Using the Segment Anything Model to expedite surficial geologic mapping

Cassandra Brigham, Chelsea P. Scott, Samuel Johnstone, Zhiang Chen, Christopher J. Crosby, & Ramon Arrowsmith

Published September 8, 2024, SCEC Contribution #13908, 2024 SCEC Annual Meeting Poster #218

Mapping Quaternary geologic units is an essential component of fault zone mapping. However, detailed manual mapping is time consuming. We propose a workflow that harnesses the expertise of a geologic mapper with the speed and repeatability of a deep learning algorithm to automate portions of the mapping process. We implement Meta AI’s Segment Anything Model (SAM), which segments unfamiliar images with mapper-generated prompts, but no additional training due to its zero-shot generalization capabilities.

Our study focuses on determining the best input datasets for mapping specific surficial geologic units. The input datasets are 1 meter resolution lidar-derived digital terrain models (DTMs) and 4-band 0.5-3 meter resolution satellite imagery. The workflow generates derivative rasters from the input datasets, creates a set of blurred datasets using a Gaussian filter, and a further set of datasets using fuzzy Gaussian membership functions to represent the degree of membership of each geologic unit and increase visual separability between units before segmentation. Prompting the model with points selected within a published map polygon, we compare the output masks to masks generated from published geologic maps, using multiple performance metrics (e.g., Intersection over Union). We ran this code on sites with different types of surficial deposits, including alluvial, colluvial, aeolian, and glacial deposits, in environments that vary in climate, degree of vegetation, relief and geologic history. Studied landforms include earthquake-triggered landslides, alluvial-fan-hosted fault scarps and fault-offset gullies and channels.

Preliminary results suggest that topographic derivatives best replicate existing mapping for channel, moraine and landslide deposits, while imagery derivatives best map alluvial fans and various glacial deposits. These findings are complicated by extensive anthropogenic modification (e.g., urban areas, agricultural fields). Higher levels of detail in the raster do not improve model performance; the Gaussian blurred datasets tend to outperform the original ones, suggesting that the segmentation algorithm over-indexes on detailed features, unless the filter scale is within an order of magnitude of the scale of the mapped deposits. Some features (e.g., river terraces) are poorly mapped by both sets of input data, suggesting that further parameter tuning is needed. These results will be integrated in a SAM-supported surficial mapping model.

Key Words
mapping, deep learning, SAM, Segment Anything Model, Quaternary, surficial

Citation
Brigham, C., Scott, C. P., Johnstone, S., Chen, Z., Crosby, C. J., & Arrowsmith, R. (2024, 09). Using the Segment Anything Model to expedite surficial geologic mapping. Poster Presentation at 2024 SCEC Annual Meeting.

Related Projects & Working Groups
Research Computing (RC)