Abstract
Earthquake-triggered large-scale landslides are considered one of the most destructive natural hazards to humanlives and infrastructure in many mountain ranges of the world (Hölbling et al. 2012). Information about theexact location of earthquake-triggered landslides is important for post-disaster humanitarian response. Althoughsome field surveying approaches are available, the remoteness of mountainous areas makes it often hard or evenimpossible to reach the affected area (Prasicek et al. 2018). Therefore, Earth observation (EO) data are widelyconsidered as the most accessible data providing up-to-date information needed to support planning and crisisresponses.Since there is a growing demand for landslide susceptibility and risk mapping, several studies have been done forlandslide detection and inventory mapping. Such inventories are an important basis for any subsequent landslidesusceptibility and risk mapping. The semi-automated, exact and fast delineation of large-scale landslides thatwere triggered by an earthquake event and separating them from older landslides that have been caused by othertriggers is a challenging task. Synthetic aperture radar (SAR) data acquired before and after a triggering event canbe helpful for delineation of such landslides, especially since SAR sensors are able to acquire data during dayand night and to penetrate clouds. This can be beneficial for rapid mapping of event landslides, since a limitedavailability of optical imagery might be compensated with SAR data to a certain extent. A surface displacementmap was generated using the differential synthetic aperture radar interferometry (DInSAR) technique andSentinel-1 images. The spectral information of high-resolution RapidEye images was used in addition to enhancelandslide detection.Several machine-learning (ML) methods have been used in different studies for landslide detection. During thepast decade, the deep learning methods and in particular the convolution neural networks (CNN) marked a newepoch in the development of ML methods. In this study, a CNN was designed and different input window sizeswere used for landslide detection. The CNN sample patches were selected within different input window sizesfrom the considered training region. Then the structured CNN was feedforwarded with the prepared samplepatches and tested in the considered testing region. The impact of using different input window sizes on thelandslide detection by the CNN method was evaluated by comparing the results to a manually mapped landslideinventory. The comparison was made using three different metrics, including precision, Recall and F1 measure,to assess the accuracy of the large-scale landslides detected by CNN. The optimal CNN input window size forearthquake-triggered large-scale landslide detection was designated based on the applied accuracy assessmentmetrics
(PDF) Detecting Earthquake-triggered Large-scale Landslides with Different Input Window Sizes Convolutional Neural Networks. Available from: https://www.researchgate.net/publication/332277182_Detecting_Earthquake-triggered_Large-scale_Landslides_with_Different_Input_Window_Sizes_Convolutional_Neural_Networks [accessed Feb 11 2020].
(PDF) Detecting Earthquake-triggered Large-scale Landslides with Different Input Window Sizes Convolutional Neural Networks. Available from: https://www.researchgate.net/publication/332277182_Detecting_Earthquake-triggered_Large-scale_Landslides_with_Different_Input_Window_Sizes_Convolutional_Neural_Networks [accessed Feb 11 2020].
Original language | English |
---|---|
Pages | 1 |
Number of pages | 1 |
Publication status | Published - Apr 2019 |
Externally published | Yes |
Event | EGU General Assembly 2019 - Austria Center Vienna (ACV), Vienna, Austria Duration: 7 Apr 2019 → 12 Apr 2019 https://meetingorganizer.copernicus.org/EGU2019/picos/30465 |
Conference
Conference | EGU General Assembly 2019 |
---|---|
Country | Austria |
City | Vienna |
Period | 7/04/19 → 12/04/19 |
Internet address |