An assessment on the use of logistic regression and artificial neural networks with different sampling strategies for the preparation of landslide susceptibility maps

Forward logistic regression for earth-flow landslide susceptibility assessment in the Platani river basin (southern Sicily, Italy)

“…At the same time, using LIPs represents a good compromise between adequateness and objectivity: these are, in fact, automatically generated by picking the cells at the top of the depletion zones which are the ones that, according to basic morphodynamic models, typically show site conditions similar to those responsible for past activations (Carrara et al 2008;Nefeslioglu et al 2008;Rotigliano et al 2011). The use of a single point to represent a diagnostic area reduces problems of spatial autocorrelation (Van den Eeckhaut et al 2009), while the automatic generation of the LIPs speeds up the modeling procedure and avoids the subjectivity of the operator.…”

“…In spite of the striking criticism which arises when selecting only a very limited subset of the mapped areas, a number of papers, which exploit logistic regression methods to produce grid cell susceptibility models, optimizes very sophisticated statistic procedures but disregards the real spatial representativeness of the fitted models; in these studies, models are trained solely on very limited part of the mapped basins, which typically stretch for hundreds of square kilometers, without verifying if changes in the random extraction of negative cases result in modifying the selected factors or their regression coefficients (Akgün 2012;Chauan et al 2010;Erener and Düzgün 2010;Mathew et al 2009;Nefeslioglu et al 2008;Ohlmacher and Davis 2003;Süzen and Doyuran 2004). At the same time, some other papers in literature deal with the estimation of robustness in terms of stability of the statistical procedure, disregarding the problem of the geologic representativeness of the subset on which regression is applied (e.g., Carrara et al 2008;Vorpahl et al 2012), using totally boot strapping-based procedures.…”

“…Some researches deal differently with this topic by adopting diagnostic areas that are defined on an analytical basis (Carrara et al 2008;Nefeslioglu et al 2008;Rotigliano et al 2011;Van den Eeckhaut et al 2009). The most adopted diagnostic areas are selected based on landslide typology: scarps, areas uphill from crowns, and landslide area for rotational slides; source areas and landslide areas for flows; and scarps and areas uphill from crowns for falls.…”

“…) Binary logistic regression (BLR) Atkinson and Massari (1998), Ayalew and Yamagishi (2005), Bai et al (2010), Can et al (2005), Carrara et al (2008), Chauan et al (2010), Conforti et al (2012), Dai and Lee (2002), Davis and Ohlmacher (2002), Erener and Düzgün (2010), Mathew et al (2009), Nandi and Shakoor (2009), Nefeslioglu et al (2008, Ohlmacher and Davis (2003), Van den Eckhaut et al (2006 Classification and regression trees (CART) Felicísimo et al (2012), Vorpahl et al (2012) Artificial neuronal networks (ANN) Aleotti and Chowdhury (1999), Ermini et al (2005), Lee et al (2004), Pradhan and Lee (2010) Original Paper exploited to compare the fitting of the model having only the constant term (all the β p are set to 0) with the fitting of the model that includes all the considered predictors with their estimated non-null coefficients so as to verify if the increase in likelihood is significant; in this case, at least one of the p coefficients is to be expected as different from zero (Hosmer and Lemeshow 2000). By exponentiating the β's, odds ratios (OR) for the independent variables are derived: these are measures of association between the independent variables and the outcome of the dependent, and directly express how much more likely (or unlikely) it is for the outcome to be positive (unstable cell) for unit changing of the considered independent variable.…”

“…Logistic regression requires a dataset composed of a near-balanced number of positive (unstable) and negative (stable) cases (Atkinson and Massari 1998;Bai et al 2010;Frattini et al 2010;Nefeslioglu et al 2008;Süzen and Doyuran 2004;Van Den Eeckhaut et al 2009). When using grid cell-based models, which exploit very small cell size so as to increase the spatial resolution of factors and diagnostic areas, positive cases are typically fewer than negatives; in fact, the diagnostic areas include only a few more neighboring cells around the LIPs, while all the stable areas (outside the landslide areas) correspond to hundreds of thousands of stable cells, from which a negative subset is randomly extracted.…”