Using Classification Trees to Evaluate the Performance of Pedotransfer Functions

Abstract

One issue in understanding the performance of pedotransfer functions (PTFs) is knowing the soil properties that contribute most to PTF errors. Classification trees provide a means for identifying these properties. The objective of this study was to use a classification tree to identify patterns in PTF residuals. The analysis was applied to PTFs developed by Vereecken and coworkers in 1989 to estimate water contents at −10 and at −1500 kPa. Errors below and above certain threshold values were defined as acceptable and unacceptable, respectively. The threshold acceptability values were assumed to be 0.04 and 0.01 cm3 cm−3 at −10 and −1500 kPa, respectively. The sensitivity of the trees to these values was studied by reducing them by 10%. Tentative predictors of the trees were sand, silt, and clay contents, bulk density (BD), soil organic C (OC) content, and horizon position (HP) in the soil profile (topsoil or subsoil). At −10 kPa, unacceptable data sets had very low clay contents or low BDs, given the poor representation of soil structure in the PTFs. At −1500 kPa, the unacceptable data sets had high clay and low OC contents with a HP of subsoil, or had high clay and high OC. Results were found to be moderately sensitive to small variations in the threshold. The least accurate estimates occurred for conditions in which the attributes used in the PTF were probably insufficient to unambiguously determine soil water retention data. Our study showed that classification trees can be helpful in identifying most optimal predictors for a PTF.

Pedotransfer Functions in Earth System Science: Challenges and Perspectives

“…Typically not as a metric to be published with a PTF, but as a diagnostic tool while deriving PTFs, one should evaluate patterns in the estimation residuals and correlations between residuals and input or output properties (Boschi et al, , ). This has helped, for instance, to diagnose and improve models (e.g., Nemes, Rawls, Pachepsky, & Van Genuchten, ; Nemes et al, ) or to help understand sources of errors and differences between models (Nemes et al, ).…”

Toward Developing a Generalizable Pedotransfer Function for Saturated Hydraulic Conductivity Using Transfer Learning and Predictor Selector Algorithm

“…More detailed descriptions of these algorithms can be found in Text . The above ML algorithms are executed in the Weka Version 3.8.3 environment (Boschi et al., 2014; Frank et al., 2004; Hall et al., 2009; Pandey et al., 2016; Ramcharan et al., 2018). While training the ANN model for predicting the K s in Weka, the “multi‐layer perceptron” package was used with the number of the hidden layers equal to ½(number of predictors + class variable) (Yadav & Chandel, 2015; Yadav et al., 2014).…”

Effects of grain-size distribution and hysteresis on soil-water characteristic curve (SWCC)

“…(iii) A split imposed on the node produces children with less than 1 observation; or (iv)The number of nodes exceeds five. Boschi and Rodrigues (2014) recommended that the number of nodes should not exceed five for ease of explanation and implementation.…”

“…The maximum split numbers for these three types of tree are 4, 20, and 100, respectively. As recommended by Boschi and Rodrigues (2014), the number of nodes should not exceed five for ease of explanation and implementation. Hence, only the "Simple Tree" is developed in this study.…”