Regionalizing soil properties in a catchment of the Bavarian Alps

Abstract

Regionalizing information on soil hydraulic properties is a fundamental task for solving many soil related ecological problems at the landscape scale. The objective of this study was to derive this information from readily available topographical and geological data for the watershed of the Halbammer (42 km2) in the Bavarian Alps. The database consisted of 297 profiles taken during a 1 month field survey in summer 2001 and the topographical and geological maps (1:25,000) of the area. Soil profiles were grouped into six pedogenetic units taking into account the water regime of the profile and its texture. Another classification was done that relied only on texture and bedrock material as to describe the soil hydrological properties independently from their position in the landscape. In order to regionalize these classes, the tree based CART-Algorithm was used. This algorithm creates optimal classification trees based on a classified training sample and available parameters. The decision rules of these trees then served as basis for the derivation of soil maps that were created using the geographical information system ArcGIS 9.2. The accuracy of the regionalization was estimated using mathematical criteria as well as human expertise. Six to ten units proved to be a good number concerning the available data, but even then not all of them could be separated by CART. The performance of the CART algorithm is measured by the decrease of the misclassification error of an optimally pruned classification tree. In our study, the decrease of the misclassification reached 32 resp. 26% compared to a uniform classification of all data into the most frequent soil unit. The distributions of soil classes in the terminal nodes of the classification trees were useful for the interpretation of the soil units in the maps’ legends and also helped to identify typical soil patterns in the study area. According to these results, CART seems to be a suitable tool for regionalizing soil-related data, but human expertise is still needed to create plausible soil maps out of the information the CART algorithm provides.