Comparing Bivariate and Multivariate Methods in Landslide Sustainability Mapping: A Case Study of Chelchay Watershed

In the last decades, due to human interventions and the effect of natural factors, the occurrence of landslide increased especially in the north of Iran, where the amount of rainfall is suitable for the landslide occurrence. In order to manage and mitigate the damages caused by landslide, the potential landslide-prone areas should be identified.
In landslide susceptibility mapping, using the independent conditioning factors, the probability of the spatial occurrence of landslide in an area is estimated (1, 2). There are different qualitative and quantitative approaches to prepared landslide sustainability maps. Quantitative approaches can be divided into three categories: Statistical, probabilistic and distribution-free methods (3). Statistical methods include bivariate and multivariate methods. In bivarite statistical methods, each individual thematic data layer is crossed with the landslide inventory maps, and the weight values, indicating the importance of each parameter class in the landslide occurrence, are assigned to each factor class (4). In contrast, in multivariate methods, the relative contribution of each conditioning factor to landslide occurrence is calculated (5). Each method of mapping has advantages and disadvantages, and there is no one method accepted universally for the effective assessment of landslide hazards.
In this study, artificial neural network, Logistic regression, frequency ratios, statistical index and Dempster–Shafer methods were used for landslide susceptibility mapping in Chehel Chay watershed in Golestan province. This watershed covers an area of about 256.83 km2 between longitude 36°59′ and 37°13′E and between the latitude 55° 23′ and 55° 38′ N, with the elevation ranging from 179.3 in the northern part to over 2928.3 in the southern part. The mean annual precipitation is 766.5 mm and the dominant land use in this watershed is forest.
The first step in the land-slide susceptibility assessment is mapping the existing landslides. In this study, using air photograph, as in previous studies, Google Earth and field surveys landslide inventory map were constructed. As landslides inventory maps constructed, using geology, topographic and land use maps thematic layers of 12 landslide conditioning factors including slope angle, slope aspect, curvature, profile curvature, plan curvature, altitude, distance from roads, distance from rivers, lithology, distance from faults, land use and topographic wetness index were prepared. To train and validate different methods, the landslide inventory was randomly split into a training dataset of 80% (73 landslide locations), for estimating the artificial neural network and logistic regression parameters and bivariate models weights, and a testing dataset of 20% (18 landslides locations). By translating bivariate methods weights to thematic layers and implementing the artificial neural network and logistic regression to all the study area, pixels landslide sustainability maps were prepared. Additionally, to evaluate landslide susceptibility maps areas under the ROC curve, the percentage of observed test landslide in each landslide susceptibility class and the area of very high susceptibility class were used.
Results showed that the area under the prediction curve for artificial neural network, logistic regression, Dempster–Shafer, frequency ratio and statistical index were 0.86, 0.77, 0.77, 0.72, and 0.71, respectively. Frequency ratio, artificial neural network, Logistic regression, statistical index, and Dempster–Shafer had the least area of very high susceptibility class, respectively. The percentage of landslide pixels coincided with the sites falling in the very high susceptibility classes for Dempster–Shafer, Artificial neural network, Logistic regression, statistical index and frequency ratio, were 0.72, 0.52, 0.32, 0.22, 0.09 respectively. With respect to the area under prediction curve, the percentage of landslide pixels coincided with the sites falling in the very high susceptibility class; multivariate methods including artificial neural network and logistic regression outperformed the other bivariate methods; also Dempster–Shafer had better performance than the other bivariate models. A similar result was obtained by Kavzoglu et al. (2015) and Pradhan and Lee (2010). On the contrary, Ozdemir and Altural (2013), Lee and Pradhan (2007) and Park (2011) concluded that bivariate models had better performance than multivariate methods. Using forward logistic regression, the factors of slope angle, plan curvature, elevation, distance from roads, distance from rivers, lithology and land use were selected as the most important factors. As distance from the road, fault and river increased, the occurrence of landslide and the weight of bivarite methods decreased.
CONCLUSIONS & SUGGESTIONS: In this study, the capability of bivarite and multivariate methods in landslide sustainability mapping in Chel-Chay watershed was evaluated. Results showed that with respect to the area under the prediction curve, the percentage of landslide pixels coincided with the sites falling in the very susceptibility class, and regarding the area of very high susceptibility class, multivariate methods had better performance.