Potential contribution of residuals for better prediction of soil salinity from remote sensing data

Abstract

Soil salinity predictions derived from Ikonos and Landsat satellite images are compared with field-collected soil salinity data for a study area in Colorado's lower Arkansas River Basin. The accuracy of the predictions is compared and issues of price, resolution, and coverage area are considered. Stepwise regression is used to select the combination of bands in the satellite images that best correlate with the field data. The Ordinary Least Squares (OLS) model is used to predict soil salinity using the combination of bands that resulted from the stepwise regression. The residuals for the OLS model are checked for whether they are roughly normal and approximately independently distributed with a mean of 0 and whether there is some constant variance or not. If the residuals do not meet these conditions it means that there is some kind of autocorrelation among them. The SAR model is used to remove some of the autocorrelation among the residuals. If the SAR model does not give satisfactory results, then a modified kriging model is used. The residuals of the OLS model which proved to have autocorrelation can be interpolated using kriging. The final predicted surface results from combining the surface produced from the OLS model with the surface produced by the kriged residuals. The results of this methodology to predict soil salinity from remote sensing data while taking into account the importance of residuals are promising.