Comparing Linear Regression and Regression Trees for Spatial Modelling of Soil Reaction in Dobrovăţ Basin (Eastern Romania)

Abstract

Our study compares the performances of two statistical methods, namely multiple linear regression and classification and regression trees, for deriving spatial models of soil reaction in the surface horizon. The applications were carried out within a 186 km2 hydrographic basin situated in eastern Romania. Statistical models were computed from a sample of 235 soil profiles, scattered in the eastern half of the basin. An independent sample of 237 expeditionary pH measurements was used to validate the results within the interpolation area, whereas an independent sample of 50 soil profiles was used to validate the results within the extrapolation area (the western half of the basin). The predictors included geomorphometrical parameters, derived from a 10x10 m digital elevation model, X and Y coordinates of soil profiles and the main soil types for the regression trees approach. The stepwise selection procedure indicated Y coordinate, digital elevation model, wetness index and surface ratio as the best predictors for soil reaction. The correlation between observed and predicted pH values for the training sample suggests a much higher quality of the regression trees spatial model. However, the validation using the two independent samples points out the instability of this model and recommends the regression model more reliable.

Visitors’ Perception Regarding Floating Treatment Wetlands in an Urban Green Space: Functionality and Emotional Values

“…The CART selects the explanatory variable for which the response variable may be best separated into two groups and identifies the optimum break point. The two resulting groups are further separated into two sub-groups based on another, or the same, explanatory variable [54]. Through such binary recursive partition, the groups, generated in a dichotomous and hierarchical manner, show values for the response variable with the maximum internal homogeneity and the maximum external differentiation [54][55][56].…”

“…The two resulting groups are further separated into two sub-groups based on another, or the same, explanatory variable [54]. Through such binary recursive partition, the groups, generated in a dichotomous and hierarchical manner, show values for the response variable with the maximum internal homogeneity and the maximum external differentiation [54][55][56].…”

A Hybrid Water Balance Machine Learning Model to Estimate Inter-Annual Rainfall-Runoff

“…Then, each of the two resulting groups is further separated into two other subsets. Following this logic, the method generates a tree structure in which the dependent variable is optimally divided into a certain number of groups, which are characterized by maximum internal homogeneity and maximum external differentiation [28]. In modeling, CART uses a set of techniques for structuring data clusters, such as AID and CHAID [29].…”