Comparison of statistical prediction methods for estimating field-scale clay content using different combinations of ancillary variables

The need for spatial information on soil properties at the field level is increasing, particularly for its applications in precision agriculture and environmental management. One important soil property is clay content; however, costs involved with obtaining soil data at the field scale are prohibitive. Geostatistical techniques have been used with some success to improve the accuracy of spatial prediction of soil properties, especially those based on easy-to-obtain ancillary information. There is also, however, the need to determine optimal spacing for generating the ancillary data for spatial prediction. In this paper, we used ancillary variables along with spatial prediction models to determine an optimal method for estimating clay content at the field scale. We also determined the optimal spacing for generating the ancillary data for spatial prediction. The ancillary variables used were apparent soil electrical conductivity (ECa) obtained with EM38 and EM31 and digitized hands (red, green, and blue) of aerial photographs of the bare soil. The spatial pre diction models tested are generalized additive models using various combinations of ancillary data (e.g., ECa and red, green, and blue data) and the geostatistical methods of ordinary-, regression- and co-kriging. The results suggest that the linear regression of average clay content with ECa (EM38) data used in combination with kriging of regression residuals was most accurate (RMSE = 3.03). The generation of ECa data on 24-m transect spacing was optimal for prediction. Doubling and tripling the transect spacing (i.e., 48 and 72 m) cause relative reductions in precision of 17% and 12%, respectively.