EZ-InSAR: An easy-to-use open-source toolbox for mapping ground surface deformation using satellite interferometric synthetic aperture radar

The processing chain of MTI analysis can be divided into two workflows: (1) the generation of differential interferograms and (2) the generation of displacement time series (Fig. 1). Most modern satellite SAR missions, e.g., Sentinel-1, provide data that is pre-processed from unfocused raw data (Level 0) to a focused Single Look Complex image (SLC, Level 1), in which each pixel contains both amplitude and phase of the backscattered radar signal. The interferometric phase resulting from the differencing of the phase information in two SLC images is a lumped sum of contributions from ground deformation, SAR imaging geometry, surface topography, atmospheric delay, and InSAR decorrelation noise (Rosen et al. 2000).

The first workflow starts with the download of SLC and orbit data. This is followed by the co-registration of the SLC stack, image pair configuration, interferogram generation, and phase unwrapping (conversion of phase to displacement). The SAR satellite's orbital parameters and an external digital elevation model (DEM) are used to compute the phase components in an interferogram that are related to the SAR imaging geometry and the surface topography. These phase components can then be subtracted from the interferometric phase to produce the differential phase, which ideally represents the ground displacement between each SLC acquisition. However, other non-displacement components of the differential phase can include atmospheric delay, as well as residual errors from inaccurate orbital parameters and DEM data. Therefore, an important goal of MTI analysis is to isolate the deformation signal from the remaining error phases and atmospheric delays by analysing a stack of differential phase observations (Osmanoğlu et al. 2016).

The second workflow starts with selection of pixels (i.e., persistent scatterers and distributed scatterers) with high phase quality for displacement time series analysis. This is usually done by statistically analysing the temporal stability of SAR backscatter intensity and/or interferometric phase, or by using an a priori model (e.g., linearity) of the temporal evolution of displacements. Therefore, depending on the rules for defining them (Osmanoğlu et al. 2016), high-quality pixels may be those that in time display a relatively high and stable backscatter intensity or a relatively consistent to gradually changing phase. Using the differential phase of the selected pixels as observations, we can then construct an inversion model to obtain relative phase changes (i.e., displacements) for each SAR acquisition with respect to the reference epoch and a reference point. The phase inversion step also accounts for the correction of errors in the interferometric phase (i.e., phase contributions unrelated to ground motion), although the correction strategies differ for different MTI methods. For example, the atmospheric phase error can be estimated either by spatio-temporal filtering of the interferometric phase, or by using an external weather model or by using an a priori model of the temporal evolution of displacements (Agram et al. 2013; Li et al. 2022; Werner et al. 2003). Note that phase unwrapping in the first workflow can be bypassed for some MTI methods that use the differential phase between the neighbouring pixels as observations in the inversion model by assuming no phase ambiguity exists between them (Zhang et al. 2012).

PSI and SBAS are the two main types of MTI approaches. Differences between them include: (1) the method of forming InSAR image pairs, (2) the criteria of selecting high-quality pixels, and (3) the phase inversion model. The SAR image pairs in the PSI approach are formed by using a single reference scene; those in the SBAS approach are formed by using multiple reference scenes. The PSI technique selects pixels representing scatterers on the ground that persistently yield high amplitude and/or low phase variance in those pixels. The SBAS method deals with pixels representing more distributed and/or lower amplitude scatterers on the ground by using interferometric coherence as a selection criterion. The interferometric coherence is the correlation between two SAR images: as such the coherence is a direct estimate for the accuracy of the determination of the interferometric phase (Zebker and Villasenor 1992). Additionally, the SBAS approach generally involves multi-looking (i.e., pixel averaging) on the interferograms to reduce noise levels, although it is not mandatory, and it can be replaced or augmented by spatial or temporal filtering. For the phase inversion model, PSI approaches commonly use assumptions about the temporal change of ground deformation (e.g., linear, or seasonally oscillating), although some modified PSI techniques, such as the StaMPS method (Hooper et al. 2004), do not need such assumptions. In the SBAS approach, the greater number of redundant interferometric observations arising from the multi-reference method of forming image pairs helps to constrain the phase inversion model without needing an a priori assumption about the displacement behaviour in time. Both have advantages (and disadvantages) that depend largely on the target area’s spatiotemporal features. Generally, PSI is suitable for areas with localised, high-reflectivity radar back-scatterers, such as urban regions and man-made infrastructure. SBAS is suitable for areas with distributed, less reflective back-scatterers, such as in rural regions.

The main imaging mode of Sentinel-1 is the Interferometric Wide Swath Mode (IW). This uses a progressive scan (TOPS) antenna beam steering technique to provide a 250 km wide image swath consisting of three sub-swaths at c. 3 m × 22 m pixel size (Potin et al. 2019). Each swath consisting of several bursts (Torres et al. 2012). ESA released the Sentinel-1 TOPS SAR data in the format of SLC (L1), a standard format that is for interferometric processing, which can be fed into a standard InSAR processing chain after co-registration. Because of variable azimuth spectral properties of TOPS SAR data, an Enhanced Spectral Diversity (ESD) method using the phase of the burst overlay region of Sentinel-1 SAR data is usually employed to make sure a co-registration error smaller than 1/100 pixel (Yague-Martinez et al. 2016). The ESD assisted co-registration method is now supported by nearly all the popular InSAR processing tools, which will be introduced in the next section. Several international and national agencies, such as ESA and the Alaska Satellite Facility (ASF), support the distribution of Sentinel-1 data through web Application Programming Interface (API) services, which make it feasible to develop an automatic processing chain.

Sentinel-1 satellites also operate in Stripmap mode, which is the conventional acquisition mode for most SAR sensors. This mode can be less challenging as the co-registration accuracy for obtaining a suitable interferometry image at a spatial resolution of several metres can be < 1/10 pixel. However, the swath width is 80 km for the Sentinel-1 Stripmap mode, compared to 250 km for the IW mode. Current X-band satellites, such as COSMO-SkyMed, TerraSAR-X and PAZ, also operate in Stripmap-like mode.

EZ-InSAR consists of three modules: (1) Preparation of SAR Data, (2) ISCE Processing, and (3) Time Series Analysis. Each module is accessed through one of three main panels in the user interface (Fig. 2). A button to define an EZ-InSAR working directory path is on the top left of the interface. At the base of the GUI, there is a bar showing the progress of each InSAR processing step and an information box showing the progress state and additional useful information (i.e., errors) during data processing.

The Preparation of SAR Data module includes functions for searching and downloading Sentinel-1 SAR data. EZ-InSAR supports the use of a KML file exported from Google Earth to define a study region. Once the study region is determined, the API provided by the ASF is then used to search the archive of available Sentinel-1 SAR images based on the input filtering keywords, such as the path number, flight direction of satellite (i.e., ascending or descending), and the desired time span. EZ-InSAR can then download the SAR data automatically after confirming the returned data list (see Fig. 3). A DEM covering the study region is required before running interferometric processing. Here EZ-InSAR provides options of automatically downloading either the SRTM DEM (Farr et al. 2007) or the Copernicus DEM from Amazon Web Services (ESA and Sinergise 2021). Both DEMs have a spatial resolution of 30 m and are saved in GeoTIFF format. EZ-InSAR also supports the input of a third-party DEM in GeoTIFF format as specified by the user (See Fig. 3). To help check the quality and coverage of the DEM, we provide a display option allowing the user to visualize a shaded relief map of the downloaded DEM overlapping onto a geographic base map.

The SAR Interferometry module includes the functions for interferometric processing. We selected the ISCE package, initially developed by a team from NASA’s Jet Propulsion Laboratory and from Stanford University for this purpose (Rosen et al. 2012). ISCE is a highly hierarchical package with Shell and Python scripts to control the different applications based on specified tasks, such as the task of generating a stack of coregistered SLC images, a stack of interferograms without phase unwrapping, and a stack of unwrapped interferograms. EZ-InSAR here gives the options to generate an “SLC stack” or an “Unwrapped interferogram stack”, depending on whether the user wishes to run a StaMPS-based or a MintPy-based time series analysis, respectively. EZ-InSAR also allow bidirectional conversions between these two data stacks (Fig. 11), so that one can more easily conduct and cross-check the MTI analysis with different approaches. EZ-InSAR supports the generation of the data stack either step-by-step or by batch processing in a parallelised computation.

The Time Series Analysis module provides the option of using either the StaMPS or the MintPy time-series processors. It also undertakes initial checks on input data suitability for these. The StaMPS package is mainly written in MATLAB and can perform both PSI and SBAS analyses at full image resolution. StaMPS also enables integration of the PSI and SBAS results to enhance the density of measurement points (Hooper 2008). MintPy is written in Python and provides several different SBAS processing algorithms for correcting interferometric artifacts (e.g., phase unwrapping error, tropospheric delay, and topographic residual) and for obtaining displacement time series (Zhang et al. 2019). The StaMPS and MintPy processors can be activated by clicking the corresponding tabs in the panel (Fig. 2). The GUI panels enable control of the parameters that will control the MTI analysis. Again, the time-series processing with either StaMPS or MintPy can be done either step-by-step or by batch processing.

The location of the reference point for displacement has a great influence on the final InSAR results. For example, EZ-InSAR allows the users to select the reference point, for MintPy-based processing, interactively on based on optical satellite images provided by MATLAB and hosted by Esri. Finally, users can export the velocity field and the displacement time series from EZ-InSAR into several formats, such as GeoTIFF, KMZ, and QGIS, such that they can be further exploited by the other software for visualization and analysis. Figure 4 shows a flow chart of functionality and data processing within each of these modules, which collectively facilitate a full MTI processing chain.

EZ-InSAR also has several unique tools to optimize the SAR data processing and to obtain high quality ground displacement measurements. Firstly, we designed a tool to automatically select the SAR acquisition in the geometric centre of the perpendicular baseline vs temporal baseline network as the optimal reference image (Fig. 5). Coregistration accuracies between the reference and secondary SAR images should thereby be improved since the temporal and perpendicular baselines of image pairs greatly influence the interferometric coherence. Secondly, we developed a display window allowing users to easily visualize all SAR interferograms, unwrapped phase maps, and coherence maps within a stack. Using this tool, the user can inspect the quality of the interferograms and drop bad ones from the MTI analysis (Fig. 12). Thirdly, we developed a tool that allows the user to manually modify the SAR image connection network by removing or adding specific image pairs in the MTI analysis (Fig. 6). For example, if the initial MTI analysis has revealed that some images are strongly contaminated by atmospheric delay, because they exhibit large residuals in the initially retrieved displacement time series, then we can drop these un-desired image pairs and re-run the MTI analysis to improve the results.

Campi Flegrei Caldera, Italy, is a 13-km-wide depression that hosts a set of young volcanic vents and active geothermal areas. It is partly submerged beneath the Gulf of Pozzuoli, and it is highly urbanised, encompassing part of the city of Naples (Pepe et al. 2019) (Fig. 7a). The most recent eruption was from the Monte Nuovo vent in 1538 (Di Vito et al. 1987). Episodes of significant uplift and subsidence within caldera have occurred since Roman times, and inflation has occurred at an increasing rate since 2005 (De Martino et al. 2021). Here we can test EZ-InSAR’s performance in an urban to sub-urban area with large ground motion and medium/high InSAR coherence.

Long Valley Caldera, USA, is a 32 × 16 km depression that hosts a resurgent dome and several active geothermal areas (Fig. 7b). It lies at the eastern edge of the Sierra Nevada mountains and occupies a mostly rural area with several small towns. Two younger volcanic systems on the margins of the caldera, Mammoth Mountain, and Mono-Inyo Craters, have most recently erupted about 600–700 years ago. Long Valley Caldera has exhibited relatively small-magnitude deformation related to hydrological forcing and volcanic activity since 1980s (Silverii et al. 2020, 2021). Here we can evaluate the performance of EZ-InSAR in revealing subtle ground deformation in a rural setting with lower InSAR coherence.

The black rectangles in Fig. 7a and b show the footprints of the KML files that we used in EZ-InSAR to download Sentinel-1 SAR images for the two test sites (Table 2). We applied both the StaMPS-based PSI and MintPy-based SBAS approaches at the Campi Flegrei volcano (see parameter settings in Figs. 13 and 15). Given the dense vegetation (forests) there, we employed the SBAS approach in StaMPS and MintPy at Long Valley caldera (see parameter settings in Figs. 14 and 15). Note that we used the default threshold values of StaMPS to select the high-quality pixels. We adopted the elevation-phase correlation method to remove the atmospheric phase delay (Li et al. 2019), and we used a linear fitting model to deramp the velocity fields. Note that StaMPS provides options of implementing spatial and temporal filtering to the results, while MintPy does not. Unlike the data processing based on full resolution SLC data in StaMPS, MintPy works by exploiting the unwrapped phase of multi-looked interferograms. Multi-looking factors of 2 in SAR azimuth direction and 10 in range direction were used to generate the interferogram stacks.

The processing time is mainly related to the SAR data amount, the area coverage of the study region, the selected MTI approach. On our server, which possesses an AMD EPYC™ 7662 CPU with 64 cores (2.0 GHz, max. 3.3 GHz) and a RAM memory of 1 TB (2,933 MHz), for example, the total processing times were 4 and 6 h for the PSI and SBAS analyses for Campi Flegrei caldera, respectively; the SBAS analyses ran for a couple of days for Long Valley Caldera.

Figure 8 shows the mean ground displacement velocities at Campi Flegrei caldera as mapped by using the StaMPS-based PSI and MintPy-based SBAS approaches, according to default parameters of processing. Both the velocity fields show surface uplift of up to ~ 120 mm·yr⁻¹ along the satellite Line-of-Sight (LOS) direction in 2021. The main deforming region covers an area of about 18 km². The two maps mainly differ in the number and spatial coverage of velocity measurements, which are greater with the SBAS method. The InSAR displacement time series agree well with GNSS observations at two local monitoring sites provided by De Martino et al. (2021): RITE and SOLO (Fig. 8c and d). We converted the west–east, south-north, and vertical components of the GNSS displacement time series into a component in the satellite LOS direction. The calculated LOS displacement velocities based on the GNSS data are 110 and 54 mm·yr⁻¹ at RITE and SOLO, respectively. The mean displacement velocities from the PSI and SBAS InSAR are 97 and 102 mm·yr⁻¹ at the GNSS site RITE, and 58 and 69 mm·yr⁻¹ at the GNSS site SOLO. The largest discrepancies are ~ 15 mm·yr⁻¹. The RSME values between the GNSS data and InSAR displacements are 5.7 mm and 7.7 mm for the StaMPS-based PSI approach, 6.3 mm and 7.6 mm for the MintPy-based approach, at RITE and SOLO, respectively.

Figure 9 shows the ground motion velocity maps derived for the Long Valley caldera from StaMPS-based SBAS and MintPy-based SBAS, respectively, according to default parameters of processing. The two SBAS approaches reveal local subsidence at a rate of about -15 mm·yr⁻¹ near the south edge of the resurgent dome, where a geothermal plant (named Casa Diablo) is located. Previous InSAR investigations using ENVISAT SAR data from 2003 to 2010 have also revealed subsidence near this geothermal plant, indicating that the subsidence has persisted in recent decades (Hetland et al. 2012). To the east of the resurgent dome, we observe an apparent widespread LOS subsidence within the caldera (clearer in the StaMPS result).

Figure 9c and d show the displacement time series at the two continuous GNSS stations, RDOM in the centre of the resurgent dome and CA99 near the geothermal exploration site. The GNSS measurements, provided by the Nevada Geodetic Laboratory, were calibrated by using the P653 station. They show that both sites were relatively stable from January 2020 until January 2021 but subsided from January 2021 until December 2021. The LOS displacement velocities at RDOM and CA99 are -6.0 and -4.9 mm·yr⁻¹, respectively, over the two years. The MintPy-based SBAS measurements are -0.2 and -1.5 mm·yr⁻¹, which underestimate the GNSS-derived values by 5.8 and 3.4 mm·yr⁻¹, respectively (i.e., by about 97% and 70%). The StaMPS-based SBAS measurements at the two stations are -3.8 and -7.4 mm·yr⁻¹, which respectively overestimate and underestimate the GNSS velocity magnitudes by about 37% and 51%. However, the StaMPS-based approach successfully captures the temporal evolution of the deformation at Long Valley caldera, with initial stability in 2020 succeeded by subsidence at both sites in 2021. The RSME values between the GNSS data and InSAR displacements are 4.8 mm and 3.8 mm for the StaMPS-based SBAS approach, and 7.0 mm and 4.9 mm for the MintPy-based approach, at CA99 and RDOM, respectively.