REMOTE SENSING OF ENVIRONMENT 12:509-518 (1982) 509

The Use of Large-Area Spectral Data in Wheat Yield Estimation

T. L. BARNETT and D. R. THOMPSON

National Aeronautics and Space Administration,* Lyndon B. 1ohnson Space Center, Houston, Texas 77058

Large-area relations between satellite spectral data and end-of-season crop yield were investigated. Green Index
Number (GIN) values from Landsat MSS data of sample segments throughout the U.S. Great Plains winter wheat belt
in 1978 were correlated to county USDA-SRS reported yields. A linear relation between GIN and yield appeared to
exist up to GIN values of 40 or 50, covering cases of severe to moderate stress. In a test on 1978 Texas winter wheat at
the county level, GIN values for sample segments in the counties were used in conjunction with an agronomic-meteoro-
logical yield model. The combined fit explained significantly more of the observed yield variation at the county level
than the agromet model alone.

1. I n t r o d u c t i o n cations to major crops--for example,

Strommen et al. (1978), Feyerherm and
Satellite imagery has revolutionized Paulsen (1981), T h o m p s o n (1962),
many aspects of large-area surveys of the Williams et al. (1975), and Baier (1973).
earth's surface for a variety of applica- All of these models are based on
tions. Data from the Landsat satellites weather-related variables that depend on
were used successfully in the 1975-1977 two observables: temperature and pre-
Large Area Crop Inventory Experiment cipitation. These two observables are the
(LACIE) to estimate wheat acreage over only ones that are at the same time
the U.S. Great Plains (MacDonald and strongly related to crop yields and
Hall, 1980). Improved techniques are routinely available world-wide, both his-
being implemented in the current Agri- torically and in terms of real time. De-
culture and Resources Inventory Surveys pending on the model, inputs may be on
Through Aerospace Remote Sensing a daily or monthly basis, and the observa-
(AgRISTARS) program. Use of satellite bles may be used directly, accumulated
spectral data in estimation of crop yields over longer periods, or transformed into
is also an attractive prospect since yield is other, presumably more meaninghd vari-
related (through indirectly in the case of ables such as potential evapotranspiration
grains) to crop vigor, which is in turn or moisture deficit. Spectral data from
related to the spectral response of the satellites represents a potential third ob-
crop measured by a satellite sensor. servable, providing additional informa-
In recent years a variety of agronomic- tion, for use in yield estimation.
meteorological (agromet) yield models In recent years several efforts have been
have been developed for large-area appli- directed either at building a yield model
based on Landsat spectral data or at ex-
*Contribution from Earth Resources Applications Di-
vision and Earth Resources Research Division, Space and
amining the usefulness of spectral data
Life Sciences Directorate. for possible yield estimation. A fairly
©Elsevier Science Publishing Co., Inc., 1982
52 Vanderbilt Ave., New York, NY 10017 0034-4257/82/060509-10502.75
510 T. L. BARNETT AND D. R. THOMPSON

rigorous attempt to build a yield model drought conditions for wheat, corn, and
based directly on Landsat data was re- soybeans in the major U.S. crop regions.
ported by Colwell et al. (1977) and de- The GIN value represents the percentage
scribed in more detail by Nalepka et al. of Landsat resolution elements (about
(1977). They correlated a variety of spec- one-hal/hectare in size) in a 9.3-by-ll.1
tral p a r a m e t e r s - - L a n d s a t spectral kilometer (5-by-6 n.mi.) sample segment
bands and transformations of the band (test sites selected during the LACIE pro-
values--against measured yields on a field gram) with a "greenness" (Kauth and
basis at two test sites in Kansas. They Thomas, 1976) greater than 15, a value
obtained strong correlations, but these did representing healthy, green vegetation.
not extend either temporally or geo- Using the GIN parameter, the 1977
graphically. Researchers at Kansas State drought areas for wheat and the 1978
University have been working for several drought areas for corn and soybeans in
years on estimating leaf area index from the U.S. were delineated and tracked as
Landsat data as an input to estimating the season progressed, showing good
evapotranspiration or as an input to a agreement with the ground data-based
large-area yield model for wheat; see, for Crop Moisture Index provided by the
example, Heilman et al. (1977). Schubert U.S. Department of Commerce. During
and Mack (1977) used Landsat data to 1977, the GIN monitoring program was
calculate percent cover indices, the per- used in real time in the wheat regions
centage of pixels within an area of a of the U.S.S.R. and Australia to detect
particular vegetation type (wheat or all and monitor drought in these areas
vegetation) that were classified as having (Thompson and Wehmanen, 1979).
a closed green canopy, for three pheno- The performance of GIN in the drought
logical stages labeled " e m e r g e d , " studies suggested that it might be a good
"headed," and "ripe." In regressions of parameter to use in a study of wheat
these indices to yields over 3 years yield as a function of spectral appearance
(1973-1975) at four test sites in different of combined wheat and background vege-
soft-climatic regions of western Canada, tation averaged over fairly large
the correlations were quite high for the areas--more than 100 km 2 in the case of
first two crop stages. A test of predicted sample segments. This represented a dif-
vs. published yields in a later year (1976) ferent approach than the yield modeling
indicated satisfactory results only for the investigations referenced above which
wheat index. Investigators have also been used Landsat high resolution capabilities
working for several years to establish rela- to focus on fields that were known to be
tionships among spectral information from wheat. Such a broad-area relation would
Landsat and field measurements and the be much simpler, more widely applicable,
agronomic properties of the crop canopies and might average out some of the local
as eventual inputs into yield estimation; effects that made extension of earlier rela-
e.g., Wiegand et al. (1979), Idso et al. tions to other cases difficult. Environmen-
(1978), and Tucker et al. (1980). tal satellites with much coarser resolution
Thompson and Wehmanen (1979, might become useful data collection tools
1980) developed the green index number for many applications, providing much
(GIN) from l_andsat data to monitor more frequent coverage than Landsat.
SPECTRAL DATA IN WHEAT YIELD ESTIMATION 511

2. Materials and Methods sponded closely to the peak value of the

GIN vs. day-of-year (DOY) curve. Figure
The GIN data base, furnished by the 1, which shows segment 1998, Jones Co.,
Lyndon B. Johnson Space Center, Texas, is a typical plot for winter wheat.
National Aeronautics and Space Adminis- Both the highly stressed wheat for 1978
tration, consisted of GIN values for 719 and the healthy wheat for 1979 are shown.
sample segmentsm9.3-by-ll.1 kilometer The 1978 wheat was so badly drought-
sites for which Landsat data were acquired stressed that almost no green vegetation
in the LACIE and subsequent pro- is discernible until well past heading when
gramsmlocated throughout the U.S. Mid- rains caused some general greenup.
west and Great Plains crop regions during To see if a general relation existed be-
1978 and 1979. Insofar as possible, the tween yield and GIN across the Great
segments were representative of the agri- Plains winter wheat region, the GIN val-
culture of each region. During 1978 and ues at heading for each segment were
1979 the operation of two Landsat satel- plotted against the USDA-SRS (U.S.
lites effectively provided a 9-day coverage Department of Agriculture-Statistical
cycle. Reporting Service) county yields for the
Winter wheat for 1978 was selected for counties containing the sample segments
the study because environmental condi- (Figure 2). Most of the points at the
tions ranged from severe drought in parts lower end represent the heavily stressed
of western Texas and Oklahoma to excel- regions of Texas and Oklahoma. Although
lent weather throughout the eastern and there is a good deal of scatter, a clear
northern portions of the winter wheat relation of yield to GIN at heading ap-
region. The winter wheat data base con- pears to exist, linear up to GIN of 40 or
sisted of GIN values at 219 sample seg- 50 and uncertain thereafter. Considerable
ments in Texas, Oklahoma, Kansas, scatter in this plot is not surprising--the
Colorado, Nebraska, South Dakota, North entire county is represented by the single
Dakota, and Montana for which Landsat 9.3-by-ll.1 km segment; all fields are not
data had been acquired several times dur- wheat; there can be wide variations in the
ing the growing season, permitting track- development rates even among wheat
ing of the spectral response as a function fields; some areas have substantial
of time. This spectral data base was con- amounts of irrigation; county yield figures
sidered to represent spectral data that one are susceptible to errors; the value of GIN
would gather across the U.S. winter wheat is taken at heading while the yield figure
belt in a sampling mode from a satellite is end-of-season; and the effects of weather
system with a spatial resolution of about on yield are not explicitly taken into
10-by-10 km. account.
To reduce the effects of these sources
3. Results and Discussion of scatter, the GIN values at heading
were averaged over the CRDs (Crop Re-
Based on the work of Colwell et al. porting Districts) in which the segments
(1977), the most appropriate spectral fell for CRDs in Texas, Oklahoma, and
parameter to consider initially appeared Kansas; in the case of Colorado, Nebraska,
to be GIN at heading, which corre- South Dakota, North Dakota, and
512 T.L. B A R N E T T A N D D. R. T H O M P S O N

~ I.I.I
I-
i11
o~

O
0)
b.,
X itl =.
f..)
"1"

o0
o

t.kl
t~
1"---
ra
o

>.

>.

.=
oo
I'-.-

Z
Iii
-'r
>- X
I-
Z
0

X ~
I-- m

[. ! I. Q
o
or,..
~ L.~
NI9
SPECTRAL DATA IN WHEAT YIELD ESTIMATION 513

." . -..

eo 0
oe • • • oo • go CO

• .. . .
eeo o e oo •

• oo
oo ooo •

• oo oo ooo •

°° o •
• oo • •
e~o o • •

:: " ... . .

• ooo • o • oo

• o •
• o • •
• • o •

o o • •° •

"r-

° e

ov - "

o:,.
• °* e

I I l i l 0
0 t.o 0 I.~ 0 1.0 o L.O

:ll:l~)V/n8 'Q'I31A AINi'iO~)

514 T. L. B A R N E T T A N D D. R. T H O M P S O N

O
00

• X °
"7,

X • •

ua -~

~z
L~

,.Z

CO UJ

f..d

• X

! I i I i i O
I.O
C3, I.O O r,,-
,~" 03 eq

31:lOV/n8 'Q-I 31A A 1NFI O0

SPECTRAL DATA IN WHEAT YIELD ESTIMATION 515

Montana, which had relatively few seg- historical weather data. However, an in-
ments, the GIN values were averaged over teresting illustration of one-way that spec-
the states. These were plotted against tral data might contribute to estimation
CRD or state yield values in Figure 3. of yields at geographical levels smaller
This plot displays the same behavior as than those possible with existing agromet
Figure 2, but shows much less scatter. weather-based models is provided by the
These results seem to imply that lower 1978 Texas winter wheat GIN data. The
values of the spectral parameter GIN are Plains and Panhandle regions of Texas
associated with lower winter wheat yields experienced severe drought conditions
under moisture stress conditions. Under due to very low preseason soil moishtre
conditions of adequate moisture (high val- for which in-season rainfall was unable to
ues of GIN) there appears to be little compensate, giving low yields. The East
discernible relation between GIN and Central region and counties along the
yield. Testing and verification of these Red River had adequate soil moisture and
implications will require that spectral data precipitation and produced good yields.
be available at many test sites across broad Suppose we wish to estimate yield at
areas of the U.S. winter wheat region, the county level in Texas. Weather data,
under conditions ranging from no stress however, is available only at the CRD
to at least moderate stress, and at inter- level, both historically and currently.
vals more frequent than that furnished by Therefore, CRD weather must be as-
a single Landsat (18 days). Furthermore, sumed to represent the weather in each of
the data must be in a form that can be the counties in the CRD. An agromet
related directly to GIN or vice versa. The yield model was built, patterned on a
1979 winter wheat GIN data set referred Canadian cereal model described by
to earlier is smaller than the 1978 set and Williams et al. (1975). Historical
does not appear to cover a wide enough (1935-1977) CRD monthly average tem-
range of stress conditions to serve as a perature and accumulated rainfall were
valid test case. Landsat data acquired converted into estimates of potential
prior to 1978 and subsequent to 1979 evapotranspiration (Thomthwaite, 1948)
have observations only every 18 days with and moisture deficit (PET--precipita-
many acquisitions missing or con- tion). For the six Texas CRDs with sub-
taminated by clouds. Their suitability for stantial wheat production (CRDs 1N, 1S,
testing is doubthtl. Data from environ- 2N, 2S, 3, and 4), the CRD weather
mental satellites such as the Advanced variables were paired with the SRS-
Very High Resolution R a d i o m e t e r reported CRD yields. A stepwise regres-
(AVHRR) on NOAA-6 and -7 may prove sion was run on the pooled data set to
to be the most suitable for testing and select significant terms, using as trial
verification (see Gray and McCrary, terms the linear and squared monthly
1981). moisture deficit, accumulated Septem-
It is not entirely clear how one would ber-February precipitation, and seasonal
go about incorporating a verified relation moisture deficit. The nine most signifi-
of spectral data to yield into a yield model, cant terms (including intercept and two
particularly into a conventional agromet trend terms) were retained and consid-
regression model based on many years of ered to constitute the weather-based yield
516 T. L. BARNETT AND D. R. THOMPSON

TABLE 1 Regression of Reported Texas County Wheat Yields (gobs) Against County Yields Predicted by
Agromet Model (~) and Agromet Model Plus Green Index Number at Heading (Y)
AGROMET
ME~,~ AGROMET ~ (;IN
OBSERVED MODEL YIELD YIELD
YIELD PREDICTION PREDICTION
COUNTY CRD" ( gobs) (~) (;IN ( )" ) )" Yob~
bu/aere bu/acre

Hunt 4 32.8 15.1 75 30.9 1.9

Fannin 4 40.0 15.1 75 309 9.1
Cochran 1S 5.6 13.4 10 9.5 3.9
Dawson IS 7.5 13.4 0 6.6 0.9
Terry IS 8.8 13.4 2 7.1 1.7
Gaines 1S 7.5 13.4 0 6.6 0.9
Castro IN 11.0 13.3 14 10.5 0.5
Nolan 28 7.8 13.5 6 8.5 0.7
Randall A IN 8.4 13.3 12 9.9 - 1.5
Parker 3 20.3 15.1 65 28.0 7.7
Palo Pinto 3 10.0 15.1 :37 19,9 9.9
Motley 2N 7.0 13.5 4 7,9 0.9
Cottle 2N 9.2 13.5 4 7.9 1.3
Archer 3 15.4 15.1 25 16.4 l.O
Baylor 28 15.0 13.5 15 11. l 3.9
Haskell 28 9.0 13.5 0 6.7 2.3
Wilbarger 2N 20.0 13.5 30 15.4 4.6
Wichita 2N 19.2 13.5 66 25.9 6.7
Stephens 3 9.2 15.1 10 12.0 2.8
Ymmg 3 12.0 15.1 30 17.8 5.8
Concho 28 t, 9.3 13.5 0 6.7 2.6
Grayson 4 34.0 15.1 7~) 2:3.6 10.4
Coleman A 28 8.1 13.5 6 8.5 0.4
Runnels 28 8.1 13.5 7 S.7 - 0.6
Hnasford IN 8.0 13.3 11 9.6 1.6
Crosby IS 9.7 13.4 1 6.9 2.8
Donley 2N 12.3 13.5 5 82 4.1
Fisher 28 8.7 13.5 6 8.5 0.2
Colenlan B 28 8.1 13.5 16 11.4 3.3
Lubbock IS 7.2 13.4 5 8.0 0.8
Williamson 4 24.2 15.1 36 19.6 4.6
McClennan 4 22.9 15.1 46 22.5 0.4
Randall B IN 8.4 13.3 20 12.2 3.8
Oldham IN 7.3 13.3 5 79 ~0.6
Jones 28 8.0 13.5 6 85 0.5

"S = south, N = North.

t'CRD 7, assigned to CRD 28.

model for prediction of 1978 yields, 0. agromet model yield did not produce very
These were assumed to represent esti- good estimates (0) of the observed county
mates of the county yields, i.e., the yields yields, Yob~-We then tried to determine if
predicted for a CRD were assumed to the addition of a spectral parameter at
represent all counties in the CRD. As the county level would explain more of
seen from columns 3 and 4 of Table 1 this the variation in observed yields than the
SPECTRAL DATA IN WHEAT YIELD ESTIMATION 517

agromet model estimates. Observed for sample segments across the U.S. Great
county yields were regressed against the Plains winter wheat belt to SRS-reported
r) estimates and GIN values at heading for county and CRD yields are promising
each county, that is, and encouraging. A linear relation ap-
pears to exist up to GIN values of 40 or
:~ = a + bO + c GIN, 50, covering cases of severe to moderate
stress. This is the region where conven-
tional agromet yield models need the most
to estimate the coefficients a, b, and c,
help. The relation at higher GIN values,
and the regression estimates of county
indicating low levels of stress, is not read-
yields, Y. The improvement of ~r versus 0
ily discernible with this data. Further
as an estimator of wheat yield indicates
analysis on other areas and other years is
the degree of improvement in yield fitting
required to determine whether the ap-
ability that resulted from incorporation of
parent relations hold true for other sets of
spectral data. The results, shown in Table
data. Hopefully, it will soon be possible to
1, indicate that the use of GIN values as a
discern the real relationships between
variable in the yield estimation produced
spectral data and yield at large scale, and
a substantial improvement in the fit for
to quantify these relationships for use in
1978 Texas winter wheat. Statistically, the
crop yield estimates.
root mean square error (RMSE) for 0
versus Yobs Was 7.9 bushels per acre References
( b u / A ) with a correlation (r) for the fit
of 0.68, while the RMSE for Y versus Yobs Baier, W. (1973), Crop-weather analysis
was 4.1 b u / A with an r of 0.87. An F-test model: Review and model development, J.
of the null hypothesis that the two vari- Appl. Meteorol. 12:937-947.
ances (RMSE 2) were equal indicates that Colwell, J. E., Rice, D. P., and Nalepka, R. F.
the statistic F = S zl / S z2 = 7.92//4.12 = 3.7 (1977), Wheat yield forecasts using Land-
is greater than the values of F~ (for de- sat data. Proceedings o f the Eleventh Inter-
grees of freedom 34 and 33, respectively): national Symposium on Remote Sensing of
F0.05 = 1.8 and F0.01 = 2.4. In other words, Environment, pp. 1245-1254.
the inclusion of spectral data provided Feyerherm, A. M., and Paulsen, G. M. (1981),
significantly better fit in this case. The Development of a wheat yield prediction
inclusion of a spectral parameter may thus model, Agron. J. 73:277-282.
be expected to improve estimates of yield Gray, T. I., and McCrary, D. G. (1981), The
at various levels smaller than those for environmental vegetation index--a tool
which weather data is available. potentially useful for arid land manage-
ment. Fifteenth Conference on Agriculture
and Forest Meteorology and Fifth Con-
4. Conclusions ference on Biometeorology (Anaheim,
California). Published by the American
This paper investigates the large-area Meteorology Society, Boston, Massachu-
relations between spectral data from setts.
satellites and crop yield. The correlations Heilman, J. L., Kanemasu, E. T., Bagley,
of Landsat spectral data in the form of J. O., and Rasmussen, V. P. (1977),
GIN values at the time of wheat heading Evaluating soft moisture and yield of winter
518 T. L. BARNETTAND D. R. THOMPSON

wheat in the Great Plains using Landsat yield models. Technical sessions, LACIE
data, Remote Sens. Environ. 6:315-326. Symposium, Johnson Space Center,
Idso, S. B., Hat_field, J. L., Reginato, R. J., and Houston, Texas, July 1979.
Jackson, R. D. (1978), Wheat yield esti- Thompson, D. R., and Wehmanen, O. A.
mates by albedo measurement, Remote (1979), Using Landsat digital data to detect
Sens. Environ. 7:273-276. moisture stress, Photogrammetric Eng. Re-
Kauth, R. J., and Thomas, G. S. (1976), The mote Sens. 45:201-207.
tasselled cap--A graphic description of the Thompson, D. R., and Wehmanen, O. A.
spectral-temporal development of agricul- (1980), Using Landsat digital data to detect
tural crops as seen by Landsat. Proceedings moisture stress in corn-soybean growing re-
of the Symposium on Machine Processing gions, Photogrammetric Eng. Remote Sens.
of Remote Sensing at Purdue University, 46:1087-1093.
pp. 4B41 to 4B-51. Thompson, L. M. (1962), Evaluation of
MacDonald, R. B., and Hall, F. G. weather factors in the production of wheat,
(1980), Global crop forecasting, Science J. Soil Water Conserv. 17:147-156.
208:670-679. Thornthwaite, C. W. (1948), An approach
Nalepka, R. F., Colwell, J. E., and Rice, D. P. toward a rational classification of climate,
(1977), Forecasts of winter wheat yield and Geogr. Rev. 38:55-94.
production using Landsat data. Final re- Tucker, C. J., Holben, B. N., Elgin, J. H., and
port for Contract NAS 5-22389, NASA, McMurtrey, J. E. (1980), Relationship of
Goddard Space Flight Center, Greenbelt, spectral data to grain yield variation,
Maryland. Photogrammetric Eng. Remote Sens.
Schubert, J., and Mack, A. R. (1977), Global 46:657-666.
agricultural productivity evaluation from Wiegand, C. L., Richardson, A. J., and
Landsat data. Proceedings of the W. Kanemasu, E. T. (1979), Leaf Area Index
Nordberg Memorial Symposium of the estimates from Landsat and their implica-
Twentieth Plenary Symposium of COSPAR tions for evapotranspiration and crop mod-
(Tel Aviv, Israel). Published by Pergamon eling, Agron. I. 71:336-342.
Press for the Committee on Space Re- Williams G. D. V., Joynt, M. I., and
search, pp. 95-99. McCormick, P. A. (1975), Regression anal-
Strommen, N. D., Sakamoto, C. M., LeDuc, ysis of Canadian Prairie crop-district cereal
S. K., and Umberger, D. E. (1978), Devel- yields, 1961-1972, in relation to weather,
opment of LACIE CCEA-1 weather/wheat Soil, and trend, Can. 1. Soil Sci. 55:43-53.

Received 18 lanuary 1982.