Water erosion susceptibility mapping by applying Stochastic Gradient Treeboost to the Imera Meridionale River Basin (Sicily, Italy)

• Published in: Geomorphology (Amsterdam, Netherlands) Vol. 262; pp. 61 - 76
• Main Authors: Angileri, Silvia Eleonora, Conoscenti, Christian, Hochschild, Volker, Märker, Michael, Rotigliano, Edoardo, Agnesi, Valerio
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-06-2016
• Subjects: Classification; Erosion; Erosion mechanisms; Freshwater; Gullies; Landforms; Mathematical models; Stochasticity; Water erosion; Italy, Sicilia
• ISSN: 0169-555X; 1872-695X
• DOI: 10.1016/j.geomorph.2016.03.018

Soil erosion by water constitutes a serious problem affecting various countries. In the last few years, a number of studies have adopted statistical approaches for erosion susceptibility zonation. In this study, the Stochastic Gradient Treeboost (SGT) was tested as a multivariate statistical tool for exploring, analyzing and predicting the spatial occurrence of rill–interrill erosion and gully erosion. This technique implements the stochastic gradient boosting algorithm with a tree-based method. The study area is a 9.5km2 river catchment located in central-northern Sicily (Italy), where water erosion processes are prevalent, and affect the agricultural productivity of local communities. In order to model soil erosion by water, the spatial distribution of landforms due to rill–interrill and gully erosion was mapped and 12 environmental variables were selected as predictors. Four calibration and four validation subsets were obtained by randomly extracting sets of negative cases, both for rill–interrill erosion and gully erosion models. The results of validation, based on receiving operating characteristic (ROC) curves, showed excellent to outstanding accuracies of the models, and thus a high prediction skill. Moreover, SGT allowed us to explore the relationships between erosion landforms and predictors. A different suite of predictor variables was found to be important for the two models. Elevation, aspect, landform classification and land-use are the main controlling factors for rill–interrill erosion, whilst the stream power index, plan curvature and the topographic wetness index were the most important independent variables for gullies. Finally, an ROC plot analysis made it possible to define a threshold value to classify cells according to the presence/absence of the two erosion processes. Hence, by heuristically combining the resulting rill–interrill erosion and gully erosion susceptibility maps, an integrated water erosion susceptibility map was created. The adopted method offers the advantages of an objective and repeatable procedure, whose result is useful for local administrators to identify the areas that are most susceptible to water erosion and best allocate resources for soil conservation strategies.
•We explore the ability of Stochastic Gradient Treeboost in predicting the spatial occurrence of soil erosion processes.•The overall accuracy of the susceptibility models is excellent.•The relationships between erosion processes and predictors is analyzed.•We design a methodological approach to create combined erosion susceptibility maps.