Estimation of Soil Fertility Using Collocated Cokriging by Combining Aerial Hyperspectral Imagery and Soil Sample Data

Abstract

This article focuses on estimation of four soil fertility variables including K, OM, p1, and pH using collocated cokriging by combining soil sample data and aerial hyperspectral imagery. A soil fertility map is a key component in precision farming practices using variable rate technologies. The common approach for soil fertility mapping is univariate spatial interpolation based on soil sample data. Another potential approach is reflectance-based method using remotely sensed images. Both methods use soil sample and image data separately. On the other hand, cokriging estimators provide a way to combine the two data sets for improving soil fertility estimation. Compared to cokriging, collocated cokriging does not require the calculation of a large number of unstable cross variograms between soil fertility variables and spectral variables from image data. In addition, the aerial hyperspectral image provides soil surface reflectance information with fine spatial and spectral details. In this study, one hyperspectral image band was selected for each soil fertility variable based on the correlation between the soil fertility variables and the image bands. For comparison, a traditional regression modeling was also conducted. Pearson product-moment correlation coefficients and root mean square errors (RMSE) between the estimated and observed values were calculated. The results showed that although only moderate correlations between the soil fertility variables and hyperspectral image bands were obtained, the collocated cokriging led to accurate estimates of four soil fertility variables. Compared to the regression modeling, the collocated cokriging increased the correlation by 7% to 105% and decreased the RMSE by 7% to 41% depending on the soil fertility variable.