Publication date: February 2014
Source:Geoderma, Volumes 214–215
Author(s): Brandon Heung , Chuck E. Bulmer , Margaret G. Schmidt
In this study, we evaluate the application of a Random Forest (RF) classifier as a tool for understanding and predicting the complex hierarchical relationships between soil parent material and topography using a digital elevation model (DEM) and conventional soil survey maps. Single-component soil polygons from conventional soil survey maps of the Langley–Vancouver Map Area, British Columbia (Canada), were used to generate randomized training points for 9 parent material classes. Each point was intersected with values from 27 topographic indices derived from a 100m try parameter was optimized using multiple replicates of 5-fold cross validation and parent material predictions were made for the region. Predictive parent material maps were validated through comparisons with legacy soil survey maps and 307 field points. Results show that predictions made by a non-optimized RF resulted in a kappa index of 89.6% when validated with legacy soil survey data from single-component polygons and a kappa index of 79.5% when validated with field data. Variable reduction and m try optimization resulted in minimal improvements in RF predictions. Our results demonstrate the effectiveness of RF as a machine learning and data mining approach; however, the need for reliable training data was highlighted by less reliable results for polygon disaggregation in portions of the map where fewer training data points could be established.
Source:Geoderma, Volumes 214–215
Author(s): Brandon Heung , Chuck E. Bulmer , Margaret G. Schmidt