Search

We describe an application of machine learning to a real-world computer assisted labeling task. Our experimental results expose significant deviations from the IID assumption commonly used in machine learning. These results suggest that the common random split of all data into training and testing can often lead to poor performance., Published as a workshop paper at ICML 2020 Workshop on Real World Experiment Design and Active Learning

Poster presented at the 2018 AGU Fall Meeting.Abstract GPS/GNSS measurements of deformation have high accuracy and temporal resolution but are spatially sparse. Conversely, InSAR provides great spatial resolution but is limited by the satellite look angle, atmospheric noise, and the delay between repeat passes. The sparse GPS data often need to be interpolated on regular grids to be used as constraints during InSAR processing or to calculate strain rates. The interpolation is routinely done separately for each component of the velocity field using minimum curvature or specialized geostatistical algorithms. Recently, a joint interpolation of the horizontal components has been proposed. It estimates forces on a thin elastic sheet that fit the observed data and subsequently uses the estimated model to predict data on regular grids or arbitrary points. The Green’s functions for the physical model serve as a coupling between the two vector components through elasticity theory. We propose an extension of this method to 3D, using the elastic Green’s functions to couple the horizontal and vertical components. This enables the inclusion of vector data projected in arbitrary directions, such as InSAR line-of-sight velocities. The degree of coupling can be controlled through the Poisson’s ratio of the medium. We apply damping regularization to smooth the model and avoid instabilities in the inverse problem. Furthermore, we automatically select optimal values for the Poisson’s ratio and regularization parameter through cross-validation, which is common in machine learning applications. We compare the performance of the coupled model with uncoupled alternatives to grid 2- and 3-component GPS velocities and calculate derivatives through finite-differences approximations. We will present preliminary results from applications to GPS data from the Himalayas and the calibration of InSAR data products. A future goal is to integrate InSAR line-of-sight velocities in a joint interpolation with GPS velocities.

Poster presented at the 2018 AGU Fall Meeting.Abstract GPS/GNSS measurements of deformation have high accuracy and temporal resolution but are spatially sparse. Conversely, InSAR provides great spatial resolution but is limited by the satellite look angle, atmospheric noise, and the delay between repeat passes. The sparse GPS data often need to be interpolated on regular grids to be used as constraints during InSAR processing or to calculate strain rates. The interpolation is routinely done separately for each component of the velocity field using minimum curvature or specialized geostatistical algorithms. Recently, a joint interpolation of the horizontal components has been proposed. It estimates forces on a thin elastic sheet that fit the observed data and subsequently uses the estimated model to predict data on regular grids or arbitrary points. The Green’s functions for the physical model serve as a coupling between the two vector components through elasticity theory. We propose an extension of this method to 3D, using the elastic Green’s functions to couple the horizontal and vertical components. This enables the inclusion of vector data projected in arbitrary directions, such as InSAR line-of-sight velocities. The degree of coupling can be controlled through the Poisson’s ratio of the medium. We apply damping regularization to smooth the model and avoid instabilities in the inverse problem. Furthermore, we automatically select optimal values for the Poisson’s ratio and regularization parameter through cross-validation, which is common in machine learning applications. We compare the performance of the coupled model with uncoupled alternatives to grid 2- and 3-component GPS velocities and calculate derivatives through finite-differences approximations. We will present preliminary results from applications to GPS data from the Himalayas and the calibration of InSAR data products. A future goal is to integrate InSAR line-of-sight velocities in a joint interpolation with GPS velocities.

The Mw 7.2 El Mayor-Cucapah (EMC) earthquake ruptured a complex fault system in northern Baja California that was previously considered inactive. The Cerro Prieto Geothermal Field (CPGF), site of the world's second largest geothermal power plant, is located approximately 15 km to the northeast of the EMC hypocenter. We investigate whether anthropogenic fluid extraction at the CPGF caused a significant perturbation to the stress field in the EMC rupture zone. We use Advanced Land Observing Satellite interferometric synthetic aperture radar data to develop a laterally heterogeneous model of fluid extraction at the CPGF and estimate that this extraction generates positive Coulomb stressing rates of order 15 kPa/yr near the EMC hypocenter, a value which exceeds the local tectonic stressing rate. Although we cannot definitively conclude that production at the CPGF triggered the EMC earthquake, its influence on the local stress field is substantial and should not be neglected in local seismic hazard assessments. Key Points Geothermal energy production causes surface subsidence and crustal stressingProduction at CPGF generates positive Coulomb stresses in EMC rupture zoneAnthropogenic stresses exceed the tectonic loading rate at EMC hypocenter