**3. Methodology**

#### *3.1. Sentinel-1 Data Processing*

We evaluated the sensitivity of S1 polarizations for land and water discrimination [52,53]. The overall backscatter differences resulting from the sensitivity analysis are summarized in Table 4. In the case of the IW imaging mode, the greatest difference between terrestrial and open water signals was obtained for VH polarization (10.5 dB). Data were acquired with an incidence angle of 39.5–47◦. In the EW mode, the differences were the greatest for HV polarization (9.6 dB). The data were acquired with an incidence angle of 37.5–46◦.

**Table 4.** Summary of sensitivity analysis for land and water discrimination for different S1 polarizations [52].


We also evaluated the sensitivity of S1 for the discrimination of flooded forests from dry forest areas. The greatest differences were observed in the case of HH polarization (a difference of 4.5 dB). The data were acquired with an incidence angle of 37.5–46◦.

Relying on the sensitivity analysis, the open water mapping algorithm was developed for IW VH and EW HV datasets, and water under vegetation was mapped from EW HH data in the current study.

Previous studies have shown the advantage of using multiple incidence angles for water mapping [54]. Therefore, we established an empirical relationship between the local incidence angle and surface water backscattering collected at our test sites. The dataset of open water backscattering from known waterbodies was collected from 7 IW and 12 EW mode images acquired at the time of flooding in October, November, December, and April in three consecutive years (2017–2019). The relationships between the local incidence angle (θ) and surface water backscattering (σ0) for polarizations with the greatest sensitivity (Table 4 in the previous section) are shown in Figure 2a (original source of data [52]). An established relationship was used for the water mapping algorithm (Table 5) dependency. The algorithms for open water mapping from IW VH and EW HV polarizations are summarized in Table 5.

**Figure 2.** (**a**) Relationship between open-water backscattered signal (σ0) and incidence angle (θ). IW mode data (VH) were acquired from 80 and 58 orbit overflights, and EW mode data (HV) were acquired from 51, 58, 87, 153, and 160 orbit overflights [52]. (**b**) Relationship between flooded forest backscattered signal (σ0) and incidence angle (θ). EW mode data (HH) were acquired from 51, 58, 87, 153, and 160 orbit overflights.

**Table 5.** Water mapping threshold conditions for different polarizations, imaging modes, and flood types (FUV and OWF).


A study by Lang et al. [55] showed the relationship between the incidence angle and backscatter of water under vegetation. The study demonstrated a decrease in backscatter by 2.45 dB at the incidence angle between 23.5◦ and 47◦, in the case of Radarsat data. For the evaluation of the dependence of incidence angle on the backscattering in the case of a flooded forest, the data were collected at the time of flooding from images acquired on 08 November 2019, 13 November 2019, and 11 April 2018. However, analysis of our dataset did not confirm the relationship between the incidence angle and backscattered signal in a flooded forest (Figure 2b). Relying on our analysis, a threshold condition of HH > −4.21dB was set for flooded forest mapping. The threshold was estimated on an averaged backscattered signal (σ0) +1 standard deviation (SD) in flooded forest areas determined from visual observations.

The data processing scheme was set up in a cluster computing environment. The data processing setup is schematically shown in Figure 3. Pre-processing included the following steps: radar signal calibration, noise filtering, terrain correction, and the image processing technical processes of reading, cutting, and extracting data (Figure 3). Pre-processing was performed using the processors from SNAP (Sentinel Application Platform) software. Water mapping was performed according to Equations (1)–(3) presented in Table 5. The automatic water mapping processes were set up in a cluster computing environment using SHELL script to download the imagery from the Sentinel Open Data Hub and to run SNAP based GPT for water mapping. A combination of the DIST and HAND approaches was applied for the elimination of water lookalikes. The auxiliary data from the Estonian Land Board, namely, the DEM dataset with 5 × 5 m resolution and the official inland water body map (from ETD), were used to improve the mapping accuracy. In the first step of post-processing, the data were polygonised. In the case of inland water bodies, open water polygons (mapped from S1) that intersected with the inland waters map (ETD) with a buffer zone of 100 m were extracted for further analysis. At the coastal zone of the Baltic Sea, the open water polygons that intersected with a coastal area of up to a one meter elevation were extracted for further analysis. GDAL (Geospatial Data Abstraction Library) software processors were used for the post-processing of the data.

**Figure 3.** Data processing setup.

In the post-processing of the water mapped under vegetation, the noise (false-positives) from buildings was extracted by removing the polygons that intersected with the buildings map (from ETD). After removal of the areas with elevated backscattering caused by buildings, the water polygons related to the wooden area map (ETD) were extracted for future analysis.
