The Forest Service, United States Department of Agriculture (USDA), has developed this information for the guidance of its employees, its contractors, and its cooperating Federal and State agencies, and is not responsible for the interpretation or use of this information by anyone except its own employees. The use of trade, firm, or corporation names in this document is for the information and convenience of the reader, and does not constitute an endorsement by the Department of any product or service to the exclusion of others that may be suitable. The U.S. Department of Agriculture (USDA) prohibits discrimination in all its programs and activities on the basis of race, color, national origin, sex, religion, age, disability, political beliefs, sexual orientation, or marital or family status. (Not all prohibited bases apply to all programs.) Persons with disabilities who require alternative means for communication of program information (Braille, large print, audiotape, etc.) should contact USDA’s TARGET Center at (202) 720-2600 (voice and TDD). To file a complaint of discrimination, write USDA, Director, Office of Civil Rights, Room 326-W, Whitten Building, 1400 Independence Avenue, SW, Washington, D.C. 20250–9410, or call (202) 720-5964 (voice and TDD). USDA is an equal opportunity provider and employer. Ann Suter Statistician USDA Forest Service Technology and Development Program Missoula, MT 9E92P32—Technical Service, Aerial Delivery September 2002 Estimating Methods, stimating Methods, Variability, and Variability, and Sampling for Sampling for Drop-Test Data Drop-Test Data Estimating Methods, stimating Methods, Variability, and Variability, and Sampling for Sampling for Drop-Test Data Drop-Test Data United States Department of Agriculture Forest Service Technology & Development Program 5700 Aviation September 2002 0257-2826-MTDC
33
Embed
Estimating Methods, Variability, and Sampling for Drop-Test ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
The Forest Service, United States Department of Agriculture (USDA), has developed this information forthe guidance of its employees, its contractors, and its cooperating Federal and State agencies, and is notresponsible for the interpretation or use of this information by anyone except its own employees. The useof trade, firm, or corporation names in this document is for the information and convenience of the reader,and does not constitute an endorsement by the Department of any product or service to the exclusion ofothers that may be suitable.
The U.S. Department of Agriculture (USDA) prohibits discrimination in all its programs and activities onthe basis of race, color, national origin, sex, religion, age, disability, political beliefs, sexual orientation, ormarital or family status. (Not all prohibited bases apply to all programs.) Persons with disabilities whorequire alternative means for communication of program information (Braille, large print, audiotape, etc.)should contact USDA’s TARGET Center at (202) 720-2600 (voice and TDD).
To file a complaint of discrimination, write USDA, Director, Office of Civil Rights, Room 326-W, WhittenBuilding, 1400 Independence Avenue, SW, Washington, D.C. 20250–9410, or call (202) 720-5964 (voiceand TDD). USDA is an equal opportunity provider and employer.
Ann SuterStatistician
USDA Forest ServiceTechnology and Development ProgramMissoula, MT
9E92P32—Technical Service, Aerial Delivery
September 2002
EEstimating Methods,stimating Methods,Variability, andVariability, andSampling forSampling forDrop-Test DataDrop-Test Data
Estimating Methods,stimating Methods,Variability, andVariability, andSampling forSampling forDrop-Test DataDrop-Test Data
Sampling ______________________________________ 15Spacing in the Downrange Direction __________________________ 15Spacing in the Crossrange Direction __________________________ 15Staggered Spacing _______________________________________ 16
Appendix—Details on Cups, Error Variance, and Grid Spacing ___ 22Cups __________________________________________________ 22Error Variance Estimate for GPC _____________________________ 22Analysis of Variance_______________________________________ 23Grid Spacing ____________________________________________ 24
Contents
3
Introduction
FF or nearly six decades, the ForestService has been using a procedureknown as drop testing to analyze the
ground patterns made by aerial drops offire retardants or suppressants (Suter2000). The procedure involves droppingfirefighting chemicals from an airtankerflying over open cups arranged in a regu-larly spaced grid (figure 1). The groundpatterns allow operators and managers to:
❏ Compare the performance of aerialdelivery systems
❏ Compare the drop characteristics offirefighting chemicals
❏ Determine whether an aerial deliverysystem is suitable for contracting
❏ Investigate the effect of changes in dropheight, drop speed, volume, flow rate,and other factors
After more than 25 aerial delivery systemswere tested during the early 1990s, threeconcerns arose:
❏ Estimation methods
❏ Variability
❏ Sampling
The first concern deals with the processof making estimates using the data col-lected from the grid. It is not feasible tocollect every gram of retardant that hitsthe ground. Instead, the drop is sampledat regular intervals with estimates madefor points in between. Historically, linearinterpolation was used to estimate be-tween sample values. Linear interpolation
assumes uniform change between points,an assumption that may be inadequate fordrop data.
The second concern relates to the varia-bility of estimates and of the test. Any timea quantity is estimated, the variabilityassociated with the estimate should beprovided.
Replicate drops can help investigatorsobtain a measure of the variation inherentin the test. Replication also reduces thevariability of mean line length for eachdrop type. Because of the cost of othertesting constraints, replicate drop tests areusually not conducted, making it impos-sible to estimate reliably the error variancedue to the test.
The third concern, sampling, pertains togrid arrangement and cup placement.Although hundreds of drops have been
conducted over grids, little testing hasbeen done to determine the appropriatecup spacing and grid dimensions (Suter2000). Usually, the length and width ofthe grid are estimated based on flow rate,volume, and ground speed. Some stepshave been taken to achieve greater con-sistency. For instance, cups are placed ata uniform height and spaced in a regularpattern. For most drop tests, a denserarea has been constructed in the middleof the grid where the majority of retardantis expected to fall. Constraints on time,budgets, and labor must be taken intoconsideration when developing a samplingscheme.
This report uses data collected from sixairtanker drops to investigate estimationmethods, variability, and sampling.
Figure 1—TBM (Avenger) aircraft dropping fire retardant over a test grid.
4
Figure 2—Diagram of the test grid.
Flight path
Test Grid (600 by 155 feet)Outer rows have 10- by 20-foot spacing.
Middle three rows have 5- by 20-foot spacing.544 cups, 31 rows
Figure 2 illustrates a 600- by 155-foot grid.A total of 544 stakes were driven into theground so their tops were 4 feet high. Thestakes were staggered to reduce thedistance between known and unknownpoints.
Examination of hundreds of past dropsshowed that the rate of change in cover-age level was often greater crossrange(perpendicular to the flight path) thandownrange (in the direction of the flightpath), especially for drops at high speeds.How close should the cups be placed tocapture this feature of ground patterns?To answer this question, the spacing wasdecreased from 10 feet to 5 feet for threecrossrange rows (figure 2). The effective-ness of the 5-foot spacing was comparedto the 10-foot spacing.
Grid Collection Method
After a drop is made, any cup with moist-ure in it is capped. The row and columnnumbers are written on the lid, identifyingthe location of the cup in the grid. All thecups with lids on them are removed andtaken to be weighed. Clean cups are putback out on the grid for the next drop(figure 3).
Weighing andCalculating
During the weighing process, the weightand coordinates of each cup are enteredinto a computer. The weight of the emptyplastic cup and lid is subtracted from thetotal weight and the weight of the liquidin grams is converted to gallons perhundred square feet (gpc) using the
density of the liquid. For retardant with adensity of 1.095 grams/cubic centimeter,the equation is:
where x is the amount of retardant inthe cup in grams.
Procedures
5
Figure 4 is an example of the computeroutput after drop samples have been
Figure 4—Data array with cup position indicatedin columns 1 and 2 and gpc values in column 3.
Table 1—Summary statistics (gallons per 100 square feet, gpc) for five point-estimation methodsfor drop 201. MAE is mean absolute error and MSE is mean squared error. Triangulation cannotpredict points in the corners of the grid, which is why the triangulation data only include 537 ofthe 543 points in the grid.
Creating Contour Plots
The contour plot in figure 5 is generatedby computer software using an internalinterpolation method to estimate gpcvalues. The software has been found tobe inconsistent. For example, if you lookat the value in the small circle (2.0), you’llnotice that the 2.0 contour line does notinclude this cup, as it should.
These plots help determine line lengths,usually expressed in feet, at differentcoverage levels within a ground pattern.In an attempt to improve the contour plotsand line length estimates, five interpola-tion methods were examined and com-pared. The five methods are polygonaldeclustering, triangulation, inversedistance weighting, local sample mean,and ordinary kriging (Kaluzny and others1998). These five methods are pointestimators that use distance (and otherfactors) as a basis for estimation. Whenestimating points in space, it is generallyassumed that points closer together aremore alike than points farther apart.Under this assumption, more weight isgiven to points that are closer together.
Cross Validation
Cross validation was used to assess theperformance of each of the five methods.Cross validation is a technique wherethe observed sample data are used tomake estimations and the estimates arecompared to the observed sample data.For example, 543 sample values makeup the observed data set in drop 201.One observed value is removed and theremaining 542 values are used to predicta gpc value for the removed value. Oncethat calculation is complete, the observedvalue is put back and another observedvalue is removed. The remaining 542values are used to predict gpc for theremoved value. This process is repeateduntil a prediction has been made at eachof the 543 locations. The result is 543original observed sample values and 543
estimated values at the same locations.The estimates are compared to the ob-served data to determine how well theestimation method performed.
Table 1 shows the cross-validation resultsfor drop 201. Triangulation depends onthree points to make a prediction, so itcannot predict points in the corners of thegrid. For this reason, cross validationproduces fewer predicted values whentriangulation is used. The observed gpcvalues at those sites were removed forcomparison purposes.
The method that produces estimates thatmost closely resemble observed data isconsidered the best. Both triangulation andordinary kriging have means identical tothe observed data. The local sample meanhas the least amount of variability, indicat-ing that it smooths the most. Smoothingis similar to averaging. It provides an
overview of underlying trends, but informa-tion can be lost with excessive smoothing.Examining the five-number summary(minimum, first quartile, median, thirdquartile, and maximum) gives an idea ofthe spread of the predicted values com-pared with the observed. Overall, thepredictions have less spread than the truevalues except when polygonal decluster-ing is used. All of the prediction methods,except for polygonal declustering, smoothdata to some extent. Of the other fourmethods, local sample mean smoothsthe most and triangulation smooths theleast. Triangulation has the highestcorrelation coefficient, while local samplemean has the lowest.
The second part of table 1 displays thesummary statistics for the error of thefive-point estimators. Error (also calledresidual) is the difference between thepredicted value and the true value. The
7
A Comparison of Five Estimation Methods
Summary statistics for five point-estimation methods for drop 203 (gpc)
Inverse LocalOrdinary Polygonal distance sample
TRUE Triangulation TRUE kriging declustering squared mean
Table 2—Summary statistics (gallons per 100 square feet, gpc) for five point-estimation methodsfor drop 203. MAE is mean absolute error and MSE is mean squared error. Triangulation cannotpredict points in the corners of the grid, which is why the triangulation data only include 538 of544 points in the grid.
Summary statistics for error distribution of point-estimation methods (gpc)
Ordinary Polygonal Inverse distance Local sampleTriangulation kriging declustering squared mean
Table 3–Summary statistics (gallons per 100 square feet, gpc) for five point-estimation methodsfor drop 205. MAE is mean absolute error and MSE is mean squared error. Triangulation cannotpredict points in the corners of the grid, which is why the triangulation data only include 538 of544 points in the grid.
(Continued —>)
table of summary statistics for error showsextreme residuals as well as the mean ab-solute error (MAE) and the mean squarederror (MSE). The MSE is the mean of thesquared residuals. Residuals are squaredto eliminate negative numbers. The MAEis the mean of the absolute value of theresiduals. Taking the absolute valueremoves negative signs to provide a moremeaningful statistic. A good predictionmethod would produce low MAE and MSEvalues (Isaaks and Srivastava 1989).
The residual means closest to zero wereproduced by triangulation and ordinarykriging. Triangulation produces the lowestMAE and MSE with ordinary krigingproducing the second lowest.
After examining three drops (tables 1, 2,and 3), triangulation appears to performthe best as a prediction method, withordinary kriging performing second best.These findings indicate that either triangu-lation or ordinary kriging could be usedas a reliable estimator for drop-test data.
Ordinary Kriging
Ordinary kriging is a weighted linear com-bination of the observed data. The weightsare based on a model called the vario-gram. The variogram is the variance ofthe difference between two cups at thedistance between the two cups. Throughmodeling, kriging attempts to minimize theprediction error variance to produce anunbiased estimate (Isaaks and Srivastava1989). Because the findings showed thattriangulation was the best predictionmethod, ordinary kriging was not used.
8
Summary statistics for error distribution of point-estimation methods (gpc)
Ordinary Polygonal Inverse distance Local sampleTriangulation kriging declustering squared mean
Figure 7—Triangles constructed from sample points.
0 100 200 300 400 500 600
150
100
50
0
Feet
Fee
t
Figure 6—Three triangles constructed to estimate an unknown point.
Area1
Point 3Point 2
Point 1
VArea
2Area
3
Triangulation
The Delaunay triangulation method thatwas used is a weighted linear combina-tion. The result is that closer points receivemore weight. Delaunay triangulation usespolygons to determine triangles. In figure6, the known points are points 1, 2, and3. The unknown point is V. Point 1 isweighted from area 1, which is the areaof the largest triangle. This gives point 1the most weight, because it is the closestpoint. Figure 7 illustrates the trianglesgenerated from drop 201.
Triangulation was used to estimate gpcvalues between observed points. Plottingthe estimated points with the observedpoints created the 10- by 5-foot grid infigure 8.
9
A Comparison of Five Estimation Methods
Figure 8—Contour plot redrawn after triangulated gpc values were added to observed gpc values.
150
100
50
0
Drop 201 after interpolationFlight direction: north to south; Aircraft elevation: 150 feet
Windspeed: 5 to 6 miles per hourLow-flow rate: 250 gallons of water per second
The contour plot generated by computersoftware was overlaid as in figure 8. Oncethe triangulation was complete, an algo-rithm was used to calculate line length.
Lengths of retardant line at different cover-age levels are calculated by searchingcrossrange rows for values above athreshold. Line segments begin at the
point of the first downrange value abovethe threshold and end at the point of thelast value. The points immediately uprangeand downrange of the starting and endingpoints are used to perform a linear inter-polation between the two. This techniqueallows reporting lengths with accuracygreater than the grid spacing. Lengthsfor each coverage level of interest are
reported as both longest continuous seg-ment and total length. This provides anindication of overall continuity of the line.Uncertainty in coverage level is appliedas a single estimated value to all pointswhen checking for the threshold condition.A coverage level value of 3.98 will be atthe threshold of 4.00 if the estimateduncertainty is 0.02.
10
Results
Target Height1 Wind Wind DensityVolume flow rate (feet, Height2 GS1 (knots, GS3 (miles direction Relative Fire- (grams perin tank (gallons per Forest (feet, Forest (knots, per (0 degrees Temp humidity fighting cubic
Drop (gallons) Time second) Service) military) Service) military) hour) headwind) (°F) (percent) chemical centimeter)
1 The Forest Service measured height and ground speed using video analysis.2 The military measured height using a radar altimeter (first value) and a self-contained navigation system (second value).3 The military measured ground speed using a pitot tube (first value) and a self-contained navigation system (second value).4 The actual volumes for drops 201 and 202 were not measured. The target volume was 900 gallons.
Length of the retardant line at specific coverage levels (in gallons per 100 square feet, gpc).
Drop 0.5 1 2 3 4 6 8
201 600 521 218 122 105 26 25
202 600 485 375 100 73 14 0
203 600 570 457 185 80 38 21
204 593 524 275 202 118 39 28
205 513 479 373 174 97 68 12
206 511 477 356 190 73 28 0
Table 4—Results from six drop tests over the grid. Wind from 0° would be a head wind. GS is ground speed and RH is relative humidity. GTS-R isa fire retardant.
Tabular
Six drop tests numbered 201 to 206 weremade over the grid and the results col-lected on July 12, 2001. Table 4 showsthe data collected and the calculated linelengths (longest continuous segment).The actual volumes for drops 201 and202 were not measured. The measured
volumes were obtained using a flow meterwhen the tank was being filled. There wasno way to measure volume dropped or theactual flow rate. For this reason, percentrecovery was not included in this table.
The flow rates shown are also a target.No usable flow-rate data were recoveredfrom this drop test. The two numbers for
height and speed were provided by themilitary. For height, one number is froma radar altimeter and the other is from theself-contained navigation system. Forspeed, one number is from a pitot tubeand the other is from the self-containednavigation system. The Forest Servicemeasured height and speed using videoanalysis.
Graphical
Figures 9 through 14 show contour plotsof the six drops after triangulation. Theouter, dark contour represents the lightestcoverage level of 0.5 gpc. The innerhachured contour indicates one of theheaviest coverage levels, 8 gpc.
Drop 206 (3:38 p.m., July 12, 2001)Flight direction: north to south; Aircraft elevation:160 feet; Windspeed, 4 miles per hour;
Medium-flow rate: 500 gallons of retardant per second; GTS-R density: 1.095 grams per cubic centimeter
Cu
ps
Rows
2.0
3.0
0.5
1.04.0
4.0
Flight direction
2.03.0
4.0
0.5
0.5 gpc1.0 gpc2.0 gpc3.0 gpc4.0 gpc6.0 gpc8.0 gpc
14
Variability
Summary of three drop types
Flow rateHeight Speed (gallons per Volume
Drop tests (feet) (knots) second) (gallons) Material
201 and 202 150 to 160 131 250 900 H2O
203 and 204 145 to 160 130 to 133 250 871 to 877 GTS-R
205 and 206 155 to 160 133 to 137 500 821 to 958 GTS-R
Table 5—Summary of the three drop types. The retardant used in the drop tests was GTS-R.
Figure 15—Comparison of line lengths at different gpc levels for drops 201 and 202.
Replicate Drops
To understand the variability betweendrops and within the experiment, replicatedrops were made where the height, flowrate, speed, volume, and material droppedwere constant. The effects of humidity,wind, and temperature were low enoughto be assumed to be negligible. Basically,three drop types were tested with tworeplicates each (table 5).
The winds were equal to or less than 5miles per hour, the temperature wasbetween 70 and 90 °F, and the relativehumidity was between 29 and 56 percent.Lids were placed on all the cups within10 minutes, minimizing the liquid lost toevaporation.
Analysis of Variance(ANOVA) Results
The ANOVA results (appendix A) demon-strate how differences between factorscan be compared. For instance, the meancontinuous line length at 0.5 gpc for wateris greater than the mean, continuousline length for GTS-R retardant (p-valueof 0.000388). It also shows that the linelength associated with the low flow rateof 250 gallons per second is longer at 0.5gpc than the length associated with thehigh flow rate of 500 gallons per second(p-value of 0.0000955).
Graphical Results
Figure 15 illustrates some differenceswithin the replicates. The scatterplotsshow gpc levels by row, presenting ahorizontal profile of the drop. Drops 201and 202 are fairly similar. They have twodistinct peaks, tapering off on either end.The line length charts indicate similar linelengths except for 2 and 8 gpc.
Graph data represent the longest continuous line of water.
Figure 16—Comparison of line lengths at different gpc levels for drops 203 and 204.
The scatterplots for drops 203 and 204(figure 16) are not quite as similar as thescatterplots for drops 201 and 202, butthe line length chart for drop 203 is similarto that for 204, except for 2 gpc.
Figure 17 shows similar profiles for drops205 and 206. Drop 205 has more pointsat coverage levels higher than 4 gpc. Theline lengths are similar with discrepanciesincreasing as the gpc values increase.
Graph data represent the longest continuous line of retardant.
Figure 18—The original 20-foot cup spacing was increased to 40 feet in the north-south directionto evaluate sampling density.
140
60
0 100 200 300 400 500 600
140
60
0 100 200 300 400 500 600
40-foot cup spacing in the north-south direction
20-foot cup spacing in the north-south direction
Co
lum
ns
Co
lum
ns
Rows
Rows
Sampling
Summary statistics for error distributionTriangulation Triangulation
Drop 201 TRUE (gpc) Drop 201 (gpc)
Mean 0.81 0.82 Mean –0.01195
Standard deviation 1.08 1.24 Standard deviation 0.793
Minimum 0.00 0.00 Minimum –7.065
1st quartile 0.02 0.03 1st quartile –0.110
Median 0.33 0.39 Median 0.000
3rd quartile 1.27 1.16 3rd quartile 0.085
Maximum 5.49 10.80 Maximum 3.535
Correlation 0.78 MAE 0.341
n 241 241 MSE 0.627
n 241
Table 6—Comparison of observed gpc values with predicted gpc values from a 40- by 10-footspacing. MAE is mean absolute error and MSE is mean squared error.
Spacing in theDownrange Direction
The original grid design shown in figure2 was compared to a design with widerspacing through cross validation. Everyother row was removed in the observeddata set to obtain the 40- by 10-footspacing (figure 18). This subset of 256values was used to predict, by triangu-lation, values at the 288 sites that wereremoved.
A comparison (table 6) of the predictedgpc values from the 40-foot spacings tothe observed values shows a correlationof 0.78, a mean error of –0.01195 gpcand a median error of zero.
Spacing in theCrossrange Direction
Spacing in the crossrange direction wasalso examined (figure 19). The 20- by 5-foot spacing was compared to a 20- by10-foot spacing.
Figure 20 shows quantile-quantile (QQ)plots. These plots are used to comparethe distribution of estimated to observedvalues. The distributions are equal when x= y or when the data fall on the 45-degreeline. In both plots the distributions aresimilar. The 40-foot QQ plot shows thatthe wider spacing is not going to pick upunusually high cup weights. In effect,widening the spacing smooths the results.Going from a 20-foot spacing to a 40-footspacing is probably too great a jump.
18
Sampling
Figure 19—The original 5-foot spacing was increased to 10 feet in the crossrange direction toevaluate sampling density.
Comparison of observed and predicted values (figure 18)
Ob
serv
ed g
pc
Predicted gpc0 1 2 3 4 5 6 7 8
Predicted gpc
10
9
8
7
6
5
4
3
2
1
0
Comparison of observed and predicted values (figure 19)
Ob
serv
ed g
pc
150
100
50
0
280 310
20- by 5-foot spacing 20- by 10-foot spacing
Center section
Co
lum
ns
at 5
-fo
ot
spac
ing
s
150
100
50
0
Co
lum
ns
at 1
0-fo
ot
spac
ing
s
Rows280 310
Rows
However, a 25-foot spacing may be appro-priate. More information on the comparisonof spacing in the direction of flight can befound in the appendix.
The 10- versus 5-foot comparison showsalmost identical distributions. The corre-lation is high (0.97, table 7) meaning thechange in spacing is not producing a bigchange in the results. Considering thetime and costs, the 5-foot spacing wouldprobably not be necessary.
Staggered Spacing
If drop-test data are viewed as spatialdata, it is assumed that two cups closetogether are more likely to have similarvalues than two that are far apart (Isaaksand Srivastava 1989). To reduce thedistance between cups, the stakes canbe staggered. Without staggering, thegreatest distance between two stakes is22.36 feet. This distance can be reducedto 20.62 feet with staggering. Even thoughthe difference in distance is less than 10percent, this small step can help improveaccuracy.
19
Sampling
Summary statistics for error distributionTriangulation Triangulation
Drop 201 TRUE (gpc) Drop 201 (gpc)
Mean 0.74 0.72 Mean –0.02375
Standard deviation 0.85 0.79 Standard deviation 0.204
Minimum 0.00 0.00 Minimum –0.965
1st quartile 0.01 0.03 1st quartile –0.043
Median 0.33 0.39 Median 0.003
3rd quartile 1.50 1.37 3rd quartile 0.048
Maximum 3.09 2.64 Maximum 0.430
Correlation 0.97 MAE 0.110
n 48 48 MSE 0.041
n 48
Table 7—Comparison of observed gpc (gallons per 100 square feet) values from the centersection of cups with 5-foot spacing with predicted gpc values from the center section of cupswith 10-foot spacing. MAE is mean absolute error and MSE is mean squared error.
20
Conclusions
Cross validation showed that triangulationand ordinary kriging were the two bestestimation methods for drop-test data.Because replicate drops were made, ananalysis of variance was performed todetermine whether differences in linelengths due to the firefighting chemicaland flow rate were significant. Also, crossvalidation helped to determine whetherchanging the grid spacing improved theaccuracy of the results. Either triangula-tion or ordinary kriging are the recom-mended interpolation methods. If the gridspacing is changed, cross validation canbe performed again to see which of thetwo methods is superior.
Replicate drops should be made wheneverinvestigators need to know whether dif-ferences in line length are due to changesin factor levels or whether they are just areflection of the inherent variability in thetest. An analysis of variance can determine
how much variability is due to changesin factor levels versus the variability in-herent in the experiment. Many sourcesof variability are associated with droptesting. For instance, variability exists inhow we measure wind, height, speed, flowrate, and volume. There may also beunknown variation in retardant cloudformation and deposition. The varianceassociated with predicting gpc values mustalso be considered. For more informationon calculating the prediction variance ofa triangulated gpc value, see appendix A.
The investigation into the samplingscheme reveals that increasing the spac-ing reduces the accuracy of the estimates.This fact must be weighed against theadded time and cost of tighter spacing.While going from a 20-foot spacing to a40-foot spacing is probably too large anincrease, the cups could be spaced a littlefarther apart in the downrange direction
without losing much information. In thecrossrange direction, the present 10-footspacing is recommended. A 5-foot spacingwouldn’t give that much more accuracy,but it would cost much more in time andmoney. The appendix examines the pre-dictive capabilities of 20- and 30-footspacings.
Overall, drop testing gives us a relativelygood idea of the performance of an air-tanker in a controlled setting. Drop testswould be even more accurate if a perma-nent grid could be set up. This would allowgreater consistency in the experiment.
Because gpc values from a drop test areused to calculate line lengths, it isimportant to remember that the gpc valuesare simply estimates. Specifications basedon these estimates should probably beexpressed in a range that reflects thevariability around the estimate.
21
Isaaks, E.H.; Srivastava, R.M. 1989. Anintroduction to applied geostatistics.New York: Oxford University Press.561 p.
Kaluzny, Stephen P.; Vega, Silvia C.;Cardoso, Tamre P.; Shelly, Alice A.1998. S+ spatial stats: user’s manualfor Windows and UNIX. New York:Springer-Verlag. 327 p.
Ott, Lyman. 1993. An introduction tostatistical methods and data analysis.California: Wadsworth Publishing Co.1,051 p.
Suter, Ann. 2000. Drop testing airtankers:a discussion of the cup-and-grid method.Tech. Rep. 0057-2868-MTDC. Missoula,MT: U.S. Department of Agriculture,Forest Service, Missoula Technologyand Development Center. 14 p.
References
22
Aerial delivery system—A fixed- or rotary-winged aircraft capable of deliveringfirefighting chemicals.
Aerial drop—A release of firefightingchemical from an aerial delivery systemin flight.
Algorithm—A rule for solving a certaintype of problem.
Analysis of variance—A statistical tech-nique by which the total variation in a setof data may be reduced to componentsassociated with the possible sources ofvariation, allowing the relative importanceof each source to be assessed.
Contour plot—A graphical picture onwhich the characteristics of a surface areshown by contour lines. In drop testing,the isopleths join points of equal coveragelevel on a surface.
Correlation coefficient—A number be-tween –1 and 1 that measures the degreeto which two variables are linearly related.
Coverage level—A recommended amountin gallons of retardant applied to a specificarea (100 square feet) of surface. Cover-age level 2 represents 2 gallons per 100square feet (gpc).
Crossrange—Perpendicular to the direc-tion of flight.
Cross validation—A method of comparingpredicted and observed values.
Data array—Data arranged in a matrixwith columns and rows.
Distribution (frequency)—A frequencydistribution shows the number of obser-vations falling into each of several rangesof values. Frequency distributions aresometimes displayed as histograms.
Downrange—Parallel to the direction offlight.
Glossary
Error (residual)—The difference betweenthe predicted value and the observedvalue.
Firefighting chemicals—Chemical productssuch as long-term retardants and waterenhancers (chemicals containing ingre-dients designed to alter the physicalbehavior of water) used in firefighting.
Fire retardant—Any substance, exceptplain water, that reduces the flammabilityof fuels or slows their rate of combustion.
Fire suppressant—An agent that extin-guishes the flaming and glowing phasesof combustion when applied directly tothe burning fuel.
First quartile—The 25th percentile. Aftera set of values has been arranged in orderof magnitude, the first quartile is the valuethat has 75 percent of the values below it.
Flow rate—The rate at which retardantexits a tank or bucket, usually expressedin gallons per second.
GPC—A unit for measuring coverageexpressed in gallons per 100 square feet.
Grid—A physical array incorporatingcontainers set in a regular, definedpattern to measure deposition patternscreated by the aerial release of firechemicals.
Ground pattern—The characteristics ofground deposition from aerially deliveredliquid.
Histogram—A graph of a frequency dis-tribution table in which rectangles withbases on the horizontal axis are givenwidths equal to the class intervals. Theheights of the rectangles are equal tothe corresponding frequencies.
Isopleth—A line drawn on a map throughall points having the same numericalvalue.
Line length—The length, usually ex-pressed in feet, of a ground pattern. Linelength is used to relate the length of dif-ferent coverage levels within a groundpattern.
Linear interpolation—Estimation of avalue of a variable between two knownvalues when it is assumed there is uniformchange between the two known values.
Mean absolute error (MAE)—The averageof the absolute value of a set of residuals.
Mean square error (MSE)—The averageof a set of residuals after each one hasbeen squared.
Median—The 50th percentile. After a setof values has been arranged in order ofmagnitude, the median is the value thathas 50 percent of the values below it.
P-value—In a hypothesis test, the proba-bility of observing an outcome “morecontradictory to the null hypothesis thanthe observed sample result” is called thep-value (Ott 1993).
QQ plots—Quantile-quantile plots. Agraph comparing the distributions of twovariables.
Replicates—Duplicates. A replicate dropor a duplicate drop is one that has thesame factor levels, specifically, the sameheight, speed, volume, flow rate, and soforth.
Residual (error)—The difference betweenthe predicted value and the observedvalue.
Sampling—The process of selecting asample for testing.
Sampling density—The number of sam-ples in a fixed area.
Tare—The weight of the empty container.
23
Third quartile—The 75th percentile. Aftera set of values has been arranged inorder of magnitude, the third quartile isthe value that has 25 percent of the valuesbelow it.
Glossary
Triangulation—A weighted linear combi-nation used for estimating values atspecific locations. The weights dependon the distance and location.
Variability—Data variability refers to thespread of values along the scale of meas-urement and the extent to which the dataare grouped.
24
[ ]V (tgpc) = V (triangulation) + V (cups) + V (lids) *0.1240872 nc nl
ˆ ˆ ˆ ˆ
Appendix—Details on Cups, Error Variance, and Grid Spacing
Weight range Average Total(grams) (grams) cups
26.85 to 26.95 26.90 5,000
26.75 to 26.85 26.80 4,000
26.65 to 26.75 26.70 3,000
Table 8—Weight of cups used in the six droptests.
Weight range Average Total(grams) (grams) lids
16.45 to 16.55 16.50 2,500
16.35 to 16.45 16.40 8,500
Table 9—Weight of the lids used in the sixdrop tests.
ˆˆ
ˆ
Cups
The following cups (table 8) and lids (table9) were used and their weight recorded.
❏ Average cup weight = [(26.9*5) +(26.8*4) + (26.7*3)]/12 = 26.816667grams
❏ Standard deviation = 0.07993 grams.Variance = 0.0063888049 grams
❏ Average lid weight = [(16.4*8.5) +(16.5*2.5)]/11 = 16.422727 grams
❏ Standard deviation = 0.04191 grams.Variance = 0.0017564481 grams
❏ Tare (average weight of cup and lid)= 43.23939 grams
Combined standard deviation:
√(0.07993)2 + (0.04191)2 = 0.09025
The lowest possible cup and lid weightwas 43.00 grams and the highest was43.50. If a cup with retardant in it weighedless than 43.23939 grams, the computerprogram automatically switched to a tareweight of 43.00 to avoid negative gpc.
At a 99-percent confidence level (CI), themargin of error for the tare weight of
43.2393 grams is ± 0.23249(2.576*0.09025 = 0.23249)
At a 95 percent CI, the margin of errorfor the tare weight is ± 0.17689 grams.(1.960*0.09025 = 0.17689 grams)
Error VarianceEstimate for GPC
The error variance estimator for triangu-lated gpc values is:
Where V (triangulation) is the triangulationvariance. V (cups) is the variance forempty cups, and V (lids) is the variancefor empty lids. nc and nl are the numberof cups and the number of lids, respec-tively. 0.124087 is a constant that convertsgrams of retardant with density 1.095grams per milliliter into gpc.
Mean square error (MSE) is an estimateof the triangulation variance. The threeMSEs are 0.215, 0.262, and 0.256, whichis an average MSE of 0.244.
An example of an analysis of variance(ANOVA) model (figure 21).
26
Appendix—Details on Cups, Error Variance, and Grid Spacing
Grid Spacing
Gpc data collected from a previousdrop test, which used a grid with cupsin a 10- by 10-foot spacing, were usedfor the following comparisons. The 10-by 10-foot spacing provided a data setthat could be divided into subsets forcross validation. Two subsets werecreated with the points in a 20- by 10-foot spacing, and three subsets werecreated with the points in a 30- by 10-foot spacing. Figures 22 and 23 showexamples of these subsets. Tables 10 to14 display the cross validation tabularresults. Figures 24 and 25 display theQQ-plots comparing distributions.
Figure 22—The original 10-foot spacing was increased to 20 feet to evaluate sampling density.
Figure 23—The original 10-foot spacing was increased to 30 feet to evaluate sampling density.
150
100
1,200 1,300 1,400 1,500 1,600 1,700 1,800
Original points, 10- by 10-foot spacing
20- by 10-foot spacing
Feet
1,200 1,300 1,400 1,500 1,600 1,700 1,800Feet
Fee
t
150
100
Fee
t
150
100
1,200 1,300 1,400 1,500 1,600 1,700 1,800
Original points, 10- by 10-foot spacing
30- by 10-foot spacing
Feet
1,200 1,300 1,400 1,500 1,600 1,700 1,800Feet
Fee
t
150
100
Fee
t
27
Table 10—First comparison of observed gpc (gallons per hundred square feet) values from a10- by 10-foot spacing with predicted values from a 20- by 10-foot spacing. MAE is meanabsolute error and MSE is mean squared error.
Summary statistics for error distributionTriangulation Triangulation
TRUE (gpc) (gpc)
Mean 1.534 1.509 Mean 0.02483
Standard deviation 0.905 0.819 Standard deviation 0.386
Minimum 0.037 0.068 Minimum –1.856
1st quartile 0.950 0.962 1st quartile –0.144
Median 1.380 1.419 Median –0.021
3rd quartile 2.008 2.052 3rd quartile 0.150
Maximum 5.081 5.428 Maximum 2.206
Correlation 0.905 MAE 0.246
n 319 319 MSE 0.149
n 319
Table 11—Second comparison of observed gpc (gallons per 100 square feet) values from a 10-by 10-foot spacing with predicted values from a 20- by 10-foot spacing. MAE is mean absoluteerror and MSE is mean squared error.
Summary statistics for error distributionTriangulation Triangulation
TRUE (gpc) (gpc)
Mean 1.524 1.550 Mean –0.02628
Standard deviation 0.922 0.805 Standard deviation 0.416
Minimum 0.040 0.093 Minimum –1.044
1st quartile 0.909 0.994 1st quartile –0.201
Median 1.412 1.411 Median –0.035
3rd quartile 2.012 2.071 3rd quartile 0.096
Maximum 8.183 4.060 Maximum 4.122
Correlation 0.893 MAE 0.241
n 308 308 MSE 0.173
n 308
Table 12—First comparison of observed gpc (gallons per 100 square feet) values from a 10- by10-foot spacing with predicted values from a 30- by 10-foot spacing. MAE is mean absoluteerror and MSE is mean squared error.
Summary statistics for error distributionTriangulation Triangulation
TRUE (gpc) (gpc)
Mean 1.553 1.534 Mean 0.01857
Standard deviation 0.962 0.736 Standard deviation 0.574
Minimum 0.051 0.053 Minimum –1.337
1st quartile 0.947 1.044 1st quartile –0.217
Median 1.400 1.481 Median –0.020
3rd quartile 1.978 2.069 3rd quartile 0.147
Maximum 8.183 3.676 Maximum 5.265
Correlation 0.803 MAE 0.334
n 396 490 MSE 0.329
n 396
Appendix—Details on Cups, Error Variance, and Grid Spacing
28
Appendix—Details on Cups, Error Variance, and Grid Spacing
Table 13–Second comparison of observed gpc (gallons per 100 square feet) values from a 10-by 10-foot spacing with predicted values from a 30- by 10-foot spacing. MAE is mean absoluteerror and MSE is mean squared error.
Summary statistics for error distributionTriangulation Triangulation
TRUE (gpc) (gpc)
Mean 1.520 1.534 Mean –0.01385
Standard deviation 0.904 0.787 Standard deviation 0.450
Minimum 0.037 0.081 Minimum –1.426
1st quartile 0.914 0.980 1st quartile –0.218
Median 1.374 1.468 Median –0.017
3rd quartile 2.046 2.018 3rd quartile 0.137
Maximum 8.183 3.956 Maximum 4.259
Correlation 0.867 MAE 0.279
n 418 418 MSE 0.202
n 418
Table 14–Third comparison of observed gpc (gallons per 100 square feet) values from a 10- by10-foot spacing with predicted values from a 30- by 10-foot spacing. MAE is mean absoluteerror and MSE is mean squared error.
Summary statistics for error distributionTriangulation Triangulation
TRUE (gpc) (gpc)
Mean 1.531 1.538 Mean –0.00609
Standard deviation 0.881 0.846 Standard deviation 0.480
Two comparisons of predicted and observed valuesThe predicted values come from using a 20- by 10-foot grid to triangulate values to form a 10- by 10-foot grid.
Three comparisons of predicted and observed valuesThe predicted values come from using a 30- by 10-foot grid to triangulate values to form a 10- by 10-foot grid.
30-f
oo
t ro
ws,
10-
foo
t co
lum
ns
Observed
8
7
6
5
4
3
2
1
00 1 2 3 4 5 6 7 8
30-f
oo
t ro
ws,
10-
foo
t co
lum
ns
Observed
7
6
5
4
3
2
1
00 1 2 3 4 5 6 7
30-f
oo
t ro
ws,
10-
foo
t co
lum
ns
Observed
THIRD COMPARISON
SECOND COMPARISONFIRST COMPARISON
Appendix—Details on Cups, Error Variance, and Grid Spacing
30
Notes
31
Notes
32
Notes
33
Additional single copies of this doc-ument may be ordered from:USDA Forest Service, MTDC5786 Hwy. 10 WestMissoula, MT 59808-9361Phone: 406–329–3978Fax: 406–329–3719E-mail: [email protected]
An electronic copy of this report willbe available on the Internet at:http://fsweb.mtdc.wo.fs.fed.us/cgi-bin/enter.pl?link=pubs/htmlpubs/htm02572826/
For further technical information,contact Ann Suter at MTDC.Phone: 406–329–4815Fax: 406–329–4763E-mail: [email protected]
Suter, Ann. 2002. Estimating methods,variability, and sampling for drop-testdata. 0257-2826-MTDC. Missoula, MT:U.S. Department of Agriculture, ForestService, Missoula Technology andDevelopment Center. 30 p.
Discusses the testing process the ForestService has used for the past six decadesto analyze the ground patterns made byaerial drops of fire retardants or suppres-sants. The process involves droppingfirefighting chemicals from an airtankerflying over open cups arranged in aregularly spaced grid. This report uses
data collected from six airtanker drops toinvestigate estimation methods, variability,and sampling. Five estimation methodswere compared: triangulation, ordinarykriging, polygonal declustering, inversedistance squared, and local sample mean.Cross validation showed that triangulationand ordinary kriging were the two bestestimation methods for drop-test data.Replicate drops should be made when-ever investigators need to know whetherdifferences in line length are due tochanges in factor levels or whether theyare just a reflection of the inherent vari-ability in the test. Investigation of the
sampling scheme shows that increasingthe spacing of the cups reduces the ac-curacy of the estimates. In the crossrangedirection (perpendicular to the flight path),a 10-foot spacing is recommended. In thedownrange direction (in the direction of theflight path), spacing could be increasedslightly from the present 20 feet withoutseriously affecting the accuracy of theestimates.
Ann Suter is a statistician for theMissoula Technology and DevelopmentCenter’s Wildland Fire Chemical Systems
About the Author
group. She joined the Forest Service in1997 after serving 2 years as a PeaceCorps volunteer in Jamaica, where sheworked on reforestation and soil erosion
control. She holds a master’s degree ininternational development from theAmerican University.