Blog | Descartes Labs

Sentinel-1 Technical Series Part 3 | Global scale InSAR

Written by Descartes Labs | Sep 8, 2022

This is Part 3 of a technical series focusing on Descartes Labs’ global SAR processing capabilities. SAR provides a valuable remote sensing tool, and this series will dive into detail about how we process SAR data globally and build SAR and InSAR-derived products. Prior articles in this series can be found here: Overview, Part 1, and Part 2.

Introduction

In Part 1 and Part 2 of this series, we described the fast data access and geocoding mechanisms for Sentinel-1 SAR data that we have built at Descartes Labs. These mechanisms are part of pipelines that provide rapid global scale radar backscatter and InSAR analytics on our platform. In this part of the series, we will describe the salient features of our global InSAR pipeline. Our global InSAR dataset is designed to be readily usable in coherent change detection and deformation monitoring analytics pipelines.

Interferogram generation with geocoded SLCs

In Part 2 of this series, we described our approach to generating geocoded SLCs on a fixed 10m Northing x 2.5m Easting grid. Our processing pipeline ensures that the same projection system is consistently used for all bursts corresponding to a given footprint. Interferogram generation is as simple as cross-multiplication of geocoded SLCs and smoothing with a spatial kernel where z1 and z2 represent complex valued observations from two different aligned geocoded SLCs and K[.] represents a spatial convolution with a given kernel.

The ease with which we can compute interferograms relies on the fact that our geocoding mechanism removes the propagation phase. In our production pipeline, we use a Gaussian kernel that accounts for the difference in sampling in Easting and Northing and has a resolution of ~50m. The data is subsampled to a standard grid of 20m and stored as images with two bands - coherence (𝝆) and wrapped phase(πš«π“). We generate all burst interferograms for VV polarization (for most land masses except Antarctica and Greenland) with a temporal baseline less than or equal to 24 days (See Figure 1 below).

Figure 1: Monthly count of the number of burst-based backscatter (blue) and interferometric (red) images in our catalog from Oct 2014 to Sep 2022.

Note that, because InSAR products and backscatter are generated consistently, we can use the backscatter to reconstruct the original complex interferogram.

There are a number of pipeline design aspects that ensures that the interferometric products generated are of high quality:

  1. Even though the original SLC imagery has a ground posting of ~14m along-track x ~5m ground range, we generate geocoded SLCs at an oversampled posting of 10m Northing x 2.5m Easting to preserve features in both phase and amplitude even in steep terrain.
  2. Cross-multiplication in the spatial domain of two cross-coregistered images results in convolution in the spectral domain. Our oversampling of the SLC data assists in suppressing aliasing effects and is a technique that is also employed in traditional interferometric stack processing.

History of our global InSAR product

The quality of interferometric phase and coherence observations directly depends on the quality of relative coregistration of the geocoded burst SLCs. As mentioned in Part 2, Sentinel-1 mission has the following characteristics that help us in this regard:

  1. Good quality orbit data with an accuracy of a few cm
  2. Good clock synchronization enabling accurate time tagging of underlying SAR imagery
  3. Slant range delay variance on order of few cm to few 10’s of cm, which is an order of magnitude smaller than slant range resolution ~2.5 m

With these data characteristics, we are able to rely on a purely geometric approach, akin to GMTSAR, to generate global scale interferometric products very efficiently.

We would like to note that our current pipeline is the third iteration of a global scale InSAR product. The very first version was based on ESA’s SNAP software and our second iteration, which we presented about in the AGU Fall meeting of 2019, was based on GMTSAR. Both these approaches presented limitations which led us to build our current pipeline. The main issues we encountered were:

  1. Lack of data structures and granular support at the burst level. Both SNAP and GMTSAR operate at the swath level at a minimum.
  2. Neither SNAP, nor GMTSAR adequately handle changing SLC granule footprints with time, often leading to frequent, time-consuming and expensive interventions.
  3. SNAP and GMTSAR were primarily designed to run on large desktop computers and not designed ground-up for operationalizing workflows in a cloud computing environment.

We used products from our GMTSAR-based iteration to verify the geometric processing of our current implementation which focuses on making single burst-based processing efficient and scalable.

Comment on phase unwrapping

Parts 4 and 5 of this technical series will describe some of the capabilities built on top of our global InSAR product in detail. However, it is useful to summarize some of our observations from working with the global InSAR related to phase unwrapping. In summary, effective phase unwrapping is often AOI and application dependent. In detail:

  1. It is hard to reliably unwrap every burst interferogram, i.e over all types of land cover types, at 20m (~50m resolution) posting with S1 IW data in an automated fashion without manual intervention. By reliability, we refer to the quality of the results produced by the unwrapping software. Performance of conventional 2D phase unwrapping (i.e, each interferogram independently) relies heavily on the amount of spatial averaging and filtering applied to the data, and our global product is designed to support numerous applications. In our experience, currently available 2D phase unwrapping methods could be made to work in some geographies at resolution greater than about 100m with heavy filtering and averaging as is done in the tectonics community, but such a dataset would preclude a large number of other applications.
  2. Most efficient large scale InSAR time-series implementation (e,g. Methods used in generating "EGMS ground deformation map) identify pixels that are stable over time and use edge-based differential time-series unwrapper that scale better than traditional 2-D single interferogram unwrappers. In most cases, phase unwrapping in a time-series sense produces better results than unwrapping interferograms individually.
  3. A number of InSAR time-series methods now estimate the wrapped time-history of phases and unwrapping these cleaner estimates also produces fewer phase jump errors than unwrapping the original interferograms.

Hence, conventional 2D phase unwrapping is better approached as a post-processing operation where users determine the amount of additional spatial averaging or filtering and mosaicking extent first.

Sample screenshots from our global InSAR product

Coherence mosaic of 24-day pairs over Nile Delta from May 3-26, 2021. Coherence mosaic of 12-day pairs over Nepal from May 3-15, 2021. The coherent area to the North is the Tibetan plateau and the region to the south is the Gangetic plains in India. The bright coherent blob in the middle is Kathmandu. Snapshot of wrapped phase for "active volcanoes from our Global InSAR product in our workbench environment (Left) Iceland over last 24 days, (Middle) Goma, DRC from May 2021 and (Right) Wolf Volcano in the Galapagos from May 2015.

Conclusions

In this blog post, we described the salient features of our global InSAR product. This global product is ready to use for coherent change detection and quicklook deformation estimation applications.

πŸ’‘ Interested in using this technology?

"Contact our team to discuss how we can integrate our global InSAR product into your processes.

Check out the next post in this series, "Sentinel-1 global velocity layer: Using global InSAR at scale".

 

Reference:
Agram PS, Warren MS, Calef MT, Arko SA. An Efficient Global Scale Sentinel-1 Radar Backscatter and Interferometric Processing System. Remote Sensing. 2022; 14(15):3524. "https://doi.org/10.3390/rs14153524