Mapping soil salinity in the Yangtze delta: REML and universal kriging (E-BLUP) revisited

Rice farmers in China need accurate maps of soil salinity to make rational decisions for management. Modern sensors such as the Geonics EM38 conductivity meter, which records the apparent electrical conductivity, ECa, allied to geostatistics to convert sparse punctual measurements into digital maps can provide them.

We have explored the combination in reclaimed land in the Hangzhou Gulf of the Yangtze delta in Zhejiang Province in south-east China. The ECa was measured at 525 points in a 2.2-ha field that was reclaimed in 1996. The data, transformed to logarithms, were treated as the realization of a mixture of strong quadratic trend and correlated random residuals. We estimated the coefficients of the trend and the parameters of the covariance of the residuals by residual maximum likelihood (REML). We then kriged the log10ECa on to a fine grid by universal kriging (UK), and transformed the predictions back to ECa for mapping. For comparison we also include regression kriging using the estimated variogram of the ordinary least squares residuals from the trend.

We compared the results by cross-validation and calculated the mean errors (MEs), mean squared errors (MSEs) and mean squared deviation ratios (MSDRs). All combinations of technique gave small MEs, as expected—kriging is unbiased. The MSEs varied somewhat. The MSDRs, which ideally should equal 1, varied more. The combination with an MSDR closest to 1 was UK with the spherical variogram estimated by REML; its MSDR was 0.993.

We matched the predictions to the US Department of Agriculture’s classes of soil salinity and found that approximately half of the field fell into its slightly saline and moderately saline classes, where rice yields would yield a profit to the farmer, and half were in the very saline and extremely saline classes where rice yields were so poor that the farmer would lose money by attempting to grow rice.