We present a new digital terrain analysis framework for digital soil mapping, referred to as contextual elevation mapping (ConMap). In contrast to common regression approaches based on features from digital terrain analysis, ConMap is not based on standard terrain attributes, but on elevation differences from the centre pixel to each pixel in circular neighbourhoods only. These differences are used as features in random forest regressions. We applied and validated the framework by predicting topsoil silt content in a loess region of 1150 km2 in Rhineland-Palatinate and Hesse, Germany. Three hundred and forty-two samples and a 20-m resolution digital elevation model were used for this illustration and validation. We compared ConMap with standard and multi-scale terrain analysis approaches as well as with ordinary kriging interpolations. Cross-validation root mean square error (RMSE) decreased from 16.1 when the standard digital terrain analysis was used to 11.2 when ConMap was used. This corresponds to an increase in variance explained (R2) from 15 to 61%. Even though ordinary kriging out-performed standard terrain analysis as well, the variance explained was 6% smaller compared with that using ConMap. The results show that the geomorphic settings in the study area must have induced the spatial trend, which can be accounted for by ConMap over different scales. We conclude that ConMap shows great potential for digital soil mapping studies.
展开▼
