ABSTRACT

Landsat TM imagery data has been used for the classification of crops in small areas, however NOAA AVHRR imagery is more appropriate for regional and continental scales for few specific categories of vegetation. The USDA is interested in assessing crop acreage at the County and State levels. The objective in this study was to conduct the feasibility of using AVHRR to classify spring wheat in North Dakota. Two methods were developed to categorize large areas using AVHRR data and a minimal amount of Landsat TM. Areas of intense agriculture were used to perform an unsupervised classification with AVHRR data. Differences in precipitation and climatic conditions between the eastern and western parts of the State created some difficulties in proper classification and to improve classification accuracy, additional ancillary data was needed. The number of Landsat TM spring wheat pixels in the overlapping AVHRR pixels provided a means of predicting the percentages of spring wheat for each AVHRR class. The accuracy of the spring wheat acreage at the State level closely matched the USDA State report for 1994.

INTRODUCTION

The use of satellite imagery to classify and identify crops is one of the primary applications of remote sensing in agriculture. The U.S. Department of Agriculture is engaged in operational and research projects that are focused on improving the methods of crop inventory in the U.S. The National Agricultural Statistical Service (NASS) has routinely conducted crop inventory and classification projects ( Bellow and Graham, 1992). In these projects, NASS has traditionally used Landsat TM data to develop their supervised classification of crops in selected areas for specific evaluation. The 30m resolution data in the visible and near-IR wavelengths is suitable for classification of crop areas. In general, about two or three images are available for a growing season, and cloud cover may be a problem that reduces the coverage area. Development of an annual and operational assessment of crop classification and inventory for the entire agricultural crop land in the U.S. using Landsat TM data is not feasible. Both the cost of Landsat TM data and the large computing facilities required to conduct such a task are prohibitively expensive.

The Advanced Very High Resolution Radiometer (AVHRR) on board the National Oceanic and Atmospheric Administration (NOAA) polar orbiting satellites provide daily coverage of remotely sensed data at a spatial resolution of 1.1 km. Approximately ten to fifteen near-nadir scenes are available for a growing season with limited cloud cover. The large area coverage was suitable to perform classification of landscape at regional scales and several investigators have attempted to provide crop classification using AVHRR data. Brown, et al., (1993) provided a very reasonable assessment of vegetation distribution in the entire conterminous U.S. The use of the crop-specific masks is limited to very specific assessments such as crop acreage estimation at the accuracy level that NASS conducts its inventory.

Therefore, various studies ( Cross et al., 1991; Ediriwickrema, 1996) have established methodology to relate Landsat TM classifications for small areas to AVHRR data in an effort to extend the Landsat TM classification over a much larger area. The primary advantage of such procedures is the cost savings that occur. The usual disadvantage is that the accuracy and number of categories selected with the AVHRR data is much lower than the number of categories within the Landsat TM data.

Besides the lower cost and greater availability of AVHRR data, having AVHRR data available at nearly daily overpass dates is a great advantage. Compositing the AVHRR scenes using the maximum Normalized Difference Vegetation Index (NDVI) value is one way to create nearly cloud free composite images over a fourteen-day period for the U.S. (Edenshink, 1992). When examining a small area such as the state of North Dakota, selecting cloud-free imagery is possible by selecting single cloud-free dates. This process of using single date imagery permits greater control over the selection of the "best" data for a specific areaTo use Landsat TM classified data for acreage assessment at the county and state levels, USDA/NASS (Hanushak and Morrissey, 1997; Bellow et al., 1993) have developed sampling regression procedures. The use of these procedures requires randomly selected ground truth segments within each county. However, using this methodology with AVHRR data would not be suitable. The AVHRR pixel's size of 1.1 km prevents associating each AVHRR pixel with a unique crop type and/or land cover because the field sizes are sometimes smaller than the pixel resolution. These limitations with AVHRR data prevented use of the sampling regression estimator that NASS has established. Thus, the purpose of this study is to investigate the feasibility of using AVHRR to classify spring wheat in North Dakota. Successful applications may lead to a methodology that can be used to categorize large areas using AVHRR data and a minimal amount of Landsat TM.

STUDY AREA

Spring wheat is the predominant crop grown in the State of North Dakota, located in the Northern Midwest region of the U.S. The total spring wheat in the State of North Dakota reported by USDA/NASS was 9.1 million acres. The soil and climatic conditions in the eastern part of the State are less harsh for cultivating crops than those in the western part of the State. The initial study was conducted using Landsat TM data in three southeastern counties of the State. Spring wheat in this area is grown in soils that are generally dominated by loams and clay loams with dark to black soil surface, limey subsoils, sandy loams and loams with gravelly substrata. The total seasonal rainfall in the eastern region (April to September) ranges from 35 to 45 cm. 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 primarily, spring barley, sunflower and corn. Pasture is generally found in the non-productive soil areas.

DATA

Two sets of satellite imagery were used in this study, Landsat TM and NOAA AVHRR. The Landsat TM images that covered the southeastern corner of North Dakota were categorized by USDA/NASS into 111 categories (Cook.P et al., 1996). The categorization was completed by using randomly selected ground truth segments within each county (Hanushak and Morrissey, 1997). This categorization was used as a template to assign the percentages to the AVHRR classification. The AVHRR imagery was collected for multiple dates: May 7, May 26, June 28, June 29, July 2, July 11, July 17, August 4, August 23, August 25, and Sept 6, 1994. This data was registered to within a 1 Km resolution, screened for clouds and corrected for atmospheric aerosol and water vapor using a procedure described by Che et al. (1995). Two Arc/Info coverages, one of the USDA/NASS strata and the other of North Dakota acreage estimates by counties were acquired from USDA/NASS. These stratifications are created according to the percentage of agriculture within the state ( Hanuschak and Morrissey, 1997). USDA/NASS divides the strata into the following groups: Stratum 11 - most intensive agriculture ( > 75% cultivated), Stratum 12 - medium intensive agriculture ( 51-75% cultivated), Stratum 20 - least intensive cultivated ( 15-50% cultivated), Stratum 31-33 - urban/city categories, Stratum 40 - grassland/pasture ( < 15% cultivated), Stratum 50 - National Parks, Stratum 62 - water. Low cropping intensity areas ( usually, stratum other than 11 or 12 ) contribute little to the state's crop acreage and so can generally be ignored.

METHODOLOGY

Classifications are groups of pixels that are similar in their spectral response. In order, to create a classification it is necessary to have at least one band of imagery. In general, it is preferable to have more than one band of data, by either having numerous different bands, such as visible and infrared or by having a few bands over multiple dates. In this study, visible and infrared bands were stacked for multiple dates through the growing season.

A "signature" is a range of values for each band created for each date of available imagery. Each signature file comprised of a mean and standard deviation for the visible and infrared bands. In the supervised classification performed on the entire state, each pixel is assigned a category based on how well it matched with this "signature". If the pixel does not fall within existing categories, it is assigned to the category which is closest in terms of the Mahalanobis distance (eq. 1) (ERDAS, 1994). A given pixel is assigned to the category for which D is the smallest.

D = (X - M_c)^T (Cov_c^-1) (X - M_c)
(equation 1)

D = Mahalanobis Distance
c = a particular class
X = the measurement vector of the candidate pixel
M_c = the mean vector of the signature of the category c
Cov_c = the covariance matrix of the pixels in the signature of the category c
Cov_c^-1 = the inverse of Cov_c
T = transposition function

Following AVHRR classifications, it was necessary to calculate the percentage of crops for each AVHRR category. Unlike Landsat TM images where each pixel can usually be assigned to a landuse/landcover type, this is generally not possible with AVHRR scenes. This is primarily because AVHRR pixels are larger than the size of the crop fields. To determine the composition of each AVHRR category, the area where the AVHRR and TM scenes overlapped was used. The 111 TM categories were used as a template to determine the composition of each AVHRR category. The number of pixels was then divided by the total number of TM pixels that the AVHRR category overlapped so that percentages could be calculated. For example, AVHRR category X might contain 990 Landsat TM pixels. Of those 990 pixels, 100 could be Landsat TM category A, 50 could be in category B and so forth. Thus, AVHRR category X, is 10.1%, Landsat TM category A, 5.05% category B. Each AVHRR category would have a different total number of pixels and a different composition of Landsat TM categories.

The AVHRR categories based on the 111 Landsat TM categories resulted in categories that were not easily discernable. Therefore, the Landsat TM categories were consolidated into six categories: woods, non-spring wheat crops, pasture, non-agriculture, spring-wheat and water. The percentages of the above categories were recalculated and a more discernable pattern emerged. For example, AVHRR category 10 was 55.6% Spring wheat and 20% non-agriculture.

Method 1

For Method 1, a multi band image was created using AVHRR bands 1 and 2 from the following dates: May 7, May 26, June 29, July 2, July 11, July 17, August 4, August 25 and September 6 of 1994. By creating a temporal multi band image it is hypothesized that the classification methodology will be able to distinguish each crop type by the changes in seasonal reflectance of the visible and infrared bands. Using ERDAS/Imagine, five Areas of Interest (AOI) were created from the Landsat TM study area. The AOI's were created only in areas that were in Strata 11. An unsupervised classification using these AOI's were created and sets of 40 "signatures" were created.

Method 2

After preliminary investigation, some modifications to the Method 1 were performed. Two AVHRR images from June 29th and August 25th possibly had some cloud contamination. They were merged with AVHRR images from June 28th and August 23rd respectively, to create an image that was cloud free. The new AVHRR stack contained imagery for bands 1 and 2 for the following dates: May 7, May 26, June 28/29, July 2, July 11, July 17, August 4, August 23/25, and September 6. In addition, AOI's were selected to cover most of the area from the Landsat TM image, regardless of strata. In this way, "signatures" for agricultural areas and non-agricultural areas were created. Once the unsupervised classification was performed for the Landsat TM area, the rest of the state was classified using a supervised classification using the "signatures" from the unsupervised classification. The complete listing of the composition of AVHRR categories from method 2 is shown in Table 1. Based on this information, estimations of spring wheat at the state were determined.

Table One

RESULTS & DISCUSSION

Categorization accuracy is best evaluated when ground truth data can be used. This is the usual procedure followed when using Landsat TM data and other satellites with pixel resolution that are smaller than a field sizes. This is not feasible with AVHRR data that has a pixel resolution of 1.1 km. To have ground truth of this magnitude it would be necessary to observe the entire 1.1 km2 area and calculate the percentage area of woodlands, non-spring-wheat, spring wheat and others. Therefore, we chose to use the acreage of spring wheat for each county as our test of accuracy for the AVHRR categorization.

Since the first methodology, used only AOI's in intensive agriculture areas, spring wheat acreages were calculated using both Strata 11 and 12. While the results were reasonable at the state level, the accuracy at the county level was not consistent. The second methodology, using merged images to reduce cloud contamination, and AOI's which included all landuse types, was found to be more accurate than the first method. However, spring wheat acreages in the western counties were not as accurate.

To improve, the classification accuracy at the county levels different sources of ancillary data were applied to the classification. The initial attempt was in the use a threshold value for the mahalanobis distance. A map of the mahalanobis distance is displayed in Figure 1 where the values increase from east to west. Due to climatic and ecological differences, crops in the western portion of the state do not match the eastern area from where the AVHRR signatures were developed. This resulted in misclassification for the western portion of the state. Three mahalanbois distance thresholds values were used; namely 39.4, 60 and 200. Even with these threshold values, the accuracies of the acreage estimations were not at acceptable levels.

Figure One

Another attempt to improve acreage estimation was investigated by eliminating all areas that were not in strata 11 or 12. It was assumed that spring wheat outside these areas were most likely a misclassification. This procedure improved the classification accuracy of Method 2. Figures 2 and 3 show the percentages of wheat before and after strata 11 and 12 were applied. In Figure 2, large areas of 30-40% wheat can be seen in the western part of the state, this area is most likely pasture or grassland rather than spring wheat. However, the eastern portion of the state, resembles the classification done with Landsat TM imagery.

Figure Two

Figure Three

Since categorization usually overestimates acreage, a linear relationship between estimated acreage and reported acreage for the five southeast counties ( Richland, Sargent, Ransom, La Moure, Dickey ) was calculated and applied to the final results of method 2. Table 2 shows a complete listing of NASS acreage estimates compared to the various methods that were investigated in this study. The southeast counties of the State are best represented since the signatures were created here for the classification of the AVHRR scene. The east-central and northeast areas also match the NASS estimates fairly closely. Categorization of the central and western counties are not as accurate, but they are farther away from the area that was used to create the AVHRR signature.

A map displaying method 2 using areas within strata 11 and 12 divided by NASS spring wheat estimates is shown in Figure 4. Without exception, the eastern counties were 80-110% of the USDA/NASS estimates. In the western part of the state, a few counties were < 70% of USDA/NASS estimates.

Figure Four

CONCLUSION

This study achieved its goal of assessing spring wheat crop acreage in North Dakota using AVHRR data. It should be noted, there is room for improvement in the methods presented. Because the Landsat TM scene was in the southeast corner, land cover in other areas of the state were not well represented within the "signature" files. As a result of these limitations, it was necessary to use ancillary data such as the USDA/NASS strata to improve acreage estimates. A possible solution for regions with large climatic and ecological variability, would be to select multiple Landsat TM scenes that represent the diversity of the landcover.

ACKNOWLEDGMENTS

The authors wish to thank Mr. Rick Mueller, USDA/NASS, for assistance with the classification of original TM images and review of the final manuscript. Appreciation also to Dr. Pedro Zara and Dr. Ann Hsu, SSAI, for a critical review and suggestions to improve the paper.

Table Two

REFERENCES

Beard, Larry W. and William Hamlin G. 1995. North Dakota Agricultural Statistics 1995. Fargo, North Dakota: North Dakota State University. p 171

Bellow, M.E. and M.L. Graham. 1992. Improved Crop Area Estimation in the Mississippi Delta Region Using Landsat TM Data, ASPRS/ASCM/RT Convention, Washington, D. C. , August 3-7, 1992, pp 423-432.

Bellow, M.E. 1993. Application of satellite data to crop area estimation at the county level. Proceedings of the International Conference on Establishment Surveys, Buffalo, NY, June 28-30, 1993, pp 1-6.

Bellow, M., M. Graham, and W. Iwig. 1993. County Estimation of Crop Acreage Using Satellite Data - Indirect Estimators in Federal Programs. Prepared by sub-committee on Small Area Estimation. Federal Committee on Statistical Methodology. Organized by OMB. Statistical Policy Working Paper 21, Chapter 6:1-28.

Brown, J.F., T.R. Loveland, J.W. Merchant, B.C. Reed and D.O. Ohlen. 1993. Using multispectral data in global landcover characterization: Concepts, requirements, and methods. Photogrammetric Engineering and Remote Sens. 59:809-813.

Che, N., P.C. Doraiswamy and W. Wie. 1995. An operational method of correction for atmospheric and scan angle effects. Proceedings of the International Geoscience and Remote Sensing Symposium, July 10-14, 1995. Firenze, Italy, Vol 3: 1955-1958.

Cook P. ,R. Mueller and P. Doraiswamy, 1996. Southeastern North Dakota landsat TM crop mapping project. Proceeding of the ASPRS, 62 Annual Convention, April 22-25, 1996. Baltimore, MD, Vol. 1: 600-614.

Cross, A.M. J.J. Settle, N.A. Drake and R.T.M. Paivinen. 1991. Subpixel Measurement of Tropical Forest Cover Using AVHRR Data. Int. J. of Remote Sens. 12(5):1119-1129.

Ediriwickrema, J., S. Khorram, M. Gumpertz, and J. Brockhaus. 1996. Estimation of Class Fractions from the AVHRR Data Using Constrained Generalized Least Squares Regression Based Linear Mixture Modeling. Proceeding of the ASPRS Annual Convention, April 22-25, 1996, Baltimore MD, Vol. 1:381-390.

Eidenshink, J.C. 1992. The 1990 Conterminous U.S. AVHRR Data Set. Photogrammetric Engineering and Remote Sens. 58 (6):809-813.

ERDAS Field Guide. 1994 - Third Edition: Atlanta, Georgia: ERDAS, Inc. p.268-269

Hanuschak, G and K. Morrissey. 1977. Pilot Study of the Potential Contributions of Landsat Data in the Construction of Area Sampling Frames. Research and Development Branch, Research Division. USDA/SRS, Washington, D.C. 66 p.

Houseman, E.E. 1975. Area Frame Sampling in Agriculture. Research and Development Branch, Research Division. USDA/SRS, Washington, D.C. SRS-20, 79 p.