The influence of Sentinel-1 SAR sub-swath on the recorded backscatter time-series over managed hemiboreal forests

. The view angle range of Sentinel-1 SAR in the Interferometric Wide swath (IW) measurement mode is 29.1° – 46.0°. The dependence of backscatter on the arbitrary local incidence angle is usually corrected using a linear regression model where the incidence angle is a predictor variable. We analysed the whole time series of Sentinel-1 SAR VV-polarised backscatter over the flat 15×15 km test site in Laeva, Estonia (26° 26' 43'' E; 58° 31' 56'' N). Time series containing measurements from three different orbits were constructed for 3,159 stands from nighttime data and for 1,105 stands from daytime data. We can confirm that daytime backscatter is systematically greater than nighttime backscatter. We found a significant deviation from linearity in the backscatter dependence on local incidence angle. The empirical finding may be caused by the microwave scattering dependence on local incidence angle or by the influence of Sentinel-1 SAR sub-swath configuration in the Terrain Observation with Progressive Scans SAR (TOPSAR) method that is used for the measurements.


Introduction
The Sentinel-1 SAR is a C-band radar on sun-synchronous polar orbit (Fletcher, 2012).Sentinel-1A has been on the orbit since 3 April 2014 and Sentinel-1B since 25 April 2016, while the 1B is not functional since December 2021 due to technical problems (ESA, 2022).The radar measures microwave backscatter using the Terrain Observation with Progressive Scans SAR (TOPSAR) method by electronically steer-ing the radar antenna over target area while the satellite is moving in azimuth direction.Measurements are done using one of four configurations.The Interferometric Wide swath (IW) mode is commonly used for land surface targets.The full ground range swath of Sentinel-1 SAR in the IWmode is divided into three sub-swaths (Fletcher, 2012;Hajduch et al., 2022).
The long Sentinel-1 SAR time series are used for prediction of soil moisture dynamics, for mapping of forests and other vegetation phenology (Van doninck et al., 2012;Dostálová et al., 2018;Rüetschi et al., 2018).One crucial step before using SAR backscatter (σ 0 ) data for studies of vegetation properties is the correction or normalization of the signal for the local incidence angle (θ) that is a combination of the view angle, terrain slope and azimuth of the slope (Gauthier et al., 1998).
A frequently used method for normalizing SAR backscatter is to assume a linear relationship and convert the σ 0 measured at angle θ to a value at some reference angle θ ref where β is the slope of the linear model fitted to σ 0 data using the difference as a predictor variable (Gauthier et al., 1998;Van doninck et al., 2012;Dostálová et al., 2018;Schaufler et al., 2018).The  � � ��� � =  � () − � −  ��� �, could be then used as a predictor variable forming dense time series for a particular target object from pixel values extracted from all overlapping image swaths viewed from different orbits.
The orbits of Sentinel-1 converge towards Earth poles and at higher latitudes it is possible to observe target objects from different orbits.In Estonia the image swaths from four to six orbits intersect on the ground.With the Sentinel-1A/1B pair on the orbit it is therefore possible during 72 hours to get three nighttime measurements and three daytime measurements for a forest stand so that an almost fullview angle range of Sentinel-1 SAR is covered in the triplets.We selected a 15 by 15 km test site near Laeva village, Estonia, in managed hemiboreal forests to study the applicability of the linear normalization (1) for the compilation of dense SAR time series.

Test site
The 15×15 km test site centre coordinates are 26° 26' 43 '' E and 58° 31' 56'' N. Within the area are two larger forested regions separated by river Laeva (map in Appendix 1).The topography of the area is flat.Forests in the northern bank of the river are regularly managed, but the southern region has common management history until 2019 when Raja-Kärevere nature reserve was established.Most soils in the forest land are fertile corresponding to Aegopodium and Filipendula site types (Lõhmus, 2004), however, lowland mires, transitional and raised bogs are also present.The area is influenced by long-time drainage with ditches.Dominant tree species are European aspen (Populus tremula L.), silver birch (Betula pendula Roth), Norway spruce (Picea abies (L.) H. Karst.), grey alder (Alnus incana (L.) Moench) and black alder (Alnus glutinosa (L.) Gaertn.).On the raised and transitional bogs dominates Scots pine (Pinus sylvestris L.).More details about the Laeva test site can be found in Lang et al. (2014).For the study we employed a database of 4,995 forest stands located in the state-owned forest land.

Sentinel-1 SAR data preprocessing
We used a stack of analysis-ready radar satellite data provided by the company Kap-paZeta (Kastani 42, 50410 Tartu, Estonia).The stack included interferometric coherence: 6-or 12-day repeat pass for VH and VV polarization, backscatter in VH and VV polarization mode and VH/VV backscatter ratio and local incidence angle data.Data were in raster files; pixel size 5 m.Backscatter values were computed as sigma0 following the latest Sentinel-1 product specification (Vincent et al., 2020) for calibration and thermal noise correction.The terrain correction and orthorectification from radar cylindrical SLC coordinates to local L-EST 97 (EPSG:3301) coordinates was done using Copernicus 30 m resolution digital eleva-from the VV-polarized band of each image.Out of the 4,995 stands we found triple SAR measurements, i.e. sequences from three orbits for 3,159 stands at nighttime and for 1,105 stands at daytime.

Fitting of the incidence angle normalization model
We compared data from nighttime {153;080;007} and daytime {087;160;058} orbit triplets and found that daytime VV-polarized backscatter  �� � is systematically greater than during nighttime (Figure 1).A similar note has been published by Schaufler et al. (2018) who found greater backscatter in the case of measurements made from ascending orbits.

Local incidence angle
The local incidence angle map was included in the data stack.It is based on satellite ephemeris data and Copernicus 30 m digital elevation model (DEM) (Fahrland et al., 2020).However, at forest edges that existed at the time of construction of the DEM occurred substantial errors that propagated into θ (Appendix 1).Within the small 15×15 km flat test site the θ varied at forest edges more than 10 degrees around its mean value.Therefore, we used constant mean incidence angle in data analysis for all stands according to relative orbits of Sentinel-1.

SAR measurement data
We used all Sentinel-1 SAR measurements that were available from the beginning of operation until the end of March 2022.From each orbit there were more than 200 images available (Table 1) for the entire study period.In this study we analysed VV polarised backscatter.

Sampling of SAR backscatter from raster data
The average value of pixels found within stand border was calculated for each stand We separated daytime and nighttime measurements and then combined measurements according to data from orbit pairs {153-080; 080-007; 087-160; 160-058} and fitted each combination with the linear model  � () =  + � −  ��� � + , (2) where α and β are found using least squares regression and ε is the model residual error.

Results and Discussion
By comparing the values of β estimated for orbit pairs we found that almost in all forest stands the mean value of  �� � decreases more per unit angle in pairs with greater incidence angles (Figure 2).This was found in daytime as well as in nighttime measurements.If we assume that backscatter of microwave pulse follows a linear monotonous decrease through the usual SAR incidence angle range, then the application of the normalization model ( 1) would be justified.Based on this assumption, Van doninck et al. (2012) used the model (1) to normalize Advanced Synthetic Aperture Radar (ASAR) data from 80 descending passes and noted that all the observations for each image pixel were used for the model fitting.Dostálová et al. (2018) and Schaufler et al. (2018) used the model (1) for Sentinel-1 SAR data that covered large areas.In the case of Laeva Sentinel-1 SAR data set, however, (1) there is a systematic difference of backscatter from ascending and descending passes and (2) the angular correction coefficient β depends on the incident angle.
The reason for the dependence of the angular correction coefficient β on the incident angle is not clear.It is possible that forward scattering increases with θ and relatively less energy is reflected back in the case of larger local incidence angles.However, it cannot be excluded that the empirical finding is related to the Sentinel-1 SAR construction and the usage of TOPSAR measurement method.In Laeva test site the stands were located on different sub- swaths in the image triplets (Appendix 2).Sentinel-1 SAR calibration and signal restoration for the image construction is done by the sub-swaths (Hajduch et al., 2022).
Our empirical finding may have importance for the construction of the normalized incidence angle of Sentinel-1 SAR backscatter time series for vegetation mapping and phenology studies.Fitting just a linear model on all Sentinel-1 SAR backscatter measurements over target to correct for incidence angle is not sufficient to remove the signal dependence on the influence of scanner subswath.In mapping and change detection applications the processing and decision-making accuracy may be increased if the effect of Sentinel-1 SAR subswath characteristics and the nighttime and daytime backscatter difference on the registered microwave pulse backscatter will be taken into account.Also, the 1B is not functional since December 2021 and this can create artefacts in long time series as the combination of sub-swaths over the targets is now different compared to the operational pair of Sentinel-1A and Sentinel-1B.