Comparison of logistic regression and neural network models in assessing geomorphic control on alluvial fan depositional processes (Calabria, southern Italy)

At the outlet of steep catchments, depositional processes ranging from stream flow to debris flow usually lead to alluvial fan development. Apart from geological and tectonic factors controlling basin sediment availability, several authors highlighted the control role of basin/fan morphometry on fan sedimentary processes. In this framework, the paper was aimed to identify the feeder basin variables that can best differentiate fan processes in Southern Italy. To evaluate the effect of the statistical method on variable selection, we compared logistic regression (LR) and artificial neural network (ANN), the latter not commonly used in fan studies. Alluvial fans were mapped at the mouth of steep V-shaped valleys dissecting the Tyrrhenian coast of northern Calabria, where crystalline-metamorphic and subordinate carbonate rocks crop out. Fans were classified through field survey into two groups: those with (F1), and those without (F0) any debris-flow evidence. Morphometric variables were derived for each basin/fan system. Percentage of lithological units cropping out in the catchments was also considered. Non-parametric statistics revealed that F0 and F1 significantly differ in fan size (area, perimeter and length), main channel slope, lowermost valley width, Melton’s number and geologic index. The relationships between morphometric variables were stronger for F0 than F1. The LR and ANN highlighted the primary control of basin lithology on fan dynamics, followed by basin mean slope. Although ANN outperformed LR in model calibration, both the approaches correctly classified most of the validation samples (> 87%). Alluvial fans with unknown depositional process were classified as belonging to the same group.