Spatial modeling of soil salinity using remote sensing, GIS, and field data

Abstract

In this study a new methodology was developed to generate accurate predicted soil salinity maps using remote sensing data. The techniques used include integrating field data, geographic information systems (GIS), remote sensing, and spatial modelling techniques. Com and alfalfa crops were selected as indicators of soil salinity during 2001 and 2004 respectively. Five images were acquired from Aster, Ikonos, and Landsat to check the correlation between measured soil salinity and remote sensing data. Observed data from four com fields during 2001 and four alfalfa fields during 2004 were used in conjunction with the Aster, Ikonos, and Landsat images. Three subsets of 75%, 50%, and 25% were randomly selected from each main set of observed data to be used in conjunction with the Ikonos and Landsat images. Three models were applied to predict soil salinity from remote sensing: the ordinary least squares model (OLS), spatial autoregressive model (SAR), and modified kriging model. The combination of satellite imagery bands that had the best correlation with measured soil salinity was used to predict soil salinity. A number of criteria were used to select the best model. The results show that the modified kriging model provides the best results over the OLS and the SAR models. The OLS model meets the model selection criteria, but, in most cases, it involves some autocorrelation among the residuals. The SAR model was able to remove some of the autocorrelation among the residuals, but the R2 was reduced. The R2 values of the OLS model were 0.34, 0.47, 0.52, 0.26, and 0.37 for the 2001 Aster, Landsat, Ikonos images for com, the 2004 Landsat and Ikonos image for alfalfa respectively. The R2 values of the SAR model were 0.05, 0.18, 0.25, 0.03, and 0.15 for the same images. The R2 values of the modified kriging model were 0.81, 0.83, 0.91, 0.60 and 0.68 for the same images. Also, the mean absolute error (MAE) improved significantly when using modified kriging over the OLS and SAR models for all data sets. When the modified kriging model was applied to the subsets of data it showed encouraging results as well.