The Effects of Spatial Resolution on Digital Soil Attribute Mapping.

Although realistic and useful soil data can be obtained through digital soil mapping, the environmental variables commonly used are scale dependent, and the scale of study needs to be considered for the success of the prediction model. In hierarchical systems, there is often a range of scales over which a process remains stable as scale changes. By changing the raster cell size of digital terrain models, various scales of data can be created and examined to determine the scale which matches the studied process. Five statistical methods were used to examine the scaling behavior of the environmental variables, and models of soil depth were developed at various scales using linear regression, additive regression and regression-kriging. Analysis revealed that altitude and aspect do not change significantly with resolution, and that terrain curvature tends to become more extreme as grid cell size increases. RMS analysis suggested that an optimum for catchment area, MRVBF, PRR, TWI, VRM and the curvature variables may occur in the 20-35 m range. This range was also indicated by the maximum RMS and local variance measures for MRVBF, NDVI, PRR, and VRM variables. Altitude and PRR were the most frequently used variables in the best fitting models, along with GEOL, NDVI and Gaussian curvature. The most precise soil depth prediction model was found to be a GAM at 40 m resolution. Although a direct relationship between model accuracy, environmental variables and spatial resolution was not found, models which included predictors at a resolution which maximizes their local variance and RMS values tended to perform well. It was also shown that the processes affecting soil depth do not display linear scaling behavior, suggesting that it should not be assumed that a simple aggregation of local soil data will provide accurate results at a regional scale.