Soil salinity mapping using spatio-temporal kriging and Bayesian maximum entropy with interval soft data

Abstract

We used geostatistical tools to identify the spatio-temporal variability of soil salinity with both field and laboratory measurements. This analysis used kriging and Bayesian maximum entropy (BME) to predict soil salinity at unobserved spatial locations and time instants. We compared the accuracy of the mapped predictions from BME using soft interval data, kriging with either hard and soft data (HSK), or hard data only (HK). A large spatio-temporal database on soil salinity data was available. It consists of 413 sites where the apparent or bulk soil electrical conductivity (EC a) was measured with electrical probes over an area of 25 ha. These measurements are our ‘data set to be calibrated’. On a limited subset of these sampling sites (13–20), electrical conductivity was determined by laboratory analysis from 1 : 2.5 soil–water suspensions (EC 2.5), which is a simple representation of the electrical conductivity of the water-saturated soil-paste extract (EC e). These are our ‘calibration data set’. The whole procedure was repeated 19 times between November 1994 and June 2001. The methods of prediction were compared quantitatively by mean error (ME) and mean squared error (MSE). The errors are the differences between the measured electrical conductivity in the laboratory on samples from March and June 2001 (which were not used in previous computations) and their cross-validation estimates. The MSE was divided further into three components revealing different aspects of the discrepancy between the observed and the estimated values of electrical conductivity. The BME predictions were less biased and more accurate than those from kriging. The MSE decomposition showed that the kriging with soft data (HSK) provided more biased estimates and failed to reproduce the magnitude of fluctuation in the observed soil salinity. The difference in incorporating soft data into the analysis was confirmed and was more acute when only the largest intervals from the soft data were used. In this situation BME produced very reliable estimates whereas HSK failed to predict soil salinity accurately.