Submitted to: Intnl Conference On Geospatial Information In Agriculture And Forestry
Publication Type: Proceedings
Publication Acceptance Date: 9/1/2001
Publication Date: 11/4/2001
Citation: Interpretive Summary: Quantitative maps of the variability of crop foliage density would provide useful information that could be used for precision agriculture management strategies. The variability of spectral vegetation index (SVI) maps calculated from multispectral remote sensing imagery captures the variability of crop foliage density patterns. A generalized least squares regression was tested as a means of defining a foliage density to SVI relationship by accounting for autocorrelation errors. The resulting relationship was found to provide better results than ordinary least squares regression. The spatial information of imagery was further explored by calculating a semivariogram from the imagery and then substituting it for the surface-based semivariogram during kriging of surface-based foliage density measurements. The patterns of the resulting map appeared very similar to a map produced by kriging the surface-based foliage density measurements using the semivariogram calculated from the surface measurements. Further development of these techniques will lead to improved abilities to map foliage density and could provide guidance for surface-based sampling.
Technical Abstract: Accurate maps of within-field crop foliage density would greatly assist crop condition monitoring, yield estimation, and assessment of crop response patterns for management decisions. Spectral Vegetation Indices (SVI), derived as linear combinations of remotely sensed crop spectral reflectances, are correlated with foliage density expressed as leaf area index (LAI). Ground measurements of LAI are traditionally used to calibrat an ordinary least squares (OLS) regression relationship between a SVI and LAI. An OLS equation is then used to convert a SVI image into a LAI map. This approach suffers from correlation of regression errors, and does not fully exploit the inherent spatial information of imagery. A generalized least squares analysis (GLS) was compared to OLS to investigate the presence of error correlation as a function of image pixel size. Use of image spatial information was explored by using the semivariogram of the SVI image to krige the LAI measurements and then compare kriging with the semivariogram from surface-based LAI measurements. Correlation of regression errors was found to be evident at 1 m pixel size, thus making GLS more appropriate at this scale. Substitution of the image semivariogram produced an LAI map comparable to that of the LAI map made solely from the surface measurements.