Exploiting Maximum Entropy method and ASTER data for assessing debris flow and debris slide susceptibility for the Giampilieri catchment (north‐eastern Sicily, Italy)

Abstract

This study aims at evaluating the performance of the Maximum Entropy method in assessing landslide susceptibility, exploiting topographic and multispectral remote sensing predictors. We selected the catchment of the Giampilieri stream, which is located in the north‐eastern sector of Sicily (southern Italy), as test site. On 1 October 2009, a storm rainfall triggered in this area hundreds of debris flow/avalanche phenomena causing extensive economical damage and loss of life. Within this area a presence‐only‐based statistical method was applied to obtain susceptibility models capable of distinguishing future activation sites of debris flow and debris slide, which where the main source of failure mechanisms for flow or avalanche type propagation. The set of predictors used in this experiment comprised primary and secondary topographic attributes, derived by processing a high resolution digital elevation model, CORINE land cover data and a set of vegetation and mineral indices obtained by processing multispectral ASTER images. All the selected data sources are dated before the disaster. A spatially random partition technique was adopted for validation, generating 50 replicates for each of the two considered movement typologies in order to assess accuracy, precision and reliability of the models. The debris slide and debris flow susceptibility models produced high performances with the first type being the best fit. The evaluation of the probability estimates around the mean value for each mapped pixel shows an inverted relation, with the most robust models corresponding to the debris flows. With respect to the role of each predictor within the modelling phase, debris flows appeared to be primarily controlled by topographic attributes whilst the debris slides were better explained by remotely sensed derived indices, particularly by the occurrence of previous wildfires across the slope. The overall excellent performances of the two models suggest promising perspectives for the application of presence‐only methods and remote sensing derived predictors. Copyright © 2016 John Wiley & Sons, Ltd.

Evaluation of Recent Advanced Soft Computing Techniques for Gully Erosion Susceptibility Mapping: A Comparative Study

“…Based on the AUPRC, accuracy, and kappa results, ME was shown to have the best overall performance. This reflects similar findings in [98][99][100][101]. ME exceeds other ML models because it uses search-based optimization to determine the relative importance of factors [101].…”

“…Arabameri et al [88] used three data-mining models for gully-erosion assessment in the Shahroud watershed in northeastern Iran, and found that RF performed best. When there is considerable noise in data, this method is less sensitive to ANNs and can better assess factors compared with others [97,98]. The most important advantages of RF models are their capacities to learn nonlinear relationships, that they have high predictive accuracy, they are able to determine relative factor importance, they can deal with distorted data, and they have high categorization ability.…”

Space-Time Landslide Predictive Modelling

“…Even for this simple twofold division, several approaches are found in the literature with authors arbitrarily setting the probability cutoff. For presence-absence balanced datasets, examples exist where the cutoff is set to 0.5 (Dai and Lee, 2002) without providing an explanation (e.g., Süzen and Kaya, 2012), or because it corresponds to the mean value between the two extremes of the theoretical probability range (e.g., Lombardo et al, 2016a). The approach is problematic, because it sets the cutoff where the model is most uncertain (Rossi et al, 2010a;Reichenbach et al, 2018).…”

“…Concerning data, to inform our models we used covariates that are known to represent geomorphic conditions that favour or hamper the formation of landslides in our study area (Guzzetti et al, 2006a,b;Ardizzone et al, 2007;Galli et al, 2008), in similar geologic, physiographic, and climatic settings (Carrara et al, 1991(Carrara et al, , 2003Carro et al, 2003;Guzzetti, 2005;Marchesini et al, 2014), and even in very different landscapes (Budimir et al, 2015;Goetz et al, 2015;Lombardo et al, 2016a;Reichenbach et al, 2018). With this respect, we maintain that our morphometric, lithological, and structural covariates (Table 2) are sound, accurate, and meaningful landslide predictors, and that they contribute to explain the known spatio-temporal distribution of landslides in our ( Figure 1) and in similar study areas.…”

Using Machine Learning to Predict Geomorphic Disturbance: The Effects of Sample Size, Sample Prevalence, and Sampling Strategy

“…The ML tools used to implement SDMs have been applied to develop geomorphic disturbance models (GDMs) for events including landslides (Convertino et al, ; Yao et al, ), debris flows (Lombardo et al, ), gully erosion (Gutiérrez et al, ; Rahmati et al, ; Svoray et al, ), and phenomena such as patterned ground (Hjort & Marmion, ; Miska & Jan, ) and permafrost (Leverington & Duguay, ). As an example, Dickson and Perry (), using three ML methods, determined that around the Auckland coastline in northern New Zealand, the key predictors of cliff failure at a site were distance to the closest active fault line and unfailed slope angle.…”