Advancing deep learning approaches to facilitate rapid post-earthquake fault mapping

Mark Hu, Zhiang Chen, Devin McPhillips, Katherine M. Scharer, Leigh A. Tucker, Alba M. Rodriguez Padilla, & Zachary E. Ross

Submitted September 8, 2024, SCEC Contribution #13738, 2024 SCEC Annual Meeting Poster #094

In post-earthquake response, field geologists fan out across a rupture to map the extent and characteristics of ground deformation. Although lidar and high-resolution orthophotos from Unmanned Aerial Vehicles (UAVs) provide base maps and help to efficiently guide the field survey, a large rupture can require weeks of personnel time to document before any analysis can occur. An automated mapping protocol available immediately following an event could reduce the time necessary to produce a rupture map, an important input for downstream products and provide time for geologists to focus on scientific questions. To facilitate the quick generation of a rupture map, we use machine learning methods to automatically identify and extract characteristics of the rupture and surface deformation. To accomplish this process, deep learning neural networks perform semantic segmentation on high-resolution UAV orthomosaics, capturing the details of the surface fractures. Using a combined dataset that includes data from earthquakes in Türkiye, China, and California (see accompanying poster by Tucker et al.), as well as bedrock data and crack data from CrackSeg9k, we train the UNet architecture model to segment the fractures. (In image processing, segmentation refers to the process of labeling pixels for subsequent analysis.) Results of the model indicate good mapping of the major fault trace on both training, validation, and test datasets. We additionally provide examples of neural network results, post-processed by edge detection, aiming to capture fine-scale features. Future work will extract fracture orientation, density, and width from the model results.

Key Words
machine learning, segmentation, earthquake, fracture, mapping

Citation
Hu, M., Chen, Z., McPhillips, D., Scharer, K. M., Tucker, L. A., Rodriguez Padilla, A. M., & Ross, Z. E. (2024, 09). Advancing deep learning approaches to facilitate rapid post-earthquake fault mapping. Poster Presentation at 2024 SCEC Annual Meeting.

Related Projects & Working Groups
Earthquake Geology