Building pedotransfer functions for estimating soil erodibility in southeastern China

Abstract

Soil erodibility (K) reflects the sensitivity of soil to detachment and transport and is a key factor for estimating the loss of soil. Most of the models for estimating K are complex and experiential, simple and local estimation model in the hilly and mountainous southeastern China is rare. This study aims to build local pedotransfer functions (PTFs) for soil erodibility estimation and evaluate the performance of the built PTFs, i.e. multiple linear regression (MLR), MLR with deformed forms of variables (MLR-DFV) and artificial neutral network with deformed forms of variables (ANN-DFV) to estimating K in southeastern China. The local true K values were obtained by a comprehensive method that considering the optimization prediction model and runoff-plot monitoring data. The best predictive variables were determined using correlation analysis, principal component analysis, importance evaluation and minimum variable-set determination. Mean K in the study area was 0.043 t ha h ha−1 MJ−1 mm−1, ranging from 0.019 to 0.060 t ha h ha−1 MJ−1 mm−1, showed a moderate spatial variability. Soil organic-matter content (SOM) was the most important factor influencing K and accounted for 17.5 % of the total importance. Soil sand content, geometric mean diameter of aggregates, SOM and synthetic curvature were identified as the best predictive variables representing soil physical properties, aggregate characteristics, nutrient and topographical conditions, respectively. The accuracies of MLR-DFV and ANN-DFV were high and similar but higher than the accuracy of MLR. K estimated using ANN-DFV was more similar in magnitude, distribution, and spatial variability to the true K data than K estimated using MLR-DFV. We developed the first local PTFs for estimating K in the hilly and mountainous southeastern China, which could provide empirical basis and method support for studying K in similar regions.