← Back to Sessions

Post-Earthquake Liquefaction Mapping by Semi-Supervised Machine Learning Using Partially Labeled Imagery

Description: 

This study explores a new method to identify and map soil liquefaction areas from aerial images after earthquake events. Traditionally, liquefaction is recorded through field visits as geographic points, leading to incomplete data. Comprehensive mapping of affected areas is crucial for developing accurate prediction models. The research introduces a machine learning approach that leverages partially labeled data from expert visual assessments to detect and map unlabeled liquefaction areas. This method aims to increase mapping's spatial accuracy, and it was tested on aerial images from the 2011 Christchurch earthquake, using existing partial data as a reliable ground-truth source. First, the study compares a new semi-supervised learning technique (self-training classification) with traditional supervised learning. This approach is designed to overcome the challenges of limited and spatially incomplete labeled data. Second, it evaluates the effectiveness of various image features (color transformations, statistical indices, and texture components) in improving classification accuracy. Building footprints were also used to refine the results by excluding building roofs from the analysis. Additionally, the study employed the Fuzzy C-Means clustering algorithm to categorize liquefaction samples into two distinct classes (dry and wet). The final results showed that the semi-supervised method, especially when using selected high-ranked features, outperforms traditional methods and provides more complete spatial maps of liquefied areas.

Session: Creating Actionable Earthquake Information Products [Poster Session]

Type: Poster

Date: 5/1/2024

Presentation Time: 08:00 AM (local time)

Presenting Author: Adel

Student Presenter: No

Invited Presentation: 

Authors