Deep Learning Based Potential Landslide Detection

Keywords: Potential Landslide, Deep Learning, Open Access Data
Session Type: Virtual Paper Abstract
Day: Monday
Session Start / End Time: 2/28/2022 03:40 PM (Eastern Time (US & Canada)) - 2/28/2022 05:00 PM (Eastern Time (US & Canada))
Room: Virtual 17

Abstract

Landslide is a type of widespread geological disaster that risks the safety of people's lives, properties, and decreases the resilience of the affected local community. The detection of potential landslides at their early stage plays an important role in landslide disaster mitigation and management. Traditional potential landslide-related studies like landslide susceptibility mapping (LSM) consider mostly topographical factors of potential landslides that have a very coarse spatial resolution and are difficult to update. Thus, traditional LSM can only produce a vague result of approximate regions that are prone to landslides, failing to detect individual potential landslides, especially the small-scaled landslides. The development of deep learning technologies brings opportunities for more precise detection of individual potential landslides by taking into consideration multi-source input of high-resolution remote sensing imagery along with topography factors and ground deformation. A designated convolutional neural network structure that can efficiently and accurately extract information from multi-source input is proposed in this study. The remotely sensed data used in this study is from freely accessible data sources like Google Earth Image, Sentinal-1, and ASTER-GDEM, enabling the quick and easy update of data that requires minimal cost. The proposed method was demonstrated in the strong earthquake zone in Sichuan, China. However, the low cost (both computational and data acquisition) and high-performance characteristics of the proposed method make it suitable to be applied worldwide, enhancing the disaster mitigation ability globally, especially for the less developed regions with limited resources.