*4.2. Spatiotemporal Dynamics or Inundation Detection by CyGNSS*

The Γ(*θ*) normalization results obtained for the Lv. 2 product indicated two peaks annually. The first peak was generally detected in the latter half of the dry season from April to June (Figure 6b), and the second peak was detected in the latter half of the rainy season from August to October. These findings indicate that the inundation status over the entire Mekong Delta is primarily controlled by double-/triple-rice-cropping irrigation activities. Approximately 57.4% of the rice-cropping area in 2012 was estimated to be triplecropped [33]. Interestingly, the northwestern region where the most intensive triple rice cropping is conducted (i.e., the An Giang and Dong Thap Districts) showed significantly greater reflectivity Γ values in the rainy season than in the dry season. Interestingly, the southwestern coastal wetlands consisted of mangroves and peatlands surrounded by acid-sulfate soils [21]. The spatially high reflectivity values found in such coastal regions, even in the dry season, might have been the result of aquacultural activities, including prawn-rice cropping rotations [34]. Because the delta receives greater attention for being

exposed to salinity intrusions exacerbated by rising sea levels [35], increased upstream dam construction [36] and groundwater depletion [37], further long-term observations over the delta are necessary for future assessments of the freshwater inundation status and the salinity intrusion succession status.

One of the novel features of our work that is presented in this paper is that our methodology realized the generation of spatiotemporally continuous data sets with a finer resolution (500 m spatial resolution, 15-day temporal resolution) than commonly used methods (that mostly have 3 km and 30-day resolutions, e.g., [13]), even though we did not use any spatiotemporal interpolation methods. Simple gridding without considering the size/shape of specular points cannot spatially rasterize the continuous CyGNSS GNSS-R data even with a lower resolution due to the data quantity limitation [13].

## *4.3. Comparison with Quadruple Polarimetric L-Band SAR Backscattering Signals*

Statistically significant Pearson correlations were confirmed through the precisionindex-based comparison between the CyGNSS reflectivity Γ and the PALSAR-2 backscattering intensity σ<sup>0</sup> or the spatial inundation percentage. We defined the inundation status based on PALSAR-2's 3–6 m resolution quadruple polarimetric data and ground truth observations [10]. To compensate for the spatial footprint size difference between the GNSS-R data and the inundation status observations with a finer spatiotemporal resolution in this study, we employed the product based on SAR data. There was still a discrepancy between the CyGNSS observations and the SAR-based inundation status product due to the heterogeneity surrounding the rice paddies over the Mekong Delta (e.g., buildings, forests, dykes), which was contaminated in the GNSS-R specular observations. The relationship was highly nonlinear, and its convexity was highly dependent on incidence angle differences. However, correlations were still found between these different microwave remote sensing methodologies even with the different observation resolutions over the heterogenous ground objects in this study.

As with other error-causing factors, notably, there are various factors causing geometric errors. For example, to propose a methodology that was independent from external data in this study, the ellipsoidal height that was derived from DEM and geoidal height information was not used for the rasterization process of the specular points. Most importantly, the grid-based rasterization of specular points was conducted by assuming that the velocity at each specular point was constant throughout each integration time (i.e., the acceleration of each specular point was assumed to be 0). It is still expected that the cross-validation performance could be better improved by rasterizing each specular point without the velocity-constant assumption. Regarding the geometric error correction, we also conducted a tuning experiment of the Gaussian function parameter of the precision index model [i.e., a value of 3.0 was used in this study, as described in Equation (2)] without downsampling. However, the tuning of the model parameter did not significantly differentiate the validation performance with the PALSAR-2 product (data not shown). Hence, rasterization with the consideration of acceleration was more important for tuning this model parameter. Without the acceleration information, the model tuning did not reliably improve the validation performance. For the current data interpretation, we also have to note the temporal differences between the observation times/dates of the CyGNSS and PALSAR-2 products. Due to the quantitative limitation of available specular points in this study, a low effective scattering area specular point group occasionally showed the opposite correlation with the PALSAR-2 backscatters (i.e., 30–35◦ incidence angles, 0–6 km square root values of the effective scattering area, Table S2). One of the causes of this result is the limited availability of quantitative specular points from CyGNSS over the Mekong Delta and the limited observable swath data of the quadruple PALSAR-2 observations (only 40–50 km widths were used to avoid incidence-angle-difference-derived biases in the polarimetric decomposition analysis). These data quantity limitations might have only partially caused the local optimization of the nonlinear function. Further observations are expected to enable the global optimization of the nonlinear function when estimating the

spatial inundation percentage or backscatter intensity from CyGNSS specular reflectivity data. The most importantly, we need to reshape this gaussian function along with the specular points velocity vector (i.e., shifting the gaussian center considering the doppler frequency, and reshape the skewness regarding its delay time in DDM information [32]).

Unlike the specular points with incidence angles wider than 10◦, a positive relationship was found between the CyGNSS reflectivity Γ and PALSAR-2 backscattering intensity σ<sup>0</sup> series for specular points with incidence angles narrower than 10◦. Since the difference between the microwave-energy-advancing vector directions of the backscatters and specular reflection values decreased as the incidence angle decreased, these positive relationships could have been found for specular points with such low incidence angles. This indicates that the reflectivity is highly dependent on the dielectric properties, particularly for lowincidence-angle specular points. Because such specular points tend to have low effective scattering areas (i.e., fine spatial resolutions), the incidence angle bias correction on such low-incidence-angle specular points is necessary to enable high-quality information on land surface properties to be derived. For most specular points with incidence angles ranging from 15 to 60◦, the CyGNSS reflectivity Γ and PALSAR-2 backscatter intensity σ<sup>0</sup> tended to show downwardly convex relationships (Table S3). This indicated that wetlands on relatively dry ground with a relatively low dielectric constant do not activate multi-time scattering (e.g., the double/triple bounce effect). Hence, the negative correlations between Γ values and σ<sup>0</sup> values tended to appear to be simply controlled by the specular reflection or single scattering effect. However, wetlands on wet ground, which have high dielectric constants at a certain level, also enhance multi-time scattering to emit relatively strong power levels not only oriented forwards but also backward. The specular points with incidence angles wider than 60◦ tended to show upwardly convex negative relationships between the CyGNSS reflectivity Γ and PALSAR-2 backscattering intensity σ<sup>0</sup> series. These findings indicated that if the incidence angle was wider than a certain level, the groundvolume interactions between inundated soil and wetland vegetation would be more prone to occur than if the specular points had lower incidence angles.

The three domains classified in the 2D scatter plots between the CyGNSS reflectivity Γ values and PALSAR-2-based spatial inundation percentages indicated a microwave scattering status difference among each domain (Figure S2g,m). The specular points in the first domain with Γ values lower than approximately −20 dB (Figure S2g,m; domain shown with the green arrow) tended to reflect relatively high odd/double bounce values. This finding indicated that the ground-volume interaction between inundated soil and the land-covering vegetation in wetlands plays a dominant role in the scattering process in this domain. Because positive correlations tended to appear between the CyGNSS reflectivity Γ and spatial inundation percentage series in this domain, this domain would be more sensitive to inundation than to soil moisture. In the second domain, where the Γ values were between approximately −20 dB and 0 dB, specular points with a 0% spatial inundation percentage were detected (Figure S2g,m; domain shown with the red arrow). In this domain, the Γ values mostly showed negative correlations with the spatial inundation rate and backscattering intensities. This indicates that the multi-time scattering effect would not have played a major role in this domain. Instead, the single scattering effect would have played a major role in such dry ground areas with relatively low dielectric constants. The negative correlations also indicate the possibility that the soil moisture and the vegetation water content may have greater roles than the spatial inundation percentage in such non-inundated wetland ROIs. In the third domain, where the specular points had Γ values greater than approximately 0 dB (Figure S2g,m; domain shown with the blue arrow), the Γ values tended to become significantly high, although the PALSAR-2 backscatter intensity values (including the odd scattering and double bounce values) tended to be low. These results indicate that the contribution of multi-time scattering was negligible and that the surface roughness in this domain would also be low. The presence of a water body without vegetation would have enabled such strong specular reflection conditions under weak backscattering effects. Consistently, Arai et al. [7,8,10] also reported

three similar domains from HH/HV backscatter 2D distribution plots. Thus, this might be a common characteristic of L-band active microwave scattering signals collected over tropical wetlands.

For further development, the application of a precision index to a finer-spatialresolution GNSS-R product (e.g., the CyGNSS interferometric coherence ratio product [38]) would be desirable to improve the spatial resolution of the resulting reflectivity Γ product. Since the differentiation of multi-time scattering processes using the phase information of scattered microwaves is mandatory to improve the inundation detection performance, the Stokes vector-based pseudo-3-component decomposition approach [39] or multi-polarimetric reflectivity/phase information (e.g., HydroGNSS) also need to be addressed for use with the GNSS-R data. To prepare for a robust comparison between SAR data and such polarimetric GNSS-R data, further improvement must be made to the precision index model.

In this study, effective scattering area was employed as the footprint size for the following two reasons. The L-band SAR polarimetric decomposition study of the rice paddies revealed that the SAR backscattering intensity is mainly controlled by ground vegetation and is sensitive to both canopy structure and ground inundation status and coherence was mostly low, impeding the possibility of using polinsar approach [10,21,23]. From this study, we also detected that most of the rice paddies whose L-band SAR backscattering intensity is relatively high showed low GNSS-R reflectivity (Figure 9). This indicated that the GNSS-R signal over the lowland wetlands/rice paddy is sensitive to ground-vegetation interaction and that the reflective property is incoherent rather than coherent. In subsequent studies study, the First Fresnel zone [40] should be considered as the footprint size, particularly for non-vegetated wetlands or paddies with immature rice paddies whose number of days since sowing is shorter than three weeks.

Regarding the nonlinear relationships between the CyGNSS reflectivity Γ and PALSAR-2 backscattering intensity σ<sup>0</sup> and between the CyGNSS reflectivity Γ and the inundation percentages as affected by incidence angle differences, a model parameterization scheme with an improved precision index model is desirable if both SAR data and GNSS-R data are to be used cooperatively to overcome their observation scale differences.
