← Back to Sessions

A Machine Learning Approach for Landslide Mapping of the 2016 Kumamoto Earthquakes From Geospatial and Image Data

Description: 

A series of earthquakes, with 7.3 Mw highest intensity, hit Kumamoto, Japan, over a period of two days in April 2016. The earthquakes caused numerous landslides and surface ruptures in the steep volcanic geological environment. In this study, pre- and post-event sets of high-resolution aerial (Geospatial Information Authority of Japan) and satellite (DigitalGlobe) imagery, paired with the USGS preferred geospatial model for landslide probability and individual geospatial inputs including elevation, surficial geology, slope, precipitation and landcover layer provided by National Mapping Organization of Japan, were used to develop a machine learning (ML) approach for landslide mapping. The ML approach was trained and validated using an inventory of human-drawn landslide occurrence polygons of the area as ground-truth labels provided by Kyoto University’s Disaster Prevention Research Institute and the Japanese National Research Institute for Earth Science and Disaster Resilience. The goal of this work is to improve automated image-based landslide mapping by adding data of physical parameters as well as temporal difference indices calculated from pre- and post-event imagery using ML algorithms. The selection of geospatial effective parameters was done via the supervised Bhattacharyya feature ranking method. The pixel-based ensemble classification algorithm used in this study not only learns from the color channels of the imagery, but also analyzes additional RGB-derived parameters, data of selected geospatial variables, and change information (vegetation and grayscale difference) derived by comparing pre-event and post-event imagery. Different combinations of input parameters were tested, and the results showed that adding data of selected geospatial parameters plus change indices to the imagery lead to the highest classification accuracy. Landslides in these formations are common throughout Japan and have been triggered by heavy precipitation as well as seismic events. Therefore, such data-driven, efficient and fast mapping methods can be effective for landslide prevention, mitigation and reconstruction operations.

Session: Coseismic Ground Failure: Advances in Modeling, Impacts and Communication [Poster]

Type: Poster

Date: 4/20/2023

Presentation Time: 08:00 AM (local time)

Presenting Author: Adel Asadi

Student Presenter: Yes

Invited Presentation: 

Authors