SCEC2024 Poster #096: Developing a machine learning dataset to facilitate fine-detail post-earthquake fault rupture mapping

Group B, Poster #096, Earthquake Geology

Developing a machine learning dataset to facilitate fine-detail post-earthquake fault rupture mapping

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

Poster Image:

Poster Presentation

2024 SCEC Annual Meeting, Poster #096, SCEC Contribution #13741 VIEW PDF

Many earthquake science goals, including understanding dynamic rupture processes and developing probabilistic fault displacement hazards, require data from earthquake fault rupture maps. These maps are typically generated by scientists mapping the rupture on foot, losing valuable information to weathering and erosion and lengthening time before compilation and analysis can occur. While satellite and unpiloted aerial vehicle (UAV) imagery have become increasingly useful data sources over the last decade, rupture maps are still compiled manually. Our goal is to develop a benchmark training dataset consisting of fine-detail post-earthquake annotation on UAV imagery to facilitate machine lear...ning applications in ground rupture and deformation mapping (see accompanying poster by Hu, et al). This dataset is motivated by the understanding that final products could help advance a variety of earthquake science studies. Imagery from earthquakes in California and Türkiye are used to inform the model of multiple surface rupture expressions due to differing geologic settings and land use. To capture a broad range of ground deformation efficiently, we annotate the imagery with both polylines and polygons that define the faults and fault zones, respectively. Annotations are at a scale of approximately 1:50 and across a variety of rupture densities and fault zone widths, providing fine detail while also being able to locate off-fault deformation. The dataset includes annotations at three confidence levels: high, mid, and null. The null confidence level is an important means of informing the model of features such as shadows and plough furrows, which can resemble fractures. Our goal is to have 2000+ unique 512x512 image tiles as a base set which can be increased through augmentation. Future work could include adding annotated imagery from additional locations and/or ruptures through built environments.
SHOW MORE