*2.4. DInSAR Processing*

The GAMMA software package developed by GAMMA Remote Sensing and Consulting AG, Bern, Switzerland [74] was used for S1 Level-1 SLC product processing. The processing chain was divided into three stages: pre-processing, conventional repeat-pass DInSAR, and stacking. Pre-processing and conventional repeat-pass DInSAR were performed following the standard workflow used for processing S1 TOPS mode image pairs. This workflow is comprehensively explained in [75].

The pre-processing stage consisted of importing SLC data; updating of image metadata with precise orbital state vectors; S1 TOPS splitting, which included polarization selection; selection of sub-swaths and bursts covering the AOI; and selection of suitable S1 SLC image pairs and coregistration. Here, suitable S1 SLC image pairs were selected within the thresholds of the perpendicular and temporal baselines, which were set to 200 m and 18 days, respectively. The connection graph for ascending and descending datasets generated using these thresholds is presented in Figure 4a,b, respectively. A total of 321 ascending and 290 descending interferograms were generated and used in the stacking procedure. However, for coherence analysis, one additional interferogram from a descending orbital pass with a temporal baseline of 24 days was generated.

The conventional repeat-pass DInSAR stage included the formation of interferograms, multi-looking, simulation of topographic phases, differential interferogram generation, coherence calculation, phase filtering, phase unwrapping, orbital error correction, atmospheric correction, phase to displacement conversion, and interferometric product (i.e., coherence, differential interferograms, displacements map) geocoding (Latitude/Longitude WGS84 coordinate system).

Topographic phases were simulated using the precise orbits and an external DEM. Differential interferograms were formed at a default 2 looks in azimuth and 10 looks in range, to obtain a pixel size of ~40 × 40 m2. To improve the quality of differential interferograms and optimize the phase unwrapping procedure, the differential interferograms were filtered using an adaptive Goldstein filter [76], with an optimal filter strength of 0.7 being employed in this study, after a number of trials. After phase filtering, a minimum-cost flow (MCF) algorithm [77,78] was used for phase unwrapping. Areas with a coherence smaller than 0.2 were masked out before unwrapping. The linear trend was estimated and subtracted from the unwrapped differential interferograms, to correct the residual linear ramp caused by orbital errors. Differential atmospheric delay in the interferometric phase, which is correlated with the topography, was reduced using the empirical phase-based

method, for which the linear correlation between the unwrapped phase and the elevation of DEM was calculated [79,80].

**Figure 4.** Temporal and spatial (perpendicular) baseline connection diagram of the Sentinel-1 SLC image pairs from (**a**) ascending and (**b**) descending orbital pass used in the DIS approach.

During the stacking stage, the unwrapped differential interferograms of each set were summed and divided by the total (cumulative) time interval of all interferograms of the set in years, to obtain an average annual LOS displacement rate. Before stacking, the interferograms were referenced to a common (32 × 32 pixels) area and were shifted accordingly, to set the reference phase to zero. The common reference area was in the center of Villahermosa.

After stacking, phase to displacement conversion was performed, and the resulting LOS displacement rate maps were geocoded.

### *2.5. Identification of Flooded Areas*

Sentinel-1 GRD images were pre-processed by implementing radiometric calibration, spot filtering, and geometric correction of the data, to identify flooded areas. Radiometric calibration was initially applied, as it is an essential step in SAR GRD image pre-processing. The pixel values of the images could directly represent the radar backscatter [81], achieving results in dB. Image pixels representing bodies of water have a lower radar backscatter coefficient than other features [82], such as land or vegetation. The effects of thermal noises were also removed, and a precise orbit file was applied to the images. Lee Sigma filtering was applied to reduce the speckle noise caused by random effects of multiple backscattering within each resolution cell, which is best suited for this processing [81], leading to better results, with a filter size of 7 × 7. Finally, atmospheric correction was performed, to compensate for topography variations caused by the satellite sensor's viewing angle [81,82].

To obtain the flooded areas, the thresholding method was used, which is the simplest method of image segmentation [83]. Here, the areas affected by flooding were identified for two flood events: February 2018 and November 2019. The binary images (water/nonwater) were created using thresholds estimated from intensity (in dB) histograms of preprocessed Sentinel-1 GRD images. The used threshold values varied for the analyzed images between −12 and −10 dB; the water areas being those with an intensity below the applied threshold's value. To separate the permanent water bodies from the areas affected by floods, the permanent water bodies were identified using pre-flood event images (dry conditions) and then masked out in co-flood event binary images, so that, as a result, binary images of areas affected/non-affected by floods were obtained.

A permanent water bodies mask was also used to exclude water bodies from interferometric coherence analysis, as water bodies generally have a low coherence (near zero).

SNAP software (Sentinel Application Platform) [84] was used for pre-processing, whereas GIS software was used to obtain the flooded areas from the pre-processed images.
