INTEGRATING MACHINE LEARNING INTO SURFICIAL GEOLOGIC MAPPING: A CASE STUDY FROM THE BIG MARIA MOUNTAINS PIEDMONT, CALIFORNIA

High-resolution remote sensing data is invaluable in geological mapping. With the increased availability of commercial drones, the volume of remotely sensed spectral and topographic datasets is rapidly increasing. Machine Learning (ML) techniques provide a unique opportunity to synthesize large datasets to predict surficial geological map units from remote sensing data alone. Ultimately, these predictions assist geologists by serving as an initial hypothesis in field areas without prior mapping, or by providing an alternative hypothesis to help geologists identify biases or blind spots in their own maps.

Here we explore the benefits and drawbacks of using ML to automatically generate surficial geological maps from remote sensing datasets using a new high-resolution drone survey. We conducted this survey in a 13 km² area on the eastern piedmont of the Big Maria Mountains in SE California, where a suite of Quaternary alluvial fans and river terraces are well preserved. We used structure from motion (SfM) photogrammetry to construct a 2.35 cm per pixel orthorectified image mosaic and digital elevation model (DEM) from 17,996 drone images of the mapping area. We then derived slope, topographic position index (TPI), vector ruggedness (VRM), and relative elevation maps from the original DEM at 10 m, 1 m, 10 cm and full (2.53 cm) per pixel resolutions. Finally, we predicted the number and extents of surface geological units solely from remote sensing datasets using unsupervised hierarchical clustering algorithms (both k-means and gaussian mixture modeling).To understand the effect of data type and resolution on ML predictions, we compared different combinations of datasets as well as different pixel resolutions. We determined the accuracy of ML predictions by comparison to key contacts mapped by USGS and AZGS scientists.

We found that the number and type of datasets influence ML predictions, and that dataset resolution influences contact location. Spectral datasets can be expressed as color band ratios to avoid biases from aspect-dependent shadowing and an a-priori determination of the number of mapping units may be a key piece of information for developing ML algorithms. We conclude that ML is a promising, low-cost approach to supplement geologic mapping projects and may also be a useful tool for mapping extraterrestrial settings.