Quantifying Vegetation Change in Semiarid Environments: Precision and Accuracy of Spectral Mixture Analysis and the Normalized Difference Vegetation Index

Because in situ techniques for determining vegetation even when applied to areas of low vegetation cover, the SMA approach correctly determined the sense of change abundance in semiarid regions are labor intensive, they usually are not feasible for regional analyses. Remotely (i.e., positive or negative) in 87% of the samples. SMA results are superior to NDVI, which, although correlated sensed data provide the large spatial scale necessary, but their precision and accuracy in determining vegetation with live cover, is not a quantitative measure and showed the correct sense of change in only 67% of the samples. abundance and its change through time have not been quantitatively determined. In this paper, the precision Elsevier Science Inc., 2000 and accuracy of two techniques, Spectral Mixture Analysis (SMA) and Normalized Difference Vegetation Index (NDVI) applied to Landsat TM data, are assessed quantiINTRODUCTION tatively using high-precision in situ data. In Owens ValRegional measurements of semiarid vegetation abunley, California we have 6 years of continuous field data dance are of great importance for identifying the effects (1991–1996) for 33 sites acquired concurrently with six of climate variability and other natural or anthropogenic cloudless Landsat TM images. The multitemporal reeffects on the environment (Tueller, 1987; Woodwell et motely sensed data were coregistered to within 1 pixel, al., 1984). Field measurements of abundance, while of radiometrically intercalibrated using temporally invarihigh quality, are limited in scope and scale, which limits ant surface features, and geolocated to within 30 m. the feasibility of making regional assessments of change. These procedures facilitated the accurate location of Satellite imagery provides the large spatial and temporal field-monitoring sites within the remotely sensed data. scales necessary to address this fundamental perspective. Formal uncertainties in the registration, radiometric However, there are some basic difficulties in using realignment, and modeling were determined. Results show motely sensed data to study vegetation change. The first that SMA absolute percent live cover (%LC) estimates involves extracting vegetation abundance from measures are accurate to within 64.0%LC and estimates of change of radiance, which is seldom measured in the field by in live cover have a precision of 63.8%LC. Furthermore, vegetation specialists. The two leading methods, Vegetation Indices (VIs) and Spectral Mixture Analysis (SMA), attempt to overcome the inherent difficulties in using ra*Department of Geological Sciences, Brown University, Providiance data to quantify vegetation abundance in compadence, RI rable units to field measures. †Inyo County Water Department, 163 May Street, Bishop, CA ‡Department of Applied Math, Brown University, Providence, RI Previous ecological assessments have utilized vegetaAddress correspondence to A. J. Elmore, Department of Geologition indices such as the Normalized Difference Vegetacal Sciences, Brown University, Providence, RI 02912, USA. E-mail: tion Index (NDVI) to measure vegetation from satellite Andrew [email protected] Received 12 March 1999; revised 21 January 2000. data (e.g., Rouse et al., 1973; Jackson, 1983; Purevdorj