The Influence of Vegetation Index and Spatial Resolution on a Two-Date Remote Sensing-Derived Relation to C4 Species Coverage

www.elsevier.com/locate/sna

The Influence of Vegetation Index and SpatialResolution on a Two-Date Remote Sensing-Derived Relation to C4 Species Coverage

Andrew Davidson* and Ferenc Csillag*

Changes in composition of plant species are expected result from the smoothing of fine-scale variation inaboveground live biomass and C4 species coverage. Ourto accompany a warming climate. In the northern mixed

grass prairie, such changes are predicted to take the form results suggest that commercially available satellite dataat resolutions of 10 m to 50 m (e.g., Landsat Thematicof shifts in the relative ground cover of C3 and C4 pho-

tosynthetic types. In this study, we explore the feasibility Mapper) may offer the potential for estimating coverageof C4 species and that the choice of vegetation index usedof using two-date remote sensing data as a potential tool

for monitoring these shifts. Our approach is based on the to estimate biomass is relatively unimportant. However,we caution that for this technique to be operationallywell-described asynchronous seasonality of C3 and C4

species. We hypothesize that the ratios of early-season to useful, statistical model performance must be strength-ened and developed to provide both temporal and spatiallate-season aboveground live biomass (Bearly/Blate) will de-

crease as sites become more C4-dominated, and that if generality. Further investigation is needed to examine theapplicability of this approach to other growing seasons,Bearly and Blate can be reliably estimated using spectral

data, it may be feasible to predict C4 species coverage community types, and grassland regions. Elsevier Sci-ence Inc., 2001(%C4) from commercially available satellite information.

Using spectral and botanical measurements from threeupland communities in the Canadian mixed grass prairie,we (a) examined the relationship between various spec- INTRODUCTIONtral vegetation indices and aboveground live biomass, (b)

In recent years, there has been a growing interest in theinvestigated the nature of the relationship between re-impact of human activities on the biosphere. Of thesemotely sensed estimates of Bearly/Blate and %C4 at multipleimpacts, the possible consequences of anthropogenicallysample resolutions (0.5 m, 2.5 m, 10 m, and 50 m), andinduced climate change have received much attention.(c) assessed whether these relationships were dependentWhile it is generally believed that climatic change willon the vegetation index used to estimate biomass. Wehave substantial effects on the functioning of the bio-found a log-linear relationship between each spectral in-sphere, the exact nature of these effects is unclear. Theredex and aboveground live biomass. Negative linear rela-is, however, a general expectation that changes in vegeta-tionships were found between %C4 and remotely sensedtion patterns will result, and that these changes will notBearly/Blate at all sampling resolutions. These relationshipsbe geographically uniform.were strongest at sampling resolutions of 10 m and 50

Zones of biological transition are expected to be es-m. The strengths and forms of relationships were foundpecially responsive to a changing climate (Neilson, 1993).to be partially vegetation index-dependent. Stronger rela-The northern extreme of the mixed grass prairie is antionships between variants at coarser resolutions likelyexample of such a region. Here, plants that fix carbonthrough the C3 photosynthetic pathway are best adapted

* Department of Geography, University of Toronto, Toronto to the cooler temperatures typical of these latitudes, andAddress correspondence to A. Davidson, University of Toronto, contribute most to the floral diversity and net primary

Department of Geography, 100 St. George St., Toronto, ON M5S 3G3, productivity of the region. Plants using the C4 photosyn-Canada. E-mail: [email protected] 9 July 2000. thetic pathway are less common and exist at their north-

REMOTE SENS. ENVIRON. 75:138–151 (2001)Elsevier Science Inc., 2001 0034-4257/00/$–see front matter655 Avenue of the Americas, New York, NY 10010 PII S0034-4257(00)00162-0

Estimating C4 Species Coverage 139

ern limit, under conditions marginal for growth. These combine two or more spectral bands to enhance the veg-etative signal while minimizing background effects. Theplants are less well adapted to the prevailing climatic

conditions. These unfavorable conditions are thought to most common of these indices utilize red and near-infra-red canopy reflectances (or radiances) in the form of ra-make C4 species extremely responsive to even subtle

changes in climate. The implications of this are twofold. tios [e.g., Normalized Difference Vegetation Index(NDVI)] or in linear combination e.g., PerpendicularFirst, shifts in the abundance and distribution of C4 pho-

tosynthetic types may indicate a changing climate. Sec- Vegetation Index (PVI), while others e.g., Modified Soil-Adjusted Vegetation Index (MSAVI) are more complexond, because C4 species are able to store more carbon

per unit nitrogen than are C3 species, large-scale and also require the derivation of soil correction factors(Chen, 1996). Although many of these VIs have beenchanges in the relative abundances of these photosyn-

thetic types may have significant feedback effects on car- found to be well correlated with various biophysical pa-rameters, including vegetation cover (Purevdorj et al.,bon sequestration and therefore future atmospheric CO2

concentrations (Peat, 1997). However, the response of 1998), aboveground biomass (Boutton and Tieszen,1983), green leaf area (Baret and Guyot, 1991), photo-C4 species to climate change is not readily predictable.

Few studies have explicitly considered the combined ef- synthetically active radiation (PAR) (Hatfield et al.,1984), and productivity (Box et al., 1989), some, such asfects of varying multiple climatic variables on C4 produc-

tion, and those that have attempted to do so have pro- the NDVI, have received more attention than others.The principal challenge in applying remote sensingvided results that are somewhat contradictory (see

Dahlman, 1993). observations to monitor the C3/C4 vegetation dynamicsof grasslands lies in inferring the relative abundances ofThe increasing availability of remotely sensed data at

various spatial and spectral resolutions offers the poten- C3 and C4 species in mixed canopies (Goodin and Hen-ebry, 1997). This is because pixel reflectances, or reflec-tial to monitor the biophysical characteristics of ecosys-

tems at various landscape scales (Tieszen et al., 1997). tance-derived estimates of biomass, are influenced by therelative contributions of both photosynthetic types. WeThis has been particularly well demonstrated by those

who have focused their attentions on grasslands. Spectral propose that the asynchronous seasonality of C3 and C4types can be used to “unmix” remotely sensed estimatesobservations over large tracts of grassland typically have

been acquired using National Oceanic and Atmospheric of total aboveground biomass into their contributingcomponents, particularly the areal coverage of C4 spe-Administration (NOAA) Advanced Very High Resolution

Radiometer (AVHRR) data (Paruelo and Lauenroth, cies. Seasonality occurs because C3 species are generallyless well adapted to conditions of high light (Tieszen,1995; Tieszen et al., 1997). While these spatial resolu-

tions (1–4 km) provide valuable information about cli- 1970), elevated temperature (Schuster and Monson,1990), and limited soil moisture (Barnes et al., 1983)matic and aggregate anthropogenic forcings on vegeta-

tion dynamics (Henebry, 1993), they are often at a than are C4 species. These factors combine to produceseasonal cycles where C3 species (cool season plants)coarser spatial resolution than some applications require.

Finer-resolution satellite observations have been derived green up in early spring and are most active under thecooler conditions of spring and fall. In contrast, C4 spe-from Landsat Multispectral Scanner (MSS) (Pickup et

al., 1993) and Landsat Thematic Mapper (TM) sensors cies (warm season plants) green up later in the growingseason and are more active under the hotter and drier(Henebry, 1993) at resolutions of 50 m and 30 m, re-

spectively. At even finer resolutions, information has summer months. As a result, peak annual biomass tendsto occur earlier in the growing season for C3-dominatedbeen provided by airborne (Walthall and Middleton,

1992) and ground-based observations (Weiser et al., sites relative to that of C4-dominated sites (see Goodinand Henebry, 1997), a trend that should also be re-1986; Goodin and Henebry, 1997), whose spatial resolu-

tions are determined by sensor height and field of view flected in both early-season (Bearly) and late-seasonaboveground live (green) biomass measurements (Blate).(FOV).

One of the greatest challenges in the remote sensing If this is generally the case, we may also expect the ratioof early-season aboveground live biomass to late-seasonof grasslands has been the reliable estimation of biophys-

ical variables, such as aboveground biomass and produc- aboveground live biomass (Bearly/Blate) to decrease as sitesbecome increasingly C4-dominated. Furthermore, iftivity, from multiresolution observations. This is largely a

consequence of the familiar “mixed pixel” problem, Bearly/Blate can be reliably estimated using some spectralVI (e.g., NDVI), it may then be feasible to predict thewhere factors other than the presence and amount of

green vegetation (e.g., senescent vegetation, soil, shadow) percentage of ground surface covered by C4 species(%C4) over wide geographical regions from various re-combine to form composite spectra (see Asner, 1998).

This spectral mixing often makes the discrimination of mote sensing platforms.However, to evaluate the potential of this remotegreen vegetation difficult and has prompted the develop-

ment of numerous spectral vegetation indices (VIs). VIs sensing approach, two sampling issues must be explicitly

140 Davidson and Csillag

addressed. First, remotely sensed data collection and tra- METHODSditional ecological research normally operate at incongru- Study Region: Grasslands National Park,ous spatial resolutions. While remote sensing observa- Saskatchewan, Canadations are usually taken at spatial resolutions ranging from

Species composition, biophysical, and spectral data were10 m Systeme Pour l’Observation de la Terre-Panchro-collected during the 1995 growing season at three sitesmatic (SPOT-P) to 4 km (AVHRR), biological researchwithin Grasslands National Park (GNP), Saskatchewan,has tended to focus on plot level (,1 m) phenomenaCanada (498159 N, 107809 W). GNP was established in(Sellers et al., 1990). This disparity is one of the most1988 to preserve a representative portion of the Cana-important limiting factors in the reliable estimation ofdian mixed grass prairie ecosystem. The climate of thegrassland ecosystem parameters from remotely sensedregion is semiarid; winters are long, cold, and dry whiledata and must be narrowed significantly. Second, many summers are short, hot, and comparatively wet (Environ-

studies fail to consider the effects of scale (sample reso- ment Canada, 1998). Mean daily temperature rangeslution) on the relationship between spectral reflectance from 158C below zero in January to 208C in July. Totaland biophysical parameters, and in doing so, ultimately annual precipitation averages 325 mm. Approximatelychoose a scale of observation that is inappropriate for the one-third of this total (110 mm) falls as snow in the win-task at hand. Relationships often vary with scale, and the ter, while the remainder (215 mm) falls as rain through-way in which pattern and process vary across scales is out the rest of the year, mostly during heavy (but infre-often not well understood. The correlation between the quent) summer thunderstorms (165 mm). The growingstate of an environmental variable and plant community season in the park is short, averaging 170 days betweencomposition may appear to be significant at some scales, killing frosts, but low moisture availability often furtherbut not at others (Lobo et al., 1998). To identify the reduces its effective length (Loveridge and Potyondi,most suitable scales of measurement, or to “scale up” 1983). However, because the average daily bright sun-plot level relationships to the scales needed for regional shine is high during the growing season (2400 h annu-studies, changes in the relationship between spectral data ally) (Environment Canada, 1998), growth rates can beand biophysical parameters must be explicitly investi- great when sufficient moisture is available. The studygated across a range of spatial scales. area is dominated by grasses, of which Stipa comata

This study explores the feasibility of using two-date (C3), Agropyron spp. (C3), and Bouteloua gracilis (C4)remote sensing data as a potential tool for the prediction are the most common. Sage, mosses, lichens, and cactiof C4 species coverage over a C3-dominated grassland. make up a significant part of the plant community inTo achieve this and to address the above sampling issues drier locations.directly, we collected C4 species cover information andspectrally derived estimates of Bearly/Blate over plots of Location and Description of Sample Sitescomparable ground resolution (approximately 0.5 m) Potential sampling locations for our 1995 field campaignfrom three field sites in the Canadian mixed grass prai- were selected based on a 1993 vegetation inventory andrie, then used a spatially nested sampling design to scale classification that was conducted on GNP lands (Michal-each property and the relationships to coarser observa- sky and Ellis, 1994). Using a Geographical Informationtional scales (2.5 m, 10 m, and 50 m). Specifically, we System (GIS), we identified sites that were representa-seek to (1) compare the forms, strengths, and similarities tive of native prairie vegetation and the surrounding Parkof eight field-derived univariate models, each describing area and were undisturbed. The only vegetation classthe statistical relationship between a spectral vegetation that met the above requirements was the “Upland Grass-index and aboveground live biomass, and (2) examine the land” type, which is characterized by “a dominant coverform of the functional relationship between C4 species of grasses or sedges, with shrub cover being low or ab-coverage and remotely sensed estimates of Bearly/Blate and sent” (Michalsky and Ellis, 1994).describe how this relationship is affected by (a) the scale We located three sampling sites in separate regions(sample resolution) of measurement and (b) the remotely of Upland Grassland at GNP. These sites were situatedsensed vegetation index utilized to estimate biomass. As along a soil moisture gradient and were chosen to repre-with many field-based experiments, sampling limitations sent areas of low and high total seasonal productivity.prevent us from validating empirical relations with inde- The “wet” and “intermediate” sites were located in topo-pendent observations derived from other grassland sites. graphic depressions, providing available soil moisture forFor this reason, the cross-validation of such relations are plant growth at each site, even during hot summerrestricted to the use of the bootstrap method. In this months (Davidson, unpublished data). In contrast, thesense, our study follows similar approaches to estimating “dry” site was located on an exposed plateau, where ex-biophysical parameters using spectral indices that do not treme water limitation was observed during the hottestexplicitly make use of separate training and validation periods of the growing season. While the dominant plant

types at each sampling location were very similar, sitesdata sets.


varied in the relative abundance of C3 and C4 species, to a barium sulfate panel (once every 7 to 12 minutes).Reflectance factors were then computed from a ratio oftotal species richness, and the abundance of rarer species.canopy radiance to that of the reference panel (afterWeiser et al., 1986). Zenith angles ranged from 39.68 toSampling Design30.98 on May 15th and from 38.48 to 30.18 on July 28th.

Estimating Bearly and Blate from Spectral We used the mean of three separate reflectance mea-Reflectance Information surements as a representative measure of canopy reflec-Plot-level relationships between aboveground live bio- tance at each subplot. These means were then trans-mass and each spectral vegetation index were derived us- formed into the suite of VIs described in Table 1ing a stratified random sampling approach. At each sam- (Richardson and Everitt, 1992; Jordan, 1969; Rouse etple site, we established eight “calibration” plots (5 m by al., 1973; Roujean and Breon, 1995; Huete, 1988; Qi et5 m), in which four circular subplots (0.5 m diameter) al., 1994; Crippen, 1990; Chen, 1996).were randomly located. Calibration plots were situated The harvesting of aboveground biomass from eachso that areas of low (5 g m22) to high (135 g m22) subplot occurred immediately after the collection ofaboveground live biomass were adequately represented spectral data. All vegetation within a subplot was clippedwithin the sampling framework. At various times during and separated into litter, live, and dead components. Allthe 1995 growing season (early May to mid-August), si- samples were dried in an oven at 608C for 48 hours andmultaneous spectral and biophysical data (litter and live, weighed. These weights (in g) were multiplied by a con-dead, and total standing biomass) were collected from a version factor (5.09) to provide biomass estimates in grandomly chosen subplot within each calibration plot. m22. Log-linear regressions were then used to character-Once a subplot was sampled, it was precluded from be- ize the relationship between each VI and abovegrounding sampled again (i.e., sampling was random without re- live biomass (see description later).placement).

Spectrally reflected radiation in Landsat TM bands Scaling Plot-Level Estimates of %C4 and RemotelySensed Bearly/Blate to Coarser Resolutions3 (0.63–0.69 lm) and 4 (0.76–0.90 lm) was measured

over each subplot using an Exotech Model 100BX radi- Estimates of %C4 and Bearly/Blate were derived directlyfrom individual plot observations (n572 at each sampleometer, Exotech Incorporated, Gaithersburg, Maryland,

USA. The radiometer has a 158 FOV and was mounted site). At this resolution, we used a square quadrat frame(0.5 m length, marked at 0.1 m intervals) to visually esti-at a height of approximately 1.5 m, giving a spatial sam-

ple resolution of approximately 0.5 m (nadir view). All mate the proportional areal coverage of individual spe-cies present in each plot. Following the sampling meth-canopy measurements were taken on cloudless days

within 2 hours of solar noon and frequently referenced odology of Michalsky and Ellis (1994), species cover was

Table 1. Vegetation Indices Used to Estimate Early-Season (Bearly) andLate-Season Biomass (Blate)

Index Formula Sources

DVI NIR2R Richardson andEveritt (1992)

RVI Jordan (1969)NIRR

NDVI Rouse et al. (1973)NIR2RNIR1R

RDVI Roujean andNIR2R√NIR1R Breon (1995)

SAVI Huete (1988)(NIR2R)(11L)(NIR1R1L)

, where L50.5

MSAVI2 Qi et al. (1994)2NIR112√(2NIR11)228(NIR2R)2

IPVI Crippen (1990)NIRNIR1R

MSR Chen (1996)NIRR

21

!NIRR

11

NIR and R correspond to reflectances in the near-infrared and red wavelengths, respectively.


each corresponding to a different spatial resolution (seeWebster, 1979). In effect, nested sampling provides agood measure of landscape “patchiness” at varying spatialscales. This approach is based on the conditions that thespatial components of the population can be divided intodistinct levels and that observations are viewed as the re-sult of the nested contributions of these levels (Belle-heumer and Legendre, 1998).

At each sample site, individual plots were spatiallynested as four levels inside a 100 m by 100 m grid (Fig.1a). Each level of nesting corresponded to a specificsampling resolution (level IV50.5 m; level III52.5 m;level II510 m; level I550 m). Plot-level measurementsof %C4, Bearly, and Blate were then spatially averaged (Fig.1b) to provide estimates of %C4 and Bearly/Blate at “upper”levels (i.e., coarser resolutions) of the hierarchy (three-plotaverages at level III; seven-plot averages at level II; eight-plot averages at level I). Thus, during the aggregationprocess, the original data were reduced to a smaller num-ber of data points, which in turn represented a greaterarea than the original units. The aggregation of plots, re-sulted in n518 estimates of each parameter at level III(2.5 m), n59 estimates at level II (10 m), and n59 esti-

Figure 1. Field sampling scheme. (a) The mates at level I (50 m). At the wettest site, however, thelocation of calibration (n58) and spatially presence of cloud during the collection of spectral datanested (n572) plots at each of our three

meant that only 64 plot measurements were usable atsample sites. (b) The spatial averaginglevel IV (n564). This resulted in n516 estimates of eachmethod through which plot-resolution (0.5

m) information is scaled to coarser parameter at level III, n58 estimates at level II, and n58observational scales (2.5 m, 10 m, 50 m). estimates at level I for this site. Linear least squares re-

gressions were used to examine the forms and strengthsof the relationship between %C4 and Bearly/Blate for each

estimated to the nearest 5% for cover values ranging vegetation index and at each sampling resolution.from 10% to 90% and to the nearest 1% for cover valuesless than 10% and greater than 90%. Individual C4 spe- Statistical Methodscies cover measurements were then summed to provide

All statistical analyses were conducted using the S-PLUSan estimate of total plot C4 species coverage (i.e., %C4).statistical package (Version 4.5, MathSoft Incorporated,The collection of plot-resolution spectral data followedSeattle, Washington, USA, 1998), and p,0.05 was usedthe sampling protocols outlined previously. Our a priorito determine significance in all tests. Simple linear re-identification of sampling dates for remotely sensed mea-gression models were used (1) to characterize the rela-surements of Bearly and Blate was based on the knowledgetionship between each spectral vegetation index andof “typical” seasonal trends in plant production across theaboveground live biomass, and (2) to assess the scale-GNP area. To capture early season and late seasondependency of the functional relationship between %C4aboveground live biomass, plot spectral measurementsand remotely sensed estimates of Bearly/Blate. Regressionwere taken on 15 May 1995 and 28 July 1995, respec-statistics for analysis (1) were generated using n592tively. For both measurement dates, these reflectancespairs of observations. Because we wished to investigatewere used to calculate the vegetation indices describedgeneral trends and because sample sizes are relativelyin Table 1, from which various estimates of abovegroundsmall at coarser resolutions, regression statistics for anal-live biomass (and subsequently Bearly/Blate) were derivedysis (2) were generated for all sites combined (n5208 atusing empirical relationships.0.5 m; n552 at 2.5 m; n526 at 10 m; n526 at 50 m).We used an unbalanced nested sampling design toWe tested all variates for normality prior to their use.scale these plot-level (0.5 m resolution) measurements ofMuch of the data were heteroscedastic and so were%C4 and remotely sensed Bearly/Blate “up” to coarser spa-transformed to meet the distributional assumptions oftial resolutions of 2.5 m, 10 m, and 50 m (Figs. 1a andthe regression models. In analysis (1), each vegetation in-1b). Spatially nested sampling applies classical analysis ofdex was loge transformed. In analysis (2), estimates ofvariance to a hierarchical design to partition the total

variance of a study variable into varying components, Bearly/Blate and %C4 were square root transformed. In all


cases, proper diagnostic checks were performed on modelresiduals. All assumptions regarding residuals were met,including those concerning spatial autocorrelation.

Because of sampling limitations, we were unable tocollect additional independent data for the validation of re-gression-derived relationships, nor to provide a sufficientlylarge sample [especially at coarser resolutions for analysis(2)] to facilitate validation through data splitting methods(i.e., the subsetting of original data into independenttraining and validation data sets). For this reason, cross-validations of the relationships derived from analyses (1)and (2) were performed using Monte Carlo bootstrap re-sampling. Bootstrap resampling is a computer-basedmethod for assigning measures of accuracy to statisticalestimates (see Efron and Tibshirani, 1993). Using thebootstrap approach, 1,000 new samples, each of the samesize of the observed data [i.e., n592 for analysis (1);n5208 at 0.5 m; n552 at 2.5 m; n526 at 10 m and 50m for analysis (2)] were drawn with replacement fromthe observed data. Regression coefficients (slopes, inter-cepts) were first derived using the observed data and

Figure 2. (a) A graphical illustration of thethen recalculated for each new set of observations, givinglogarithmic relationship between RVI anda bootstrap distribution for each statistic. These resam- aboveground live biomass (g m22), as derived

pled estimates were then used to estimate the bias of the from plot-level (0.5 m) observations. (b)Coefficients generated from linear least squaresstatistic derived from the original observations.regressions of each spectral vegetation indexWe tested the effects of vegetation index on esti-listed in Table 1 (loge transformed) onmates of Bearly, Blate, and Bearly/Blate at each sampling resolu- aboveground live biomass (g m22).

tion using single-factor analysis of variance (ANOVA).Vegetation index was treated as a fixed factor. Analysis ofvariance is robust to violations of the assumptions of nor- squares regression provides unbiased estimates of slopemality and homogeneous variance as long as sample sizes and intercept for each univariate relationship. In allare nearly equal (Zar, 1984), a condition that held true cases, bootstrapped estimates (and their associatedfor our study. Where a significant effect of vegetation in- RSEs) do not differ significantly from least squares re-dex was detected, Tukey’s pair-wise comparisons were gressions of the original data. Such results indicate thatused to contrast mean values for significant differences. these statistical models are appropriate descriptors for

the majority of our data.

RESULTSEffects of Spectral Vegetation Index on

Relationships between Vegetation Indices and Estimates of Bearly, Blate, and Bearly/BlateAboveground Live Biomass We used the previously described empirical relationshipsThe relationship between each vegetation index and above- to estimate Bearly and Blate at each scale of observation (Ta-ground live biomass was found to be logarithmic in na- ble 2) from which corresponding estimates of Bearly/Blate

ture (Fig. 2a). These relationships were described statis- were subsequently derived. Early-season biomass esti-tically using log-linear regressions of each loge-trans- mates showed a nonsignificant response to vegetation in-formed index (independent variable) on aboveground dex at each sampling resolution. In comparison, signifi-live biomass (dependent variable) (Fig. 2b). In all cases, cant responses were found for estimates of Blate atestimates of slope (b) and intercept (c) were highly sig- sampling resolutions of 0.5 m and 2.5 m and for esti-nificant (p50.0000). The variation in aboveground live mates of Bearly/Blate at 0.5 m. In all cases, Tukey’s pair-biomass was best explained using the NDVI, Soil- wise comparisons showed that these significant responsesAdjusted Vegetation Index (SAVI), Second Modified were due to lesser average biomass estimates derived us-Soil-Adjusted Vegetation Index (MSAVI2), and Modified ing the DVI (Table 2). This was especially evident forSimple Ratio (MSR) [r250.64; residual standard error DVI-derived estimates of late-season biomass, which(RSE)518.9] and least well explained by the Difference were statistically different to those generated from allVegetation Index (DVI) (r250.51; RSE522.0). Monte but one [the Ratio Difference Vegetation Index (RDVI)]

of the other indices at both 0.5-m and 2.5-m resolutions.Carlo bootstrap resampling verifies that linear least


Table 2. Spectrally Derived Estimates of Early-Season (Bearly) and Late-Season (Blate) Aboveground Live Biomass (mean6SD,g m22) at 0.5-m, 2.5-m, 10-m, and 50-m Sampling Resolutions

Estimated Bearly (g m22) Estimated Blate (g m22)

Resolution 0.5 m 2.5 m 10 m 50 m 0.5 m 2.5 m 10 m 50 m

RVI 34.9613.0 34.868.7 34.867.4 34.967.0 70.1623.8 72.1622.5 70.8621.7 70.1620.8NDVI 33.7617.2 33.6611.4 33.669.7 33.769.1 70.3619.8 72.0618.1 70.8617.7 70.3617.1IPVI 33.9614.9 33.8610.0 33.868.5 33.968.0 70.2622.2 72.1620.6 70.8620.1 70.2619.3MSAVI 33.9617.2 33.8611.3 33.869.7 33.969.1 70.4619.8 72.2618.1 71.0617.7 70.4617.0MSAVI2 33.7617.7 33.7611.7 33.7610.0 33.769.4 70.2619.0 71.9617.2 70.7616.8 70.2616.2DVI 33.0614.5 31.968.5 32.367.0 33.065.9 59.3617.4 60.4614.7 60.3614.2 59.3613.6MSR 33.8616.3 33.8610.9 33.869.2 33.868.7 70.4621.0 72.2619.4 70.1618.8 70.4618.2RDVI 32.5615.8 31.869.8 32.068.3 32.567.3 64.8618.8 66.2616.8 65.6616.4 64.8615.7

Estimates at 0.5 m were derived directly from the empirical relationships presented in Fig. 2 and subsequently scaled to 2.5-m, 10-m, and 50-mresolutions using the nested sampling scheme illustrated in Fig. 1.

Relationships between Remotely Sensed Bearly/Blate We found negative and statistically significantand %C4 (p50.0000) linear relationships between square root-

transformed %C4 and all VI-derived measures of squareSlopes (b), intercepts (c), coefficients of determinationroot-transformed Bearly/Blate at each sampling resolution(r2), and RSEs were derived from the simple linear re-(Fig. 3, Table 3). The slopes and strengths of these rela-gressions of square root-transformed Bearly/Blate on squaretionships are strongly sample resolution-dependent. Forroot-transformed %C4 at each scale of observation (Figs.each vegetation index, estimated slope coefficients de-3, 4, and 5 and Table 3). Note that a direct interscalecrease nonlinearly as sample resolution becomes coarsercomparison of correlations and regression coefficients by(Fig. 4). This indicates that linear regression model esti-means of RSEs and p values is difficult because samplemates of %C4 are more sensitive to changes in Bearly/Blatesize varies between resolutions (the exception being 10at coarser observational scales and that interscale differ-m and 50 m). We still present these statistics, but they

should be viewed as general indicators. ences in model sensitivity are maximal between sampling

Figure 3. Linear least squares regressions of RVI-derived Bearly/Blate (square root transformed) on %C4(square root transformed) at sampling resolutions of 0.5 m (a), 2.5 m (b), 10 m (c), and 50 m (d).Point data illustrate the scatter (RSE) around the regression line.


Figure 4. Sample resolution and vegetation index dependence of slope estimates derived from thesimple linear regressions of remotely sensed estimates of Bearly/Blate (square root transformed) on%C4 (square root transformed).

resolutions of 0.5 m and 2.5 m and minimal between 10 estimates of %C4 are approximately 3%. Figures 4 and5 demonstrate that relative to each other, most indicesm and 50 m [also see Fig. 5, where %C4 (as opposed to

its square root transformation) is expressed graphically as “perform” similarly across observational scales. For ex-ample, at each sampling resolution, slope coefficientsa function of square root-transformed Bearly/Blate]. The co-

efficients of determination generated from these regres- generated by regressions using the RVI are more nega-tive than those generated using the Infrared Percentagesions indicate that the strength of the above relationships

is also sample resolution-dependent (Table 3). These co- Vegetation Index (IPVI), which in turn are more nega-tive than those derived using the MSR, SAVI, NDVI,efficients show that the level of explanation of all statisti-

cal models increases nonlinearly as sample plots are ag- and MSAVI2 (i.e., bRVI,bIPVI,bMSR,bSAVI, bNDVI,bMSAVI2).In contrast, slope estimates derived using the DVI andgregated to coarser resolutions. Weak linear relationships

at 0.5-m resolutions (r250.08 to 0.14) become stronger RDVI do not show consistent relative trends across reso-lutions. These results indicate that of all the indices ana-as samples are aggregated to 2.5 m (r250.33 to 0.47) and

10 m (r250.58 to 0.68). However, further aggregation to lyzed, RVI-derived estimates of %C4 are the most sensi-tive to changes in Bearly/Blate, while in comparison,50 m does not increase the strength of the relationship

between variates (r250.57 to 0.65). MSAVI2-derived estimates are generally the least sensi-tive to such changes. Slope coefficients derived using theThe slopes and strengths of these relationships are

also vegetation index-dependent (Figs. 4 and 5, Table 3). RDVI display the largest overall “jump” across sampleresolutions, especially from 10 m to 50 m. In compari-At each sampling resolution, the effects of vegetation in-

dex are most evident at the lowest extreme of the ob- son, coefficients derived from the RVI display the leastchange across resolutions. The strength of these relation-served range in Bearly/Blate (Fig. 5). Here, these differences

correspond to estimates of %C4 ranging from (a) 38% ships is also vegetation index-dependent. The strongestrelationship (i.e., highest r2, lowest RSE) at each resolu-(MSAVI2) to 52% (RVI) at 0.5 m, (b) 56% (MSAVI2) to

72% (RVI) at 2.5 m, (c) 75% (MSAVI2) to 89% (RVI) tion was produced using the RVI, while the weakest rela-tionship was produced using the MSAVI2 (Table 3).at 10 m, and (d) 84% (MSAVI2) to 96% (DVI) at 50 m.

At resolutions coarser than 2.5 m, the effects of vegeta- Monte Carlo bootstrap resampling indicates thatleast squares regression provides unbiased estimates oftion index decrease to a minimum as observed Bearly/Blate

increases to a maximum. Here, the interindex ranges in slope (Fig. 6) and intercept for each of the previously


Figure 5. Regression lines showing the influence of sampling resolution and vegetation index on the predictability of %C4(nontransformed). For the sake of clarity, where the responses of two or more vegetation indices are similar, trendsare presented as a single line.

described relationships. For all indices, and at all sample tionship (Friedl et al., 1994). We have shown that theresolutions, bootstrapped estimates (and their associated relationship between each spectral vegetation index andSEs) do not differ significantly from least squares regres- aboveground live biomass is best described by a logarith-sions of the original data. Again, such results illustrate mic (loge) curve. Our results differ from other grasslandthat these regression models are appropriate descriptors studies, such as those of Weiser et al. (1986) and Roy etof the bulk of our data. al. (1991) (Fig. 7). In contrast to these studies, our re-

sults show a greater sensitivity of RVI to changes inaboveground live biomass for much of the observedDISCUSSIONrange in both parameters (i.e., for sites where live bio-

Relationships between Vegetation Indices and mass ,95 g m22 and RVI,2.9). Above this threshold,Aboveground Live Biomass our observed relationships are similar to those derived

by Weiser et al. (1986) for sites where the previous year’sOf the many studies that have attempted to estimate bio-senescent material had been removed by burning.physical characteristics from remote sensing data, surpris-

We are unable to provide a physically based explana-ingly few have concentrated their efforts on the predictiontion for a logarithmic relationship between variates, andof aboveground live biomass. Those that have attemptedindeed we know of no other studies where similar resultsto do so (e.g., Weiser et al., 1986) have demonstratedhave been reported. For this reason, we also use linearthat the form of the relationship between parameters isleast squares regression models to describe our originalhighly site- and time-dependent and that a large number

of samples is needed to adequately characterize the rela- (i.e., non-loge-transformed) data (Fig. 7). A comparison


of each logarithmic relationship with its correspondinglinear fit reveals that both statistical models provide sta-tistically similar predictions of Bearly and Blate from spectralinformation. As a result, we speculate that the logarith-mic nature of our observed relationships may simply bean artifact of sampling from within a relatively narrowbiomass gradient and that the “true” biomass-spectral in-dex relationship is actually linear in nature.

Our results also indicate that the strength of eachlogarithmic relationship is spectral index-dependent. Aspreviously noted, variations in aboveground live biomassare best explained by the NDVI, SAVI, MSAVI2, andMSR and are least well explained by the DVI. The in-ability of the “soil-adjusted” indices (SAVI, MSAVI2) toexplain more of the variability in aboveground live bio-mass than some of the “non-soil-adjusted” indices (e.g.,NDVI, MSR) is possibly due to the minimal soil back-ground contributions of each sample plot. Only a fewplots contained bare patches of soil, and in most casesthese patches comprised less than 10% of the total plotarea. This was even true of the most arid plots that weresparsely populated with grass, shrubs, or forbs, becausethese sites were almost completely covered by Selaginelladensa (Dense Club Moss). The similarities between thenon-soil-adjusted and soil-adjusted indices are consistentwith the results of Huete (1988), who found very similarresponses between NDVI and SAVI under conditionstypical of our grassland sites [i.e., light-colored (dry) soiland low leaf area index (,1.5)]. We are unable to explainthe comparatively poor performance of the DVI.

Relationships between Remotely Sensed Bearly/Blate

and %C4

The Effects of Sample ResolutionThe evidence presented indicates that the form andstrength of the relationship between remotely sensedBearly/Blate and %C4 is highly sample resolution-dependentand that this relationship scales nonlinearly. Similar non-linear scaling relationships have also been reported be-tween NDVI and other terrain characteristics, such as el-evation (Walsh et al., 1997; Bian, 1997) and soil depth(Lobo et al., 1998). However, these relationships werefound to be strongest at much coarser resolutions thanthose identified in our study (e.g., .2,200 m for NDVIand elevation) (Bian, 1997).

The weak linear relationships between variates indi-cate that factors other than C4 species cover influenceBearly/Blate at the plot level. These factors likely include thediversity of plant species and plant functional types (i.e.,grasses, shrubs, and forbs) present and the varying com-petitive abilities of these plant forms. The strong nonlin-ear relationship between plant functional diversity andproductivity at the plot level has been well described(Tilman et al., 1997), as have the differing abilities ofT

able

3.E

stim

ated

Slop

es(b

),In

terc

epts

(c),

Coe

ffici

ents

ofD

eter

min

atio

n(r

2 ),an

dR

SEfo

rSi

mpl

eL

inea

rR

egre

ssio

nsof

Rem

otel

ySe

nsed

Bea

rly/B

late

(Squ

are

Roo

tT

rans

form

ed)

on%

C4

(Squ

are

Roo

tT

rans

form

ed)

at0.

5-m

(n5

206)

,2.5

-m(n

552

),10

-m(n

526

),an

d50

-m(n

526

)Sa

mpl

eR

esol

utio

ns

Squa

reR

oot

Tra

nsfo

rmed

(Bea

rly/B

late)

Res

olut

ion

Reg

ress

ion

(m)

Para

met

ers

DV

IR

VI

ND

VI

RD

VI

IPV

IM

SRSA

VI

MSA

VI2

0.5

b,c

22.

3,6.

92

3.9,

8.0

22.

2,6.

72

2.2,

6.7

23.

0,7.

32

2.4,

6.9

22.

2,6.

72

2.0,

6.6

r20.

090.

140.

080.

080.

110.

090.

080.

07R

SE1.

961.

901.

971.

961.

931.

961.

971.

972.

5b,

c2

6.5,

10.2

27.

4,10

.72

5.6,

9.2

25.

9,9.

52

6.5,

9.9

25.

6,9.

42

5.7,

9.3

25.

4,9.

1r2

0.44

0.47

0.35

0.40

0.41

0.37

0.35

0.33

RSE

1.19

1.16

1.29

1.24

1.22

1.27

1.28

1.30

10b,

c2

8.4,

11.6

29.

5,12

.32

8.5,

11.3

28.

4,11

.32

9.1,

11.8

28.

8,11

.52

8.5,

11.3

28.

4,11

.2r2

0.63

0.68

0.60

0.62

0.65

0.62

0.60

0.58

RSE

0.91

0.85

0.94

0.92

0.88

0.92

0.94

0.97

50b,

c2

9.7,

12.8

210

.2,1

2.8

29.

3,11

.92

9.7,

12.4

29.

9,12

.52

9.6,

12.2

29.

4,12

.02

9.2,

11.9

r20.

630.

650.

580.

640.

630.

600.

590.

57R

SE0.

870.

840.

920.

860.

870.

890.

910.

93

All

slop

ean

din

terc

ept

estim

ates

are

high

lysi

gnifi

cant

(p5

0.00

00).

these functional types to compete for limiting resources


Figure 6. Comparisons of the regression-derived slope estimates (and their associated RSE) for our original data and thosederived from Monte Carlo bootstrap resampling at each sampling resolution.

such as water or nitrogen (Golluscio and Sala, 1993). to reduce the variance of both variables while increasingthe covariation between the two.Less well studied, however, are the constraints that spa-

tial structure (pattern) imposes on the above relation- Effects of Vegetation Indexships. The ability of pattern to constrain ecological proc- Although we found the relationship between Bearly/Blateesses has been discussed (Wu and Levin, 1994), and its and %C4 to be partially vegetation index-dependent, theinfluence on plant productivity at multiple scales de- lack of an independent data set limits our ability to iden-serves further attention (Czaran and Bartha, 1989). tify the indices that provide the most accurate absolute

At coarser resolutions (2.5 m, 10 m, 50 m), remotely estimates of %C4 at each sampling resolution. We ac-sensed estimates of Bearly/Blate explain variations in %C4 knowledge this limitation and the need for further de-more fully. r2 values reach their maxima at a sample res- tailed field investigation. However, we believe that ourolution of 10 m, indicating that any further aggregation approach may still provide valuable information regard-does not “improve” the performance of the statistical ing C4 species distributions, even in the absence of fur-model. We explain this trend as follows. Variations in ther field validation. Our results clearly show that irre-Bearly/Blate and %C4 corresponding to processes operating spective of the vegetation index used, remotely sensedat finer scales are filtered out, and like the previously estimates of Bearly/Blate have the potential to provide infor-cited studies of Bian (1997), Walsh et al. (1997), and mation on the relative C4 species coverages of various

sample plots within a given dataset.Lobo et al. (1998), the aggregation of sample plots serves


Figure 7. A comparison of various field-derived relationships between RVI and aboveground live biomass (g m22).We present our results as both logarithmic and linear trend lines.

IMPLICATIONS OF RESULTS, LIMITATIONS used. A number of existing satellite systems, such asOF APPROACH, FUTURE DIRECTIONS, SPOT-P (10 m) and Landsat TM (30 m), operate at com-AND CONCLUSIONS parable spatial resolutions to those where the relationships

between Bearly/Blate and %C4 were found to be strongest.The goal of this study has been to assess the feasibility ofAlthough our nested field sampling scheme did not allowestimating the percentage of ground surface covered bythe scaling of these relations to the spatial resolution ofC4 species using two-date remotely sensed data of vary-sensors that are usually utilized in large-scale assessmentsing spatial resolution. Our approach is based on the con-(e.g., National Oceanic and Atmospheric Administrationditions that various remotely sensed indices can be used(NOAA) Advanced Very High Resolution Radiometerto estimate aboveground live biomass, and that the ratio(AVHRR)), it is possible that the heterogeneity of suchof early season aboveground live biomass to late seasonlarge pixels will mask the spectral-temporal features of theaboveground live biomass (Bearly/Blate) will be negativelyseasonal aboveground biomass curves that are crucial tocorrelated with the areal coverage of C4 species (%C4).the diagnosis of C4 abundance (Goodin and Henebry,This multiresolution approach distinguishes our study1997). Thus, of the commercial satellite information that isfrom others that have attempted to estimate the contri-readily available to environmental scientists, Systeme Pourbutions of C3 and C4 species from remotely sensed sig-l’Observation de la Terre-High Resolution Visible (SPOT-nals (e.g., Goodin and Henebry, 1997) and allows us toHRV) and Landsat TM may be the most suitable databetter assess the operational usefulness of our technique.sources for this monitoring tool.Our results support the existence of a negative linear

It is also clear from our results that this monitoringrelationship between remotely sensed estimates of Bearly/technique produces relatively similar results for all of theBlate and %C4, especially at coarser sample resolutionsvegetation indices used in this study, especially at coarser(i.e., 10 m to 50 m) where these relationships are strong-sample resolutions. While at present we are unable toest. However, we have used fine-resolution field radiom-identify which indices best relate to %C4 at each sam-etry to derive these relationships, and for this technique

to be operationally useful, satellite observations must be pling resolution, our results suggest that computationally


faster indices (e.g., RVI, NDVI) may be utilized in pref- In conclusion, although there are several unresolvederence to those that are more cumbersome (e.g., issues, our results do not contradict our original expecta-MSAVI2), without significant losses or changes in statisti- tion, and they are encouraging for the prospect of moni-cal model performance. It is important to note that these toring C4 species coverage using commercially availableconclusions only apply to the indices listed in Table 1. two-date remotely sensed data (such as Landsat TM).We have not attempted to assess the relative perfor- While the methods presented here have the advantagemances of indices that require the derivation of a soil of simplicity over other techniques, the operational appli-line [e.g., PVI (Richardson and Wiegand, 1977), Trans- cability of our approach depends largely on the improve-formed Soil-Adjusted Vegetation Index (TSAVI) (Baret ment of statistical model performance and the ability toand Guyot, 1991), MSAVI (Qi et al., 1994)] or those that provide a statistical model that is temporally and spatiallyutilize additional wavebands [e.g., GVI (Crist and Ci- robust. We are currently refining our approach in re-cone, 1984)]. sponse to these issues.

One of the most attractive features of our two-dateapproach is its minimal data requirements in comparison We thank Grasslands National Park for their help and coopera-to other techniques. For example, while the discriminant tion with this project. We thank H. Peat, L. Rissling, and M.mixture modeling approach of Goodin and Henebry Wynn for their field assistance. We gratefully acknowledge the

comments and suggestions made by R. Fernandes, F. Gadallah,(1997) required intensive field sampling (every 10 daysD. Jackson, and S. Mitchell, and those of the anonymous re-during the 1995 growing season), our approach relies onviewers. This research was supported by a Natural Sciencesthe acquisition of aboveground biomass information fromand Engineering Research Council (NSERC) of Canada grantonly two sampling dates. These requirements signifi- to F. Csillag.

cantly reduce the amount of data that needs to be col-lected, processed, and analyzed.

However, our approach also has its limitations. Even REFERENCESthough a strong linear relationship between remotelysensed estimates of Bearly/Blate and %C4 has been identi- Asner, G. P. (1998), Biophysical and biochemical sources offied at coarser spatial resolutions, statistical model per- variability in canopy reflectance. Remote Sens. Environ. 64:formance must be improved for this technique to be de- 234–253.

Baret, F., and Guyot, G. (1991), Potentials and limits of vegeta-veloped into a reliable and robust predictive tool that istion indices for LAI and APAR assessment. Remote Sens.equally applicable through time and space (i.e., acrossEnviron. 35:161–173.seasons and other community types or grassland regions).

Barnes, P. W., Tieszen, L. L., and Ode, D. J. (1983), Distribu-Temporal generality might be achieved through thetion, production and diversity of C3- and C4-dominatedidentification of sampling dates where the relationshipcommunities in a mixed prairie. Can. J. Botany 61:741–751.between Bearly/Blate and %C4 is strongest and the charac-

Belleheumer, C., and Legendre, P. (1998), Multiscale sourcesterization of these dates in terms of objective and pheno-of variation in ecological variables: Modeling spatial disper-logically meaningful criteria (e.g., degree day; time aftersion, elaborating sampling designs. Landscape Ecology 13:

green up). This technique could then be tested using 15–25.data from subsequent years and further adapted if neces- Bian, L. (1997), Multiscale nature of spatial data in scaling upsary. Although the tallgrass prairie data (Konza, Kansas, environmental models. In Scale in Remote Sensing and GISUSA) of Goodin and Henebry (1997) suggest that strong (D. A. Quattrochi and M. F. Goodchild, Eds.), CRC Press,relationships between Bearly/Blate and %C4 may also exist Boca Raton, pp. 13–26.in other grassland environments, total geographical gen- Boutton, T. W., and Tieszen, L. L. (1983), Estimation of plant

biomass by spectral reflectance in an east African grassland.erality will be difficult to achieve because seasonal trendsRange Management 36(2):213–221.in productivity (i.e., the timing of Bearly and Blate) are often

Box, E. O., Holben, B. N., and Kalb, H. (1989), Accuracy ofspatially variable. At the macro-scale, such variability isthe AVHRR vegetation index as a predictor of biomass, pri-largely a result of factors relating to latitude (such as daymary productivity and net CO2 flux. Vegetatio 80:71-89.length and temperature), while at local scales, variability

Chen, J. M. (1996), Evaluation of vegetation indices and ais a result of local micro-climatic and topographic condi-modified simple ratio for boreal applications. Can. J. Remotetions. If the metrics used in this study can be linked toSens. 22(3):229–242.phenological stage (see previous discussion on temporal

Crippen, R. E. (1990), Calculating the vegetation index faster.generality), it may be possible to apply our approach to Remote Sens. Environ. 34:71–73.other communities and, with a greater understanding of Crist, E. P., and Cicone, R. C. (1984), Application of the tas-latitudinal grassland productivity patterns, other grass- seled cap concept to simulated Thematic Mapper data. Pho-land biomes. However, additional independent field data togramm. Eng. Remote Sens. 50:343–352.collection from other plant communities at GNP as well Czaran, T., and Bartha, S. (1989), The effect of spatial patternas other geographical regions is needed to address these on community dynamics: A comparison of simulated and

field data. Vegetatio 83:229–239.issues directly.


Dahlman, R. C. (1993), CO2 and plants: Revisited. Vegetatio Richardson, A. J., and Everitt, J. H. (1992), Using spectral veg-etation indices to estimate rangeland productivity. Geocarto104/105:339–355.

Efron, B., and Tibshirani, R. J. (1993), An Introduction to the International 1:63–69.Bootstrap. Monographs on Statistics and Applied Probabil- Richardson, A. J., and Wiegand, C. L. (1977), Distinguishingity, 57, Chapman and Hall, New York. vegetation from soil background information. Photogramm.

Environment Canada (1998), Canadian Climate Normals 1961– Eng. Remote Sens. 43:1541–1552.1990, Swift Current A, Saskatchewan. Roujean, J.-L., and Breon, F. M. (1995), Estimating PAR ab-

Friedl, M. A., Michaelsen, J., Davis, F. W., Walker, H., and sorbed by vegetation from bidirectional reflectance mea-Schimel, D. S. (1994), Estimating grassland biomass and surements. Remote Sens. Environ. 51:375–384.leaf area index using ground and satellite date. Remote Sens. Rouse, J. W., Haas, R. H., Schell, J. A., and Deering, D. W.Environ. 15(7):1401–1420. (1973), Monitoring vegetation systems in the Great Plains

Golluscio, R. A., and Sala, O. E. (1993), Plant functional types with ERTS-1, 3rd Earth Resources Technology Satelliteand ecological strategies in Patagonian forbs. J. Veg. Sci. 4: Symposium, pp. 309–317.839–846. Roy, P. S., Jonna, S., and Pant, D. N. (1991), Evaluation of

Goodin, D. G., and Henebry, G. M. (1997), A technique for grasslands and spectral reflectance relationship to its bio-monitoring ecological disturbance in tallgrass prairie using mass in Kanha National Park (M.P.). Indi. Geocarto Inter-seasonal NDVI trajectories and a discriminant function mix- national 1:39–45.ture model. Remote Sens. Environ. 61:270–278. Schuster, W. S., and Monson, R. K. (1990), An examination of

Hatfield, J. L., Asrar, G., and Kanemasu, E. T. (1984), Inter- the advantages of C3-C4 intermediate photosynthesis incepted photosynthetically active radiation estimated by spec- warm environments. Plant Cell Environ. 13:903–912.tral reflectance. Remote Sens. Environ. 14:65–75. Sellers, P. J., Hall, F. G., Strebel, D. H., Asrar, G., and Mur-

Henebry, G. M. (1993), Detecting change in grasslands using phy, R. E. (1990), Satellite remote sensing and field experi-measures of spatial dependence with Landsat TM data. Re- ments. In Remote Sensing of Biosphere Functioning (R. J.mote Sens. Environ. 46:233–234. Hobbs, and H. A. Mooney, Eds.), Springer-Verlag, New

Huete, A. R. (1988), A soil-adjusted vegetation index (SAVI). York, pp. 169–202.Remote Sens. Environ. 25:295–309. Tieszen, L. L. (1970), Photosynthetic properties of some

Jordan, C. F. (1969), Derivation of Leaf Area Index from qual- grasses in eastern South Dakota. In Proceedings of the Da-ity of light on the forest floor. Ecology 50:663–666. kota Academy of Science, Vol. 49, pp. 78–89.

Lobo, A., Moloney, K., Chic, O., and Chiariello, N. (1998), Tieszen, L. L., Reed, B. C., Bliss, N. B., Wylie, B. K., andAnalysis of fine-scale spatial pattern of a grassland from re-

DeJong, D. D. (1997), NDVI, C3 and C4 production, andmotely-sensed imagery and field collected data. Landscapedistributions in Great Plains grassland land cover classes.Ecology (13):111–131.Ecological Applications 7(1):59–78.Loveridge, D. M., and Potyondi, B. (1983), From Wood Moun-

Tilman, D., Knops, J., Wedin, D., Reich, P., Ritchie, M., andtain to the Whitemud: A Historical Survey of the GrasslandsSiemann, E. (1997), The influence of functional diversityNational Park Area, History and Archaeology 67, Nationaland composition on ecosystem processes. Science 277:Historic Parks and Sites Branch, Parks Canada, Environ-1300–1302.ment Canada, Ottawa.

Walsh, S. J., Moody, A., Allen, T. R., and Brown, D. G. (1997),Michalsky, S. J., and Ellis, R. A. (1994), Vegetation of GrasslandsScale dependence of NDVI and its relationship to moun-National Park, DA Westworth and Associates, Calgary.tainous terrain. In Scale in Remote Sensing and GIS (D. A.Neilson, R. P. (1993), Transient ecotone response to climaticQuattrochi, and M. F. Goodchild, Eds.), CRC Press, Bocachange: Some conceptual and modelling approaches. Eco-Raton, pp. 27–56.logical Applications 3(3):385–395.

Walthall, C. L., and Middleton, E. M. (1992), Assessing spatialParuelo, J. M., and Lauenroth, W. K. (1995), Regional patternsand seasonal variations in grasslands with spectral reflec-of normalized difference vegetation index in North Ameri-tances from a helicopter platform. J. Geophys. Res. 97:can shrublands and grasslands. Ecology 76(6):1888–1898.18905–18912.Peat, H. C. L. (1997), Dynamics of C3 and C4 Productivity

Webster, R. (1979), Nested classification and sampling. Inin Northern Mixed Grass Prairie, MSc Thesis, University ofQuantitative Numerical Methods in Soil Classification Sur-Toronto, Toronto.vey (R. Webster, Ed), Oxford University Press, Bristol, pp.Pickup, G., Chewings, V. H., and Nelson, D. J. (1993), Esti-90–106.mating changes in vegetation cover over time in arid range-

Weiser, R. L., Asrar, G., Miller, G. P., and Kanemasu, E. T.lands using Landsat MSS data. Remote Sens. Environ. 43:(1986), Assessing grassland biophysical characteristics from243–263.spectral measurements. Remote Sens. Environ. 20:141–152.Purevdorj, T., Tateishi, R., Ishiyama, T., and Honda, Y. (1998),

Wu, J., and Levin, S. A. (1994), A spatial patch dynamic model-Relationships between percent vegetation cover and vegeta-ing approach to pattern and process in an annual grassland.tion indices. Int. J. Remote Sens. 19(18):3519–3535.Ecological Monographs 4(64):447–464.Qi, J., Chehbouni, A., Huete, A. R., Kerr, Y. H., and Soroo-

Zar, J. H. (1984), Biostatistical Analysis, Prentice-Hall, Inc.,shian, S. (1994), A modified soil adjusted vegetation index.Remote Sens. Environ. 48:119–126. Englewood Cliffs, NJ.

Documents

The Influence of Vegetation Index and Spatial Resolution on a Two-Date Remote Sensing-Derived Relation to C4 Species Coverage