Integrating remote sensing and geostatistics to estimate woody vegetation in an African savanna

Woody vegetation in a tropical savanna was predicted using an approach that integrates geostatistics and remote sensing. Field measurements of woody tree density and canopy cover were correlated with several texture indices and band ratios derived from SPOT5 image data. Next, variogram models were developed: one for the woody vegetation variables; one for the best SPOT5-derived vegetation index; and a cross-variogram between woody variables and SPOT5-derived index. The variogram parameters derived were then used in co-kriging model. The results obtained demonstrated that through co-kriging: (1) tree density interpolation error can be reduced (RMSEcv=19 trees/400m(2), with cross validation R-2=0.74) when compared to ordinary kriging (RMSEcv=28 trees/400m(2), with cross validation R-2=0.47) and stepwise linear regression (RMSE=62 trees/400m(2), with R-2=0.41, calculated on an independent test dataset) and, (2) tree canopy cover interpolation error can be reduced (RMSEcv=13 percent, with cross validation R-2=0.79) as compared to ordinary kriging (RMSEcv=21 percent, with cross validation R-2=0.48) and stepwise linear regression (RMSE=25 percent, with R-2=0.56, calculated on an independent test dataset). This study confirms the importance of considering spatial auto-correlation of regionalised variables, a widely overlooked aspect in remote sensing applications to vegetation studies.