Removing Infrastructure Noise in Airborne Electromagnetic Data with Deep Learning

Abstract:

Noise in airborne electromagnetic data often comes from infrastructure such as powerlines and pipelines. Infrastructure noise needs to be removed before the data can be used as input in inversion algorithms in order to avoid spurious model artifacts. While infrastructure often has human-recognizable characteristics in individual soundings, along flight-line profiles, and/or on auxiliary magnetic or powerline-monitor data channels, it does not always manifest with the same characteristics throughout a survey area, making automated methods for removal difficult.

We present an application of deep learning to automatically remove infrastructure-impacted airborne frequency domain electromagnetic data that is usually done manually and takes an enormous amount of time and labor. Our method is based on deep convolutional neural networks (CNN) and is trained to predict removal of impacted data points for all 12 in-phase and quadrature channels for each sounding independently. We classify each of the 12 channels; the geophysicist can then choose to remove either the impacted channels for the sounding or the entire sounding. We apply this method on a data set acquired in late 2018 – early 2019 over an area of nearly 100,000 square kilometers in the lower Mississippi River Valley as part of the Mississippi Alluvial Plain regional water availability study. The raw data contain more than five million sounding locations. About one million data points were labelled manually. In our experiments, training data set had 144,000 points, and validation and test sets each had 30,000 data points. We train our models using the USGS supercomputer Tallgrass on V100 GPU. We achieve accuracy above 90% on the holdout test data set.