AI with PS-InSAR

With the development of InSAR technology, remote sensing satellites with high resolution and different bands, such as Terra-SAR-X, Sentinel-1, and ALOS-2, have brought massive amounts of data to geodesy. Artificial intelligence, machine learning, and computer vision technologies have shown considerable potential in analyzing these data.

The work at the Geodetic Institute focuses on exploring methods of deep learning to analyze ground deformation patterns. Typically, in addition to the deformation information, the radar signal is affected by the atmospheric delay, topographic errors, orbit errors, and different sources of noise. Persistent Scatterer Interferometry (PSI) is a common process for monitoring deformation, which uses the phase-stable pixels in a stack of interferograms, named as PS pixels, to avoid decorrelation, which degrades most of the pixels. However, present PSI approaches high demand on computing resources is challenged by ever bigger amounts of data that need to be processed. Furthermore, the growing periods covered by data sets cause a drop of numbers of PS and make it necessary to use the information of partially persistent scatterers, that existing algorithms rarely are able to use. Artificial neural networks have strong feature extraction capabilities and have achieved successes in computer vision, natural language processing, and other fields. By constructing an artificial neural network to learn PS pixel features by using GPU parallel computing, the deep learning method provides the possibility to quickly analyze the information from massive InSAR data.

Main goals of the project:

1. Development of a PSI approach based on deep learning to quickly analyze massive amounts of InSAR data.
2. Development of atmospheric correction and DEM correction approaches based on deep learning to improve the accuracy of PS-InSAR.
3. Development of a deformation monitoring approach using partially persistent scatterers.