Validation of digital soil maps at different spatial supports

Abstract

In this study we validated digital soil maps of clay content at different spatial supports; point, 48m blocks and soil–land use complexes (SLU). The aim being to examine the change in prediction quality with different prediction supports. Digital soil maps of clay content at depths of 0–10cm, 10–30cm and 30–50cm were created using linear mixed models, legacy soil data and readily available covariates such as digital terrain attributes, landsat and gamma radiometrics. A random stratified sampling design was used to collect an independent validation dataset. There was little change in the accuracy between the point (RMSE=15.7% for 0–10cm, 17.4% for 10–30cm, 15.2% for 30–50cm) and 48m block supports (RMSE=15.5% for 0–10cm, 17.1% for 10–30cm, 13.5% for 30–50cm). However the prediction quality was much improved at the support of the SLUs (RMSE=9.5% for 0–10cm, 10.1% for 10–30cm and 5.1% for 30–50cm). It is a standard practice in digital soil mapping studies to validate at the point support but our results show that this is likely to represent the worst case scenario for assessing their prediction quality. Further work should consider finding optimal prediction supports for creating digital soil maps.