Geological maps convey different and multifaceted information including lithology, age, tectonism and so on. This complex information is not fully exploited in landslide susceptibility (LS) studies, as a single parameter is usually derived from the geological map of the study area (e.g. the area is divided into lithological or lithostratigraphic or geological units). The aim of this work is testing different approaches to extract significant information from geological maps, creating different parameterizations, and analyzing the sensitivity of a LS model to these variations. Our test site is a 3100 km2 wide area in Tuscany (Italy) characterized by a very complex geological setting. A 1:10000 scale geological map subdivides the area into 194 different lithostratigraphic units. This map was reclassified according to different criteria, creating 6 different parameters derived from the same geological map: lithology (6 lithological classes), age of deposition (the area was subdivided into 6 chronological units), paleogeography (6 units were differentiated on the basis of their environment of formation), genesis of the bedrock (5 classes accounted for the mechanism of formation of the outcropping rock/terrain), broad tectonic domain (the mapped elements were grouped into 5 broad structural units accounting for their tectonic history), detailed tectonic domain (same as before but with a more detailed subdivision into 10 classes). Some of these parameters have already been used in LS studies, others have been used here for the first time; however, all of them have some connections with landslide predisposition. These parameters were used (one by one and altogether) to run seven times a landslide susceptibility model based on the widely used random forest machine learning algorithm. The model configurations and resulting maps were evaluated in terms of AUC(Area Under Curve) and OOBE(out of bag error): while the former expresses the forecasting effectiveness of each configuration, the latter expresses, among a single configuration, the importance of each input parameter. We discovered that the results are very sensitive to the approach used to consider geology in the susceptibility assessment, with AUC values ranging from 63.5% (using chronological units) to 70.0% (using genetic units) and 75.2% (using all the geology-derived parameters simultaneously). These results are in line with OOBE statistics, which showed a similar relative importance of the geologically-driven parameters. These outcomes can to assist future landslide susceptibility studies for different reasons: (i)at least in our study area, lithology, which is commonly used in LS, did not provide the best results; (ii)as geological maps provide multifaceted information, a single classification approach cannot fully grasp this complexity; therefore, the best results can be obtained using different geology-based parameters simultaneously, because each of them can account for specific features connected to landslide predisposition (to our knowledge, a similar approach has never been attempted before in LS literature). (iii)When using thematic maps to feed LS models, it is important to fully understand the nature and the meaning of the information provided by the geology-related maps: results are very sensitive to this kind of information and the interpretation of the results should take it into account.

An attempt to increase the geological information in landslide susceptibility mapping and sensitivity to different geological parameters

Segoni, Samuele;Tamburini, Bimla;Pappafico, Giulio;
2020

Abstract

Geological maps convey different and multifaceted information including lithology, age, tectonism and so on. This complex information is not fully exploited in landslide susceptibility (LS) studies, as a single parameter is usually derived from the geological map of the study area (e.g. the area is divided into lithological or lithostratigraphic or geological units). The aim of this work is testing different approaches to extract significant information from geological maps, creating different parameterizations, and analyzing the sensitivity of a LS model to these variations. Our test site is a 3100 km2 wide area in Tuscany (Italy) characterized by a very complex geological setting. A 1:10000 scale geological map subdivides the area into 194 different lithostratigraphic units. This map was reclassified according to different criteria, creating 6 different parameters derived from the same geological map: lithology (6 lithological classes), age of deposition (the area was subdivided into 6 chronological units), paleogeography (6 units were differentiated on the basis of their environment of formation), genesis of the bedrock (5 classes accounted for the mechanism of formation of the outcropping rock/terrain), broad tectonic domain (the mapped elements were grouped into 5 broad structural units accounting for their tectonic history), detailed tectonic domain (same as before but with a more detailed subdivision into 10 classes). Some of these parameters have already been used in LS studies, others have been used here for the first time; however, all of them have some connections with landslide predisposition. These parameters were used (one by one and altogether) to run seven times a landslide susceptibility model based on the widely used random forest machine learning algorithm. The model configurations and resulting maps were evaluated in terms of AUC(Area Under Curve) and OOBE(out of bag error): while the former expresses the forecasting effectiveness of each configuration, the latter expresses, among a single configuration, the importance of each input parameter. We discovered that the results are very sensitive to the approach used to consider geology in the susceptibility assessment, with AUC values ranging from 63.5% (using chronological units) to 70.0% (using genetic units) and 75.2% (using all the geology-derived parameters simultaneously). These results are in line with OOBE statistics, which showed a similar relative importance of the geologically-driven parameters. These outcomes can to assist future landslide susceptibility studies for different reasons: (i)at least in our study area, lithology, which is commonly used in LS, did not provide the best results; (ii)as geological maps provide multifaceted information, a single classification approach cannot fully grasp this complexity; therefore, the best results can be obtained using different geology-based parameters simultaneously, because each of them can account for specific features connected to landslide predisposition (to our knowledge, a similar approach has never been attempted before in LS literature). (iii)When using thematic maps to feed LS models, it is important to fully understand the nature and the meaning of the information provided by the geology-related maps: results are very sensitive to this kind of information and the interpretation of the results should take it into account.
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11576/2732511
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact