Application of multiple regression and neural network approaches for landscape-scale assessment of soil microbial biomass

Abstract

Previous soil surveys across the north-east German lowland have reported significant correlations of soil microbial biomass (SMB) contents and organic carbon and total nitrogen contents as well as texture. Using these data sets obtained from 89 arable sites along a regional-scale transect, a linear full-factorial regression model and a neural network model were constructed and evaluated for landscape-scale assessment of SMB. The validation by means of an additional data set consisting of 30 long-term soil observation sites located in the federal state of Brandenburg was within a confidence range of 95%. Using existing models from other regions with our data sets resulted in underestimation of SMB, while using data sets from another region with our model led to overestimation of SMB. It was concluded that a linear full-factorial regression model approach, as well as neural network modelling are promising tools for the prediction of SMB at the landscape scale but need to be validated for the respective region.