Deep Transfer Learning Framework for Regional Landslide Mapping Using Post-Event Imagery
Description:
Landslides, often occurring in mountainous regions and triggered by earthquakes or heavy rainfall, are a major natural disaster. Traditionally, identifying landslides involves manually analyzing optical remote sensing imagery, a process that is both slow and labor-intensive. This study proposes an automatic landslide detection method using advanced deep learning techniques. Previous research in this area has focused on using RGB channels from both before and after the event, along with other data sources like digital elevation models, for change detection. However, the effectiveness of deep learning in this context depends heavily on the availability of high-quality, large datasets, which can be a challenge in the immediate aftermath of a landslide. This study specifically explores the transferability of a deep learning model trained on data from the 2016 Kumamoto Earthquakes. It tests the model's effectiveness in detecting landslides caused by different events: the 2018 Hokkaido earthquake and the 2017 Asakura Rainfall, both in Japan. These events were chosen for their regional similarities, including terrain, vegetation, urban areas, geology, and landcover. The proposed deep convolutional neural network model uses a ResNet50-based DeepLabV3+ framework, which facilitates automated landslide feature extraction from post-event images. This approach does not require fine-tuning for each specific event or the development of unique training data sets. The model demonstrated high accuracy in identifying landslides from the 2016 Kumamoto Earthquakes and performed well in the two unseen events. The impacts of different tile sizes, number of tiles, imagery resolution, and data augmentation techniques are investigated to provide instructions on best practices.
Session: Detecting, Characterizing and Monitoring Mass Movements [Poster Session]
Type: Poster
Date: 5/2/2024
Presentation Time: 08:00 AM (local time)
Presenting Author: Adel
Student Presenter: No
Invited Presentation:
Authors
Adel Asadi Presenting Author Corresponding Author adel.asadi@tufts.edu Tufts University |
Laurie Baise laurie.baise@tufts.edu Tufts University |
Snehamoy Chatterjee schatte1@mtu.edu Michigan Technological University |
Magaly Koch mkoch@bu.edu Boston University |
Babak Moaveni babak.moaveni@tufts.edu Tufts University |
|
|
|
|
Deep Transfer Learning Framework for Regional Landslide Mapping Using Post-Event Imagery
Session
Detecting, Characterizing and Monitoring Mass Movements