. AgDRIFT uses an evaporation module that re-
quires the wet bulb temperature depression as input. That sin-
gle number (collapsing temperature and relative humidity ef-
fects) is calculated with an algorithm based on the Carrier
equation [25], with the assumption that the ambient pressure
was one standard atmosphere in each trial.
The spraying height of the aircraft was obtained by locating
the center of the wheels (or the helicopter skids) from vid-
eotapes taken during the SDTF field trials. These heights were
analyzed and summarized for use as model input by Stewart
Agricultural Research (Macon, MO, USA). For the fixed-wing
aircraft, ground measurements were made of the vertical dis-
tance from the center of the wheels to the tip of the trailing
edge of the wing; for the helicopter, the assumption was made
that the spray boom rested on the skids. The swath widths of
the three test aircraft were all assumed to be 13.72 m (45 ft),
the distance between flight lanes. In all spray trials, the edge
of the field was set as one-half swath width downwind of the
farthest downwind flight line.
Calculations and analysis
Both field and modeled deposition were normalized to the
ideal (zero drift) in-field application rate and reported as a
fraction of this rate. The application rate was calculated using
the measured average flight speed and flow rate, lane sepa-
ration of 13.7 m as the swath width, and a tank mix concen-
tration based on the field mixing recipe. Although tank mix
sampling and analysis was done for each application, doubts
as to the accuracy of this measurement [1] led to the use of
the concentration based on the mixing recipe rather than on
the tank mix analysis.
One of the issues we face when comparing model predic-
tions to observations is the presence of model input errors. If
we can assume that these input errors are random, that the set
of incorrect model inputs at least represents a population of
values similar to the population of correct model inputs, and
that the inputs are somewhat independent, then the errors be-
come less important if we examine the differences in the dis-
tributions of predicted and observed deposition values. Ven-
katram [26] demonstrates the usefulness of comparing distri-
butions. An informative method for comparing distributions
is the empirical quantile-quantile (Q-Q) plot [27]. A Q-Q plot
is a pairing of the predicted concentrations ranked highest to
lowest against the observed concentration ranked in the same
way. If the two distributions are identical, all the points would
lie exactly along the
y ϭ x (or 1:1) line. Departures from the
line give us information about how the distributions differ. The
Q-Q plots provide no information about the temporally paired
relationships of predictions to observations.
Deposition in the SDTF trials varied approximately five
orders of magnitude. Model performance is of interest over
this entire range. The over- or underprediction ratio is a more
important measure of model performance than is the absolute
difference between predictions and measurements. A log trans-
formation of the model predicted/observed ratio is centered
on zero, is symmetrical for the same relative factors of over-
or underprediction, and is approximately normally distributed.
The transformed quantity
log10( log10(DA
is the model predicted/observed ratio, DA is the de-
position predicted by AgDRIFT, and F is the deposition mea-
sured in the field. If the model and data are in agreement
e
is defined as
e
ϭ
E
)
ϭ
/
DF log10(DA
)
ϭ
) Ϫ log10(DF)
where
E
D
E
ϭ
1 and
e
ϭ
0. While both the mean (e¯) and variance (
a ϵ 10e¯
a is a measure of the average
is a multiplicative factor, and
␴e) of
e
are of direct interest in this analysis, the quantities
E
and ␳ ϵ 10 a are also computed.
E
␴
over/underprediction ratio.
when ranges through e¯ Ϯ ␴ , E will vary from ␳ ␳Ea
1Ea to
e
␳
Ϫ
e
and provide an estimate of the expected range of over/under-
prediction ratios. Another measure of model performance rec-
ommended for the evaluation of regulatory models [19] is the
fraction of cases that a model predicts within a specific factor
(either over or under) of the field observation. The fraction
that predicts within a factor of two (f2ϫ), i.e. E ranging from
0.5 to 2.0, is calculated in this analysis. From a regulatory
safety perspective, the fraction of the time the observed is less
than the model prediction multiplied by a safety factor is a
more pertinent statistic of interest and a factor,
the fraction of the time the value of 0.5 was also cal-
culated, which indicates how protective the model results are
likely to be incorporating a safety factor of two.
Although downwind deposition is the primary AgDRIFT
model output, the model results are used in a number of ways
For the comparison between AgDRIFT and field measure-
ments of spray deposition and buffer zones in this study, the
model input datasets were generated independently of the field
deposition results. Neither the model algorithms nor the model
input values were altered in an effort to obtain improved com-
parisons with the field data. In the jargon of model evaluation,
this is considered a hands-off, independent evaluation.
f2t, indicating
E
Ͼ

676
Environ. Toxicol. Chem. 21, 2002
S.L. Bird et al.
Fig. 2. Quantile-quantile (Q-Q) plots at four downwind distances plot-
ting modeled deposition as a function of measured deposition (⅙) for
standard case treatments. The straight line indicates perfect agreement
in the distribution of modeled and measured values.
pendicular to the flight line, but model calculations are cor-
rected to account for the increased transport distance due to
the wind angle deviation from perpendicular. These two ex-
amples illustrate the variability of the line replicates for the
field data and the range of deposition values observed as a
function of distance. Deposition is plotted on a logarithmic
scale since the values can decrease by two to four orders of
magnitude or more from the edge of the field to the downwind
measurement station, making differences between model and
data results difficult to distinguish on a linear scale. In many
cases, as illustrated in Figure 1 for the variable equipment
case, model simulations are within the range of the line-to-
line variability at all downwind distances. In other cases, sim-
ulation results differ substantially from the field results beyond
100 m, as illustrated in the standard equipment example plot
in Figure 1.
The paired treatment design of the SDTF field trials results
in two distinct datasets. In the standard equipment (or diazi-
non) treatments, both equipment (i.e., aircraft, boom and noz-
zle characteristics, release height, and speed) and tank mix
formulation were held constant. This data subset is well suited
for analysis of the response of drift to meteorological factors,
including wind speed and direction, temperature, and relative
humidity. In contrast, the much broader design conditions used
in the variable treatments (malathion, carbaryl, or acephate
tracers) permits this data subset to be used to analyze the
impact of equipment and formulation variables on drift.
Figures 2 and 3 show the Q-Q plots for the standard and
variable cases, respectively. The data have been further seg-
regated according to the downwind distance at which the pre-
dictions and measurements were made. For both the standard
and variable treatments, the modeled distributions of deposi-
tion close to the field of application (8 and 30 m) are similar
to the observed distributions, as indicated by the closeness of
the Q-Q slope to the 1:1 line. With these similar slopes, the
range of observations matches the range of predictions. Since
these are log-log (base 10) comparisons with a slope similar
Fig. 1. Comparison of AgDrift
line data for a standard and variable equipment case field test (
௡
results (———) and individual card
).
●
for estimating exposures to sensitive aquatic and terrestrial
habitats. Defining buffer zones around these sensitive habitats
within which the deposition amounts are equal to or below
acceptable levels is one important end-use of the model. In
particular, it is important to know if the model’s response to
management variables (e.g., drop size distribution, wind speed)
is similar to that found in the SDTF studies and to what extent,
if any, the model possesses bias in estimating buffer zones.
In the context of this analysis, a buffer or no-spray zone is
defined as that distance from the downwind edge of the sprayed
field beyond which the deposition rate drops below a specified
level of concern. Four target levels were chosen for which to
calculate buffer zones and compare the model estimates to the
observations. These values as fractions of the application rate
were 0.1, 0.05, 0.01, and 0.005. Downwind deposition values
from either field measurements or model output were used to
estimate (by simple linear interpolation between measurement
location or, in a few cases, short linear extrapolations) the
distance to the selected fractional rate. For each pair (modeled
and observed) of buffer zones, a model/observed ratio was
calculated and log transformed for analysis and graphical pre-
sentation analogous to the analysis for the deposition results.
RESULTS
Example plots comparing model and field deposition for
two individual trials, a standard and a variable equipment case,
are illustrated in Figure 1. Both the field data and model pre-
dictions are normalized to the application rate and plotted as
a function of downwind distance. In these and subsequent
analyses, downwind distance is reported as the distance per-
to the 1:1 line, any constant arithmetic offset (call it c) from
that line represents a constant multiplicative (10c) relationship
between the predictions and observations. At 8 and 30 m for
the standard treatments, the model is generally predicting at
about 0.8 of the observations over the full range of deposition

Evaluation of the AgDRIFT
௡
aerial spray drift model
Environ. Toxicol. Chem. 21, 2002
677
Table 3. Summary statistics for variable equipment treatments
x
(m)
8
e¯
␴
Ea
␳
f2
f2t
e
ϫ
Ϫ
Ϫ
Ϫ
0.02
0.03
0.03
0.01
0.06
0.05
0.01
0.02
0.02
0.10
0.22
0.28
0.25
0.25
0.29
0.34
0.49
0.64
0.64
0.73
0.95
0.93
0.94
1.02
0.87
0.88
0.97
0.95
1.05
1.27
1.66
1.89
1.76
1.78
1.97
2.18
3.12
4.36
4.36
5.34
0.83
0.74
0.81
0.76
0.73
0.71
0.67
0.47
0.48
0.37
0.87
0.86
0.89
0.90
0.81
0.83
0.87
0.91
0.91
0.85
15
23
30
46
61
91
Ϫ
Ϫ
Ϫ
Ϫ
137
183
305
was in the near-field (15 and 23 m) for the repeated standard
case studies.
A subset of trials where the variable equipment treatments
had application parameters similar to the standard equipment
treatments was analyzed to examine whether the difference
between the model performance for the standard equipment
and variable equipment treatments was due to the difference
in tracers used in the two sets of data or to the range of
application variables. This subset contains all the variable
Fig. 3. Quantile-quantile (Q-Q) plots at four downwind distances plot-
ting modeled deposition as a function of measured deposition ( ) for
variable case treatments. The straight line indicates perfect agreement
in the distribution of modeled and measured values.
⅙
values. At these same distances for the variable treatments,
the highest depositions are only slightly overpredicted, with
the distributions generally matching well. At the most distant
sampling location (305 m) for all the standard cases and the
upper half of the variable treatments, the model is consistently
overpredicting the observations by about a factor of two. How-
ever, for the lower portion of the variable treatments, the model
is underpredicting deposition. These cases represent the large
droplet helicopter cases and will be discussed in more detail
subsequently.
treatments that used the D6-46 nozzle with a 45Њ release angle
on the standard, medium speed, fixed-wing airplane (18 treat-
ments) at a low release height. The variable equipment treat-
ments and the corresponding standard equipment treatments
in this subset have nearly identical application and environ-
mental conditions, differing only in the tracer used. Box plots
as a function of downwind distance are shown in Figure 4 for
both of these standard and variable equipment data subsets.
Similar trends of slight underprediction at short range followed
by overprediction far downwind are seen in both datasets. The
Summary statistics for the standard and variable treatments
as a function of distance are presented in Tables 2 and 3. The
summary statistics (as with the Q-Q plots) indicate that the
model underpredicts the standard treatment results in the near-
data were compared using a
distance. Statistically significant differences (
means only occur at 23 m, and the difference in the means
t
test on the ratio at each downwind
p
Ն
0.05) in
x
ϭ
(0.086) is of no practical importance. Thus, it appears that the
datasets are consistent with regard to the behavior of the tracers
and that the broader range of conditions in the variable treat-
ment accounts for differences in the model/data comparison
occurring in the limited standard case dataset.
field (e¯ is negative and Ea
in the far-field (e¯ is positive and Ea
Ͻ
1.0) while overpredicting results
1.0). In contrast, the
Ͼ
model appears to perform better on average with the variable
treatment dataset and not to have significant under- or over-
prediction problems based on the mean values (e¯ and
However, the measures of variance ( e and ) are much greater
for the variable than for the standard cases. Based on the f2
Ea).
The standard equipment treatments are used to examine the
␴
␳
ϫ
values, the model predicts within a factor of two of the field
data more often (at most distances) for the standard than the
variable treatments. The single-tailed analysis represented by
f2t in Tables 2 and 3 indicates that the model predictions mul-
tiplied by a safety factor of two will generally be in excess of
the observed value over 80% of the time. The only exception
Table 2. Summary statistics for standard equipment treatments
x
(m)
8
e¯
␴
Ea
␳
f2
f2t
e
ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
0.11
0.20
0.21
0.14
0.10
0.06
0.08
0.18
0.26
0.35
0.14
0.19
0.14
0.17
0.20
0.18
0.23
0.28
0.33
0.35
0.78
0.63
0.61
0.72
0.79
0.86
1.20
1.50
1.81
2.25
1.37
1.53
1.37
1.49
1.58
1.53
1.69
1.89
2.16
2.24
0.90
0.68
0.76
0.82
0.83
0.89
0.82
0.73
0.63
0.35
0.90
0.69
0.77
0.85
0.86
0.91
0.95
0.99
0.99
0.95
15
23
30
46
61
91
Fig. 4. Box plots of the ratio of model and field deposition as a function
of downwind distance for variable treatments applied with standard-
like conditions (top) and the corresponding standard case treatments
(bottom).
137
183
305

678
Environ. Toxicol. Chem. 21, 2002
S.L. Bird et al.
Fig. 6. The measured in-the-field and modeled deposition (fraction of
the application rate) at 23 and 305 m downwind as a function of wet
bulb depression.
lustrated in Tables 2 and 3 is explained to some extent by the
discrepancy between the response of the model and field-trial
results to wet bulb depression. The model does not account
for local humidity increases generated by the spray and spray
cloud, which could decrease evaporation rates. These differ-
ences are most likely due to a combination of the limitations
in both the model and the dataset.
Fig. 5. The geometric average (Ea) of the ratio of model and observed
results for standard case trials for wind speed and wet bulb depression
categories as a function of downwind distance.
sensitivity of the model/field ratio to the environmental factors.
Meteorological data collected in the field included wind speed,
temperature, and relative humidity. Wet bulb temperature is a
function of both temperature and relative humidity and serves
as a measure of the evaporative potential of the atmosphere.
We next turn to the variable equipment dataset and examine
the influence of controllable application factors (such as release
height and droplet size) on the ratio of model prediction to
field observations. The model/field ratios that deviate the far-
thest from one occur in the helicopter applications. This is
shown in Figure 7, where the log of this ratio for the helicopter
treatments is plotted as a function of downwind distance. Since
there is a total of only 10 helicopter applications in the dataset,
conclusions drawn based on this small dataset have a higher
level of uncertainty than the much larger fixed-wing dataset.
The treatments are further subdivided by nozzle type. The
largest disagreement between model predictions and field mea-
surements occurs when the D6-SS nozzle (a solid stream noz-
zle that generates a very coarse spray) is used on a helicopter.
At far downwind distances, the model underpredicts by as
much as two orders of magnitude, and the underprediction
occurs at both low and high release heights. Since the heli-
copter treatments with finer sprays have model-predicted/ob-
served values that are close to one and are in line with the
balance of the variable equipment dataset, this difficulty ap-
pears to be associated with the very coarse spray rather than
the helicopter algorithms. The question naturally arises as to
whether this strong droplet size effect also occurs on fixed-
wing aircraft. In Figure 8, the model to observed ratio is plotted
as a function of downwind distance for the fixed-wing aircraft
treatments, split into three different drop size categories based
on the volume median diameter (the diameter at which 50%
of the volume is contained in larger drops and 50% is contained
in smaller drops). The middle graph contains all of the D6-46
nozzle treatments, the top graph contains the finer sprays, and
the bottom graph shows the coarse, solid-stream nozzle treat-
ments. Unlike the helicopter results, no dramatic shift in the
ratio is observed with the coarse spray applications. However,
there appears to be a small downward shift in the ratio (moving
toward underprediction) as the droplet size increases that oc-
The response of model predicted/observed ratio (Ea) as a func-
tion of downwind distance and categories of wind speed and
wet bulb temperature depression are shown in Figure 5. While
for all categories the model/field ratio increases as a function
of downwind distance, there is a substantial difference between
low and high wet bulb conditions in the far-field. In the far-
field, the model/observed ratio is very close to one at low wet
bulb depression values, while during highly evaporative con-
ditions, this ratio is greater than four. The effect of wet bulb
depression is further illustrated in Figure 6, where the mea-
sured and modeled deposition levels are plotted as a function
of wet bulb depression for two downwind distances. At the
near-field location (
measurements nor the model predictions are correlated with
wet bulb depression. At the far-field location ( 305 m),
x ϭ 23 m), as expected, neither the field
x
ϭ
however, the modeled deposition is well correlated with wet
bulb depression whereas the field deposition is not.
The absence of correlation between far-field deposition and
wet bulb depression (as measured in the field study) was un-
expected. In high evaporative conditions, released droplets
evaporate rapidly, decrease in volume (size), and have a greater
drift potential. From this perspective, an increase in the per-
centage of far-field deposition might be expected, and the mod-
el results show this effect. However, these very small droplets
may in fact simply remain airborne and not deposit on the
collectors. However, the far-field data are nearer the detection
level and there is high analytical variability, which makes de-
tection of trends more difficult. In addition, wind deviations
from the sampling media potentially lower sample collection
at the far-field locations. The overall tendency of the model
to overpredict deposition at the far-field sampling location il-

Evaluation of the AgDRIFT
௡
aerial spray drift model
Environ. Toxicol. Chem. 21, 2002
679
Fig. 7. The ratio of modeled and field deposition as a function of
distance for helicopter applications identified by nozzle type and re-
lease height.
Fig. 9. Box plots of the ratio of model and field deposition as a function
of downwind distance for three categories of flight speed.
movement to greater overprediction at higher release heights,
no remarkable features stand out.
curs at most downwind distances. For the faster fixed-wing
application, the droplet size for the coarsest spray is signifi-
cantly finer than for the helicopter application. The measured
volume median diameter for the coarsest helicopter application
A major use of the model is expected to be in calculating
the width of buffer zones needed to limit deposition on sen-
sitive areas to some percentage of the application rate. How-
ever, when the relative performance of the model is examined
for calculating buffer zones for levels of deposition of concern,
this is no longer as significant. In Figure 11, the model/ob-
served ratio based on four target deposition rates (10, 5, 1,
and 0.5% of the application rate) are shown as box plots.
Generally, these results reflect the deposition results. The small
target values show a slight overprediction of buffer zone size
0. The median results show less than a 20% bias. Although
the tendency of the model to overpredict far-field deposition
rates is the feature that draws the most attention in the previous
analysis, this problem is relatively insignificant when pre-
dicting buffer zones protective to 0.5% of the application rate.
Table 4 (standard equipment treatments) and Table 5 (variable
equipment treatments) summarize statistics for the buffer zone
analysis of residuals. For both the standard treatment and the
variable treatment buffer zones for the four deposition levels,
the model predicts a buffer zone within a factor of two 90%
of the time.
is 811
␮m but only 546 ␮m for the coarsest fixed-wing ap-
plication.
The impact of aircraft type or, equivalently, flight speed is
explored in Figure 9. Note that the helicopter graph contains
only six treatments; box plots cannot accurately portray the
distribution of the data when the dataset is this small. This
graph should only be used to gain a sense of the location and
spread of the data; the median and central quartiles have little
meaning with so few data points. The over/underprediction
trend in the helicopter treatments is reversed from the general
fixed-wing trend, with overprediction at short distances and
underprediction farther out. When compared with the medium-
speed fixed-wing airplane, the trend in the fast fixed-wing
airplane case is more exaggerated. Finally, in Figure 10, we
look at the effect of release height on the residuals. Although
we note an increase in residual variability coupled with a slight
Fig. 8. Box plots of the ratio of model and field deposition as a function
of downwind distance for fixed-wing aircraft (variable equipment
treatments) for three drop size categories based on volume mean di-
Fig. 10. Box plots of the ratio of model and field deposition as a
function of downwind distance segmented by release height.
ameter Dv50
.

680
Environ. Toxicol. Chem. 21, 2002
S.L. Bird et al.
Table 5. Summary statistics for variable equipment treatment buffer
zone residuals
Target deposition
(% of application)
e¯
b
␴
Eab
e¯
b
f2.0b
eb
10
5
1
Ϫ
0.0071
0.0041
0.0412
0.0627
0.163
0.98
1.01
1.10
1.15
1.45
1.54
1.50
1.51
0.93
0.91
0.91
0.90
0.1868
0.1753
0.1797
0.5
mask the increase in drift with increasing evaporation. In ad-
dition, the low rate deposition field control spikes did not
bound the deposition values observed at the far-field collectors,
and it is impossible to definitively conclude that there is no
loss from the collectors at these low levels.
Fig. 11. Box plot of buffer zone residual for four deposition rates.
The number of helicopter applications during the trials was
small and all were done with a single type helicopter. Ag-
DRIFT strongly underpredicted deposition for four helicopter
applications from coarse straight stream nozzles. The results
from six other helicopter applications with two other types of
nozzles that produced finer sprays did not display this under-
predictive behavior. The fact that the results from applications
with the other two nozzles were quite good suggests that either
the helicopter algorithms do not handle very coarse droplets
well or the measurements of the droplet spectra for the coarse
nozzle is problematic. One problem may be in the extrapolation
of laboratory (wind tunnel) measured droplet spectra for the
simulation of field conditions. Very large droplets are sensitive
to changes in shear forces and will tend to fragment with
relatively little increase in shear. The wind tunnel configuration
used to simulate atomization experiments considered only the
forward speed of the aircraft and did not simulate any of the
rotary-induced downwash, which changes the effective nozzle
angle. Obtaining an adequate simulation of field-generated
droplet spectra from very coarse spray nozzles may be a lim-
itation for accurate use of the model. Additional model eval-
uation for helicopter applications needs to be pursued since
the number of helicopter trials in the SDTF was very small
and use of helicopters has potential for drift reduction relative
to fixed-wing aircraft [28] since the flight speeds and therefore
air shear are generally lower, facilitating the production of
sprays with fewer relatively small droplets.
In general, the model shows a response similar to that of
the field observations for the application variables, including
droplet size and application height as well as wind speed.
AgDRIFT was also in good agreement with field results for
estimating buffer zones. Generally, this model appears satis-
factory for regulatory evaluations, although care should be
exercised when applying the model beyond the ranges tested
here. Major limitations of the current algorithms are the as-
sumptions of flat terrain and near-neutral atmospheric stability.
Extension of the algorithms to handle stable atmospheric con-
ditions along with the capability to estimate effectiveness of
vegetative buffers and use in simulation of ground sprayer
applications are model development efforts required to in-
crease flexibility of AgDRIFT for use in spray drift exposure
assessments.
SUMMARY AND CONCLUSIONS
The SDTF field trials provided a rich database for use in
model evaluation. The field data have a substantial amount of
random variability that limits the ability to evaluate any model
on a trial-by-trial basis. However, the number of trials used in
the analysis (161) including the 82 standard case replicates of
the same equipment allow us to make statements on the per-
formance of the model in predicting average deposition. The
deposition values from both the field trials and the model
decreased by up to four orders of magnitude over the range
of downwind distances considered in this study.
Overall, AgDRIFT predicts average field deposition levels
reasonably well over the range of application conditions that
are likely to be encountered in aerial agricultural applications.
The model has a slight tendency to underpredict mean depo-
sition levels in the near-field and to overpredict deposition at
far-field distances. Buffer zones show similar over/underpred-
iction trends to the deposition patterns but with lower mag-
nitudes—bias is generally less than 20%. Over 90% of the
distances to a target deposition rate (buffer zone size) com-
puted from modeled deposition were within a factor of two of
the value calculated from field observations for the four target
deposition rates.
While both the observed and modeled deposition responded
to wind speed in a similar fashion, this was not the case for
wet bulb depression (a function of relative humidity and tem-
perature, which is a measure of evaporative potential). Ag-
DRIFT is sensitive to evaporative effects, and modeled de-
position in the far-field responds to wet bulb depression,
whereas the field observations do not. The differential response
to relative humidity is sufficient to account for most of the
overpredictive tendency of the model relative to the field trials
in the far-field. The absence in the field observations of an
evaporation effect on deposition is puzzling. The very small
particles generated under evaporative conditions may have low
collection efficiencies under turbulent conditions, which may
Table 4. Summary statistics for standard equipment buffer
zone residuals
Target deposition
(% of application)
e¯
b
␴
Eab
e¯
b
f2.0b
eb
REFERENCES
10
5
1
0.0342
0.1371
0.1270
0.0595
0.0815
0.1029
0.1168
0.0959
1.08
0.73
0.75
1.15
1.21
1.27
1.31
1.24
1.0
1. Hewitt AJ, Johnson DR, Fish JD, Hermansky CG, Valcore DL.
2000. Development of the Spray Drift Task Force Database for
aerial applications. Environ Toxicol Chem 21:648–658.
Ϫ
Ϫ
0.98
0.95
0.96
0.5
2. Crabbe RS, McCooeye M, Mickle RE. 1994. The influence of

Evaluation of the AgDRIFT
௡
aerial spray drift model
Environ. Toxicol. Chem. 21, 2002
681
atmospheric stability on wind drift from ultra-low-volume aerial
forest spray applications. J Appl Meteorol 33:500–507.
3. Maybank J, Yoshida K, Grover R. 1978. Spray drift from agri-
cultural pesticide applications. J Air Pollut Control Assoc 28:
1009–1014.
4. Yates WE, Akesson NB, Cowden RE. 1974. Criteria for mini-
mizing drift residues on crops downwind from aerial applications.
Trans ASAE 17:627–635.
5. Yates WE, Akesson NB, Coutts HH. 1966. Evaluation of drift
residues from aerial applications. Trans ASAE 6:389–397.
6. Yates WE, Akesson NB, Coutts HH. 1967. Drift hazards related
to ultra-low-volume and dilute sprays by agricultural aircraft.
Trans ASAE 9:628–635.
7. Ware GW, Cahill WP, Estesen BJ. 1974. Pesticide drift: Aerial
applications comparing conventional flooding vs. raindrop noz-
zles. J Econ Entomol 68:329–334.
oak forest and its prediction with the FSCBG model. J Climatol
Appl Meteorol 31:1457–1466.
16. Reckhow KH, Clements JT, Dodd RC. Statistical evalution of
mechanistic water-quality models. J Environ Eng 116:250–268.
17. Parrish RS, Smith CN. 1990. A method for testing whether model
predictions falls within a prescribed factor of true values, with
an application to pesticide leaching. Ecol Model 51(1–2):59–72.
18. Luis SJ, McLaughlin D. 1992. A stochastic approach to model
validation. Adv Water Resour 15:15–32.
19. Burns LA. 1992. PIRANHA: Pesticide and Industrial Chemical
Risk Analysis and Hazard Assessment, Ver 2.0. U.S. Environ-
mental Protection Agency, Environmental Research Laboratory,
Athens, GA.
20. Hardy CE. 1987. Aerial application equipment. Report 8734-
2804. U.S. Department of Agricultural Forest Service Equipment
Development Center, Missoula, MT.
21. Lambert M. 1991. Jane’s All the World’s Aircraft. Jane’s Infor-
mation Group, Surrey, UK.
22. Riley CM, Sears II, Picot JJC, Chapman TJ. 1995. Spray Drift
Task Force droplet evaporation studies. In Hall FR, Berger PD,
Collins HM, eds, Pesticide Formulations and Application Sys-
tems, STP 1234. American Society for Testing and Materials,
Philadelphia, PA.
23. Teske ME, Hermansky CG, Riley CM. 1998. Evaporation rates
of agricultural spray material at low relative wind speeds. At-
omization and Sprays 8:471–478.
24. Haugen DA. 1963. A simplified method for automatic compu-
tation of turbulent wind direction statistics. J Appl Meteorol 2:
306–308.
25. Jennings BH, Lewis SR. 1950. Air Conditioning and Refriger-
ation. International Textbook, Scranton, PA, USA.
26. Venkatram A. 1999. Applying a framework for evaluating the
performance of air quality models. Proceedings, 6th International
Conference on Harmonization within Atmospheric Dispersion
Modeling for Regulatory Applications, October 11–14, Rouen,
France, p 142.
8. Bird SL, Esterly DM, Perry SG. 1996. Off-target deposition of
pesticides from agricultural aerial spray applications. J Environ
Qual 25:1095–1104.
9. Teske ME, Bird SL, Esterly DM, Curbishley TB, Ray SL, Perry
SG. 2000. AgDRIFT௡: A model for estimating near-field spray
drift from aerial applications. Environ Toxicol Chem 21:659–671.
10. Bilanin AJ, Teske ME, Barry JW, Ekblad RB. 1989. AGDIS: The
aircraft spray dispersion model, code development and experi-
ment validation. Trans ASAE 32:327–334.
11. Beck MB, Ravetz JR, Mulkey LA, Barnwell TO. 1997. On the
problem of model validation for predictive exposure assessments.
Stochastic Hydrol Hydraulics 11:229–254.
12. Teske ME, Bowers JF, Rafferty JE, Barry WJ. 1993. FSCBG: An
aerial spray dispersion model for predicting the fate of released
material behind aircraft. Environ Toxicol Chem 12:453–464.
13. Riley MC. 1995. A sensitivity analysis and validation of the AG-
DISP 6.1, FSCBG 4.3, and PKBW2 spray drift and deposit mod-
els. B/95/0008. Environment Canada, Downsview, ON.
14. Richardson B, Ray J, Miller K, Vanner A, Davenhill N. 1995.
Evaluation of FSCBG, an aerial application simulation model.
Appl Eng Agric 11:485–494.
27. Chambers JM, Cleveland WS, Kleiner B, Tukey PA. 1983.
Graphical Methods for Data Analysis. Duxbury, Boston, MA,
USA.
28. Bayer DE, Akesson NK. 1999. Off target movement of phenoxy
herbicides from rice fields during and following aerial and ground
applications. Proc Western Soc Weed Sci 52:114–120.
15. Anderson DE, Miller DR, Wang Y, Yendol WG, Mierzejewski
K, McManus ML. 1992. Deposition of aerially applied Bt in an