Geostatistical analysis of soil texture fractions on the field scale

Abstract

Geostatistical estimation methods including ordinary kriging (OK), lognormal ordinary kriging (LOK), cokriging (COK), and indicator kriging (IK) are compared for the purposes of prediction and, in particular, uncertainty assessment of the soil texture fractions, i.e. sand, silt, and clay proportions, in an erosion experimental field in Lower Austria. The soil samples were taken on 136 sites, about 30-m apart. The validation technique was cross-validation, and the comparison criteria were the mean bias error (MBE) and root mean squared error (RMSE). Statistical analysis revealed that the sand content is positively skewed, thus persuading us to use LOK for the estimation. COK was also used due to a good negative correlation seen between the texture fractions. The autocorrelation analysis showed that the soil texture fractions in the study area are strongly to moderately correlated in space. Cross-validation indicated that COK is the most accurate method for estimating the silt and clay contents; RMSE equalling to 3.17% and 1.85%, respectively. For the sand content, IK with RMSE (12%) slightly smaller than COK (RMSE = 14%) was the best estimation method. However, COK maps presented the true variability of the soil texture fractions much better than the other approaches, i.e. they achieved the smallest smoothness. Regarding the local uncertainty, the estimation variance maps produced by OK, LOK, and COK methods similarly indicated that the lowest uncertainty occurred near the data locations, and that the highest uncertainty was seen in the areas of sparse sampling. The uncertainty, however, varied much less across the study area compared to conditional variance for IK. The IK conditional variance maps showed, in contrast, some relations to the data values. The estimation uncertainty needs to be evaluated for the incorporation into the risk analysis in the soil management.

A Comparison of Markov Chain Random Field and Ordinary Kriging Methods for Calculating Soil Texture in a Mountainous Watershed, Northwest China

“…Then, this function will be used to calculate the transition probability value in the Equation (3). In this study, the initial (or reference) soil texture type of the prediction point is determined from the existing coarse resolution soil texture map [18]. At last, for a prediction point u j , its predicted soil texture value (composition of sand, silt, and clay) Z * u j , is calculated by Formula (4).…”

“…This is because the OK method is a classical linear method, thus has a smoothing effect with overestimations of small values and underestimations of high values [18,62]. Thus, we would expect the results produced by the OK method to occupy a relatively small range of values (excluding extreme values) [18].…”

“…This is because the OK method is a classical linear method, thus has a smoothing effect with overestimations of small values and underestimations of high values [18,62]. Thus, we would expect the results produced by the OK method to occupy a relatively small range of values (excluding extreme values) [18]. In this study, because silt loam accounts for around 90% of the sampling points, the predicted values are concentrated in the ranges that are associated with silt loam, thus leading to the greater accuracy of the OK method.…”

“…Density of sampling points affects the performance of interpolation methods [18,63]. For example, Delabri et al reported that the performance of OK method, for the silt, the root mean square error was 3.35 for the study area of 0.25 km 2 with 136 sample points [18]; the performance became worse, the root mean square error was 16.09 with the study area increasing to 144 km 2 with 164 sample points [63].…”

“…For example, Delabri et al reported that the performance of OK method, for the silt, the root mean square error was 3.35 for the study area of 0.25 km 2 with 136 sample points [18]; the performance became worse, the root mean square error was 16.09 with the study area increasing to 144 km 2 with 164 sample points [63]. Thus, it appears the density of sample points affects the performance of the OK interpolation.…”