Assessing gully erosion susceptibility and its conditioning factors in southeastern Brazil using machine learning algorithms and bivariate statistical methods : a regional approach.

Abstract

Despite being a very common phenomenon worldwide, in Brazil, the factors and mechanisms that control gully erosion on a regional scale are still little known, which leads to the neglect of this environmental hazard by territorial and environmental management policies. In order to reducing this gap, we explored the potential of four common supervised machine learning algorithms, named random forest (RF), logistic regression (LR), naïve Bayes (NB) and artificial neural network (ANN) to produce gully erosion susceptibility models for two gullied watersheds, located in the state of Minas Gerais, southeastern Brazil. The modeling was based on the construction of a solid gully inventory and a database consisting of fifteen geo-environmental factors (GEF), whose influence was determined from the information gain ratio (IGR) and two bivariate statistical methods, named frequency ratio (FR) and modified information value (MIV). The predictive performance of the models was evaluated by the area under the receiver operating characteristic curve (AUC), overall accuracy (ACC) and sufficiency analysis. The results revealed that random forest achieved the highest overall performance in correct prediction of gullies and produced the most realistic gully susceptibility maps. The IGR data indicated that all GEF considered in the analysis contributed to the predictive model, although lithology, elevation and rainfall are the most influential variables. From an integrated analysis between the gully inventory, field observations, FR and MIV values, we found that gullies seem to be triggered by high annual average rainfall, but only develop where a set of specific geo-environmental conditions occur simultaneously. Finally, despite the limited land use data available, anthropogenic activities do not seem to affect the regional distribution pattern of gullies, although we have not excluded their local influence in triggering some erosive features.