Machine learning ensemble modelling as a tool to improve landslide susceptibility mapping reliability

Description

Statistical landslide susceptibility mapping is a topic in complete and constant evolution, especially since the introduction of machine learning (ML) methods. A new methodological approach is here presented, based on the ensemble of artificial neural network, generalized boosting model and maximum entropy ML algorithms. Such approach has been tested in the Monterosso al Mare area, Cinque Terre National Park (Northern Italy), severely hit by landslides in October 2011, following an extraordinary precipitation event, which caused extensive damage at this World Heritage site. Thirteen predisposing factors were selected and assessed according to the main characteristics of the territory and through variance inflation factor, whilst a database made of 260 landslides was adopted. Four different Ensemble techniques were applied, after the averaging of 300 stand-alone methods, each one providing validation scores such as ROC (receiver operating characteristics)/AUC (area under curve) and true skill statistics (TSS). A further model performance evaluation was achieved by assessing the uncertainty through the computation of the coefficient of variation (CV). Ensemble modelling thus showed improved reliability, testified by the higher scores, by the low values of CV and finally by a general consistency between the four Ensemble models adopted. Therefore, the improved reliability of Ensemble modelling confirms the efficacy and suitability of the proposed approach for decision-makers in land management at local and regional scales.

Additional details