From NetSysLab

Jump to: navigation, search

InSAR (Interferometric Synthetic Aperture Radar) is an earth-orbiting satellite-based remote sensing technology. The InSAR technique computes differences in signal components (e.g., phase) of radar echoes between multiple radar images taken at different points in time. This information can be used for tracking spatial displacement/deformation of an object on earth, such as a building and a bridge, over time at millimeter-scale accuracy. The object analysis part of the processing belongs to an offline processing chain that operates on very high resolution images consisting of terabytes of data. Various signal processing techniques are used; for example, spatial frequency interpolation such as Fourier transformation. Given the huge volume of data, most of the kernels in the processing chain have high resource demand in terms of compute power, memory bandwidth and capacity, and storage.

3vGeomatics is a Vancouver-based InSAR company seeking innovation in the area of InSAR processing efficiency. To date, most innovations in the area of InSAR processing has been driven by signal processing experts and often computational efficiency has not been the priority. The advent of high-resolution sensing coupled with advancements in InSAR analysis opened the door for novel services that can measure displacement at an unprecedented scale and resolution. However, as both the volume of the data and the complexity of the algorithms increase, the runtime for the current InSAR processing pipeline increases significantly. Together with 3vGeomatics, we are exploring opportunities for improving computational efficiency of the InSAR processing chain. A typical processing chain consists of several independent or loosely coupled operating entities or kernels. At the grass-root level, we are investigating how performance of individual kernels can be improved: using computationally efficient data structures and algorithms, and harnessing parallel processing. Many InSAR algorithms are inherently parallel, hence, can take advantage of modern multi-core CPUs and accelerators, such as GPUs, aimed at general purpose computing. There are further research opportunists for kernel auto-tuning to aid optimization and regarding tasks scheduling: the operational procedure of the processing chain falls under the many-task computing paradigm.

To date, the collaboration between NetSysLab and 3vGeomatics has resulted in the following kernels being parallelized either for multi-core CPUs or ported for GPU acceleration:

1. Goldstein-Werner Filter (GPU) - An adaptive filtering algorithm that lowers phase noise and improves measurement accuracy. 2. Automatic Motion Reconnaissance (GPU) - Determines spatial coherence over time. 3. Persistent Scatterer Selection (GPU) - Determines temporal coherence of permanent targets in a radar image. 4. Interferogram Generation (multi-core CPU) - Generates an interferogram from co-registered single-look complex images.


Tahsin Reza
Tanuj Kr Asawat
Abdullah Gharaibeh
Matei Ripeanu


[3] Accelerating Persistent Scatterer Pixel Selection for InSAR Processing Tahsin Reza', Jose Manuel Delgado Blasco, Aaron Zimmer, Parwant Ghuman, Tanuj Aasawat, Matei Ripeanu, IEEE Transactions on Parallel and Distributed Systems (TPDS), accepted May 2017. pdf
[2] Efficient GPU Techniques for Processing Temporally Correlated Satellite Image Data, Tahsin Reza, Dipayan Mukherjee, Tanuj Kr Aasawat, Matei Ripeanu, poster at SC’15, Austin, TX, November 2015 pdf.
[1] Accelerating Persistent Scatterer Pixel Selection for InSAR Processing, Tahsin Reza, Aaron Zimmer, Parwant Ghuman, Tanuj kr Aasawat, Matei Ripeanu, 26th IEEE International Conference on Application-specific Systems, Architectures and Processors (ASAP) July 2015, Toronto (acceptance rate: 21/85=24.7%) pdf slides.