Finding Active Faults in LiDAR Terrain

Overview

Reading the shape of the ground

When a fault moves over thousands of years it leaves subtle marks on the landscape: scarps, offset stream channels, and aligned notches along ridges. Bare earth LiDAR strips away the vegetation and reveals these landforms in fine detail.

This project trains a model to recognize those landforms automatically. The model reads images made from LiDAR elevation, mainly a shaded relief view and a slope map, and labels every pixel as fault or not fault. It learns purely by example from the faults already mapped in the USGS Quaternary Fault and Fold Database. The goal is a first pass map of likely active faults across large areas, far faster than a geologist can review by hand. Because a fault is only a thin line, it covers a very small share of the pixels in any image, so the real challenge is finding that rare signal without mislabeling ordinary ground.

Pipeline

From raw LiDAR to a trained model

01 / SELECT

Choose areas by fault density

A program scans the whole EarthScope Southern and Eastern California LiDAR coverage block by block, measures how much mapped fault length sits inside each block, and keeps the densest. This replaces choosing study areas by hand.

02 / TRAIN

Learn from mapped faults

For each area, a separate model learns from shaded relief and slope, in tiles a few hundred meters across, using the USGS mapped faults as examples. Each model is then tested on parts of its area it never saw during training.

03 / CHECK

Score against the database

Predictions are compared to the USGS faults. Because mapped fault lines are themselves only approximate, exact pixel overlap is a harsh measure, so scoring with a distance tolerance is the next step.

Results

Accuracy follows the terrain

Each area was scored on a test set the model never saw. On a zero to one overlap scale the scores were modest, and they tracked how strongly faulting has shaped the topography rather than anything about the model.

Study area	Slip style	Test IoU	Recall	Precision
Owens ValleyBig Pine	right lateral	0.29	0.44	0.47
Panamint ValleyPanamint Range front	normal	0.16	0.36	0.22
Quail Mountainsnear Death Valley	left and right lateral	0.09	0.14	0.17
Cantilnear the Sierra Nevada	left lateral	0.07	0.10	0.24

Bars scaled to the strongest area. These are strict overlap scores, which understate a thin line prediction that is only slightly offset.

Regions

What the model sees, side by side

For each area, the left panel is the setting and the right panel shows the model prediction over the shaded relief.

Owens Valley

Big Pine · strongest result

Owens Valley prediction over hillshade — Prediction

USGS faults Model prediction IoU 0.29

Panamint Valley

steep range front, over predicts

Panamint Valley prediction over hillshade — Prediction

USGS faults Model prediction IoU 0.16

Quail Mountains

near Death Valley

Quail Mountains prediction over hillshade — Prediction

USGS faults Model prediction IoU 0.09

Cantil

near the Sierra Nevada

Cantil prediction over hillshade — Prediction

USGS faults Model prediction IoU 0.07

Findings

Five things the data showed

Terrain decides, not the model

The same model scored 0.29 in Owens Valley and 0.07 at Cantil. Detection follows how clearly faulting has marked the land, and recall follows with it.

More input layers did not help

Five terrain channels scored about the same as two on Panamint (0.16 versus 0.14). The limit is the landscape, not the inputs.

Very steep ground backfires

The steepest area over predicts, marking ordinary slope lines as faults, so its precision was only 0.22.

Validation is optimistic

Held back test ground scored well below validation. The Quail Mountains fell from a validation 0.23 to a test 0.09.

Data preparation mattered more than the model

Fixing a missing data error in how the slope image was built, on its own, raised the Owens Valley test score from 0.08 to 0.29.