| The accuracy of soil property maps for site-specific management may be inadequate at sampling intensities recommended by commercial agriculture. Since the success of site-specific fertilizer applications depends on the quality of soil property maps, it is critical for Michigan farmers who are adopting these practices to have an understanding of the accuracy associated with different soil sampling strategies, soil sampling intensities, and interpolation techniques. This thesis evaluates how grid sampling and interpolation schemes affected map accuracy based on measures of map error. The second objective was to evaluate how different interpolation techniques that incorporate terrain attributes affects the spatial predictions of soil properties and whether relative performance of these techniques is affected by the scale of soil sampling. In addition to the soil samples used for spatial interpolation, samples were collected to assess the quality of the prediction. For each measured value in the evaluation data set, predicted values were calculated. Spatial predictions were evaluated visually by examining plots of predicted versus measured and quantitatively using a scaled measure of accuracy and precision, the root mean square error. Grid point sampling at the industry standard intensity (100 m regular grid), grid cell sampling (100 m grid cells) and directed sampling based on soil type were not adequate to produce accurate nutrient condition maps for this field even though most of the variables were spatially structured. Prediction errors generally decreased with increasing sampling intensity but soil K, the nutrient with the least spatial structure (46% relative nugget effect), was minimally affected by the scale of measurement. Prediction efficiencies were 0.5 to 10.5% greater for inverse distance weighted interpolation than for kriging using a distance exponent of 1.5 at the 30-m grid sampling intensity. At high sampling densities (30 m regular grid), interpolation methods that utilized terrain attributes had similar prediction errors to interpolation methods that did not utilize terrain attributes. There was an interaction between the scale of measurement and the most appropriate interpolation procedure. At a lower sampling intensity (61 m regular grid), methods that utilized terrain attributes, especially multiple regression, were more accurate than methods that did not. At the 61 m grid, the RMSE for multiple regression was lower (3.3 g kg-1) than the RMSE for ordinary kriging (4.1 g kg-1). Since the cost of grid sampling is inversely related to the square of the grid sampling increment, spatial prediction with terrain attributes is economically appealing. A 100 m grid was not of sufficient intensity to be used to create accurate maps of soil properties for site-specific nutrient management, enhancing spatial estimates with terrain attributes can reduced the number of samples required to create an accurate map. It was not necessary to use complex, time consuming geostatistical techniques (i.e. cokriging, kriging with an external drift, and random field analysis) to use terrain attributes because multiple regression was sufficient. |
