Quantifying the Probability of False Alarm for Automatically Detected
Features in InSAR Deformation Maps
Abstract
The expansion in spatial coverage and data volume of Interferometric
Synthetic Aperture Radar (InSAR) is prompting the need for automated
InSAR processing. To be useable by stakeholders, deformation maps
derived from InSAR must come with estimates of reliability. In this
study, we develop a new computer vision algorithm for automatic
detection of surface deformation features in InSAR deformation maps. We
estimate the atmospheric noise power spectrum directly from
interferograms, which we use to generate realistic synthetic noise
instances. This allows us to calculate a likelihood that features in a
real deformation map came from atmospheric artifacts. Because the
procedure only focuses on the probability of false alarm for candidate
features, it does not require any geophysical model for the signals of
interest. Our method is agnostic to the computer vision algorithm used,
and it can be embedded within InSAR processing frameworks to quantify
the uncertainty of machine learning detection results. We demonstrate
our algorithm using 80 Sentinel-1 SAR images covering 80,000
km2 of the Permian Basin in West Texas, where oil and
gas production activities have led to a rise in the number of low
magnitude earthquakes. Our algorithm reliably detects
millimeter-to-centimeter deformation features related with oil and gas
production, groundwater pumping, wastewater injection, and the M5.0
earthquake west of Mentone, Texas. Our method provides guidance on the
minimum number of Sentinel-1 acquisitions needed for interferogram
stacking to confidently detect the subtle deformation. A decrease in
uncertainty can be achieved by detecting and removing SAR images
corrupted by tropospheric noise, which reduces the number of required
acquisitions for mitigating tropospheric noise.