Search

Rockfalls can cause loss of life and material damage. In Northern Morocco, rockfalls and rock avalanche-deposits are frequent, especially in the Dorsale Calcaire morpho-structural unit, which is mostly formed by Jurassic limestone and dolostone formations. In this study, we focus exclusively on its northern segment, conventionally known as "the Haouz subunit". First, a rockfall inventory was conducted. Then, two datasets were prepared: one covering exclusively the source area and the other representing the entirety of the mass movements (source + propagation area). Two algorithms were then used to build rockfall susceptibility models (RSMs). The first one (Logistic Regression: LR) yielded the most unreliable results, where the RSM derived from the source area dataset significantly outperformed the one based on the entirety of the rockfall affected area, despite the lack of significant visual differences between both models. However, the RSMs produced using Artificial Neural Networks (ANNs) were more or less similar in terms of accuracy, despite the source area model being more conservative. This result is unexpected given the fact that previous studies proved the robustness of the LR algorithm and the sensitivity of ANN models. However, we believe that the non-linear correlation between the spatial distribution of the rockfall propagation area and that of the conditioning factors used to compute the models explains why modeling rockfalls in particular differs from other types of landslides. [ABSTRACT FROM AUTHOR]

Rockfalls can cause loss of life and material damage. In Northern Morocco, rockfalls and rock avalanche-deposits are frequent, especially in the Dorsale Calcaire morpho-structural unit, which is mostly formed by Jurassic limestone and dolostone formations. In this study, we focus exclusively on its northern segment, conventionally known as “the Haouz subunit”. First, a rockfall inventory was conducted. Then, two datasets were prepared: one covering exclusively the source area and the other representing the entirety of the mass movements (source + propagation area). Two algorithms were then used to build rockfall susceptibility models (RSMs). The first one (Logistic Regression: LR) yielded the most unreliable results, where the RSM derived from the source area dataset significantly outperformed the one based on the entirety of the rockfall affected area, despite the lack of significant visual differences between both models. However, the RSMs produced using Artificial Neural Networks (ANNs) were more or less similar in terms of accuracy, despite the source area model being more conservative. This result is unexpected given the fact that previous studies proved the robustness of the LR algorithm and the sensitivity of ANN models. However, we believe that the non-linear correlation between the spatial distribution of the rockfall propagation area and that of the conditioning factors used to compute the models explains why modeling rockfalls in particular differs from other types of landslides.