Application of machine learning techniques to the analysis of soil ecological data bases: relationships between habitat features and Collembolan community characteristics

Abstract

We applied novel modelling techniques (neural networks, tree-based models) to relate total abundance and species number of Collembola as well as abundances of dominant species to habitat characteristics and compared their predictive power with simple statistical models (multiple regression, linear regression, land-use-specific means). The data used consisted of soil biological, chemical and physical measurements in soil cores taken at 396 points distributed over a 50×50 m sampling grid in an agricultural landscape in southern Germany. Neural networks appeared to be most efficient in reflecting the nonlinearities of the habitat–Collembola relationships. The underlying functional relations, however, are hidden within the network connections and cannot be analyzed easily. Model trees — next in predictive power to neural networks — are much more transparent and give an explicit picture of the functional relationships. Both modelling approaches perform significantly better than traditional statistical models and decrease the mean absolute error between prediction and observation by about 16–38%. Total carbon content and measurements highly correlated with it (e.g. total nitrogen content, microbial biomass and respiration) were the most important factors influencing the Collembolan community. This is in broad agreement with existing knowledge. Apparent limitations to predicting Collembolan abundance and species number by habitat quality alone are discussed.