Geostatistics and Spatial Extrapolation Using Vegetation Indices

The combination of the automated modeling and mapping system based on sparse point analysis coupled with robust satellite imagery provides an excellent opportunity to interpolate results to areas not actively monitored. The geostatistical methods of ordinary kriging and co-kriging (Rossi et al., 1994) were explored for this interpolation analysis as a mechanism to make projections across large landscapes without intensive sampling.

In our case, the secondary variable is the NASA 10-day normalized difference vegetation index (NDVI) for continental Africa that provides a spatially rich data set of vegetation greenness across the landscape that has been correlated with plant biomass production (Tucker et al., 1985). A description of the index is provided at the U.S. Geological Survey Africa Data Dissemination Service Web site ( Paper.php).

Forage production estimates from the PHYGROW biophysical simulation model for each of the monitoring sites served as the primary variable in both the kriging and co-kriging analysis. Gridded (8 x 8 km) dekadal (10-day) NDVI was used as the covariate in the co-kriging analysis. The majority of dekads analyzed have exhibited moderate to high correlations between forage production and NDVI (r = .60-.86; Angerer et al., 2002). Cross-validation indicated that the co-kriging analysis generally does a good job of estimating forage production (r2 = .59-.80; standard error = 292-495 kg/ha). Mapped surfaces of the co-kriging output allow us to pinpoint areas of drought vulnerability (figure 22.2). During the periods of high rainfall or extended drought, we have found that the correspondence between forage production and NDVI can be low (r < .30), requiring that ordinary kriging be used for mapping forage production.

Early October 2001

Figure 22.2 Maps of forage deviation from the long-term average that are used to pinpoint areas of drought vulnerability in the Livestock Early Warning System.
0 0

Post a comment