Submitted to: Intnl Conference On Geospatial Information In Agriculture And Forestry
Publication Type: Proceedings
Publication Acceptance Date: February 12, 2002
Publication Date: March 20, 2002
Citation: Timlin, D.J., Walthall, C.L., Pachepsky, Y.A., Dulaney, W.P., Daughtry, C.S. 2002. Spatial regression of crop parameters with airborne spectral imagery. Intnl Conference On Geospatial Information In Agriculture And Forestry. Interpretive Summary: Image tone and color from energy reflectance in remotely sensed images provide useful information on plant leaf area. Mathematical relationships between image tone and color, and plant leaf area have to be calibrated from ground-based measurements of leaf area. The reflectances in remotely sensed data are likely to be spatially correlated, for example if the energy reflectance from one location is high, it is likely that the reflectance from a close neighboring location is also high. This spatial correlation is usually not accounted for when calibrating image color with plant leaf area and as a result the calibration may not be the best possible. The purpose of this work was to show how this spatial dependence could be used to improve the strength of calibrated relationships. We compared two different statistical methods to relate image properties to leaf area. This study showed that the strength of the relationship between image tone and color and leaf area is improved when spatial correlations are accounted for by using a statistical method called spatial autoregression. This improved method will provide a useful tool for researchers to relate color from a remotely sensed image to leaf area or other soil related processes. The results of this benefit specialists who work with remotely sensed data.
Technical Abstract: The spatial information in remotely sensed images is often not fully exploited to improve calibrations of image tone with plant factors. A number of statistical tools that include ordinary least squares regression (OLS) are typically used to develop equations that relate image tone to plant leaf area or biomass. Methods such as OLS ignore the spatial relationships in the data which may result in erroneous conclusions. Further, these methods generally provide a global fit to the data. Methods such as Spatial Autoregression (SAR) that correct for spatial autocorrelation in errors can be more effective in accounting for local information when developing calibration equations. We used data from an airborne image with a 1m X 1m resolution on a 4.5 ha field taken in July, 1998. Leaf area index (LAI) was measured in a corn crop at 74 locations on an irregular grid concurrent with image acquisition. An autoregressive model that uses a 2 dimensional spatial weighting matrix was used to relate the Modified Soil and Vegetation Index (MSAVI) calculated from the image pixels to LAI. MSAVI was averaged over 3, 5, 7 and 11 meter blocks. We compared the AR model to ordinary least squares (OLS) regression that does not account for spatial information. The AR model provided a better fit to the LAI data than the OLS model and increased the r2 from 29 to 34 percent. The best fit was obtained using 5-m block averages.