A Slope-Unit (SU) is a type of morphological terrain unit bounded by drainage and divide lines that maximize the within-unit homogeneity and the between-unit heterogeneity across distinct physical and geographical boundaries [1]. Compared to other terrain subdivisions, SU are morphological terrain unit well related to the natural (i.e., geological, geomorphological, hydrological) processes that shape and characterize natural slopes. This makes SU easily recognizable in the field or in topographic base maps, and well suited for environmental and geomorpholog- ical analysis, in particular for landslide susceptibility (LS) modelling. An optimal subdivision of an area into a set of SU depends on multiple factors_ size and complexity of the study area, quality and resolution of the available terrain elevation data, purpose of the terrain subdivision, scale and resolution of the phenomena for which SU are delineated. We use the recently developed r.slopeunits software [2,3] for the automatic, parametric delineation of SU within the open source GRASS GIS based on terrain elevation data and a small number of user-defined parameters. The software provides subdivisions consisting of SU with different shapes and sizes, as a function of the input parameters. In this work, we describe a procedure for the optimal selection of the user parameters through the production of a large number of realizations of the LS model. We tested the software and the optimization procedure in a 2,000 km2 area in Umbria, Central Italy. For LS zonation we adopt a logistic regression model implemented in an well-known software [4,5], using about 50 independent variables. To select the optimal SU partition for LS zonation, we want to define a metric which is able to quantify simultaneously_ (i) slope-unit internal homogeneity (ii) slope-unit external heterogeneity (iii) landslide susceptibility model performance. To this end, we define a comprehensive objective function S, as the product of three normalized objective functions dealing with the points (i)-(ii)-(iii) independently. We use an intra- segment variance function V, the Moran's autocorrelation index I and the AUCROC function R arising from the application of the logistic regression model. Maximization of the objective function S = f(I,V,R) as a function of the r.slopeunits input parameters provides an objective and reproducible way to select the optimal parameter combination for a proper SU subdivision for LS modelling. We further perform an analysis of the statistical significance of the LS models as a function of the r.slopeunits input parameters, focusing on the degree of coarseness of each subdivision. We find that the LRM, when applied to subdivisions with large average SU size, has a very poor statistical significance, resulting in only few (5%, typically lithological) variables being used in the regression due to the large heterogeneity of all variables within each unit, while up to 35% of the variables are used when SU are very small. This behavior was largely expected and provides further evidence that an objective method to select SU size is highly desirable. [1] Guzzetti, F. et al., Geomorphology 31, (1999) 181-216 [2] Alvioli, M. et al., Geoscientific Model Development 9 (2016), 3975-3991 [3] http_//geomorphology.irpi.cnr.it/tools/slope-units [4] Rossi, M. et al., Geomorphology 114, (2010) 129-142 [5] Rossi, M. and Reichenbach, P., Geoscientific Model Development 9 (2016), 3533-3543

A prototype regional early warning system for the prediction of rainfall-induced shallow landslides in Sicily, Italy, is presented. The system is prepared in collaboration with the Regional Civil Protection Department of Sicily. Rainfall thresholds and soil moisture indexes are used for the definition of the predictive tool. In particular, soil moisture indexes derived in a continuous form are used to define a first alert phase. Then, rainfall thresholds, derived as a function of the soil moisture conditions at the beginning of the day, are employed for the subsequent activation of alarm phases. Daily soil moisture indexes, representative of the moisture condition of the catchment, were derived by using a parsimonious and simply to use approach based on the IHACRES model application in a modified form developed by the authors. It is a simple, spatially-lumped rainfall-streamflow model, based on the SCS-CN method and on the unit hydrograph approach that requires, as input data, only rainfall, streamflow and air temperature data (Brigandì and Aronica, 2015). It consists of two modules. In the first a non linear loss model, based on the SCS-CN method, was used to transform total rainfall into effective rainfall. In the second, a linear convolution of effective rainfall was performed using a total unit hydrograph with a configuration of one parallel channel and reservoir, thereby corresponding to 'quick' and 'slow' components of runoff. In the non linear model a wetness/soil moisture index, varying from 0 to 1, has derived to define daily soil moisture catchment conditions. A catalogue of 229 rainfall conditions that have resulted in 265 shallow landslides in Sicily in the period 2002-2012 (Gariano et al., 2015) is used to define rainfall thresholds at different exceeding probabilities and at different soil moisture conditions, for the prediction of slope failures and the activation of the first alert level is based on soil moisture condition analysis. The system can be applied at regional and sub-regional scale (e.g., considering one or more alert zones defined by Civil Protection Department), and can be validated and updated with recent data.