Kansas Academy of Science

Mapping land cover in a High Plains agro-ecosystem using a multi-date Landsat Thematic Mapper modeling approach

Kevin P. Price,1 Stephen L. Egbert,1 M. Duane Nellis,2 Re-Yang Lee,1 and Ryan Boyce.1

  1. Kansas Applied Remote Sensing Program and Geography Department, University of Kansas, Lawrence, Kansas 66045 (k-price@ukans.edu).
  2. Department of Geography, Kansas State University, Manhattan, Kansas 66506.

This article is published in the Transactions of the Kansas
Academy of Science, vol. 100, no. 1/2, p. 21-33 (1997).

Table of Contents
Introduction Conclusions
Methodology Acknowledgements
Results & Discussion References

ABSTRACT

The objective of this study was to develop a repeatable procedure for modeling land use and land cover (LULC) within one of the most agriculturally developed and economically significant areas of the High Plains region: Finney County in southwest Kansas. The technique involved the use of Landsat Thematic Mapper (TM) images for three seasons for each of three years (1987, 1989, and 1992). Through a series of image preprocessing and automated classification procedures we were able to discriminate between grassland and croplands more than 95% of the time (previous to this study, less than 70% classification accuracy was common). As we further refined the approach, we were also able to identify crop types: wheat, grain sorghum (milo), corn, and alfalfa, and fallowed lands with greater than 80% accuracy for all five classes, with most crop types mapped at more than 90% accuracy. We also developed a technique that correctly mapped U.S. Department of Agriculture (USDA), Conservation Reserve Program (CRP) lands more than 90% of the time. An analysis of our Finney County crop-type maps for the three study years showed little change in acreage of wheat, grain sorghum, corn and alfalfa, but large increases in grassland, primarily agricultural lands that were converted to the CRP. Lands in fallow declined by about 36,400 ha (90,000 acres) between 1987 and 1992.

Return to Table of Contents

INTRODUCTION

The High Plains region of the United States is characterized by relatively flat topography, a subhumid climate, and shortgrass prairie--see Figure 1. One of the most distinguishing features of the High Plains is that much of it is underlain by the High Plains aquifer, a vast underground reservoir that has transformed the High Plains from a grazing and dryland farming economy to one where irrigated agriculture plays a substantial role (Kromm and White, 1990, 1992). Because of the region's sensitivity to changes in groundwater, climate, and economic and policy factors, it is essential to both inventory and monitor the nature of its land use. The purpose of this study was to develop a reliable, repeatable, and economically feasible protocol for mapping land use/land cover (LULC) within one of the most intensively developed and economically significant areas of the High Plains region -- Finney County in southwestern Kansas.

See physiographic map.
See county map.

There have been a number of large-area LULC mapping projects for regional and statewide areas, primarily using single-date imagery. These include, for example, statewide maps for South Dakota (Tessar et al., 1975), Ohio (Baldridge et al., 1975), Alaska (Fitzpatrick-Lins et al., 1987), Maryland (EOSAT, 1992), and Georgia (ERDAS, 1992). Nearly all of the land cover maps being produced for the National Biological Service GAP project were, or are being, produced using a single-date classification approach. The use of single-data imagery, while cost-effective from an image acquisition standpoint, necessitates making compromises in thematic detail and classification accuracy (Whistler et al., 1995).

Studies using multi-date satellite data to map LULC, although less common, in most cases have reported improved classification accuracies over single-date techniques (Mergeson, 1981; Hill and Megier, 1986; Mauser, 1989). The State of South Carolina Land Resources Conservation Commission developed a detailed land cover map for 19 land cover classes using leaf-on and leaf-off TM imagery (EOSAT, 1994). Fuller et al. (1994) created a land cover map for Great Britain using bi-temporal (summer and winter) scenes. Landsat TM bands 3 (red), 4 (near infrared), and 5 (middle infrared) were combined from summer and winter scenes, resulting in 6-band images that were then submitted to a supervised classification approach that used a maximum likelihood classifier in an iterative process. The authors reported the combined summer-winter scenes offered substantial improvement over single-date classifications.

In the Kansas Land Cover Mapping Project recently completed by the Kansas Applied Remote Sensing Program (Whistler et al., 1995), the state of Kansas was mapped for land cover, county by county, using computer classified single-date Thematic Mapper imagery. Although high accuracies were achieved for the project (>85%), it was found that single-date imagery was limited in the number of cover types that could be identified reliably. For example, in Finney County, grassland could not be discriminated from croplands better than 70% of the time using a single-date automated classification approach. Using the single-date approach, a substantial amount of manual digitizing was required in this county before a minimum classification accuracy of 85% was achieved. It was hypothesized for the current project that the use of multi-date satellite imagery would increase classification accuracies, reduce the manual input required to produce an accurate map, and increase the number of mappable classes.

The specific goals for this study were to develop an improved method for discrimination between croplands and grasslands, and develop the methods for mapping individual crop types and U.S. Conservation Reserve Program (CRP) lands.

Return to Table of Contents

METHODOLOGY

Mapping Grasslands and Crop Types In Finney County, the land cover types of primary concern are cropland and grassland, which together comprise approximately 98% of the county's area (76% crop and 22% grassland, as of 1989) (Whistler et al., 1995). Water and woodland, although present, represent a very small component in the landscape and were classified separately. Although cropland and grassland are relatively coarse classes, due to similar spectral characteristics they proved difficult to separate using the July 1989 image which was used in the Kansas Land Cover Mapping Project. Two particularly intractable problems were the confusion of subirrigated riparian grasslands with growing crops and weedy fallow fields with grassland (Whistler et al., 1995).

Cloud-free multi-temporal Landsat TM images (Path 30 / Row 34) were obtained from the EROS Data Center under NASA's Global Climate Change Program for three years: 1987, 1989, and 1992. Images for each year were acquired for three times during the growing season (April/May, July, and September) (Table 1). The ground resolution of TM data is approximately 30 x 30 m (0.09 ha) per picture element (pixel). The spectral data were converted to radiance values and the optimal index factor (OIF) described by Jensen (1996) was used to select the least inter-correlated bands with the greatest variance. Bands 3, 4, 5, and 7 were selected because they were consistently ranked as having the highest OIF values. Subscenes of Finney County were extracted, registered to each other, and transformed to a Universal Transverse Mercator (UTM) projection (30 x 30 m pixel size) using a nearest neighbor resampling algorithm (rms error < 0.5 pixels). the data were next adjusted for atmospheric scatter using the improved dark object subtraction method developed by chavez (1988).

Table 1. Dates of Landsat TM imagery that were selected for the analysis for three different years.
1987 1989 1992
09 May 87 28 April 89 06 May 92
28 July 87 01 July 89 25 July 92
30 Sept. 87 19 Sept. 89 27 Sept. 92

The resulting 12-band datasets (3 dates per year with 4 bands per date) were classified using an Iterative Self-Organizing Data Analysis technique (ISODATA) (ERDAS, 1994) to generate spectral statistics. These statistics were then submitted to a maximum likelihood classifier that assigned each pixel in the image to one of 100 spectral classes. These 100 classes were regrouped into two classes, cropland or grassland. The spectral classes whose pixels did not fit well into either group (confusion classes) were reclassified into 15 classes and the more refined spectral classes were again assigned to either cropland or grassland. This iterative process, sometimes called "cluster busting," (Jensen et al., 1987) was repeated until an acceptable level of classification accuracy was obtained.

The crop types, for randomly selected fields, were determined from the records of the Finney County Farm Service Agency (FSA - formerly the U.S. Agricultural Stabilization and Conservation Service - ASCS). This information was used to identify field training sites for five major crop and land cover types (winter wheat, grain sorghum (milo), corn, alfalfa, and fallowed lands). The training sites were located on the TM imagery as it was displayed on a computer monitor. The UTM coordinates defining the outer boundaries of each site were extracted through a process called "heads-up digitizing," which involves the use of a computer mouse to delineate the boundaries on the displayed TM image. The summary statistics (minimum, maximum, mean, and standard deviation) for all the pixels within the training site of each cover type were generated. Using a maximum likelihood classifier, all the pixels, except those falling in grassland areas, were classified as one of the five cover types listed above. The number of training sites used for each cover class was proportional to the percentage cover represented by each land cover category. Approximately one third of the training sites that were not used to generate the classification statistics were later used to assess classification accuracy.

Modeling Conservation Reserve Program Lands The CRP represents one of the most profound changes in LULC in southwest Kansas (and in many other regions of the United States). Initiated in 1985, one of the major objectives of the CRP was to decrease crop production and plant croplands to alternative cover types, usually native grasses, for the primary purpose of reducing soil erosion. Over 98% of the CRP contracts were issued to farmers between 1986 and 1991 (FSA, 1994). Southwest Kansas experienced a high degree of CRP participation, with some counties enrolling upwards of 20-25% of their total acreage in CRP (Bair, 1991). Finney County farmers, for example, have enrolled approximately 23,877 ha (59,000 acres) in CRP. Since our first TM dataset was for 1987, and our latest was for 1992 (6 growing seasons), we were able to use a post classification change detection approach (Jensen, 1996) to identify lands that were converted from cropland to grassland during this 6-year period. From field observations and conversations with FSA Officials, we determined that such a land cover change was most often associated with the CRP (Egbert et al., in review; Nellis et al. 1996).

The classification results were compared to field crop type data provided by the Finney County FSA. Verification sites were randomly selected from all geographic areas within the county and represented about a 5% sample of the total area. Within the grassland areas, verification sites were visually identified on Natural Resource Conservation Service (NRCS - formally the Soil Conservation Service) 35 mm natural color aerial photographs.

Return to Table of Contents

RESULTS AND DISCUSSION
Classification Results

Prior to this study, overall classification accuracy of single-date imagery for grassland and cropland in Finney County was 70%. A comparison between the multi-date classification and the FSA field data show percentage agreement ranging from 92.0% to 99.5%, with an average agreement of 96.7% (Table 2). The map for crop type and land cover that was produced through the classification of the 1992 multi-date approach is shown in Figure 2. A comparison between crop types and FSA field data show percentage agreement ranging from 82.5% to 99.2%, with an average agreement of 90.5% (Table 3).

Table 2. Percentage agreement between cropland classes and the USDA field crop type information, and grasslands and aerial photo interpretation for Finney County, Kansas (using unsupervised approach).
Cover Type 1987 1989 1992
Crop 99.3 99.2 99.5
Grassland 93.1 92.0 97.3

Table 3. Percentage agreement between crop cover types and USDA field crop type information for Finney County, Kansas (using supervised approach).
Crop Type 1987 1989 1992
Wheat 90.5 90.0 99.2
Sorghum 84.6 90.8 90.0
Corn 94.3 90.2 84.2
Alfalfa 87.2 99.0 92.8
Fallow 82.5 93.2 88.6

Figure 3 compares acreages as reported by Finney County FSA and acreages as estimated using the multi-date classification approach. For the most part, the reported and estimated acreages are very similar. There are, however, apparent discrepancies between the 1989 estimates for winter wheat and milo. The differences observed for 1989 are actually due to differences in inventory and reporting methods. Acreage estimates by FSA for winter wheat, which is planted in the autumn of the year prior to the year in which it is harvested, are made from aerial photography flown during late autumn following planting and emergence. In this case, wheat acreage was estimated in the fall of 1988. During January, 1989, a warm spell caused the winter wheat to come out of dormancy and approximately 30% of the wheat was killed by hard freezes during the following months. During the following summer, many farmers reseeded their damaged fields to a "ghost crop," which was mostly milo. This is why in 1989, FSA reported more winter wheat, and less milo, than predicted by the multi-date classification approach. The strong overall similarity between the two sets of acreage estimates suggests that both FSA's and the multi-date image-derived estimates are accurate, but because of differences in the reporting procedures, year-to-year differences in acreage estimates may occur, especially if a crop like winter wheat does not reach maturity.

Discussion

The USDA CRP, enacted by Congress in 1985, makes payments to farmers who convert environmentally sensitive cropland to grassland for a period of 10 years. Some of the objectives of the CRP program are to: 1) reduce soil loss from croplands that are on highly erodible soils, 2) reduce the number of acres in crop production so that farmers can make more money on their crops, and 3) improve wildlife habitat. Figure 4 shows the multi-date classification map of the CRP superimposed on the cropland/grassland map. According to this map, the majority of CRP lands are located in the northeast panhandle region of the county. This area is dominated by shortgrass prairie and dry farm lands. Little irrigated land is located in this area due to an insufficient supply of ground and surface water. An initial comparison of a CRP map provided by FSA to the change detection map showed 88% agreement. We believe, however, that this percentage should actually be over 90%. This is because several large CRP fields were located in the study area that were not recorded on the Finney County FSA map. We later learned that if a land owner's residence is in another county, his/her CRP acreage may be recorded in the county of residence.

Areal comparisons showed little change in wheat, grain sorghum (milo), corn, and alfalfa between 1987 and 1992--see Figure 5. As expected, there was a marked increase in the area of grassland due to CRP. This figure shows that the increase in grassland is offset by a comparable decrease in fallowed lands. The data show that between 1987 and 1992, Finney County farmers reduced the lands in fallow by about 36,400 ha (90,000 acres). This change in land use affected over 10% of the total study area, yet the personnel working with the Natural Resource Conservation Service and Farm Service Agency in Finney County could provide no explanation for the dramatic change.

Given our new ability to accurately classify land cover, some questions that might be asked include: (1) how is land use/land cover changing in southwest Kansas, (2) can crop yield models be more effectively extended over larger areas, (3) can irrigated and non-irrigated croplands be distinguished accurately (some preliminary findings suggest that such discrimination is possible), (4) can land cover change be modeled in relationship to factors such as groundwater depletion, government agricultural policy, economic factors, and climate variation, and (5) how is wildlife habitat affected by changes in land cover and government policies that affect land use?

With the majority of 10-year CRP contracts expiring between 1995 and 1999, there is an urgent need to evaluate the status and success of this program in ways that were not possible before the location of these lands could be accurately mapped and evaluated in the context of their surrounding environment. The necessity to rapidly evaluate current CRP lands implies that remote sensing mapping techniques, such as the one described here with multi-seasonal imagery, may be essential tools in the hands of conservationists and agricultural policy planners.

Return to Table of Contents

CONCLUSIONS

Classification accuracy of land cover types and the level of classification detail were substantially improved by using the multi-date dataset derived from Landsat TM imagery. The multi-date approach improved our ability to discriminate between grassland and cropland by over 25%. We found over 90% agreement between the multi-date crop type classification and the crop types reported by FSA. A question that remains unanswered is which estimate is correct when there are disagreements between the multi-date classification and the FSA information. To our knowledge, this study produced the first automated classification of CRP lands. This demonstrates the ability of satellite remotely sensed data to identify lands under different management practices, which is critical to a better understanding such things as: land use, vegetation productivity, biodiversity and changing quality of wildlife habitat, CO2 gas flux, and potential impacts of such changes on the social and economic conditions of a region.

From these results, we learned that in Finney County, the CRP did not reduce crop production, and that major changes in land use did take place within the time frame of this study. We also learned that even major changes in land use may not be detected using existing land use inventory methods.

Preliminary findings from a similar study in northwestern Kansas has shown the multi-date analysis approach can be extended to other areas within the High Plains. The mapping accuracy in northwestern Kansas exceeded 90% for nine cover types which included: alfalfa, cane sorghum, milo, oats, soybeans, sunflowers, winter wheat, fallow, and grasslands. Methods we are now developing also show promise for discriminating irrigated from nonirrigated agriculture by crop type. Similar methodologies as described above are now being adapted to nonagricultural lands in Kansas as well. By the end of 1999, a vegetation alliance level (similar to plant community level) map will be completed for the State of Kansas. Preliminary findings from this project again suggest that the multi- date analysis approach will significantly improve our ability to discriminate among natural vegetation types of Kansas.

Return to Table of Contents

ACKNOWLEDGMENTS

The authors wish to thank the National Aeronautical and Space Administration for financial support provided by a grant through the Remote Sensing Applications section of the Mission to Planet Earth program (NAGW 3810). We also thank Brad Rundquist, Rich Lissitschenko, and Eric White at Kansas State University for their work in acquiring ground truth data for training sites and accuracy assessment. The personnel working for the: Finney County Farm Service Agency, Groundwater Management District #3, Natural Resource Conservation Service, and Kansas State Agricultural Experiment Station were most helpful.

Return to Table of Contents

REFERENCES

Return to Table of Contents

Return to Geospatial Analysis Symposium

Comments to k-price@ukans.edu