An expert-based landslide susceptibility mapping (LSM) module developed for Netcad Architect Software

Highlights

• An expert-based LSM module was developed for Netcad Architect Software.

• Application of M-AHP for landslide susceptibility mapping is first attempted.

• A rule generation algorithm based on M-AHP for fuzzy inference system is proposed.

• Mamdani type FIS completes its task in 3 h and 39 m whereas M-AHP requires 47 s.

Abstract

The main purpose of this study is to introduce an expert-based LSM module developed for Netcad Architect Software. A landslide-prone area located at the eastern Black Sea region of Turkey was selected as the experimental site for this study. The investigations were performed in four stages: (i) introducing technical details of LSM module and theoretical background of the methods implemented in the module, (ii) experiments; landslide susceptibility evaluations by applying the methods M-AHP and Mamdani type FIS by using the expert-based LSM module, (iii) map similarity assessments and evaluations for the generalization capacities of the expert-based models, and (iv) performance assessments of the LSM module. When considering the areal distributions of matching ratios obtained from the map similarity evaluations, it is revealed that M-AHP is more pessimistic and covers a greater area in higher hazard classes, whereas the Mamdani type FIS behaves more optimistically and restricts the area of higher hazard classes in the experimental site. According to the Receiver Operating Characteristics (ROC) curve analyses, the value of Area Under the ROC Curve (AUC) was obtained as 0.66 for the resultant map produced with Mamdani type FIS and 0.82 for the map produced with M-AHP. To compare the time consumptions of the expert methods, experiments were implemented. Mamdani type FIS completes its task in 3 h and 39 min, whereas M-AHP only requires 47 s. As a consequence, (i) the LSM module developed for Netcad Architect Software presents full-featured expert-based landslide susceptibility mapping abilities, and (ii) M-AHP is a useful method for obtaining an expert opinion and modeling landslide susceptibility.

Introduction

Landslide susceptibility mapping constitutes one of the core procedures in quantitative landslide hazard and risk assessments. According to the recommendations published by Corominas et al. (2014), the methods for landslide susceptibility assessments are grouped as follows: knowledge-driven methods, data-driven methods, and physically based methods. Knowledge-driven methods are actually expert-based techniques; to construct a knowledge-driven model, a responsible expert should be included during model construction (Muthu et al., 2008, Akgun et al., 2012, Bourenane et al., 2015, Osna et al., 2014, Zhu et al., 2014, Saponaro et al., 2015). Bivariate and multivariate statistics and data mining methods are the most well-known data-driven techniques in landslide susceptibility mapping (Suzen and Doyuran, 2004, Gokceoglu et al., 2005, Lee, 2005, Duman et al., 2006, Kanungo et al., 2006, Nefeslioglu et al., 2008, Yilmaz, 2009, San, 2014, Polykretis et al., 2015). Physically based methods mainly include geographical information-based limit equilibrium techniques (Baum et al., 2002, Van Beek and Van Asch, 2004, Yilmaz and Keskin, 2009, Nery and Vieira, 2015). The effect of landslide inventory depends on the method applied to infer landslide susceptibility values. The least sensitive techniques are in the category of expert-based approaches; this means that landslide inventory information is not mandatory to construct a predictive model while using knowledge-based methods; this information is particularly essential for model validation. On the other hand, a complete landslide inventory is required in data-driven techniques, and material strength properties and hillslope hydrology information are required in physically based methods. It is possible to note that data space in medium-scaled landslide susceptibility evaluations is commonly incomplete due to various reasons; additionally, the degree of this incompleteness could not be measured. Moreover, if this incomplete part exhibits a random distribution, i.e., landslide information is only acquired from part of the region, the data are known to be skewed, which decreases the prediction and generalization capabilities of landslide susceptibility mapping. Therefore, the data-driven techniques try to model the likelihood of occurrence by using only some of the data space that already exists in the database. In other words, recent landslides can serve as a guide for only similar types of occurrences; however, any circumstance that has not been concluded with a landslide cannot be learned because of the nonexistence of samples belonging to that incident. On the other hand, knowledge-driven/expert-based methods attempt to describe the phenomenon by considering the overall information about the hazard that is already possessed by a responsible expert. For this reason, even though apparent prediction performances of the knowledge-driven methods seem to be lower than those of the data-driven techniques, generalization capabilities of the expert-based methods in landslide susceptibility evaluations could be considered as higher than those of others. Perhaps for this reason, an increase in expert-based landslide susceptibility mapping studies has been observed in the literature in recent years (Akgun et al., 2012, Bui et al., 2012, Demir et al., 2013, Kayastha et al., 2013, Osna et al., 2014, Roodposhti et al., 2014, Zhu et al., 2014, Ahmed, 2015, Saadatkhah et al., 2015, Yang et al., 2015, Althuwaynee et al., 2016, Kumar and Anbalagan, 2016, Myronidis et al., 2016).

However, some difficulties occur in expert-based landslide susceptibility mapping methods. For example, depending on the number of conditioning factors, the number of rules in a fuzzy inference system increases exponentially. Additionally, it may be impossible for responsible experts to determine the consequential parts of fuzzy rules. The application of expert systems such as fuzzy logic in landslide susceptibility mapping was commonly limited to fuzzification of conditioning factors, determination of fuzzy index values, and simple overlying of index maps to evaluate landslide susceptibility. Landslide susceptibility values were only estimated by using a fully enabled fuzzy inference system by applying some additional MATLAB routines in the studies published by Akgun et al. (2012) and Osna et al. (2014). The Analytical Hierarchy Process (AHP) is another expert-based method that is commonly applied in landslide susceptibility mapping. Similar to the dimensionality problem in fuzzy inference systems, depending on the number of conditioning factors, the preparation of factor comparison matrices becomes difficult for responsible experts in landslide susceptibility evaluation. Additionally, depending on the number of conditioning factors, the preparation of comparison matrices for decision points almost turns out to be impossible (Nefeslioglu et al., 2013). Perhaps for this reason, landslide susceptibility is commonly evaluated by using only the weight vectors calculated in AHP in the literature. Additionally, fuzzy AHP extends AHP by using numbers in fuzzy form and all operations are implemented in a fuzzy manner (Chang, 1996). Fuzzy AHP argues that assignment of an exact value to a parameter is difficult for an expert. Critical criticism of fuzzy AHP was given by Zhü (2014) in terms of the possible inconsistencies because of improper use of membership functions. In fact, Zhü (2014) indicates that (i) the arithmetic operation of fuzzy AHP disrupts the basic principles of AHP; (ii) fuzzy AHP cannot give a generally accepted method to rank fuzzy numbers; and (iii) the definition of fuzzy numbers in fuzzy AHP is also improper. The other version of AHP is proposed as the Modified Analytical Hierarchy Process (M-AHP) by Nefeslioglu et al. (2013) to compensate for expert subjectivity encountered in factor comparisons. In other words, the responsibilities of an expert in M-AHP are the determination of the total scores of factors and instant assignments on the basis of relevant cases; hence, the rest of the procedure continues without expert intervention. Therefore, this version of AHP becomes more appropriate for LSM than its original form because comparison matrices can be prepared automatically without expert subjectivity. Considering the limitations given thus far, it can be realized that there is an obvious need for a software to perform full-featured expert-based landslide susceptibility mapping.

The purpose of this study is to introduce a landslide susceptibility mapping (LSM) module for Netcad Architect Software that implements M-AHP and a Mamdani type Fuzzy Inference System (FIS) as the expert methods for landslide susceptibility evaluation. Application of M-AHP for this problem is first attempted in this study. Additionally, a semi-automated rule generation algorithm based on M-AHP, which allows practical application of the fuzzy inference system in landslide susceptibility assessment, is also suggested. A landslide-prone area located at the eastern Black Sea region of Turkey was selected to be the experimental site for this study. The investigations were performed in four stages: (i) introducing technical details of LSM module and theoretical background of the methods implemented in the module, (ii) experiments; landslide susceptibility evaluations by applying the methods of M-AHP and Mamdani type FIS by using the expert-based LSM module, (iii) map similarity assessments and evaluations for the generalization capacities of the expert-based models, and (iv) performance assessments of the LSM module.