Identification of Surface Deformation in InSAR Using Machine Learning

The availability and frequency of synthetic aperture radar (SAR) imagery are rapidly increasing. This surge of data presents new opportunities to constrain surface deformation that spans various spatial and temporal scales. This expansion also introduces common challenges associated with large volumes of data, including best practices for analyzing these data. In recent years, machine learning techniques have been at the forefront of big data challenges, as an efficient methodology for automatically classifying large volumes of data. Convolutional Neural Networks (CNNs), in particular, have achieved strong levels of performance on image classification problems. Here we present SarNet, a CNN developed to detect, locate, and classify the presence of co‐seismic‐like surface deformation within an interferogram. We trained SarNet using 4 × 106 synthetic interferograms, including both wrapped and unwrapped forward modeled co‐seismic‐like surface deformation with synthetic noise representative of the atmospheric and topographic noise found in interferograms. The results show that SarNet obtains an overall accuracy of 99.74% on a validation data set. We use class activation maps (CAMs) to show that SarNet returns the location of surface deformation within the interferogram. We employ a transfer learning method to translate the accuracy of SarNet trained on synthetic data to real interferograms with manually classified co‐seismic surface displacement. We train SarNet on 32 interferograms containing labeled co‐seismic surface deformation as well as noise. The results show that, through transfer learning, SarNet obtains an overall accuracy of 85.22% on a real InSAR data set, and that SarNet returns the location of the surface deformation within the interferogram. Plain Language Summary We use machine learning algorithms to classify InSAR interferograms containing surface deformation. We train our algorithm using millions of synthetic interferograms and with modeled surface deformation and noise. Our machine learning algorithm is capable of correctly classifying surface deformation in 99.74% of synthetic interferograms and identifies the location of the surface deformation within the interferogram in these cases. We employ a transfer learning algorithm to transfer the classification capabilities of our algorithm to real interferograms. After transfer learning, our algorithm is capable of correctly classifying surface deformation in 85.22% of real interferograms and id