Landslide susceptibility mapping on the country scale with data mining techniques in Armenia

Armenia is a country strongly affected by landslides, but still not enough research has been done on landslide susceptibility in this area. Therefore, the main goal of the study was to create one of the first landslide susceptibility zonation maps covering the entire territory of Armenia. Two different data mining techniques for landslide susceptibility analysis were applied: artificial neural networks (ANN) and C5.0 decision trees (C5.0). Created landslide susceptibility maps showed that the C5.0 models are rather not applicable for susceptibility analysis due to strong overfitting and artifacts caused by pruning. The ANN models provided significantly better results than C5.0 models. Though, they still need to be improved. Another goal of the study was to investigate the influence of landslide inventory on susceptibility analysis. For that reason, two different landslide catalogs were used for the analysis: the GEORISK inventory (2004) based on aerial photographs and field surveys provided by GEORISK Scientific Research Company and MATOSSIAN inventory (2017) based on satellite imagery created by Matossian (2017). Susceptibility zonation maps created using ANN models based on MATOSSIAN catalog showed better performance than models based on GEORISK catalog. This observation was rather unexpected, because MATOSSIAN catalog contains only half as much landslides as GEORISK inventory. However, the main difference between these two inventories is the absence of landslides in a large volcanic region (the Aragats region) in MATOSSIAN inventory. In this regard, it was suggested that this area might negatively influence the results and should be excluded for further investigations. References: Matossian, A. (2017). Identification of giant mass movements in the Lesser Caucasus and assessment of their spatial relationship with major fault zones and volcanoes. Master’s thesis. University of Liège.