Analysis of geomorphological parameters for landslide susceptibility in southeastern Brazil

Abstract

In Brazil, shallow landslides are a frequent phenomenon, especially in the Serra do Mar, a mountain range extending for approximately 1,500 km in the South and Southeastern coast. Several studies aimed to analyze the controls of different conditioning factors on landslide susceptibility, mainly on the oriental front of the Serra do Mar. This article aimed to identify the most relevant conditioning factors for landslide susceptibility mapping in the Paraitinga-Paraibuna Highlands (Southeast Brazil), an area intensely affected by landslides and floods triggered by intense rainstorms in summer durring 2009-2010. Initially, a correlation analysis was performed to quantify the association between different conditioning factors (slope, aspect, Topographic Wetness Index, lithology, and land use) and the available landslide inventory. Then, the information value bivariate statistical model and a variable selection procedure, based on the contribution of each factor to the model performance and its capacity of separating unstable and stable classes, were applied. Two susceptibility scenarios were built: one using only the most relevant factor identified using the variable selection procedure (S₂) and another using all available conditioning factors (S₆). Both scenarios were validated using Receiver Operating Characteristic (ROC) curves and compared with Cohen’s kappa. Our results show that the slope gradient and the Topographic Wetness Index (TWI) are the most relevant factors for susceptibility mapping in the Paraitinga-Paraibuna Highlands. The ROC analysis showed that S₆ has a better performance and predictive capacity than of S₂. However, the better results obtained by S₆ are a function of the scarcity of detailed geographical information and do not demonstrate the inadequacy of the variable selection procedure, since Cohen’s kappa showed that exists a better agreement between S₂ and S₆ for areas classified in the very high and very low susceptibility classes. Due to the scarcity of detailed geographical data in most study areas in Brazil, we suggest that the selection of variables should be based on the operator’s knowledge on the statistical model as well as on his knowledge from the geomorphological point of view.