Artificial neural networks and decision tree classification for predicting soil drainage classes in Denmark

Abstract

Soil drainage constitutes a substantial factor affecting plant growth and various biophysical processes, such as nutrient cycling and greenhouse gas fluxes. Consequently, soil drainage maps represent crucial tools for crop, forest and environmental management purposes. As extensive field surveys are time- and resource-consuming, alternative spatial modelling techniques have been previously applied for predicting soil drainage classes. The present study assessed the use of Artificial Neural Networks (ANN) for mapping soil drainage classes in Denmark and compared it to a Decision Tree Classification (DTC) technique. 1702 soil observations and 31 environmental variables, including soil and terrain parameters, and spectral indices derived from satellite images, were utilized as input data. Based on a 33% holdback validation dataset, the best performing ANN and DTC models yielded overall accuracy values of 54 and 52%, respectively. DTC models benefited from the use of all variables, but ANN models performed better after variable selection. Notably, ANN and DTC model performances were comparable although differential costs for misclassification were only implemented for DTC modelling. Nevertheless, both methods produced predictive drainage maps in accordance with one another and demonstrated promising classification abilities over a large study area (c. 43,000km2).