ABSTRACT

The U.S. Department of Agriculture is responsible for making monthly forecasts of crop condition and yields at the state and regional levels. Supplemental spatial data from satellite observations can provide timely information on crop condition and potential yields. This paper discusses the use of a crop growth model EPIC (Erosion Productivity Impact Calculator) with certain input parameters calibrated using remotely sensed data to provide temporal and spatial integrity. A radiative transfer model provided the link between the satellite data and crop model. The study was conducted in the semi-arid region in the state of North Dakota where spring wheat is the predominant crop. The primary objective was to evaluate a method of integrating Landsat TM and NOAA AVHRR data in a crop growth model to simulate spring wheat yields at the sub-county and county levels respectively. The model simulated yields at sub-county levels agreed very well with the reported farm and county averages at selected NASS survey sites. Yield estimates using the NOAA AVHRR data were good except for the northeastern part of the state where varying levels of disease were reported.

INTRODUCTION

The monitoring and assessment of natural resources using satellite remote sensing have advanced greatly in the past decade. The technology to provide temporal and spatial coverage of the landscape in a timely manner has enabled us to apply this capability to various applications specifically, in the management of agricultural system has always been difficult because of the complexity and extensiveness of the problems that need to be addressed. Remotely sensed data when properly processed and analyzed, is well suited for resolving this complexity. Early assessment of yield reductions based, in part, on these data, could avert a disastrous situation and help in strategic planning to meet the needs.

The National Agricultural Statistics Service (NASS) of the U.S. Department of Agriculture (USDA) monitors crop condition to provide monthly estimates of major crop yields and production in the United States. NASS has developed methods to assess crop growth and development from several sources of information including surveys of farm operators, crop condition reports from field surveys and local weather information. Remote sensing technology can provide supplemental spatial data to provide timely information on crop condition and potential yields. The timely evaluation of potential yields is increasingly important because of the growing economic impact of agricultural production on world markets.

The use of remote sensing technology for monitoring vegetation condition in general, has been studied extensively during the past decade. The normalized difference vegetation index (NDVI) derived from the visible and near-infrared reflectances of the National Oceanic and Atmospheric Adminstration's (NOW, a Very High Resolution Radiometer (AVHRR) meteorological satellite has been successfully used to monitor vegetation changes at regional scales. Temporal changes in the NDVI have been shown on net primary production (Prince, 1991 and Goward et al. 1987). A theoretical background to relate primary production estimates based on the absorption of photosynthetically absorbed radiation (PAR) by the canopy was presented by Tucker and Sellers (1986). Daughtry et al. (1983), Asrar et al. (1985) provided experimental validation of biomass estimates from remotely sensed data.

The use of the NDVI parameter to estimate crop yields is a specific extension of the above general concept. The seasonal accumulated NDVI values correlate well with the reported crop yields in semi-arid regions (Groten, 1993). Doraiswamy and Cook (1995) further demonstrated that accumulating the NDVI values for spring wheat during the grain-fill period can improve the potential crop yield estimates in North Dakota.. Although the results were encouraging, the relationships appear to be valid only for the study areas and required adjustment for both differences in soil background and the mixture of crops.

Crop growth simulation models have been used successfully for predicting crop yields at the field level. However, numerous input requirements that are specific to the crop type, soil characteristics, and management practices limit their applicability for regional studies. Integrating parameters derived from remotely sensed data with a growth model provides spatial integrity as well as near real-time "calibration" of crop growth simulations (Maas, 1988, 1993). Remotely sensed data arc incorporated in simulations of agricultural crop yields to calibrate or adjust model parameters during the simulation period to ensure agreement between the modeled and satellite observed parameters (Moulin et al. 1995).

The objective of this research is to evaluate a method of integrating Landsat Thematic Mapper (TM) and NOAA A VI HRR data in a crop growth model to estimate spring wheat yields at the sub-county to the state levels.

MATERIALS AND METHODS

Study Area

Spring wheat is the predominant crop grown in the state of North Dakota, located in the northern Midwestern region of the US. The soil and climatic conditions in the eastern part of the state are less harsh than those in the western part of the state. The initial study was conducted using Landsat TM data in three south-eastern counties of the state. Spring wheat in this area is grown in soils generally dominated by loams and clay loams with dark to black soil surface, limy subsoils, sandy loams and loams with sandy or gravelly substrata. The total seasonal rainfall in the eastern region (April to September) ranges from 14 to 18 in. Because spring wheat is grown under non-irrigated conditions, the seasonal variability in rainfall patterns contributes to the variability in crop yields. Other crops cultivated in the study area include spring barley, sunflower and corn. Pasture is generally found in the non-productive soil areas. The 1994 season simulation study with Landsat TM data was conducted at the sub-county level for Sargent, Ransom and Richland in the south-eastern corner of North Dakota (Figure 1). The spring wheat acreage in Sargent, Ransom and Richland counties were respectively, 120, 120 and 2 10 thousand acres. The simulation at the county level for the entire state was conducted using NOAA AVHRR data. The total spring wheat in the state of North Dakota reported by USDA/NASS was 9.1 million acres.

Climate Data

Daily weather data collected from three first order NOAA weather stations in the three counties studied and in surrounding areas were used to extrapolate weather for a particular location in the county using a distance-weighted procedure. The daily maximum and minimum air temperatures and precipitation data was acquired from stations operated by the NOAA's National Climatic Data Center (NCDC). Climate data from a total of 79 climate stations in the entire state of North Dakota were used for estimating yields at the county level.

Soils Data

The major soil groups were identified from the General Soil Map of North Dakota and from the County Soil Survey Report published by the North Dakota Agricultural Experiment Station and the U.S. Soil Conservation Service (USDA Soil Survey Report, 1990). Soil physical and chemical properties were obtained from the EPIC Soils5 database for North Dakota (Sharpley and Williams, 1990). The digital form of the data was brought into the GIs and the general soil association polygons were identified as the basic units to run the crop growth model to obtain regional yields.

Figure 2 shows the major soil groups in the three county Landsat TM study area. The soils are generally nearly level to gently rolling with a thick black surface layer underlain by calcareous, claypan or wet subsoils. Surface texture varied from fine to coarse. The dominant soil groups in the area are the moderately well drained loams and clay loams (Forman-Aastad, Barnes-Hamerly, Barnes-Svea, Garden-Glyndon and Overly-Bearden). Embden-Tiffany and Hecla- Hamar groups are moderately well drained fine sandy loams with rapid permeability and low water holding capacity. Renshaw (fine loamy over sandy) is somewhat excessively drained and had moderately rapid permeability. The Fargo series consists of deep poorly drained fine texture soils with slow permeability and high water holding capacity. The soils data used to run simulations for the entire state using NOAA AVHRR data were assembled from the predominant soil class in each county.

Remote Sensing Data

The Landsat TM satellite data used in the study were acquired on two dates (May 28 and June 30). There were some clouds present in the imagery, however the area covering the three county study area was clear on both dates. A total of four acquisitions of NOAA AVHRR data were available for the 1994 crop season. The digital counts were calibrated to radiances to obtain surface reflectances. The normalized difference vegetation index for each pixel is calculated using the red and near infrared (NIR) reflectance as follows:

NDVI= (NIR-RED) / (NIR+RED) (1)

Landsat TM Crop Classification and Crop Data

Accurate location of the spring wheat within counties was important for obtaining accurate results. Therefore, classification of the Landsat TM data into land use and crop types was done to produce the first input to the yield model simulations conducted for the three south-eastern counties. USDA/NASS developed an accurate crop classification using four acquisition dates of 1994 Landsat TM data for path 28, row 30 in southeastern North Dakota and northeastern South Dakota (Cook, et al. 1996). Two of these acquisitions were from pre-emergence and post maturity stages of crop development.

The crop classification using Landsat TM for the south-eastern counties used ground information obtained by NASS to establish the crop categories and develop the clusters for classification. The Farm Services Administration (FSA) provided farmer- supplied field data to verify the classification accuracies. Spring wheat classification in the south-eastern counties of North Dakota had an accuracy of 87.2% for the NASS data, but 79.2% correct for the FSA data. The classification appears very maplike and does represent well the location of spring wheat in the area (Figure 3.).

Crop information (acreage, yield) for the 1994 season was obtained from the North Dakota Agricultural Statistics Statistical Report. Information on crop phenology was obtained from weekly crop-weather bulletins of North Dakota Agricultural Statistics Service. Spring wheat yields from four field sites obtained by NASS during the 1994 season were used in further validation of the model results discussed later in this paper.

Signature Extension from Landsat TM to Classify Spring Wheat in North Dakota using NOAA AVHRR

The extension of TM classification signatures to NOAA AVHRR imagery were created by selecting areas of interest (AOI) throughout the Landsat TM scene (Stem et al., 1996). In the next step, an unsupervised classification was conducted within the area of interest for the AVHRR scene resulting in a signature file for each class. A "signature" is a range of values for each band. The next step was to determine the composition of each NOAA AVHRR class. For the area where the AVHRR and TM scenes overlapped, a count of the total number of Landsat TM pixels within each AV14RR class was made. Also, the number of pixels for each of the TM classes that overlapped in each AVHRR class was calculated. The number of TM pixels for each AVHRR class in a TM class were divided by the total number of TM pixels the AVHRR class yj overlapped to produce the percentages of TM class N. of which the AVHRR class was composed.

Percent of y_j class composed of x_i class =

The crop mask of spring wheat acreage within each county is shown in Figure 1.

Data Organization

The ARC/INFO geographic information system (GIS) was used to organize, extract, analyze and perform spatial extrapolate data layers. Weather data required for the model simulation with the Landsat TM data was generated by interpolating the data from existing weather stations located with and around the three-county study area. The digitized soils association map provided the basic polygon units for model simulations at the sub-county level.

All data layers were co-registered within the GIS. The NDVI statistics of mean and standard deviation for spring wheat pixels derived from the Landsat TM images were extracted for each soil class within the county, and crop model simulations were run within each soil class. The spring wheat crop for the state was delineated from the NOAA AVHRR classified image and the NDVI statistics obtained for each county. The physical characteristics of the predominant soil class in each county was identified and therefore only one simulation rim was made for each county.

Biophysical Crop Model

Several models were examined for their ability to provide a simulation for regional assessments with a minimum number of input parameters. Since soil moisture conditions are a key factor in semi-arid areas for determining crop yields, it was necessary to have a model with a rigorous soil-water budget component The EPIC (Erosion Productivity Impact Calculator) model developed by Williams et al., (1984) has such a component and was selected to simulate the spring wheat crop growth and yield. EPIC simulates these processes using a daily time step for several different crops using generally available inputs. The EPIC model is a mechanistic growth model describing the potential growth of the crop as a function of irradiation, temperature, precipitation and crop characteristics. The potential biomass is adjusted daily as a function of five plant stress factors (water, temperature, nutrient, aeration and root growth). Its ability to simulate yields of grain sorghum and wheat was reported by Steiner et al., (1987). In southern Alberta, Canada, yields of spring wheat and spring wheat rotations were simulated accurately by EPIC (Toure et al., 1995). However, validation is critical for specific crops in the study region before simulated data are used in further analyses Addiscott and Wagenet, 1985).

Radiative Transfer Model

The one dimensional radiative transfer model, SAIL (Scattering by Arbitrarily Inclined Leaves) (Verhoef, 1984), provides simulated canopy reflectance in the direction of the sensor. The SAIL model requires information about four canopy parameters: LAI, leaf angle distribution (LAD), the single leaf reflectance and transmittance. Earlier investigators have shown how leaf optical properties differ with spring wheat varieties (Pinter et al. 1985 and Jackson and Pinter 1986), Optical properties of the spring wheat varieties grown in North Dakota were selected according to prior studies. Other parameters included solar zenith and azimuth angles, proportions of direct and diffuse shortwave radiation and the sensor view angle. Solar angles are computed as a function of latitude, date and time of satellite overpass time. The EPIC model simulated the daily LAI required as input to the SAM model.

Calibration of Model Simulations with Satellite Data

The model is calibrated by adjusting several parameters that can be measured or derived from a more precise alternate source. In this study, calibrations were limited to adjusting the maximum potential LAI and leaf area decline rate parameters. In addition, since sowing dates are not readily available, planting dates were also adjusted. The calibration procedure is shown schematically in Figure 4. The SAIL model provided the link between the remotely sensed data and the crop model. The simulation was initially made using default parameters to generate two output parameters, the LAI and crop growth. The daily LAI is an input to the SAIL model which simulates reflectances in the red and NIR spectral ranges similar to the Landsat TM and NOAA AVHRR sensors. The NDVI values calculated from the SAIL model were compared with the NDVI derived from direct satellite measurements. A graphical analyses of the daily NDVI simulation and the satellite derived NDVI provided a mechanism for adjusting the model parameters until a reasonable statistical fit was attained.

RESULTS AND DISCUSSION

Crop Simulation and Model Calibration

The crop simulation was conducted at the sub-county level using the Landsat data and county level using NOAA AVHRR data. The input was organized as data layers of climatic parameters, soil physical properties, surface reflectances and NDVI in the GIS within each of the soil types, represented as separate map unit for simulations. The EPIC model was initially calibrated to establish the number of growing degree days between crop emergence and maturity. The spring wheat season in North Dakota begins in mid April and continues through the first week of June while the crop matures by the mid July and continues through mid August. For areas in the southern part of the state where the earliest planting occurs, the crops emerge by the first week in May. Flowering occurs by the second week of June and the spring wheat crop matures by the first week in July for the earliest planting dates. Simulation of crop growth using the earliest planting date and growing degree days of 1300 from emergence to maturity agreed very well with the observed data provided by NASS reports.

Simulation of Crop Yield at the Sub-County Level using Landsat TM Data

The range and magnitude of the NDVI values for the classified spring wheat for Landsat TM acquisition on June 30 is shown in figure 3 . This image represents the post flowering stage of the spring wheat crop. There was a range of variability in vegetation condition, with higher NDVI in the eastern side compared to the western side of the study area. This variability can be attributed to differences in sowing/emergence dates, weather (rainfall and temperature) and soils.

Simulation of the crop growth calibrated with remote sensing data was carried out for all soil types within each county. The model was run at the soil association level and the planting date selected for simulating yield for a particular soil type was based on the statistical fit of simulated NDVI and the satellite derived NDVI from the May 28 acquisition. The second point of adjustment occurs past the flowering stage (June 30) since an earlier acquisition was not available. The calibration of the rate of decline in NDVI is adjusted to match the satellite derived values. The final yields are simulated using the adjusted crop parameters that provided the match for the NDVI values.

The analyses of soil moisture conditions obtained from model simulations suggest that more water stress days occurred in the western part of the study area compared to the eastern side. Barnes and Forman soils in the western part of the study area reach lower levels of available soil moisture compared to Gardena and Fargo soils in the east. This low availability of soil moisture during the vegetative phase continues into the critical stage of flowering. This resulted in a lower seasonal LAI and a more rapid rate of decline in green leaf area. Soil moisture availability is the most critical factor influencing crop yields in dryland agriculture. Although the interpolated rainfall data in the four soil types are similar, the soil water holding capacity and the ability to maintain a continuous water supply throughout the growing season determined the yield levels.

Figure 5 is a map of spring wheat yields in the three county study area simulated within the GIS. The spring wheat yields varied from 9.2 to 44.8 Bu/ac depending on soil types and seasonal patterns of rainfall. The yields were simulated for each soil type, aggregated to obtain the weighted county level average yield and compared with the USDA/NASS reported county yields. A comparison of the simulated and reported yields are shown in figure 6. There are four farmer reported yields at the farm level that are presented along with the county level aggregated yield. The results of simulated yields have shown that the technique for integrating remotely sensed data with crop models is reasonably accurate to monitor yields at field and at regional scales.

Simulation of Crop Yields at the County Level using NOAA AVHRR Data

The simulations at the county level was calibrated with the NOAA AVHRR data. The land use map delineated the spring wheat areas in each county. Although the classification of spring wheat was reasonably accurate there are limitations due to the pixel resolution. The classification of spring wheat presented in Figure I for the state of North Dakota is not intended to calculate the total spring wheat acreage, however, the map product provides the geographic distribution of the spring wheat in the state. At the one Kin pixel resolution, a mixture of summer crop and pasture were present and data from pixels estimated to have greater than 50 percent spring wheat are used in the calibration procedure.

An analyses of the percentage of spring wheat in the classified NOAA AVHRR pixels was conducted and only pixels that had greater than 50% spring wheat were used in the simulation to represent condition of the spring wheat. This selective process reduced the NOAA AVHRR data to a sampling of the pixels most representative of spring wheat. A total of four acquisition were selected to calibrate the crop model simulations. These dates were selected because they provided most cloud free imagery and represented the phenological periods that were suitable for model calibration. Figure 7 shows the results of the simulations for 38 counties in the state. The simulation yields were compared with the USDA/NASS reports and except for five counties in the northeast (closed circles), where there were reports of diseases, the agreement was very good (R2 = 0.98). The remaining 15 counties in the state that did not have a significant acreage of classified spring wheat areas (>50 %) were simulated using model parameters from nearby counties. The combined mean yields for all the counties in North Dakota was 31.99 Bu/ac (SD=3.06) while the reported mean was 31.43 Bu/ac (SD=3.47) and the difference was not significant in a paired t-test at P=0.05 level. The model does not include the effects of pests and diseases and hence an overestimates yields.

CONCLUSION

In this study, we have demonstrated that by combining the EPIC crop model and remotely sensed data, the strengths of one technology overcomes the weakness of the other. This methodology provided an ideal mechanism for monitoring crop growth and estimating yields of spring wheat in North Dakota. Model simulation calibrated with remotely sensed data obtained during the growing season predicted spring wheat yields within one bushel per acre at the Landsat TM resolution. The final adjustment to the crop simulation using remotely sensing satellite data took place about midway between the time of flowering and crop maturity. These stages of crop development are most suitable for the adjustment of the simulations.

The results at the state level using NOAA AVHRR was also very good except in the northwestern part of the state where the spring wheat classification was not at acceptable accuracy. This may be a limitation of the one-Km resolution used for classifying spring wheat where there was a lower acreage of spring wheat in the classified pixels. Additionally, there were other factors that contribute to errors in the data such as sub-pixel cloud contamination and particularly view angle effects since the NOAA- I I satellite overpass time was much later in the day (1600 hrs CDT) during the 1994 season.

The availability of cloud-free satellite data during a critical window of data acquisition is necessary to achieve optimum calibration of the crop model. Although only two Landsat TM satellite acquisitions were used to calibrate the model, data were sufficient to provide good calibration data. The four NOAA AVHRR acquisitions made during the early vegetative phase, flowering and senescence stage of crop development were suitable for performing the calibration. This research has demonstrated improvement in a crop model (EPIC) by using infrequent satellite data acquired at optimum periods during the crop growing season to estimate regional crop yields.

ACKNOWLEDGMENTS

The authors wish to acknowledge Mr. Paul Cook and Mr. Rick Mueller of USDA/NASS for assistance provided in the classification of the spring wheat using the Landsat TM and NOAA AVHRR data. USDA/NASS, Spatial Research Section assisted in acquiring the crop statistical data and their support is acknowledged. Acknowledgement to Dr. Terry Taylor and Mr. Pat Ashburn of the USDA, Foreign Agricultural Service, Production Estimates and Crop Assessment Division, for providing the Landsat TM data required for the study. This project was funded by the U.S. Department of Agriculture, Agriculture Research Service and the National Agricultural Statistical Service.

REFERENCES

Addiscott, T.M. and R.J. Wagenet. 1985. Concept of solute leaching in soils: review of modeling approaches. J. Soil Sci. 36:411-424.

Asrar, G., Kanemasu, E.T., Jackson, T.D., Pinter, JR. 1985. Estimation of total above-ground phytomass production using remotely sensed data. Remote Sens. Environ. 17:211-220.

Cook, P.W., R. Mueller and P. C. Doraiswamy. 1996. Southeastern North Dakota Landsat TM Crop Mapping Project. Proc. Annual Convention & Exhibition, ASPRS/ACSM, April 22-25, Baltimore MD, vol 1:600-614.

Daughtry, C.S. , K.P. Gallo, M.E. Bauer. 1983. Spectral estimation of solar radiation intercepted by corn canopies. Agron. J., 75:527-53 1.

Doraiswamy, P.C. and P.W. Cook. 1995. Spring wheat yield assessment using NOAA AVHRR data. Can. J. Remote Sens. 21 (1): 43-5 1.

Goward, S.N., D. Dye, A. Kerber and V, Kalb. 1987. Comparison of North and South biomass from AVHRR observations. Geocarto International. 1:27-39.

Groten, S.M.E. 1993. NDVI- crop monitoring and early yield assessment of Burkina Faso. Int. J. Remote Sens. 14(8), 1495-1515.

Jackson, R.D., and P.J. Pinter Jr. 1986. Spectral responses of architecturally different wheat canopies. Remote Sens. Environ. 20:43-56.

Maas, S.J. 1988. Using satellite data to improve model estimates of crop yield. Agron. J. 80:655-662

Maas, S.J. 1993. Parameterized model of gramineous crop growth: 11, Within-season simulation calibration. Agron J. 85: 354:358.

Moulin, S., A. Fisher, G. Dedieu and R. Delcolle. 1995. Temporal variation in satellite reflectances at field and regional scales compared with values simulated by linking crop growth and SAIL models. Remote Sens. of Environ. 54:261-272.

Pinter, P.J. Jr., R.D. Jackson, C.E. Ezra and H.W. Gausman. 1985. Sun angle and canopy architecture effects on the spectral reflectance of six heat cultivars, Int. J. Remote Sens. 6:1813-1825.

Prince, S.D., 199 1. A model of regional primary production for use with course resolution satellite data. Int. J. Remote Sens. 7:1555-1570.

Steiner, J.L., J.R. Williams and OR. Jones. 1987. Evaluation of the EPIC simulation model using a dryland wheat- sorghum-fallow crop rotation. Agron. J. 79(4):732-738.

Stern, A. P.C. Doraiswamy and P.W. Cook. 1996. Classifying spring wheat in an AVHRR image by signature extension of a TM classified image. Proceeding of the PECORA- 13 Conference on "Human Interaction with the Environment: Perspectives from Space". August 20-22. 1996, Sioux, Falls, SD.

Toure, A., D.J. Major and C.W. Lindwall. 1995. Comparison of five wheat simulation models in southern Alberta. Can. J. Plant Sci. 75:61-68.

Tucker, C.J. and P.J. Sellers. 1986. Satellite remote sensing of primary production. Int. J. Remote Sens.7(11):1395-1416.

Verhoef, W., 1984. Light scattering by leaf layers with applications to canopy reflectance modeling: the SAIL model. Remote Sens. Environ., 16:125-141.

USDA Soil Survey Report, 1990. Soil Survey of North Dakota Counties. Published by the Soil Conservation Service in Cooperation with North Dakota Agricultural Experiment Station, North Dakota Conservation Extension Service, and North Dakota State Soil Conservation Committee. Washington, D.C.

Williams, JR., C.A. Jones and P.T. Dyke. 1984. A modeling approach to determining the relationship between erosion and soil productivity. Trans. ASAE. 27:129-144.

Figures One through Seven