GIS-Based landslide susceptibility mapping with validation and comparison of spatial prediction models at the basin scale
|Category||Tectonic & Seismotectonic|
|Location||International Geological Congress,oslo 2008|
|Author||Bai, Shibiao۱; Wang, Jian۱; Lu, Guonian۱; Kanevski, Mikhail۲; Pozdnoukhov, Alexei۲|
|Holding Date||23 September 2008|
Landslide susceptibility map is one of the study fields portraying the spatial distribution of future slope failure susceptibility. The purpose of this study is to evaluate and to compare the results of Machine Learning Methods and logistic regression model for basin scale landslide susceptibility mapping based on GIS.
The logistic regression (LR) approach is further elaborated on by crosstabs method, which is used to analyze the relationship between the categorical or binary response variable and one or more continuous or categorical or binary explanatory variables derived from samples. It is an objective assignment of coefficients serving as weights of various factors under considerations while expert opinions make great difference in heuristic approaches. Different from deterministic approach, it is very applicable to regional scale. Nowdays the data-driven models are becoming more and more important. Particularly, machine learning methods provide promising perspectives in the landslide susceptibility mapping, being well-suited to non-linear high dimensional data modeling problems.
This work describes the application of LR and the Support Vector Machines (SVM) for landslide susceptibility mapping and validation in a 1361 km2 study area mainly on the Bailongjiang river basin, in northwest China. The site includes 6 subbasins and the landslide area covers an area of 19.67 km2. For the study area, we prepare a multi-temporal detailed inventory database through the digital inventory map and the interpretation of aerial photographs. Some extensive field studies are used to check the sizes and shapes of landslides.
The study area was subdivided into calibration area (3 subbasins of the study zone, 924 km2 ) and validation area (2 subbasins of the study zone, 437 km2 ).And We partition each basin into geo-morpho-hydrological units, and obtain the probability of spatial occurrence of landslides by logistic regression model and the Support Vector Machines (SVM) of thematic variables, including elevation, slope, aspect, curvature, distance to river, geological formation, distance to fault, distance to road, distance to settlement, land cover, vegetation index and precipitation distribution.For the calibration area after exluded of the highly correlated dependent variables, stepwise logistic regression was carried out in SPSS in order to incorporate only the predicator variables with an important contribution to the presence of landslides. Receiver Operating Characteristic (ROC) curves and the Kappa index were used to validate the model. Both show a good agreement between the observed and predicted values of the validation dataset. The susceptibility maps produced by these two methods were validated and compared. The comparisons of the classified result susceptibility map and validation area landslide contribution were also processed.