THE LARGE AREA CROP INVENTORY EXPERIMENT_Presen… · type experiment (ref. 3) was proposed as a first agricul-tural experiment that should be conducted by such a system. In 1969,

THE LARGE AREA CROP INVENTORY EXPERIMENT

..

,

CROP INVENTORY

Presented By

R. B. MacDONALD, CHIEF

EARTH OBSERVATIONS DIVISION

NATIONAL AERONAUTICS AND SPACE ADMINISTRATION

lYNDON B. JOHNSON SPACE CENTER

HOUSTON, TEXAS 77058

To The

2ND ANNUAL

WI lLiAM T. PECORA

MEMORIAL SYMPOSIUM

Sioux Falls, South Dakota

October 25 - 29, 1976

THE LARGE AREA CROP INVENTORY EXPERIMENT

R. B. MacDonaldEarth Observations Division

National Aeronautics and Space AdministrationLyndon B. Johnson Space Center

Houston. Texas 77058BIOGRAPHICAL SKETCH

R. B. MacDonald is presently Chief of the Earth ObservationsDivision at the Johnson Space Center. He has designed andexecuted major efforts in the development and application ofremote sensing technology. The foremost example is theLarge Area Crop Inventory Experiment (LACIE). initiated in1974. This multiagency project is to develop, test. anddemonstrate in a "quasi-operational" environment the tech-nology to produce agricultural crop production informationon a global scale.Before joining NASA in 1971. Mr. MacDonald served as Tech-nical Director of the Laboratory for Application of RemoteSensing at Purdue University. He was chief architect of the1971 Corn Blight Watch experiment. which involved more than1000 participants from 17 federal and state agencies. Thisproject was described as the largest experiment ever con-ducted in agriculture and is to this day still. perhaps, themost operational and complete demonstration of what remotesensing can do.

ABSTRACTA Large Area Crop Inventory Experiment (LACIE) has beenundertaken jointly by the U.S. Department of Agriculture(USDA). the National Oceanic and Atmospheric Administration(NOAA) of the Department of Commerce. and the National Aero-nautics and Space Administration (NASA) to prove out aneconomically important application of remote sensing fromspace.The first phase of the experiment. which focused upon deter-mination of wheat acreages in the U.S. Great Plains and uponthe development and testing of yield models. has been com-pleted. The results and conclusions are pre~ented. Thesecond phase of the experiment concentrated on the estima-tion of wheat acreage, yield. and pr.oduction over severalimportant production regions. The results are discussed.A preliminary assessment to date indicates that the perform-ance goals of the experiment are being met.

INTRODUCTIONOrganizations and personnel responsible for the maintenanceand operation of food production systems. designers anddevelopers responsible for incorporating research outputsinto improved operational food production systems. andresearchers concerned with developing the basic technologyand methodology to produce ever greater quantities of food

with higher nutritional value vitally need timely and accu-rate information of the state of agricultural production andconditions of supporting natural resources on a worldwidescale. Events in recent years have clearly indicated thatimproved information of this type needs to be produced topermit the prediction of world food shortages far enough inadvance to permit governments to deal properly with foodcrop supplies.Small grains are man's most important food crops; theseinclude the wheats, rice, barley, rye, and oats. The worldgrain trade has grown from 113 million tons a year in 1971-72to approximately 154 million tons currently. The UnitedStates has been primarily responsible for this growth andcontinues to be the largest international trader of smallgrains; consequently, as a major supplier, the United Statesis dependent on the market demand around the world which isa function of the production in foreign countries. The pricefor crops the United States or any seller receives is estab-lished by the interactions of production and demand.The price tag for the United States has increased signifi-cantly in recent years. Congressional legislation in theearly 1970's removed restrictive export regulations andestablished a policy of full production. The largest U.S.exports are now in the agricultural category; U.S. farmexports increased from $8 billion in fiscal year 1972 tonearly $22 billion in fiscal year 1975; in comparison, U.S.aerospace exports in 1975 were at a new high of $7.8 billion.It is interesting to note that, on the average, one out ofevery three acres harvested in the United States and onedollar in four of farm income are currently dependent onexport sales.Wheat is the most abundant of the small grains and provides20 percent of the total food calories consumed by the world'spopulations (ref. 1). It is also by far the most interna-tionally traded small grain. Wheat is the largest nonforagecrop in the world and the second largest in the United States,which grows 14 percent of the world total and is secondonly to the U.S.S.R. as a producer. In 1975, nearly70 million acres were harvested from 75 million acresplanted in the United States. Approximately 2.1 billionbushels or 58 million metric tons were produced from thisacreage (ref. 2). Over 575 million acres are estimated tobe planted throughout the world each year.

BACKGROUNDIn the early 1960's, the NASA, in cooperation with the USDAand on the advice of the Committee on Remote Sensing forAgricultural Purposes of the National Academy of Science, ini-tiated research to investigate the feasibility of assessingagricultural conditions with automated remote sensing tech-niques. A consortium of universities, NASA, and USDAresearch groups was tasked to conduct these investigations.The principal groups were the Willow Run Laboratories of theUniversity of Michigan and the Laboratory for AgriculturalRemote Sensing at Purdue University. This consortium madesteady progress from 1965 through 1973.

Certain milestone events were posted during this interval.In 1965, multispectral measurements were collected withexisting ground-based sensors and with the first aircraftscanner developed at the University of Michigan. Dataacquired by this scanner over agricultural fields wereanalyzed by a digital data processing system which wasassembled at Purdue University in 1966, clearly showing thefeasibility of automatically identifying wheat and othermajor crop types (ref. 3). In 1967. 1968 specificationsfor the first Earth Resources Technology Satellite (ERTS;now known as Landsat) to be equipped with a multispectralscanner (MSS) were established. Interestingly, a LACIE-type experiment (ref. 3) was proposed as a first agricul-tural experiment that should be conducted by such a system.In 1969, the ERTS was simulated by a camera system (ref. 4)flown on Apollo IX and indicated the promise of the futureERTS-A (ref. 5). In 1971, the corn crops in the UnitedStates were victims of a new disease - the southern cornleaf blight. The technology was thrown into services inthe Corn Blight Watch experiment to survey the 1971 cropover a seven-state region every 2 weeks throughout the1971 corn-growing season. This effort provided considerableexperience to the LACIE (ref. 6). In 1972, the ERTS-A wassuccessfully launched. The Earth Observations Division ofthe Science and Applications Directorate at the Lyndon B.Johnson Space Center (JSC) NASA conducted joint experimentswith the USDA establishing the feasibility of surveying majorcrop types from space with the multispectral remote sensingtechnology (ref. 7). All of these events led directly tothe proposal (ref. 8) of the LACIE in 1973. Severe shortagesin wheat production occurring in 1972 stimulated the develop-ment of the experiment.The LAClE was formally initiated late in 1974. It wasannounced November 6, 1974, and described briefly bySecretary of State Kissinger at the World Food Conferencein November 1974 (ref. 9) as follows:

Our space, agriculture, and weather agencies willte~t advanced satellite techniques for surveyingand forecasting important food crops. We willbegin in North America and then broaden the projectto other parts of the world. To supplement theWorld Meteorology Organization (WMO) on climate,W~ have begun our own analysis of the relationship'Jetweer. climate patterns and crop yields over a~tatist;c?11y significant period. This is a prom-ising and Dotentially vital contribution to rationalplanning of global production.

A BRIEF DESCRIPTION OF THE LARGE AREACROP INVENTORY EXPERIMENT

The LAClE is a cooperative project of the USDA, the NASA,and the NOAA of the U.S. Department of Commerce. The majorgoals of the experiment are:a. To evaluate and demonstrate the capability of existing

technology (remote sensing, data processing and analysis,and other associated technologies) to make improved

worldwide crop production information available todecision makers in a cost-effective manner in a testconducted in a quasi-operational environment.

b. To research and develop alternate approaches and tech-niques which, upon evaluation, are qualified to beincorporated into the LACIE quasi-operational systemwhere required to meet performance goals or to improveefficiency.

The experiment is being conducted over three consecutive cropseasons in a 3-1/2 year timespan and is divided into threecorresponding phases. Each phase is designed to build on theexperience of the previous phase(s). Phase I was conductedduring the 1975 crop year and concentrated on a system testto identify and estimate the wheat acreage within selectedu.s. wheat growing regions and to test wheat recognitionanalysis in other regions selected throughout the world.During this phase, mathematical wheat yield models weredeveloped and yield feasibility determinations conductedover selected regions in the United States. Phases II andIII concentrated on bringing all elements of a systemtogether in a quasi-operational environment to test thetechnology's capability to develop area, yield, and produc-tion estimates for U.S. test regions and other major wheat-producing r~gions of the world.The LACIE is composed of two elements conducted in parallel.A major effort involves the operation of an experimentalsystem throughout the crop season to produce periodic esti-mates of wheat acreage, yield, and production in selectedregions and to evaluate the accuracy of those estimates. Asecond element is involved with researching and developingimproved approaches for possible incorporation and testingin the experimental system. It is believed that the exten-sive testing conducted in the LACIE program is essential toadvancing the technology to a level where it is ready to beincluded in a future operational system.A decision to require no ground truth in the analysis ofLand Satellite (Landsat) data is an important factor in thetechnical approach selected in LACIE. In order to have anapproach thought to be more operationally feasible and costeffective, no use of current season ground truth is made toderive LACIE acreage estimates.It is also important to note that the systems that supportthe experiment have been assembled, for the most part, fromavailable components designed for research and development.The LACIE is intended to test the techniques and functionsnecessary for crop production inventorying and not to providea streamlined, cost-effective operational system. The intentis to utilize the experimental results to support, as a con-current effort. the design of a user-oriented operationalsystem and an associated estimate of the performance andcost of such a future system.Performance is evaluated on a number of criteria. Theseinclude the accuracy. timeliness, and objectivity of esti-mates produced from the experimental processes of LACIE as

early as possible in a crop year and periodically throughoutthe season. In the United States, estimates are first madein the December-February time frame and then regularly on amonthly basis from April to the end of harvest. Landsat dataare analyzed within approximately 30 days after satelliteacquisition in such fashion that a 15-day-total operationalcapability is demonstrated. The LAGlE estimates are producedusing objective and repeatable procedures and are not modi-fied using current season intelligence from other sources.The accuracy goal of the LAGlE is a 90/90 at-harvest cri-terion for wheat production information. This specifiesthat production estimates which are made at harvest for aregion or country be within 90 percent of the true produc-tion, 90 percent of the time, i.e., 9 years out of 10. Inaddition, the LAGlE is to establish the accuracy and reli-ability of early season estimates and of estimates made atregular intervals throughout a crop season prior to harvest.The 90/90 at-harvest goal is based on the premise thatinformation of this quality, with an associated timelinessand objectivity, would be an improvement over what is avail-able from conventional sources of wheat production estimatesfor regions outside the United States.The LAGlE is designed to estimate wheat production at aregional or country level for selected major producingregions of the world. Each region to be surveyed is strati-fied into relatively homogeneous parts within which produc-tion tends to be more uniform. Wheat acreage and yield foreach stratum are estimated and multiplied together to obtainan estimate of the production for the stratum. These esti-mates of production in the strata are then added to arriveat an estimate of production at regional and/or countrylevels. Wheat acreage and yield are each treated in thestrata and aggregated to form estimates of both acreage andyield at regional and country levels.Wheat area estimates are derived from analysis of LandsatMSS data (ref. 10) collected over 5- by 6-mile sample seg-ments, statistically located over the survey region. Thesegments represent approximately 2 percent of the agricul-tural areas in each of the survey regions. Mathematicalmodels are used to calculate an estimate of the wheat acre-age for the total survey region from the wheat acreage esti-mates made for the 5- by 6-mi1e sample segments. Statisticalpattern recognition analysis of the multispectral dataacquired by Landsat over the samples is performed to iden-tify wheat on a computerized system-at NASA/JSG.Yield is estimated with mathematical models relating yieldto principal meteorological conditions and other factorswhich determine the yield of wheat. Precipitation and tem-perature are primary variables in the LAGlE models. Ini-tially, these data are being obtained from the WorldMeteorological network of ground weather stations. Inlater stages of the experiment, the use of supplementalmeteorological data from weather satellites is planned.The yield models have been programmed on digital computersat the National Meteorological Center of NOAA and are runwith current estimates of meteorological conditions to pro-duce yield estimates for all strata.

Additionally, wheat growth models have been developedrelating the physiological growth stage of wheat to theresponsible factors. These factors are primarily daylength and daily maximum and minimum temperatures. Esti-mates of crop stage, produced by these models, are beingutilized initially in the analysis of the Landsat data butwill also be important to the operation of improved yieldmodels in the later phases of LACIE.The first phase of LACIE concentrated on the hard, red wheatcrop area of the United States and was completed in April1976. The selected region is comprised of nine states inthe U.S. Great Plainsl which, on the average, accounts for90 percent of the summer and spring, hard, red wheat and is70 percent of the total U.S. wheat crop. This region hasbeen referred to as the "yardstick" area and is included inall phases of the experiment. Additionally, exploratorysegments were analyzed in other major wheat producingregions around the globe.Winter wheat is normally found in the southern portions ofwheat producing regions and is a fall-planted crop which isharvested in the late summer. Spring wheat is confined toareas with more severe winters and is a spring-planted cropwhich is harvested in late summer. Normally, the yields ofwinter wheat varieties are higher than the yields of springwheat.In Phase II, the LACIE system was augmented to make periodicestimates of yield and production throughout the crop season.Also, the wheat regions experimentally surveyed by LACIEwere logically expanded to include portions of Canada andthe U.S.S.R. Periodic estimates of acreage, yield, andproduction were developed and assessed throughout the 1976crop season. Again, exploratory segments were selected inother regions in the Northern and Southern Hemispheres.The final selection of regions for Phase III will be basedon Phase II results.Much of the detailed assessment of performance of the tech-nical approach is accomplished with ground data collectedover the yardstick area. A number of sample sites, approxi-mately 30 square miles in size, have been selectivelylocated throughout the area. These are referred to asintensive test sites and were selected to support assess-ment analysis. In addition, a number of the 5- by 6-milesample segments are randomly selected from the variousstratum after wheat acreage proportion estimates have beencompl~ted. Ground-truth data are then collected over theseblind sites to support an evaluation of the lACIE approach.

lTexas, Colorado, Oklahoma, Kansas, Nebraska, Montana,Minnesota, North and South Dakota.

2Sites are referred to as blind sites because lACIE analystsare not aware of which sites have been selected prior totheir completion of the initial analysis and development ofresultant LACIE estimates.

In addition to the intensive and blind-site segments. theLACIE relies upon data compiled by the Statistical ReportingService (SRS) of the USDA as a reference for determining theaccuracy of the LACIE estimates in the yardstick area. TheSRS data tend to be quite accurate at a national level andmarkedly degraded at smaller geographic levels. To date.estimates of errors in the SRS estimates have not beenincorporated into the LACIE performance analysis. Thus.SRS estimates are used as best estimates of the acreage.yield. and production at the national and regional levels.Only a very crude estimate of LACIE performance can be madeoutside the United States where relatively poor "truth"data is available.Several hundred other sites denoted as "exploratory" seg-ments have been selected in the major wheat producingregions of the world to be analyzed in Phases I and II inorder to begin to identify the unique characteristics ofdifferent wheat producing areas of the world.

DISCUSSION OF RESULTSThe experiment was extremely successful in piecing togethera total system capable of processing the quantities ofLandsat data scheduled in Phase I. Approximately 2604 acqui-sitions from 693 segments were analyzed in total; 411 ofthese segments form the selected sample population in theGreat Plains yardstick region. Although the satellitepasses over each segment every 18 days. the probability ofLandsat's seeing a segment is approximately 60 percent dueto cloud cover.An average of about 12 hours was required for analysis inPhase I to derive a wheat proportion estimate for a 5- by6-mi1e sample segment. An average total of 30 days wastaken to move a segment from its moment of acquisition byLandsat through to a final proportion estimate. It is esti-mated that in an operational environment. the required timecould conservatively be reduced to less than 15 days.In Phase II. a total of 9276 qcquisitions over 1720 segmentswas collected and analyzed. The system was augmented inPhase II with a higher speed computer processor to supportthe increased processing loads planned in Phases II andIII. The average time required for a computer classifiedanalysis of a sample segment was reduced from 12 hours to6 hours in Phase II. In Phase II. every cloud-free Landsatacquisition over each segment was reviewed by the analystfor computer classification rather than restricting theanalysis to one acquisition for each of four crop biowin-dows as in Phase I. This change was effected to improvethe accuracy of early season estimates and was based onPhase I experience. In order to handle the increased dataload. two additional types of analysis routines were added.A "no change" analysis routine required an analyst to over-lay a computer classification map from a previous acquisi-tion over a color-infrared image created from the newacquisition and manually determine if any significant changein wheat acreage had taken place. The average time required

for this was approximately 1 hour. A third type of routinerequired an analyst to manually interpret a color-infraredimage made from the Landsat multispectral data and countwheat pixelsl where less than 5 percent of the sample seg-ment was judged to be in wheat. The average time for thistype of analysis was approximately 2-1/2 hours. Thus far.Phase II experience indicates that considerable accuracyimprovement was realized in the early season estimates byutilizing every cloud-free Landsat acquisition. As addi-tional experience is gained regarding the optimum biowindowsfor identifying wheat. it is expected that selected acquisi-tions will be adequate.In Phase II. the LACIE system was successful in acquiringand processing the meteorol09ical data from the World Mete-orological Organization (WMO) stations through the yieldand crop growth models programmed on digital computers.Thirty-day average values of precipitation and temperaturewere utilized in the yield models in Phase II. Daily maxi-mum and minimum temperatures were collected as inputs forthe wheat growth stage model.Again. the LACIE system proved extremely successful inbeing able to acquire. process. and interpret the datavolumes and rates required. No significant problem isanticipated in the later phases. based on the experienceto date.Significant experience was acquired in Phase I in the anal-ysis of Landsat data to estimate wheat acreage. A consid-erable amount of time was required to locate and eliminatesystem and analysis bugs in the experiment. One significantanalysis bug dramatically affected the early season LACIEacreage estim&tes until it was located during the latterhalf of Phase I. Bare soil was correctly classified as suchbut was erroneously aggregated as wheat acreage in earlyestimates. This led to high overestimates of wheat acreagesin both the spring and winter wheat area reports untilearly acquisitions were replaced by later season data. Nearthe end of Phase I. this and other less significant bugswere corrected. and a final analysis of all the Landsatacquisitions was completed.Some further explanation is required to understand andinterpret LACIE results. Elementary statistical theoryindicates that a good statistical estimator has severalproperties. It should be unbiased; i.e .• the expectedvalue of the estimator should be equivalent to the truevalue; and the precision or variance of the estimatorshould be relatively small. In LACIE. a statistical indi-cator. the coefficient of variation. the standard deviationof the LACIE estimate divided by the mean of the estimate.is used to describe the variability of the LACIE estimates.The relative difference between the LACIE estimate and thetrue value. as approximated by SRS. ground truth. or some1 Each of the individual resolution elements acquired by the

Landsat MSS is referred to as a picture element or pixel.

other comparative source, and expressed as a percentage ofthe LACIE estimate, is analyzed for the presence of bias inthe evaluation of the accuracy of LACIE. Without a timeseries of LACIE estimates developed over a number of years,it is difficult to definitively and directly estimate anyexisting bias. In the LACIE, the relatively simple T-testis used to test the relative difference for statisticalsignificance, i.e., bias. As previously discussed, theaccuracy performance goal of LACIE is the 90/90 criteria.If the estimate is unbiased, the random component, as esti-mated by the coefficient of variation (c.v.), can be on theorder of 6 percent at the regional or country level. Ifthe estimate has a 10 percent bias, then the c.v. must bezero to meet the 90/90 goal. If this allowable coefficientof variation of 6 percent for unbiased estimates is equallyproportioned to acreage and yield, then the coefficient ofvariation can be on the order of 4.25 percent for acreageand for yield estimates. Thus, if the estimates areunbiased, the c.v. 's can be 4.25 percent or smaller; how-ever, in the presence of a bias, the c.v. 's must be propor-tionately smaller. Thus, the LACIE estimates for yield,acreage, and production are analyzed in terms of whether ornot relative difference from a true value is judged to bestatistically significant and thus indicative of a bias.Additionally, the coefficient of variation is analyzed interms of the aforementioned simplified allocation of errorto the acreage and yield estimations.After isolation and correction of all significant analYSisand implementation bugs and upon reworking the data, theresulting wheat area estimate at harvest for the U.S. GreatPlains was deemed marginally satisfactory in considerationof the 90/90 at-harvest goal for wheat production. However,the LACIE Great Plains area estimate was approximately46 000 000 acres, compared to the SRS end-of-year estimateof 51 000 000 acres, or about 10 percent below the baselinefigure. This relative difference is statistically signifi-cant at the la-percent level. A major contribution to thisunderestimate occurred in North Dakota and is believed to bea sampling problem. It is also believed that an improvedallocation and an increase in the number of samples on thebasis of a better geographic partitioning of agriculturallands into a more homogeneous strata would reduce any biasto a tolerable level. The use of Landsat imagery is criticalto defining adequate strata. The c.v. computed for the LACIEarea estimator was about 5.7 percent, which was slightlyabove the desired 4.25 percent needQd to meet the 90/90 pro-duction goal. Because the data loss due to early implemen-tation problems resulted in a reduction in the number ofLACIE sample segments, this random error component isexpected to be reduced to or below the target of 4.25 per-cent in Phases II and III.Additionally, accuracy was examined for selected samplesegments and the results indicated that the Landsat data andthe classification technology estimated small grains areawithin a sample segment accurately enough to meet the per-formance goals. In North Dakota, where 20 such sites wereexamined in detail, no significant difference was detectedbetween the LACIE and ground observations over the samplesegments.

The estimated c.v. of the random classification error wasalso acceptably small. These analyses indicated that bias,introduced by factors such as Landsat resolution, lack ofspectral resolution, and the classifier are not excessive.Results of these tests did indicate a difficulty in dif-ferentiating wheat from other closely related small grains.However, in those cases, wheat area estimates were obtainedthrough the reduction of the small grain area estimates inaccordance with the historic prevalence of these crops.Twelve yield regression equation models, developed inPhase I for the Great Plains region, were tested with his-toric data for the years from 1965 to 1975. SRS estimatesof wheat acreage were used to weight the meteorologicalparameters because no LACIE estimates exist for that10-year record. Initial tests indicated that the yieldestimates, combined with SRS estimates of acreage, margin-ally would not satisfy the 90/90 production criteria givenequal or greater errors in the area estimates. Additionalanalysis of the problem revealed that a significant sourceof the yield estimation error was in the form of the modelwhich resulted in unrealistically high or low yield esti-mates for extremely high or low values of temperature orprecipitation. To correct this situation, the meteorologi-cal input values were not permitted to exceed specificupper and lower bounds. The models, when retested, werefound to be acceptable. In 10 out of 11 years, the yieldestimates, when combined with acreage as estimated by SRS,were found to be within tolerance of the true production -again, as determined by the SRS historic reports.For the 1975 crop year, the estimated yield for the GreatPlains was 4.3 percent higher than the value estimated bythe SRS. The c.v. was estimated to be 2.3 percent.When the Phase I (1975 crop year) lACIE area estimates werecombined with the yield estimates for the same period, theresulting production estimates met the 90/90 goal in theGreat Plains yardstick area. This produced an at-harvestproduction estimate for the U.S. Great Plains of1 291 098 bushels compared to 1 363 400 bushels as esti-mated by the SRS. The LACIE production estimate was5.6 percent below the SRS final estimate. With a C.v. of5.9 percent and a deviation of ±76 174 bushels, a differ-ence of 5.6 percent is not significant (i.e., could likelybe a r~ndom fluctuation), and the estimator was judged tosatisfy the accuracy goal. It;s important to note thatan allocation of 637 sample segments was originally madeto provide LACIE with a capability for making U.S. acreageestimates at the country level, having approximately a2-percent sampling error, i.e., the error due to the numberand location of sample segments. Early in the experiment,the decision was made to concentrate on the nine-statehard, red wheat region as a yardstick area to permit expen-sive, in-depth accuracy assessment analysis to be conductedthroughout the duration of the experiment. The yardstickincludes approximately 70 percen~ of the U.S. wheat cropand is representative of an extremely wide range of condi-tions of weather, cropping and agricultural practices, and

field sizes. The 411 segments are considered to have anassociated sampling error in excess of 2 percent for theGreat Plains region. Thus. in Phase I. a major questionwas the adequacy of 411 segments to support the accuracygoal at the yardstick level. At the end of Phase I. it wasconcluded that the allocation was marginally acceptable withcertain exceptions. such as in the North Dakota area.Landsat imagery was analyzed to measure the variability ofthe proportion of small grains over the strata. This mea-sure. together with the correlation of the proportion ofwheat to small grains, permitted an improved estimation ofthe number and location of segments to be made. On thebasis of such analysis, the number of segments in NorthDakota was increased in Phase II. A study in Phase II alsoinvestigated the need to restratify the total survey regioninto agricultural and nonagricultural areas and furtheraugment the original 411 segments to provide a samplingerror on the order of 2 percent at the nine-state level inPhase III.On the basis of an evaluation of Phase I final results, itwas concluded that the technology and state of understand-ing was sufficient to expand the experiment to include addi-tional wheat-producing regions in Phase II.LACIE was particularly interested in the repeatability ofPhase I results in the Phase II crop year. In addition.critical attention was placed on an evaluation of how wellthe yield models would perform in foreign regions wherehistoric data was thought to be of considerable poorerquality than that of the United States. Also in questionwas an issue of how well the models might perform underabnormal weather conditions that might occur in some partsof the United States, Canada, or the U.S.S.R. Fortuitously,the 1976 crop year did provide a somewhat radical departurefrom normal weather patterns in the U.S. Great Plains yard-stick region. The monthly average precipitation for thefive winter wheat states is shown in table I. It shouldbe noted that much of the above-average November precipita-tion occurred at a time when the crop was entering dormancy.As a result of an evaluation of the Phase I experience,significant changes were made for Phase II. These includedthe following:a. A requirement was instituted to have the complete

analysis of a segment conducted"by a single analyst oranalyst team as opposed to having a series of differentanalysts perform the different functions required todevelop a proportion estimate for a 5- by 6-mi1e samplesegment. This afforded analysts an opportunity todevelop an understanding of the interactions of thevarious analysis procedures, thus leading to a moreaccurate final estimate.

b. Every cloud-free acquisition of each sample segment wasto be analyzed as opposed to utilizing one acquisitionin each of four different biowindows. This change wasrequired because of the uncertainty of estimating the

biowindow of wheat at a specific time as we11 as a lackof understanding of the best times to differentiatewhedt from other confusion vegetation.

c. Agrlcultural are6S were differentiated from nonagri-cultural areas using full-frame Landsat imagery.

In add~tion to such changes, a large experience factor wascarried forward into Phase II. It is important to notethat in Phas2 I very little experience with the refinedtechnology had beer obtained in the early season reporting.Therefore, major emphasis in Phase II was placed on evalu-ating LACIE performance in early season estimating in theU ,S. G I'eat P 'I a ins, Can a d a, and U. S . S . R . In P has e I r, jus tas in Fhase T, LACIE estimates were required to be producedprior to the availability and release by official USDAsources. Ail the evaluation of performance at the regionallevel was ~one against the official estimates after theirrelea:c

T~e ~)76 wheat j2dr turned out to be an abnormal one; theG rea ( P: a i '~:;'" par tic ui ,1 r was s t r u c k by d r 0 ugh t 'j n bot hthe ,,;);i::~J arid winter wheat areas and generally i)roved tob2 ~ d"ff1CU , challenge for USDA forecasting services aswell 05 f~r L~CIE.

A1U1(J'':C!;~, P'lose !l 11as been concluded, only the resu"ltst' r l': 0 ;, .j LJ '.( 3 0 ,1 <; 76, w i 11 be discussed in this paper. Asa' ,2: v", t :J f d a,; a 52 YIS ~ t iv itY , the C 0 lil pie t ere 5 u 1t s forPi'd 0:e 1 t.,' r o"L b I" ) L' J 1 i s red un t i 1 i 20 d a 'j s aft e r t he fin a 1S (..S .. ,r~)' :;,! !) il 1) y p p 0 r tis r e 1 e a sed i n 1 9 7 6 .

Ii _ liif1o,'caH tc iJnd2fstand the way in which the refer-'],'" y.,>r:1 I"U'lC,:i8ilS because LACIE estimates are evaluated,n part, as~ins: ,he USDA system in the Great Plains yard-

stic' recju~" Tt b~came apparent to LACIE Phase r experi-rnerot:rr, (h~_ v"j'elge and yield estimates produced by the;'e !: .c;'r:, ..•~ s; c.. t e r.; (.~ R S) d u r in 9 the fir s t ha 1 f 0 f the sea son')" e- . ': "<; " ,L € ')") f ;:he. a m € qua n t "it i e s be i n 9 estimated byl A";; _ " J. n ..1 " .: PIT.L e r; S R S n,a k e sam 2 a sur i:? men t 0 f a c l' e s.'i~W e~i ,;r,d ?;t"ii"ate5 yield for seedeo acres and production

n ("',~:o,"rorn '\:I,efive rrevious years. Current year ccneli-o " ,.c' P 'JA t 5 ;, '" F! a1 sou t i 1 i z.:: (j • LAC lEe J r 1y sea son est imat e s

~re::Or e!ii~I~~f'C. a(reage anc vieid for :lcres for harvest.'ilj;n'fjd·~Y, (It' 'jSDf\ estimate is converted to acres for har-j (; " f 'Vi .• ' r :" U G' i S 0 f m ai 1 sur v ey 5 a Ii d mea sur 2 In en t S 0 f

.' ,,:;:j'1;~ e 'ot~ .. Ii: 1976, SRS produced a spe(ialr· c ,; i : ! .. ',' C . '11.' S eve red r 0 u 9 n tin t /'e Grea t P'I,,' n s .Ir r ~ s .tuation, the SRS necessarily usee aWEdtne(' niouel to estimate yield for acres for harvest. Ina norma' yea,", objective measurements are made only on fieldvi~its !O ~ample fields each month after April through har-ie.:;i. ir: 0:' ':'1' ;:) den';e month"ly yield estimates. A mp.asure-

III e T1 -: .: S if d de::, f Ci. ere s for h a r v est d u r i n 9 1ate J un e andrepon",(] on June 30 This is the most accurate measurementof acrea !0p hapves~ up to that point in the year. Theestimates made by SRS in 1976 are show~ in figure 1. AsOPpOS2C to tnis, LACIE observes wheat tnat is sufficientlydeveloper to bE detectable in the Landsat data. It is then

assumed that these acres are acres for harvest. These acresare monitored for significant changes using Landsat datathroughout the season until after harvest. Yield is esti-mated on the basis of weather conditions from early seasonthrough to harvest.The SRS estimates in the yardstick region for the fivewinter wheat states is shown in figure 1. It is interestingto note that a major adjustment was made to acres forharvest in April. Also, a second major adjustment was madeto acres for harvest in the June 30 enumerative survey. Amajor adjustment was made by SRS in April in yield. Theseadjustments are reflected in the production estimates. Asignificant upward estimate was made for production on thebasis of the June 30 acreage and June yield updates. TheLACIE estimates for Phase II are shown in figure 2. Itshould be noted that the average age of data in the LACIEestimates is on the order of 45 days. In an operationalsystem, the LACIE estimates would, therefore, be delivereda month earlier. Considering this, it can be seen that theMay LACIE area estimate agrees well with the April SRSacreage estimate. The June 1 acreage estimate agrees quitewell with the SRS June 30 best acreage-for-harvest estimate.The LACIE yield estimates appear to have accurately pre-dicted the yieldS throughout the season. A significantadjustment in June was later substantiated by a July SRSyield adjustment. It would appear that the LACIE produc-tion estimates also accurately reflected the changing condi-tions in the Great Plains through June. LACIE was estimatingwheat production significantly higher in June than was SRS.However, the end of July SRS adjustment served to substan-tiate the accuracy of the LACIE June estimate. Again, inan operational system, the LACIE June estimate would havebeen achieved in May. In summary, the LACIE productionestimate for the southern portion of the yardstick areareached the 90/90 goal in May, as compared to the SRS refer-ence estimates, and exceeds the goal for there through thelate June-early July period. In fact, the -8.9 percentrelative difference between the LACIE May estimate and theSRS early July estimate is not statistically significant atthe 90-percent level with a C.V. of 5.9 percent. The C.V.of 5.9 percent is considerably less than the 8 percentrequired at the 5-state level to meet the 90/90 criterionat the Great Plains level. The relative difference in pro-duction in early June is on the order of 1/2 percent with ac.v. of 5.4 percent. This clearly exceeds the 90/90 goaleven at the 5-state level. The relative difference betweenthe LACIE May area estimate and the SRS June 30 acreage esti-mate is approximately -7 percent with a C.V. of 5.6 percent.The relative difference of the LACIE early June acreage esti-mate and the SRS early June 30 estimate is about -3.4 percentwith a C.V. of 5.2 percent. Thus, it appears that the LACIEearly season estimates in Phase II are actually approaching,and then exceeding, the 90/90 goal in the May-June timeframe. Estimates for the U.S.S.R. winter wheat indicatorregion were first developed in April. It should be noted,however, that the Landsat data contributing to these esti-mates could have been arrived at in February.

Table II compares at-harvest Phase I and Phase II area,yield, and production estimates for the five-state winterwheat portion of the U.S. yardstick region. The earlyseason and monthly estimates for this area are shown intable III. A monthly comparison of LACIE and SRS estimatesare charted in figure 2. The LACIE estimates for the winterwheat region of the U.S.S.R., together with estimates com-piled by the Foreign Agricultural Service (FAS) of theUSDA are shown in table IV and figure 3. It is interestingto note that the situation appears to follow the pattern ofthe Great Plains. A relative difference of 22 percent isindicative of a bia, between what LACI£ detected as wheatand what was reported to have been seeded. Again, a falldrought cleated a general reduction in fall germination.It was suspected that the June report made from laterLandsat acquisition; should indicate a Significantlygreater estimate of acres for harvest on the rationale thatmany wheat fields had not previously developed to a detect-able stage beca~se of the drought. As can be seen, theLAC r E ,! i..! nee s t I f'~ate, in cor po rat in g d a t a fro mAp r i 1 t h r 0 ughmid-May, increased significantly. Production was within2 perce"t of th·'~ FAS estimate, acreage was under the FASesc'l:l!.Hl:' by 5 p:'rcera, and yield was over the FAS estimateby some 4 percer<t. The c.v. of the LACIE April area esti-mate was dpproximately 9 percent and 7 percent for the Juneestimate. Thus, the c.v. 's are within the 9-1/2 percentrequired at the indicator region level to meet the:-J 0 / 9 0:1 Ii a 'i -3t': hen il t ion all eve 1 . The c r 0 pre p 0 r t est i -mdta~ d compG'ed to reference sources are summarized int;':l.·'? I',. :"! ~l,d IV dnd charted on figures 1, 2, and 3.T, f: ,(-'·,,1;1: !1(J ~, ( '.1. Pi) i1 S e I I can c e n t rat e don the s p r i n g w he at(~:; 'C: m'; ,,:: 5 ; (, :; n· n :' ( \: her n are d s of the U. S . Great P1 a ins,"e 9 ::':. I;' C F ..:c r,., .' a, and ani nd icae era rea in the s p r in 9,\i"',, r :-" ~; CO!: C', ': Ii e U.S. S .R . The s ere s u 1 t s will bed i s -

',' . " ,. t J ~.' , r.: t e ~0 r t .

Tr: 'ld,'lDdIY,;\.'/C years of LACIE have p,oduced resu'its which" r ; n 91 j' S L!~.;,0 t u C rj Ii ten ti 0 nth a t r e ill 0 t e s ens i n g sup po r t e d

Gy lanc~at tngether with an ayro-meteorological approachto SSUiTe: to:) yielc.. is capable of providing superior earlyseaSvi: end bt-hai'vesc estil1ates in major wheat producingreq';Cj:5;' 'n':: l'iOI'ld. l\lthough the technology needs to bef U i' t } e r ( va' u " Le J i n P has e I I I, t h u s fa r t he res u 1 t s havebee :;vi '-/ice j'<J':;ing. itis a!so in~portant to note thatsigr."icc'"jc ii';Jlov~mEnt shou'id he exp..:cted in the future.:1', i lil'l"P'I~i' 'iP:':c,i1C:ll.cC

J remo~e sensing tecnnology anaa~'.~( .:r. 'I"~ 1.rE ~irly develc.Ol;lentol stage. In addition,t: i: e i:' i s (' n I ,/ c.; i m .~ e d Lin a e r 5 tan d: n g 0 f the f act 0 r s w h i c haffect the accuracy of remote sensing crop surveys. Thetechniques developed and evaluated in LACIE do appear toprovide a suitable base at this time for the design andtes~jng of an optimized system for the USDA. The USDA,withln LACIE, is currently in the process of developingthe conceptual design for such a system.

REFERENC ES1. Quisenberg, K. S., ed.: Wheat and Wheat Improvement.

American Society of Agronomy, 1967.2. USDA: Crop Production Report, July 12, 1976. Crop

Reporting Board, SRS, USDA.3. Remote Multispectral Sensing in Agriculture. Report of

the Laboratory for Agricultural Remote Sensing, Vol-ume 3. Purdue University Agricultural Experiment Sta-tion, Record Bulletin No. 844 (Lafayette, Indiana),1968.

4. SChowengerdt, R. A.; and Slater, P. N.: Final Post-flight Calibration Report on Apollo 9 Mu1tiband Photog-raphy Experiment S065. Optical Sciences Center,University of Arizona (Tucson, Arizona). May 1972.

5. MacDonald, R. B.; Baur, M. E.; Allen, R. D.; Clifton,J. W.; and Erickson, J. D.: Results of the 1971 CornBlight Watch Experiment. Eight International Symposiumon Remote Sensing of Environment, (Ann Arbor, Michigan)Oct. 2-6, 1972.

6. r~acDonald, R. B.; Hall, F. G.; and Erb, R. B.: The Useof Landsat Data in a Large Area Crop Inventory Experi-ment (LACIE). Proceedings of the Tenth InternationalSymposium of Remote Sensing of Environment, Volume I(Ann Arbor, Michigan), Oct. 6-10, 1976.

7. Erb, R. Bryan: The ERTS-1 Investigation (ER-600).Volume I - ERTS-l Agricultural Analysis. NASATM X-58l17, 1974.

8. MacDonald, R. B.: A "Data Acquisition Only" ERTS-BProposal for a Large Area Crop Inventory ApplicationsSystem Verification Test Project. Earth ObservationsDivision, Science and Applications Directorate, LyndonB. Johnson Space Center (Houston, Texas). December1973.

9. Excerpt of a speech by U.S. Secretary of State Henry F.Kissinger at the International Food Conference in Rome,Italy. November 4, 1974.

10. Goddard Space Flight Center: Data Users Handbook - NASAEarth Resources Technology Satellite. Goddard SpaceFlight Center (Greenbelt, Maryland), September 1971.

TABLE I.- MONTHLY AVERAGE PRECIPITATION FORWINTER WHEAT STATES IN "YARDSTICK" REGION

4 c:J "NORMAL" 1931 - 74

~ 1975 76....... -

...... . ',

.....:<:j

OCT NOV DEC JAN FEB MAR APR MAY JUN

.....&..,"' ... , .. , .. . . . . . . . . ............... ..... ...... .

·········1 .............. ..... ...... ..... ...... ..... .......... I··· ..····............... ' , .}::::::J~ I' } .. :.:: i

.;.;.;.:. '';';';.;. \

-:::::< .. :« i..........~ ..... ... - .i :::;:::> j ::::;:::;:......... " ...... .

.......

.......

i.·.···.···..........

;; 3wJ:UZ:::Z0i=<tl-ii:uwa::~

MONTH OF THE YEAR

TABLE 11,·- r, COnPARISON OF PHASE I AND PHASE II RESULTS FOR1HE F!V~-STATE WINTER WHEAT PORTION OF THE

NINE-STATE YARDSTICK REGIONlComparison with SRS estimates: relative

difference ± coefficient of variationof LACIE estimate]

Period Area Yield ProductionPhase I (1975) -0.13% ± 7% 4% ± 3% 5% ± 7%Phase II (1976) -6% ± 5% -0.4% ± 2% -6% ± 5%

TABLE III. - MONTHLY COMPARISONS OF LACIE AND SRS ESTIMATES FOR U.S. SOUTHERN GREAT PLAINS WINTER WHEAT6 6[p = production x 10 bushels; A = acreage x 10 acres; Y = yield bushels/acre.]

-- ~~bH·--~~f~.-~~b~:r~-~=y-=r~~=M<~='-~--:---~~;1~-~y---.----.-. -llay June late Junf' July----- --, .-- --'p P A Y P A

yP A

yP A

y---. , ..-...... - ---

SRS 659 3.32 19 9 1 611 25,7 23 8 631 25 4 24.9 611 25,3 24 4 27,3 116 27.3 26.2-.- -- - -- ..,. --.

lACIE - 1614 24.4 21.6 60J 22.3 27 0 566 21.9 25.8 638 25.'> 2'>,a 699 26.4 26 4 26.9 695 26.2 26.S-' ..---- .- -.

R.D. 2.0 -36 27 8 -9.2 -48,9 26.a -7.9 -17.4 7 7'> 1.2 -0 -0 11.8 4.3 7.6 -1.5 -3.0 -4.0 1.1.. ~-

C. V. 9.0 9.0 2,0 7.5 8,0 3.3 7.r 7.0 3,a 5,9 5.6 2.8 S.4 5,2 2.3 5,3 5.3 S.3 2.3..R.D. = Relative DifferenceC.V. = Coefficient of Variation

TABLE' IV.- PHASE II RESULTS OF COMPARISON OF LACIE AND INDEPENDENT ESTIMATES IN USSR WINTER WHEAT INDICATOR REGION

[Area = Acres x 106i yield = bushels per acrei production = bushels x 108].•. -

~

February a April June JulySource . Area Yield Production Area Yield Production Area Yield Production Area Yield Production

1-- -.- --FAS 32.1 42.0 13.5 28.1 36.4 10.3 28.1 37.5 10.6,-..--f--- ,-.- fo---------

LACIE 20.2 39.2 8.0 26.6 37.4 10.1 29.2 37.9 11.2••......- _._--, _'_,,_' _ ••••••• -.-- __ 0 --,-.--.",.,--",,- ~~-'-_.-, "'-' ...-- .. ,.-.-.

aFAS February projection is not released as an estimate.

JO,000

Mo•...x~ 20,000a:uc(

10,000

SEEDED ACReSDEC ENUMERATIVESURVEY_____ llIiiii211:£201il

ACRES JUNE ENUMERATIVEFOR HARVEST SURVEY'HARVEST INTE'.TlONS '1(_HISTORIC ABANOONMEfH"'- .-'Y',MAIL SURVEY -_.~ __ •

ADJUSTMENTBV SPOT CHECKSIN OBJECTIVE VIElD

30·

wQ;;uc(

3II) 20o..JW

>-

10

~

..VIELD/HARVESTED ACRE, ~ __ ••• _(WEATHER MODEll ~ __ •

,(IELD/H VESTED ACRE(OBJECTIVE YIELD SURVEV)--------VIElO/SEEDED ACRE

HISTORIC TREND

TJAN FEB

IMAR APR MAY JUNE

REPORT DATE

IJULY AUG

TJAN FEB

I I I IMAR APR MA Y JUNE JUL Y AUG

MS! 800,000)(

II)..J

'"%II)::lII) 600,000zo~U::lQoa: 400,000IL

---------- a__ .•HISTORIC • __ , ~ __ •TREND

SOUTHERN GREAT PLAINS

• COLORADO

• KANSAS

• NEBRASKA

• OKLAHOMA

• TEXAS

T_JAN FEB

I I

MAR APR MAY JUNE JULY AUG

Figure 1.- SRS estimates in yardstick region (southern Great Plains).

•••

TJAN

TJAN

30 ,000

x::3 20,000a:uc(

10,000SRS SEEDED AREA

SRS HARVEST AREA+ LACIE ACRES ± lOA

I IFEB MAR APR MAY JUNE JULY AUG

REPORT DATE

30

wa:uc(

:>UI 20o..JW

>

10

I ----1-+_. -t----+--I--~--- -.-.-

--------

- - -- SRS YIELDI LACIE YIELD ± lOy

I I I IFEB MAR APR MA Y JUNE JULY AUG

-+.--L-L__a--.._-.&-_

JAN HB MAR APR

M~ 800,000Xen..Jw~en:)al 600,000zo~u:)ooa: 400 ,000CL

._.~ SRS PRODUCTION

LACIE PRODUCTION ± lOp

IMAY

.l._.._---4__ ._J.--JJUNE JUL Y AUG

SOUTHERN GREAT PLAINS

• COLORADO

• KANSAS

• NEBRASKA

• OKLAHOMA

• TEXAS

Figure 2.- Monthly comparison of LACIE and SRS estimates inyardstick region (southern Great Plains).

40,0001,000,000 50

I • _ FAS AREJl

T + LACIE AREA t lOA L_ ....L_A_C_I_E_Y_I_E.•..L_D__ --"'- ------TAPR MAY JUNE JUL Y AUG

JO

•••a:C~U•••%::;;C~za

20

-.--.-

FAS YielD.-

._ ..u._._._

30

40

'"~)(

75,000 en w••• a:a: 0<:( c(~ Ciiuw ...•% W

:2:en:;)III

50,000 0...•w>

TAUGJULYJUNEMAYAPR

3C,000M0•..Xenwa:0c(

20,000

1,400,000 .-.-.-.-.-.M

~ 1.200,000)(en...•w:2:en:;)III 1,000,000zo~o:;)o~ 800,000Go

- • - .w ••.

I' _ FAS PRODUCTION

LACIE PRODUCTION

T-----'~""""----~ TAPR MAY JUNE JULY AUG

Figure 3.- Monthly comparisons of LACIE ~~d FAS estlmat2s for theU.S.S.R. winter wheat indicator region.

NASA-JSC