A Comparative Study of Deep Learning and Conventional Neural Network for Evaluating Landslide Susceptibility Using Landslide Initiation Zones

Considerable efforts have achieved to comprehend where seismically triggered landslides may occur because they are a disastrous hazard with an extraordinarily risk component in tectonically active mountainous areas. The objectives of this research are to investigate and compare two advanced artificial intelligence models (AI), i.e., artificial neural network (ANN) and deep learning (DL) techniques to evaluate susceptible zones using landslide initiation zone polygons (i.e., scarp areas). For this, a comprehensive landslide inventory map comprising of the representative of the landslide scarp, which is constructed using high resolution aerial photographs and Lidar digital elevation models (Lidar DEM) for the 2018 Hokkaido earthquake-affected sites. Afterward, 11 causative factors were prepared, including seismic, topographic, and hydrological factors. Our results show that DL has better predictive performance than the traditional ANN obtained model. Furthermore, the importance of factor ranks indicates that topography has played the leading role in the landslide occurrences. The DL model shows a promising way for rapid response in the field of landslide hazard mitigation.