ISPRS Archives XXXVI-8/W48 Workshop proceedings: Remote sensing support to crop yield forecast and area estimates

OPERATIONAL PREDICTION OF CROP YIELDS

USING MODIS DATA AND PRODUCTS

Paul C. Doraiswamya , Bakhyt Akhmedovb , Larry Beardc , Alan Sterna and Richard Muellerc
a
USDA, ARS, Hydrology and Remote Sensing Laboratory, Beltsville, MD 20705
(paul.doraiswamy, alan.stern) @ARS.USDA.GOV
b
Science Systems and Associates, Inc. Lanham, MD 20704; - Bakhyt.akhmedov@ARS.USDA.GOV
c
USDA, NASS, Research and Development Division, 3251 Old Lee Highway, Fairfax, VA 22030-1504
(larry_beard, Rick_mueller)@nass.usda.gov

Commission VIII, WG VIII/10

KEY WORDS: Remote Sensing, Agriculture, Crop yield, MODIS algorithm, Operational, Crop Classification

ABSTRACT:

Official crop progress, condition and production estimates for the United States are responsibilities of the U.S. Department of
Agricultures, National Agricultural Statistics Service (NASS). In addition to weekly and monthly survey-based data, biweekly
composite maps of the normalized difference vegetation index (NDVI) from the NOAA AVHRR sensor (1 km resolution) are
produced by NASSs Research and Development Division (RDD) for monitoring vegetative change. This provides a qualitative
assessment of differences in crop condition that may be an indication of potential yields. There is need for a more quantitative
assessment of crop yields and spatial variability. Currently, NASS acquires crop yield indications via ground-based sample surveys
(objective plant and fruit counts, fruit weights and farmer reports) which are collectively used to develop tools for its decision
support system to assess weekly crop progress, monthly crop yield estimates for each state and the U.S, and annual county yield
estimates. This paper describes the joint research between RDD and the Agricultural Research Service (ARS) of USDA for the
development of simplified process models and algorithms to supplement the NASS field data collection. Potential advantages to
using remote sensing include integration of spatial variability into county yields, enhanced timeliness, and efficient use of resources.

In the preliminary phase, MODIS data and products for the states of Iowa and Illinois were used to develop an operational
assessment of crop yield forecasts for corn and soybeans. Spatial estimates of crop yields at county and sub-county levels offer a
major improvement of current capabilities. The timeliness in producing these estimates is a vast improvement over the present
assessment capability at the county level. Potential use of the estimates will supplement current tools and improve NASS crop
condition and yield decisions. Results of the pre-harvest forecasts developed for the 2005 and 2006 crop seasons are presented.

1. INTRODUCTION independently assess crop condition at the county and state

levels. The timely
Accurate and timely monitoring of agricultural crop conditions
and estimating potential crop yields are essential processes for evaluation of potential yields is increasingly important because
operational programs. Assessment of particularly decreased of the huge economic impact of agricultural products on world
production caused by a natural disaster, such as drought or pest markets and strategic planning.
infestation, can be critical for countries or locales where the
economy is dependent on the crop harvest. Early assessment of County statistics are noted as a driving force for rural economic
yield reductions could avert a disastrous situation and help in development, and are essential to proper management of
strategic planning to meet demands. The National Agricultural USDAs many farm, education, and natural resources
Statistics Service (NASS) of the U.S. Department of management programs. Many allocations of federal resources
Agriculture (USDA) monitors crop conditions and makes the to states and counties are determined by their production of
Official USDA production assessments in the U.S., providing farm commodities. Demand for accurate commodity estimates
monthly production forecasts and end-of-year estimates of crop at the lowest level of aggregation, and at the earliest possible
yield and production. NASS has developed methods to assess time, has and continues to increase substantially. Literally
crop growth and development from several sources of millions of business decisions rely on this basic production data
information, including several types of surveys of farm produced by USDA/NASS.
operators and field-level measurements. Field offices in each
state are responsible for monitoring the progress and health of In the early 1960s, NASS initiated objective yield surveys for
the crop and integrating crop condition with local weather crops such as corn, soybeans, wheat, and cotton in States with
information. Information on crop condition and progress is also the greatest acreages (Allen et al., 1994). These surveys
distributed in a biweekly report on regional weather conditions. establish small sample units in randomly selected fields which
NASS offices provide monthly information to the Agriculture are visited monthly to determine maturity, numbers of plants,
Statistics Board, which assesses the potential yields of all numbers of fruits (wheat heads, corn ears, soybean pods, etc.),
commodities based on crop condition information acquired and weight per fruit. Yield forecasting models are based on
from different sources. This research complements efforts to relationships of samples of the same maturity stage in
comparable months during the past four years in each State.

45
 ISPRS Archives XXXVI-8/W48 Workshop proceedings: Remote sensing support to crop yield forecast and area estimates

These indications are then compared to farmer-based survey architectural parameters in the SAIL or other radiative transfer
results to produce monthly yield forecasts. Additionally, the model for crop-specific LAI should be accurate. These
Agency implemented a midyear Area Frame Survey that parameters also change during the growing season. The
enabled creation of probabilistic based acreage estimates. For reflectance data are used to derive LAI seasonal profile, which
major crops, sampling errors are as low as 1 percent at the U.S. is used in initializing or constraining parameters in the crop
level and 2 to 3 percent in the largest producing States. yield simulation model. The temporal and spatial
Accurate crop production forecasts require accurate estimates of inconsistencies of the MODIS 8-day reflectance product data
acreage at harvest, its geographic distribution, and the (Doraiswamy, 2006) limits its application in crop yield models
associated crop yield determined by local growing conditions. at regional scales.
There can be significant year-to-year variability which requires
a systematic monitoring capability. To quantify the complex In this research we evaluated the use of the MODIS NDVI and
effects of environment, soils, and management practices, both surface temperature products to develop a multi-dimensional
yield and acreage must be assessed. A yield forecast within regression algorithm to predict the state and county level yields.
homogeneous soil type, land use, crop variety, and climate The NDVI seasonal dynamics is representative of crop growth
preclude the necessity for use of a complex forecast model. and biomass changes and thermal data is representative of the
crop moisture stress condition.
Doraiswamy, et al. (1979), provided an inventory of various
crop yield models, including statistical and deterministic The objectives of this research are to: a) develop a MODIS-
models. The performance of deterministic models for large based algorithm for operational classifications of corn and
area forecasts depended on the availability of local climatic data soybean crops in the U.S. Corn Belt; b) develop a multi-
with adequate spatial resolution. Use of remotely sensed data dimensional regression method to provide a consistent, timely
was limited to studying the temporal changes in vegetation and accurate yield prediction for potential use in NASSs
condition such as crop growth and development. operational program.

Integration of remotely sensed data in crop yield models 2. METHODOLOGY

evolved during the next decade of research based on field
experiments and advances in biophysical modeling. The need Study Region.
for linking real-time remotely sensed data initiated research on
retrieval of biophysical parameters from satellite imagery. The study areas presented in this paper are the states of Iowa
and Illinois, which cover a major part of the U.S. Corn Belt.
The potential application of remote sensing technology for The two States are intensively cultivated, with approximately
monitoring crop condition and predicting crop yields at regional 75% of the land in corn and soybean crops. Crops are grown
scales have been studied extensively during the past several under rain fed conditions where soil moisture is normally
decades. Traditionally, the empirical approach based on adequate in the growing season; however, moisture stress
statistical regression models was used with lower resolution conditions can occur in the early stages of crop development
data such as that from NOAA AVHRR. Studies have shown and more often during the latter part of the season. Seasonal
that the cumulative seasonal normalized difference vegetation rainfall ranged between 800 mm in the north to 450 mm in the
index (NDVI) values were significantly correlated with reported southern part of the region. Soil moisture is generally at field
crop yields (Groten, 1993). Doraiswamy and Cook (1995) capacity at the start of the season. However, because of spatial
demonstrated that cumulative NDVI values for spring wheat variability in the spring rainfall, planting dates across the region
during the grainfill period improved estimates of crop yields in were variable. Crop planting in the region was completed by
North Dakota. Combining a growth model with input mid-May, with corn generally planted about 2 weeks earlier
parameters derived from remotely sensed data provides spatial than soybeans. Crop maturity occurred by late September.
integrity as well as a real-time calibration of model
parameters. Moulin, et al., (1995) successfully used this Data and Re-processing
approach using SPOT/HRV data in a radiative transfer model
SAIL (Scattering by Arbitrary Inclined Leaves, Verhoef,1984) The MODIS 8-day composite surface reflectance product
to map crop leaf area index (LAI) and also inversion models to (MOD09Q1 collection 4) at 250 m resolution and land surface
that simulated canopy reflectance that were comparable with temperature product (MOD11A2) at 1 km resolution was
SPOT/HRS reflectance. These methods enabled constraining acquired through the NASA-DAAC EROS Data Center, Sioux
the model parameters with satellite observations to retrieve key Fall, SD. Imagery data for the available 5-year period (2002-
parameters of soil moisture and above ground plant biomass. 2006) was acquired for day of year (DOY) from 121 to 305.
Doraiswamy et al. (2003, 2004, 2005) simulated LAI from The images were re-projected from ISIN projection to the
NOAA AVHRR (1 km) two-band reflectance as well as standard UTM projection for zone 15 (Iowa) and 16 (Illinois)
MODIS (250 m) 2-band reflectance products, and used them in using the MODIS re-projection tool. The data was imported to
crop yield simulation models over small areas. The MODIS ERDAS Imagine format, and NDVI time series for the crop
derived LAI seasonal profile was developed in the above season were calculated.
studies using the SAIL model, and the crop yield model
parameters were initialized. The MODIS 8-day composite The cloud free MODIS data product has potentially three
reflectance product was acquired through the NASADAAC sources of errors georeference, atmospheric influence and
EROS Data Center, Sioux Falls, SD. The MODIS data has BRDF. The 8-day composite images always have patches even
already been corrected for atmospheric effects and cloud cover. on homogeneous areas which are caused by difference in single
day images used in the composites. Besides cloud cover, errors
The limitation in using satellite-based LAI methods with high in reflectance caused by atmospheric condition always reduced
(Landsat) and moderate (MODIS) spatial resolution data in crop the NDVI value. To eliminate these errors we used multiple-
yield simulation models at regional scales is that the canopy step processing of the composite data using the Savitzky-Golay

46
 ISPRS Archives XXXVI-8/W48 Workshop proceedings: Remote sensing support to crop yield forecast and area estimates

filtering technique adapted for tracking the upper envelope of The NDVI and surface temperature (Ts) from MODIS imagery
the NDVI time series profile (Jonsson and Eklundh, 2004; are parameters that are correlated with crop yields. Initial
Doraiswamy et al. 2006). The Savitzky-Golay filter uses investigations suggest that these two parameters can provide
moving 5-point window for each pixel time series profile and in spatial variability of crop growth conditions during the growing
each window, noisy values is approximated by polynomial to season. MODIS data shows better correlation with crop yields
smooth NDVI values in the window. The thermal data is compared to similar analyses using data from NOAA AVHRR.
screened using the quality assurance data provided along with The better results may be due to better spatial resolution and to
the thermal imagery, taking data with errors <= 2 degrees the specific narrow bands in VIS and NIR of the MODIS
Kelvin. sensors. The seasonal NDVI profile describes the crop growth
and development, surface temperature provides additional
Crop Classification information regarding potential crop stress conditions. Stress
conditions may include crop water stress as well as disease and
Landcover classification is an important step to assure accurate infestation. An algorithm that combines NDVI and Ts was
retrieval of crop specific data to monitor crop condition and developed that correlates with yield in a two dimensional
predict yields. Classification has traditionally been completed regression equation:
at 30-m resolution using Landsat ETM+ images. In the past
decade, USDA-NASS used Landsat data to develop crop Yield=a+b*NDVI+c*Ts. (1)
classification for crop acreage estimation over selected states
including Iowa and Illinois. However, the classification is Currently, one set of equations are developed for corn and
usually not available until about 4-5 months after the crops are soybeans for each state. The NDVI and Ts parameters are
harvested. In an operational program where remote sensing summed for a period between the mid-vegetative to mid-
data is used to predict crop yields, it is critical to have timely senescence period. The state yield estimate is obtained by
crop classification for assessing crop specific yields at specific averaging the parameters for all corn or soybean pixels in the
time periods. Additionally, the uncertainties in the availability state. The NDVI and Ts parameters extracted for each crop for
of Landsat data required the development of crop classification the 2002-2005 crop season are then used in equation (1) with
using the MODIS 8-day composite time-series data. A decision the NASS estimated yields. The coefficients a, b and c are
tree algorithm was developed (Doraiswamy et al, 2007) to map derived and the regression equation is then used as a predictor
corn and soybean fields. In a two step process the crop area in for 2006 state crop yields.
the state is first selected using a threshold of NDVI values
based on the combined crop phenology of corn and soybean The spatial variability of crop yields are assessed at the county
crops. The next step of the decision tree algorithm separates the level by extracting the mean NDVI and Ts parameters from
corn and soybean crops within the crop area. Figure 1 is the corn and soybean pixels for each county. Equation (1) is used
general NDVI time-series profile showing differences between to determine the a, b and c coefficients using data from the
corn and soybeans. The soybean crop is planted several weeks 2002 crop season. Then equation (1) was applied to predict
after corn and the maturity follows that of corn. The clear county level yields for successive years. The 2002 crop
distinction between the corn and soybean NDVI profile occurs season data was selected to develop the regression algorithm
around day of year (DOY) 177 in Iowa and Illinois. These because the MODIS data quality was better than other years and
features are used to separate the corn and soybean crops, the crop yields for corn and soybeans were spread over a wider
predominant crops in these two states. The Landsat range. County yields for the 2005 and 2006 crop seasons were
classification for Iowa and Illinois developed by the NASS predicted using the regression algorithm developed from the
Research and Development Division was used as the template 2002 data sets. The initial predictions for state and county level
for evaluating the accuracy of the MODIS-based classification. yields are made in early September prior to crop harvest and
updated after the senescence is completed in October.

3. RESULTS and DISCUSSIONS

The MODIS data acquired for this research covered five crop
seasons when complete seasonal data was available (2002 -
2006). The results from this research are compared with USDA
official estimates from NASS. Final state yield estimates are
published in January of the following year. These estimates are
based on objective yield and farmer-based survey data collected
from well planned sampling strategies and considered to be
accurate. The NASS county level estimates are generally
published in March of the following year and the data used are
from similar farmer-based surveys and from observed local
weather and crop conditions. These estimates may not be
systematic from county-to-county within the state. This
research seeks to provide additional information derived from
Figure 1. Example of the normalized difference vegetative remote sensing data to supplement spatial information that
index (NDVI) for corn and soybean crops using MODIS 8-day would strengthen the county level estimates., at a much earlier
composite data (250 m) for the 2005 crop season after the data time period.
was processed with the filtering algorithm.

Crop Yield Algorithm Crop Classification

47
 ISPRS Archives XXXVI-8/W48 Workshop proceedings: Remote sensing support to crop yield forecast and area estimates

The MODIS-based classification of corn and soybean crops was for 2006 were 165.2 b/ac and 47.8 b/ac versus NASSs estimate
developed using the NDVI time series profile, first separating of 163 b/ac and 48 b/ac respectively for corn and soybeans.
the crop area from other classes by selecting pixels with an
NDVI threshold of 0.4 at the beginning and end of the crop The state level predictions for the 2007 crop season will be
season with a value of 0.8 at mid-season. In a narrow window based on a 5-year multi-regression algorithm developed from
around DOY 177, the magnitude of NDVI for corn is greater 2002-2006 state estimates. The coefficients of determination
than soybeans as shown in an example in Figure 1. The for Iowa corn and soybeans were 0.96 and 0.88, and for Illinois
MODIS classification was compared with a NASS they were 0.98 and 0.438, respectively. Eliminating the 2003
classification developed from Landsat ETM images for the year from the regression algorithm for soybeans in Illinois
2005 crop season. Landsat 30 m pixels were aggregated to 250 increases the coefficient of determination to 0.97.
m MODIS pixels by picking up only those pixels that contained
90% of specific (corn or soybean) crop. In Iowa 357,795 Difference in corn county yields in Iowa for 2005.
MODIS pixels had 90% of corn or soybean fields (Figure 2). Iowa State Yield=173 Bushels /Acre, RMSE=10.1

These pixels were used to test the corn and soybean 50

discrimination algorithm. The results were then compared with 40

Difference (Bushels/Acre)
the Landsat classification aggregated to 250m pixel resolution. 30
20
The overall accuracy was 81.7% with a kappa coefficient of 10
difference
0.63. In Illinois the number of MODIS pixels with 90% crop 0 20%+

was found to be 409,108 pixels. The comparison with Landsat -10

20%-

-20
classification showed that the overall classification accuracy

1
4
7
10
13
16
19
22
25
28
31
34
37
40
43
46
49
52
55
58
61
64
67
70
73
76
79
82
85
88
91
94
97
-30
was 75.1% with a kappa coefficient of 0.50. The accuracy for -40
Iowa was better than Illinois perhaps because of the large -50

variation in corn and soybean phenological stages in Illinois Counties

Difference in soybean county yields in Iowa for 2005.

between the northern and southern areas of the state. State Yield=52.5 Bushels /Acre, RMSE=3.6
15

10
Difference (Bushels/Acre)

Difference
0 20%+
20%-

-5
1
4
7
10
13
16
19
22
25
28
31
34
37
40
43
46
49
52
55
58
61
64
67
70
73
76
79
82
85
88
91
94
97
-10

-15
Counties

Figure 3. Difference in county yields between the regression

algorithm predictions and NASS estimated yields for the 99
counties in Iowa for the 2005 crop season.

The county level yield predictions were completed at the same

time as the state predictions. The regression algorithm
Figure 2. The 2005 classification for Iowa using the 8-day developed from the 2002 data provided the coefficients for the
MODIS 250 m composite imagery. multidimensional regression algorithm. The multidimensional
coefficients of determination for Iowa were 0.86 and 0.71 and
Crop Yield for Illinois they were 0.82 and 0.86 respectively for corn and
Using the USDA/NASS state yield estimates for the 2002-2005 soybeans. These relatively high values indicate that the MODIS
crop seasons, the multi-dimensional regression equations were data for 2002 were good and the predictions fit well with the
developed for corn and soybean yields for Iowa and Illinois. NASS crop yield estimates. For the other crop seasons the
Based on the regression algorithm, the coefficients of coefficient of determination was lower. County level
determination for Iowa at the state level were 0.98 and 0.88 predictions are more sensitive to the quality of the remote
respectively for corn and soybean crops. These equations were sensing data, combined with the accuracy and consistency of
used to predict the Iowa state yield estimates for 2006 corn and the estimation of NASS county yields.
soybeans. The regression algorithm yield predictions for the
2006 season for corn was 170 bushels per acre (b/ac) and 52.2 Using the regression model developed from 2002 crop season,
b/ac for soybeans versus NASSs estimates of 166 b/ac and 50.5 the corn and soybean crop yields for 2005 were predicted and
b/ac respectively. In Illinois, the regression algorithm was results compared with the NASS estimates. The coefficient of
developed for the same period (2002-2005), and showed determination for corn is 0.73 and for soybeans is 0.59, and for
coefficients of determination of 0.98 and 0.36 respectively for Illinois these coefficients are 0.60 and 0.52 respectively. The
corn and soybeans. The low coefficient of determination for regression coefficients are strongly dependent on the quality of
soybeans is due to disease outbreak in parts of Illinois during satellite imagery and NASS county yield estimates for that
the latter part of the crop season in 2003 (Malvick, 2003). This particular year. Figure 3 shows the difference in county yields
occurred during the latter part of the season and the yellowing between the NASS estimate and the regression algorithm
of the foliage due to disease was not observed from imagery. predicted yields for Iowa 2005 crop season. The red lines are
Nevertheless the yield predictions from the regression algorithm the +/- 20% standard deviation from the NASS estimated
yields. The regression algorithm predicted yields are all within

48
 ISPRS Archives XXXVI-8/W48 Workshop proceedings: Remote sensing support to crop yield forecast and area estimates

20% of the NASS estimates, and the great majority are within Srinivas. Narosa Publishing House, New Delhi. Chapter 24: pp
10percent. RMSE for predicted yield for corn and soybeans in 229-240.
Iowa are 10.1b/ac and 3.6 b/ac and in Illinois are 19.3 b/ac and
5.6 b/ac respectively. Doraiswamy, P.C., S. Moulin, P.W. Cook, and A. Stern. 2003.
Crop yield assessment from remote sensing, Photogrammetric
4. CONCLUSION Engineering and Remote Sensing, 69, 665 674.

Timely and accurate prediction of crop yields is critical for Doraiswamy, P.C., J.L. Hatfield, T.J. Jackson, J.H., B.
agricultural markets, planning and development. Daily Akhmedov, and A.J. Stern. 2004. Crop condition and yield
frequency of MODIS data acquisition at 250 m pixel resolution simulations using Landsat and MODIS imagery, Remote
offers a great potential for use of the data and products in Sensing of Environment, 92: 548 559.
operational yield prediction programs. In this study, a simple
algorithm that uses near-real time MODIS imagery and Doraiswamy, P.C., T.R. Sinclair, S. Hollinger, B. Akhmedov,
products was developed to predict crop yields at county and A. Stern, and J. Prueger. 2005. Application of MODIS derived
state levels. The algorithm includes crop-specific classification parameters for regional yield assessment, Remote
and yield prediction prior to crop harvest. The crop Sensing of Environment. 97(2), 192-202.
classification was developed using a decision tree algorithm
that relied on the characteristics of crop growth phenology Doraiswamy, P.C., B. Akhmedov and A.J. Stern. 2006a.
without the need for ground-based data. The classification Improved techniques for crop classification using MODIS
accuracies were compared with the USDA NASS Landsat- Imagery. Proceedings of the International Geoscience and
based classification data and found to be acceptable for yield Remote Sensing Symposium, July 31 August 4, 2006,
predictions. The correlation between NDVI and crop yields Denver, Colorado. CD ROM.
and between surface temperature and crop yields are integrated
in a multidimensional regression model for predicting yields at Doraiswamy, P.C., B. Akhmedov and A.J. Stern. 2006b.
the county and state levels. Differences between the NASS MODIS time Series data applications in Agriculture. A CEOS
state level yield estimates and the regression algorithm Land Product Validation topical workshop: Validation of global
predictions for both Iowa and Illinois for the 2006 season was vegetation indices and their time series. University of Montana
less than 4 b/ac for corn and less than 2 b/ac for soybeans. Missoula, MT. August 7, 2006. (Abstract)
http://www.ntsg.umt.edu/VEGMTG/val-home.html
The quality of MODIS data is very critical for crop yield
predictions and this paper describes some of the steps that we Doraiswamy, P.C. and B. Akhmedov. 2007. Crop specific
achieved to enhance the quality of data for cloud cover and classification for yield predictions using MODIS imagery.
atmospheric effects. The computational scale appeared to make (Submitted for publication).
a difference in the tolerance on the imagery data quality.
Although the same algorithm was used for both state and Groten, S.M.E. 1993. NDVI- crop monitoring and early yield
county level yield predictions, the county yield predictions assessment of Brukina Faso. International Journal of Remote
appeared to be more sensitive to quality of the images and the Sensing 14, 1495-1515.
yield predictions were not as well correlated with the NASS
estimated yields. Another important factor in this lower Jonsson, P. and Eklundh, L., 2004. TIMESATA program for
coefficient of determinations at the county level was that the analyzing time-series of satellite sensor data, Computer &
NASS estimates have an error that is not reported. However, Geosciences. 30, 833 845.
assuming an error in the NASS county yield estimates, the
predictions are well within a 20% standard deviation of the Malvick, D. 2003. Charcoal Rot of Soybeans in Illinois:
estimates. Primary or Secondary Disease? Bulletin of Pest Management
and Crop Development information for Illinois. University of
REFERENCES Illinois at Urbana- Champaign. University of Illinois
Extension.http://cptc.ipm.uiuc.edu/bulletin/pastpest/articles/200
Allen, R., G.A. Hanuschak, and M.E. Craig, 1994. Forecasting 323g.html
crop acreages and yield in the face of and in spite of floods,
Crop Yield Forecasting Methods, Proceedings of the Seminar Moulin, S., Fisher, A., Dedieu, G and Delcolle, R. 1995.
Villefranche- sur-Mer, 2427 October, Villefranche-sur-Mer, Temporal variation in satellite reflectances at field and regional
France, pp. 87110. scales compared with values simulated by linking crop growth
and SAIL models. Remote Sensing of Environment 54, 261-
Doraiswamy, P.C., T. Hodges, and D.E. Phinney, 1979. Crop 272.
Yield Literature for AgRISTARS Crops Corn, Soybeans,
Wheat, Barley, Sorghum, Rice, Cotton and Sunflowers, Verhoef, V. 1984. Light scattering by leaf layers with
AgRISTARS Technical Report SR-L9-00405, Lockheed application to canopy reflectance modeling: The SAIL model.
Electronics Company, Inc., Houston, Texas, 105 p. Remote Sensing of Environment. 16, 125-145.

Doraiswamy, P.C. and P.W. Cook. 1995. Spring Wheat Yield ACKNOWLEDGEMENTS
assessment using NOAA AVHRR data. Canadian J Remote
Sens. 21:43-51. Funding for this research was provided in part by the National
Agricultural Statistics Service and the Agricultural Research
Doraiswamy, P.C., P.Zara, and A. Stern. 2000. Satellite Service of the U.S. Department of Agriculture. MODIS data
remotely sensed data application in estimating crop condition was provided free of charge from the NASADAAC EROS
and yields. In Remote Sensing Applications. Edited by M.S. Data Center, Sioux Falls, SD.

49
ISPRS Archives XXXVI-8/W48 Workshop proceedings: Remote sensing support to crop yield forecast and area estimates

50