Guo, X., K.P. Price and J.M. Stiles. 2000. Modeling Biophysical Factors for Grasslands Using Landsat TM Data in Eastern Kansas. The Second International Conference on Geospatial Information in Agriculture and Forestry, Lake Buena Vista, Florida. Jan. 10-12.

MODELING BIOPHYSICAL FACTORS FOR GRASSLANDS
USING LANDSAT TM DATA IN EASTERN KANSAS

Xulin Guo
Kevin P. Price
Kansas Applied Remote Sensing Program and Department of Geography

James M. Stiles
Radar Systems and Remote Sensing Laboratory

University of Kansas
2291 Irving Hill Road, Lawrence, KS 66045, USA

ABSTRACT

The objective of this study is to evaluate the relationships between grassland biophysical factors and spectral reflectance patterns recorded by Landsat Thematic Mapper (TM) in eastern Kansas. The grasslands were stratified by cool and warm season life-forms, as well as grazed, hayed, and Conservation Reserve Program (CRP) management practices. The strength of the relations was examined using the raw TM reflectance values, and various vegetation indices (i.e., NDVI, brightness, vegetation greenness, wetness). Results showed significant relationships between aboveground biomass and all spectral variables (all correlation values were r > 0.65, except TM4 which had an r = 0.42). Relationships between plant moisture and all spectral variables were also evaluated (all correlation values were r > 0.60 except brightness r = 0.41). Aboveground biomass was also highly correlated with plant moisture (r = 0.76). Total vegetation cover was moderately correlated with spectral data (r < 0.55). The relative cover by grasses was uncorrelated, while forb cover was correlated with spectral data suggesting that variation in forb composition among grassland types is a significant factor affecting spectral reflectance. Regression analysis results suggest that TM data are better predictors of aboveground biomass than any other biophysical factors (Adjusted R2 = 0.77). Relationships between the spectral variables and many of the biophysical factors were significantly improved when the models were developed for the individual grasslands types. The use of the TM vegetation indices did not significantly improve the ability to predict biophysical factors.

INTRODUCTION

Monitoring biophysical features of grasslands using remotely sensed data is crucial for many purposes, such as estimating productivity on grazed and hayed grasslands, wildlife habitat planning, biodiversity conservation planning, and understanding grassland processes. The 1987 NASA-sponsored First ISLSCP (International Satellite Land Surface Climatology Program) Field Experiment (FIFE) project conducted on the Konza Prairie near Manhattan, Kansas (NASA, 1987; Sellers et al., 1990) has provided an avenue for evaluating the utility of remotely sensed data to estimate biophysical properties of grassland ecosystems. Weiser et al., 1986 assessed the biophysical properties of tallgrass vegetation, and found that relationships exist between spectral reflectance measurements and grass leaf area index (LAI) and total aboveground green biomass. Asrar et al., 1986 studied the relationships between canopy light interception and leaf area.

Many studies have shown that land management practices affect the biophysical characteristics of grasslands (e.g., Gibson et al., 1993; Collins and Steinauer, 1998). Other studies have documented the relationship between grassland biophysical factors and spectral reflectance patterns (Price et al. 1993; Frank and Aase, 1994; Briggs and Knapp, 1995; Dunham and Price, 1996; Rasmussen, 1998; Zhang, 1998). Purevdorj et al. (1998) showed the relationship between cover and spectral vegetation indices for grasslands in Mongolia and Japan. Todd et al. (1998) used Thematic Mapper (TM) data to estimate biomass for grazed and ungrazed rangelands of the shortgrass steppes of eastern Colorado. Mino et al. (1998) used satellite data to monitor grassland management strategies in Japan. Briggs and Nellis (1991) studied seasonal variation in tallgrass prairie texture as measured by the SPOT satellite to identify spectral differences between grazed and burned prairie ecosystems of eastern Kansas. Dyer et al. (1991) and Turner et al. (1992) measured the effects of mowing and fertilization on tallgrass productivity and spectral reflectance patterns in eastern Kansas.

Studies have documented relationships between grassland biophysical factors and spectral reflectance patterns, few studies, however have shown how spectral and biophysical relationships change as a function of grassland management practices. The objective of this study, therefore, is to examine the relationships between spectral reflectance and grassland that have been stratified by cool and warm season life-forms, and grazed, hayed, and Conservation Reserve Program (CRP) management practices in eastern Kansas. A secondary objective is to compare raw Landsat TM reflectance values versus transformed vegetation indices to determine which are best for predicting grassland biophysical factors in eastern Kansas.

STUDY AREA

The study area is Douglas County in eastern Kansas, USA. Kansas is located in the center of the US which has a mid-continental temperate climate. The county receives an average of 900 mm of precipitation per year with 70% falling during the growing season (April through September). The average annual temperature is 13oC with a mean low monthly temperature of -2oC in January and a high of 26oC in July. The average growing season (frost-free period) is 185 days (USDA, 1977).

In 1990, grassland covered 41% of the county, which has at total area of 122,766 ha (Whistler et al., 1995). Grass species in the county can be classified as either warm season (C4 photosynthetic pathway) or cool season (C3 photosynthetic pathway) life-forms. The dominant warm season grasses include Big Bluestem (Andropogon gerardii Vitman.), Little Bluestem (Andropogon scoparius Michx.), Indiangrass (Sorghastrum nutans) and Switchgrass (Panicum virgatum L.). Based on findings at the Kansas Ecological Reserve, which is 2 km north of the study area, the species composition of the native prairie is approximately 75% forbs and 25% grasses, but the grasses produce approximately 70% of the biomass (Price et al., 1992; Dunham and Price, 1996). The dominant cool season grasses are Smooth Brome (Bromus inermis Leyss.), Tall Fescue (Festuca arundinacea Schreb.), Kentucky Bluegrass (Poa pratensis L.), and Orchard Grass (Dactylis glomerata L.).

The grassland types that were examined during this study include: cool and warm season grasslands under various land management practices (cool season CRP (CC), cool season grazed (CG), cool season hayed (CH), warm season CRP (WC), warm season grazed (WG), and warm season hayed (WH)).

METHODS

FIELD AND SPECTRAL DATA COLLECTION
Field data were collected during the late spring and mid-summer of 1997. Sixty-nine field sites were randomly located throughout the grassland areas of Douglas County. The field sampling period was designed to coincide with the May and July satellite overpasses as closely as possible (approximately two weeks before and after the overflight dates). Estimates of forb and grass cover (%) (n = 69) and aboveground biomass (hereafter, "biomass") (g/m2) (n = 40) were collected within five 0.25 m2 quadrats in each 90 m x 90 m study plot. Due to time constraints, biomass could only be collected at 40 of the 69 sites. Plant moisture content (g/m2) (n = 40) was calculated from the difference between wet and dry biomass. The geographic location of each sample site was measured using a Global Positioning System (GPS).

Three nearly cloud-free Landsat 5 TM images (path 27, row 33) that covered the study area were acquired. The coverage dates of these images were May 15, July 2, and September 4, 1997. After each image was examined for sensor errors (i.e., striping, banding, line dropout), the digital numbers were converted to reflectance and rescaled to values between 0 and 100. The reflectance values were then adjusted for atmospheric scatter using the Improved Dark Object Subtraction technique described by Chavez (1988). The images were corrected for geometric error by first transforming the July image to a Universal Transverse Mercator (UTM) projection. The geometric transformation equation was computed using 38 ground control points that produced a final RMS error of < 0.35 pixels (<10.5 m). The May and September images were then georegistered to the ground-rectified July image (Jensen, 1996). The spectral values for each pixel were interpolated using a nearest neighbor resampling approach and the data were output to a 30 x 30 m pixel size. During the geometric correction process, the images were clipped to the Douglas County political boundary and the thermal bands were eliminated from each image resulting in an 18 band dataset (6 bands for each date).

The following spectral vegetation indices were computed: Normalized Difference Vegetation Index (NDVI), brightness (BI), vegetation greenness (GVI), and wetness (WI) from Tasseled Cap transformation (Crist et al., 1986), green ratio (GR), and Middle Infrared (MIR) ratio (MR). These indices were calculated using the following formulas:

NDVI = (TM4 - TM3) / (TM4 + TM3)
GR = TM4 / TM2
MR = TM4 / TM5
BI = 0.2909(TM1) + 0.2493(TM2) + 0.4806(TM3) + 0.5568(TM4) + 0.4438(TM5) + 0.1706(TM7)
GVI = -0.2728(TM1) -0.2174(TM2) -0.5508(TM3) + 0.7221(TM4) + 0.0733(TM5) - 0.1648(TM7)
WI = 0.1446(TM1) + 0.1761(TM2) + 0.3322(TM3) + 0.3396(TM4) - 0.6210(TM5) - 0.4186(TM7)

STATISTICAL ANALYSIS
Pearson's Correlation analysis was used to determine the linear relationships between the biophysical and spectral variables. Correlation coefficients were listed using a correlation matrix table. A log transformation was used to normalize the data distribution of the biomass and plant moisture. All correlation results were tested at the p < 0.05 using a two-tail test of significance.

Multiple Regression Analysis was used to develop models to test the ability of spectral data (TM bands and vegetation indices) to predict biophysical properties of the grasslands. The dependent variables include biomass and cover. The models were run using the raw TM data and the vegetation indices, separately. A Stepwise method was used to develop the models. The strengths of the Multiple Regression relationships are reported using the Adjusted R2 values.

The correlation and Multiple Regression analyses were performed using the May and July TM imagery. All relationships between cover and spectral data were made using the July TM imagery. Some biomass measurements were made in late May, others started in late June and continued through mid-July. Biomass relationships were therefore calculated using May and July TM imagery.

RESULTS AND DISCUSSION

RESULTS FROM CORRELATION ANALYSIS
Correlation values showing the strength of the relationships between selected biophysical and spectral factors are listed in Table 1. There was a strong relationship between biomass and most spectral variables. The TM visible bands were consistently the most strongly correlated with biomass, with TM1 (blue-green) and TM3 (red) being the stronger predictor (r = 0.88 and 0.83, respectively). Todd et al. (1998) found TM3 to be the best predictor of biomass for the steppes of eastern Colorado. The poorest predictor of biomass and best predictor of total cover was the near infrared (NIR) TM4 (r = 0.42 and 0.51, respectively). Biomass was also correlated with plant moisture (r = 0.76). Since biomass explained about 60% of the variation in plant moisture, plant moisture was not used in the subsequent statistical analyses. Total vegetation cover (grass and forbs) and forb cover were weakly correlated with the spectral data, yet no significant relationships were found between grass cover and the spectral data. TM4 and GVI variables, however, moderately explained variation in forb cover (r = 0.54 and 0.52, respectively). We attribute the lack of correlation between the spectral data and grass cover to low variation in grass cover between and among grassland types. Forb cover varied considerably among grassland types, which we believe accounts for the stronger relationship between the spectral values and forb cover. The grazed sites typically had a much higher forb cover component than the other grasslands (see biophysical statistics in Table 3).

Table 1. Correlation Coefficients (r values) for Biophysical and Spectral Variables

ESTIMATING BIOMASS USING MULTIPLE REGRESSION ANALYSIS

Figure 1 illustrates the inverse relationship between biomass and TM1, which was the best spectral predictor of biomass when all sites were considered together. TM1 accounted for 77% of the biomass variation among all the grassland sites. The distribution pattern of cool season biomass relative to TM1 values shown in Figure 1 is different comparing the pattern for the warm season values. Note that the slope associated with the distribution of cool season values is steeper than the slope for the warm season values. We suspect that the difference in slope between these two grassland life-forms is attributable to less forb cover associated with the cool season grasslands (Table 3). Because forbs generally orient their leaves in a more horizontal plane and grasses tend to be more vertical in their orientation, a small increase in forb composition could have a significant influence on spectral reflectance patterns. This being the case, increases in cool season grassland biomass (which has a much lower forb component) would have less influence on spectral reflectance than warm season grasslands with a much greater forb component. The factor(s) contributing to the differences in the relationship between the spectral measurements and biomass for warm versus cool season grasslands are not completely understood at this time. Regardless of the cause, to most accurately estimate biomass within our study area, different regression equations are needed for cool and warm season grasslands.

Generally speaking, the strength of the relationship between the spectral variables and biomass increased as the grasslands were stratified into more specific grassland types (Table 2). For example, the relationships between biomass and the spectral variables were generally stronger when the six grassland management practices were evaluated separately (R2 values ranged from 0.88 - 0.99), as compared to when all the sites were combined (40 sites combined) (R2 = 0.77). The table also shows that the visible bands were most frequently selected as the best predictor variables used in the Multiple Regression equations.

Table 2. Multiple Regression Results Showing the Relationship between Biomass (Log Transformed) and Spectral Variables.

ESTIMATING COVER USING MULTIPLE REGRESSION ANALYSIS
Figure 2 illustrates the relationship between vegetation cover and TM4 for all 69 study sites. From the regression lines running through the graph, we see a positive relationship between TM4 and total cover (grass and forbs combined), as well as for forb cover. No significant correlation was found between TM4 and grass cover (Figure 2, Table 1). Note that the amount of cover contributed by grasses among the grassland types was similar (Table 3), which explains the lack of a relationship between grass cover and the spectral variables.

The results from Table 3 show that the relationships between the spectral variables and total cover were stronger for the warm season sites, and that correlations between total cover and the spectral variables were strongest for CRP sites (Table 3). Also, strength of the relationship between the spectral variables and total cover increased as the grasslands were stratified into more specific grassland types. The exception in this case was for the cool season hayed sites, which showed no correlation between total cover with any of the spectral variables. This was because, for some hayed sites, there was a different hayed status (hayed or not hayed) between the time of the satellite overflight and the time that the field data were collected. The lack of correlation between cover and spectral variables for the hayed sites also explains the low correlation ( R2 = 0.09) between total cover and spectral variables for the cool season sites compared to the warm season sites (R2 = 0.59) (Table 3).

Table 3. Multiple Regression Results Showing the Relationship between Total Cover and Spectral Variables.

Results in Tables 2 and 3 show that the use of vegetation indices did not significantly improve the correlation between spectral variables, and biomass or cover. Vegetation indices are useful for normalizing spectral variation due to Sun angle, viewing angle, atmospheric, topography, and soil variation (Running et al., 1994). Within our study area, variation due to these factors was very small because the study area was relatively small (reducing atmospheric and solar angle variation), the average slope angle was less than 5% (USDA, 1977), and the soils were mostly covered by live vegetation and litter. Therefore, it seems logical that little advantage would be realized by using vegetation indices for this particular study.

Among the three grassland treatments, generally speaking, cover and biomass were most accurately predicted for the CRP grassland type. CRP sites exhibited the least variability with respect to grass and forb cover (Table 3). These sites are sometime burned in the Spring, and are basically left unaltered throughout the remainder of the year.

CONCLUSIONS

In this study, we determined that spectral measurements from Landsat TM could be used to estimate biomass and cover for cool and warm season grasslands under three management practices. The strength of the relationships between the spectral measurements and biophysical factors generally improved as the grasslands were stratified into more specific grassland types. Biomass estimates from spectral data were more accurate than cover estimates. The visible TM bands were most frequently selected as the best predictors of biophysical factors. Predictions of the biophysical factors were most accurate for warm season as opposed to cool season grasslands. Cool and warm season biomass was most accurately predicted using different regression equations. The biophysical factors for the CRP management practice were estimated more accurately than grazed and hayed management practices. The use of vegetation indices did not improve estimates of the biophysical factors.

ACKNOWLEDGMENTS

This study was supported through the Kansas NASA EPSCoR Project and the Kansas Applied Remote Sensing (KARS) Program, University of Kansas. This work is being conducted in cooperation with the Department of Geography at Kansas State University and the Earth Science Department at Emporia State University. We thank Ina Robertson and Dana Peterson for most of the field data collection.

REFERENCES

G. Asrar, R.L. Weiser, D.E. Johnson, E.T. Kanemasu and J.M. Killeen., "Distinguishing among Tallgrass Prairie Cover Types from Measurements of Multispectral Reflectance," Rem. Sens. Environ.,Vol. 19, pp. 159-169, 1986.

J.M. Briggs and A.K. Knapp, "Interannual Variability in Primary Production in Tallgrass Prairie: Climate, Soil Moisture, Topographic Position, and Fire as Determinants of Aboveground Biomass," American J. Botany, Vol. 82, No. 8, pp. 1024-1030, 1995.

T.M. Briggs and M.D. Nellis, "Seasonal Variation of Heterogeneity in Tallgrass Prairie: a Quantitative Measure Using Remote Sensing," Photogramm. Eng. Rem. Sens., Vol. 55, No. 4, pp. 407-411, 1991.

P.S. Chavez, "An Improved Dark-object Subtraction Technique for Atmospheric Scattering Correction of Multispectral Data," Rem. Sens. Environ., Vol. 24, pp. 459-479, 1988.

S.L. Collins and E.M. Steinauer, "Disturbance, Diversity, and Species Interactions in Tallgrass Prairie," In Grassland dynamics, Oxford University Press: New York, Oxford, pp. 140-156, 1998.

E.P. Crist, R.Laurin and R.C. Cicone, "Vegetation and Soils Information Contained in Transformed Thematic Mapper Data," In Proceedings of IGARSS 1986 Symposium, Zurich, Vol. 2, pp. 1465-1470, September, 1986.

J.W. Dunham and K.P. Price, "Comparison of Nadir and Off-nadir Multispectral Response Patterns for Six Tallgrass Prairie Treatments in Eastern Kansas," Photogramm. Eng. Rem. Sens., Vol. 62, No. 8, pp. 961-967, 1996.

M.I. Dyer, C.L. Turner and T.K. Seastedt, "Mowing and Fertilization Effects on Productivity and Spectral Reflectance in Bromus Inermis Plots," Ecol. Appl., Vol. 1, No. 4, pp. 443-452, 1991.

Frank, A. B., and J. K. Aase, "Residue effects on radiometric reflectance measurements of northern great plains rangelands," Rem. Sens. Environ., 49:195-199, 1994.

D.J. Gibson, T.R. Seastedt and J.M. Briggs, "Management Practices in Tallgrass Prairie: Large- and Small-scale Experimental Effects on Species Composition," J. Appl. Ecol., Vol. 30, pp. 247-255, 1993.

J.R. Jensen, Introductory Digital Image Processing: a Remote Sensing Perspective, Prentice-Hall: Upper Saddle River, New Jersey, p.316, 1996.

N. Mino, G. Saito and S. Ogawa, "Satellite Monitoring of Changes in Improved Grassland Management," Int. J. Rem. Sens., Vol. 19, No. 3, pp. 439-452, 1998.

NASA, The FIFE Experiment Plan, eds P.J. Seller and F.G. Hall, Goddard Space Flight Center, Earth Resources Branch, Greenbelt, MD 20771, 1987.

K.P. Price, V.C. Varner, E.A. Martinko and D.C. Rundquist, "Analysis of Multitemporal Narrow-band Spectroradiometer Measurements from Six Prairie Treatments in Kansas," In Proceedings of ASPRS/ACSM, Albuquerque, New Mexico, pp. 372-385, February 29 - March 5, 1992.

K.P. Price, V.C. Varner, E.A. Martinko, D.C. Rundquist and J.S. Peake, "Influences of Land Management and Weather on Plant Biophysical and Hyperspectral Response Patterns of Tallgrass Prairies in Northeastern Kansas," In Proceedings of PECORA12, Sioux Falls, South Dakota, pp. 441-450, August 24-26, 1993.

Ts. Purevdorj, R. Tateishi, T. Ishiyama and Y. Honda, "Relationship Between Percent Vegetation Cover and Vegetation Indices," Int. J. Rem. Sens., Vol. 19, No. 18, pp. 3519-3535, 1998.

M.S. Rasmussen, "Developing Simple, Operational, Consistent NDVI-vegetation Models by Applying Environmental and Climatic Information: Part I. Assessment of Net Primary Production," Int. J. Rem. Sens., Vol. 19, No. 1, pp. 97-117, 1998.

S.W. Running, C.O. Justice, V. Salomonson, D. Hall, J. Barker, Y.J. Kaufmann, A.H. Strahler, A.R. Huete, J.P. Muller, V.Vanderbilt, Z.M. Wan, P. Teillet and D. Carneggie, "Terrestrial Remote Sensing Science and Algorithms Planned for EOS/MODIS," Int. J. Rem. Sens., Vol. 15, No. 17, pp. 3587-3620, 1994.

P.J. Sellers, F.G. Hall, D.E. Strebel, G. Asrar and R.E. Murphy, "Satellite Remote Sensing and Field Experiments," In Remote Sensing of Biosphere Functioning, eds. R.J. Hobbs and H.A. Mooney, Springer-Verlag, NY, pp. 312, 1990.

S.W. Todd, R.M. Hoffer and D.G. Milchunas, "Biomass Estimation on Grazed and Ungrazed Rangelands Using Spectral Indices," Int. J. Rem. Sens., Vol. 19, No. 3, pp. 427-438, 1998.

C.L. Turner, T.R. Seastedt, M.I. Dyer, T.G.F. Kittel and D.S. Schimel, "Effects of Management and Topography on the Radiometric Response of a Tallgrass Prairie," J. Geophysical Research, Vol. 97, No. D17, pp. 18,855-18,666, 1992.

USDA, Soil Survey of Douglas County, United States Department of Agriculture, Soil Conservation Service in Cooperation with Kansas Agricultural Experiment Station, Washington, D. C., 1977.

R.L. Weiser, G. Asrar, G.P. Miller and E.T. Kanemasu, "Assessing Grassland Biophysical Characteristics from Spectral Measurements," Rem. Sens. Environ., Vol. 20, pp. 141-152, 1986.

J.L. Whistler, S.L. Egbert, M.E. Jakubauskas, E.A. Martinko, D.W. Baumgartner and R. Lee, "The Kansas State Land Mapping Project: Regional Scale Land Use/land Cover Mapping Using Landsat Thematic Mapper Data," In Technical Papers of ACSM/ASPRS Annual Convention and Exposition, Charlotte, NC, 1995.

X. Zhang, "On the Estimation of Biomass of Submerged Vegetation Using Landsat Thematic Mapper (TM) Imagery: a Case Study of the Honhu Lake, PR China," Int. J. Rem. Sens., Vol. 19, No. 1, pp. 11-20, 1998.

Back to top.

Return to Proceedings.