The importance of input data on landslide susceptibility mapping

Landslide detection and susceptibility mapping are crucial in risk management and urban planning. Constant advance in digital elevation models accuracy and availability, the prospect of automatic landslide detection, together with variable processing techniques, stress the need to assess the effect of differences in input data on the landslide susceptibility maps accuracy. The main goal of this study is to evaluate the influence of variations in input data on landslide susceptibility mapping using a logistic regression approach. We produced 32 models that differ in (1) type of landslide inventory (manual or automatic), (2) spatial resolution of the topographic input data, (3) number of landslide-causing factors, and (4) sampling technique. We showed that models based on automatic landslide inventory present comparable overall prediction accuracy as those produced using manually detected features. We also demonstrated that finer resolution of topographic data leads to more accurate and precise susceptibility models. The impact of the number of landslide-causing factors used for calculations appears to be important for lower resolution data. On the other hand, even the lower number of causative agents results in highly accurate susceptibility maps for the high-resolution topographic data. Our results also suggest that sampling from landslide masses is generally more befitting than sampling from the landslide mass center. We conclude that most of the produced landslide susceptibility models, even though variable, present reasonable overall prediction accuracy, suggesting that the most congruous input data and techniques need to be chosen depending on the data quality and purpose of the study.