Evaluation of Spire GNSS-R reflectivity from multiple GNSS constellations for soil moisture estimation

ABSTRACT Spaceborne Global Navigation Satellite System-Reflectometry (GNSS-R) has emerged as a pivotal tool with different land applications, prominently encompassing soil moisture estimation. In contrast to conventional radiometer satellites commonly used for this purpose, GNSS-R offers higher spatiotemporal coverage while maintaining cost-effectiveness. The potential of using Global Positioning System (GPS) reflections measured by the Cyclone Global Navigation Satellite System (CYGNSS) mission to retrieve soil moisture has been previously demonstrated. In 2019, Spire Global Inc. launched their first GNSS-R satellites, which now comprise a constellation of four CubeSats. These satellites track reflections from multi-constellation, encompassing GPS, Galileo, BeiDou Navigation Satellite System (BDS), and Quasi-Zenith Satellite System (QZSS). In this study, an analysis and validation of Spire GNSS-R L1B surface reflectivity for soil moisture retrieval within east Australia during an eight-month period in 2021 is presented. A comparison of the estimated Spire surface reflectivity to that of CYGNSS is performed, unveiling analogous behavioural patterns and biases across both missions. Soil moisture is estimated using observations from Spire GPS-only, Spire multi-constellation, and CYGNSS. The Soil Moisture Active and Passive (SMAP) retrievals are used as the reference, presuming a linear relationship between changes in soil moisture and changes in reflectivity. Our results indicate that the Spire GNSS-R mission can detect variations in soil moisture with a performance comparable to that of CYGNSS. A median unbiased root-mean-square difference (ubRMSD) of 0.062 m3.m−3 is found for both Spire GPS and multi-constellation when using 9-km products and SMAP as the reference.


Introduction
The understanding and quantification of large-scale near-surface soil moisture storage are vital for scientific and socioeconomic applications (McColl et al. 2017).Soil moisture, an essential climate variable, affects ecosystems, food security, and human health, and influences land-atmosphere interactions, rainfall-runoff processes, and the occurrence and persistence of floods and droughts (Ochsner et al. 2013;Peng et al. 2021;SU, Singh, and Shojaei Baghini 2014).Microwave remote sensing has traditionally been the primary tool for retrieving soil moisture at global scales, as it is highly sensitive to the water content in soil and can penetrate clouds and vegetation cover (Karthikeyan et al. 2017).The Soil Moisture Active Passive (SMAP) (Entekhabi et al. 2010) mission offers global soil moisture information at 36 km with revisits every 2-3 days; moreover, an enhanced 9-km soil moisture product is made available.However, certain hydrology and climate applications demand a more refined spatiotemporal resolution (Brocca et al. 2017).
An innovative alternative to microwave remote sensing lies in leveraging signals from Global Navigation Satellite System (GNSS) as a bistatic radar (Garrison et al. 2020), known as GNSS-Reflectometry (GNSS-R).This approach involves the reception of L-band GNSS signals that have been reflected and scattered by the Earth's surface (Zavorotny et al. 2014).These reflections carry information about the reflecting surface, including soil moisture, topography, surface roughness, and overlying vegetation canopy (Chew and Small 2018;Guerriero et al. 2013).
In 2016, NASA launched the Cyclone Global Navigation Satellite System (CYGNSS), the first mission fully dedicated to spaceborne GNSS-R (Ruf et al. 2013(Ruf et al. , 2016)).Aiming to estimate ocean wind speed in the tropics, CYGNSS comprises eight small satellites.Motivated by the extensive dataset collected by the CYGNSS over land, which offers finer temporal and/or spatial resolutions compared to traditional soil moisture microwave remote sensing missions, several studies have developed methods to retrieve soil moisture from reflected signals, often assuming coherent reflections.Correlation assessments by Carreno-Luengo et al. (2018) demonstrated higher correlation for wetter soils and homogeneous vegetation, especially croplands.An algorithm proposed by Clarizia et al. (2019) coupled CYGNSS surface reflectivity data with SMAP data, achieving a global rootmean-square difference (RMSD) of 0.07 m 3 .m−3 .A further study by Chew and Small (2020) proposed a linear algorithm for CYGNSS surface reflectivity and soil moisture changes, while Yueh et al. (2020) developed a semiempirical model.Conversely, Al-Khaldi et al. (2019) considered incoherent scattering, proposing a soil moisture time-series retrieval algorithm using CYGNSS normalized bistatic radar cross section (NBRCS) and meansquare slope (MSS).Additionally, various machine learning and deep learning approaches have been developed and validated, e.g. in Eroglu et al. (2019); Jia et al. (2021); Roberts et al. (2021); Senyurek et al. (2020aSenyurek et al. ( , 2020b)), leveraging ancillary data.
Recently, Spire Global, Inc. has initiated the development of their own spaceborne GNSS-R mission.In December 2019, the first batch of two GNSS-R CubeSats were launched, followed by an additional two in January 2021 (Freeman et al. 2021).These satellites have the capability to track and record up to 30 (Batch-1) and 45 (Batch-2) simultaneous reflections (Jales et al. 2021).In this study, a thorough analysis and validation of Spire GNSS-R data for soil moisture estimation in East Australia is performed, with a focused exploration of different GNSS reflected signals.To the best of our knowledge, this study represents the first instance of such comprehensive analysis using GNSS-R data from four different navigation systems.It is pertinent to clarify that our aim is not to propose a new method for soil moisture estimation.Rather, our objective revolves around presenting and thoroughly examining the inherent potential, advantages, and limitation associated with Spire GNSS-R data.Section 2 elaborates on the Spire GNSS-R mission and underlying procedure for surface reflectivity estimation.The subsequent Section 3 delves into the methodology and the dataset used in this study.Section 4 describes and discusses the results, providing insights into the sensitivity of Spire surface reflectivity to variation in soil moisture.The validation process involves comparison against SMAP and in-situ retrievals, as well as with CYGNSS-derived estimates.The potential of multi-constellation observations is also discussed.Finally, the conclusions of this work are summarized in Section 5.

Spaceborne GNSS-R missions
This section introduces the Spire GNSS-R mission, which is the focus of this study.Additionally, we provide an overview of the main characteristics of CYGNSS mission for comparative purposes.

CYGNSS mission
The CYGNSS mission encompasses eight satellites, each equipped with four channels.This constellation orbits at an altitude of approximately 510 km and maintains an orbital inclination of 35 � .Left-hand circularly polarized (LHCP) signals from Global Positioning System (GPS) L1 C/A (1575.42MHz) are recorded.The sampling occurs pseudo-randomly due to the constant movement of both CYGNSS and GPS satellites.In a coherent reflection scenario, the footprint extends approximately 0.5 km across-track.Meanwhile, the alongtrack footprint varies based on the satellite motion and integration time; up until July 2019, the use of a 1 Hz data rate yielded an approximately 7.0 km along-track footprint.Following this, the data rate was increased to 2 Hz, resulting in a reduced alongtrack footprint of , 3.5 km (Chew and Small 2018;Gleason et al. 2020).

Spire GNSS-R mission
Spire Global, a commercial satellite constellation operator, engages in diverse GNSSrelated scientific endeavours, including radio occultation and ionosphere and space weather measurements (Buendía et al. 2023;Buendía, Tabibi, and Talpe 2022;Freeman et al. 2020).Spire is actively in the process of developing its first GNSS-R mission, currently consisting of two batches of four satellites in orbit.The first two satellites were placed in a 37 � inclination orbit at an altitude of 571 km.The subsequent pair were positioned at an altitude of 530 km in a sun-synchronous polar orbit with a local time of descending node (LTDN) at 9:30.This placement signifies that the Batch-1 satellites effectively cover the latitudinal band ranging from approximately + 37 � to −37 � , while the Batch-2 satellites provide coverage for the entire Earth.The mission's overarching objective lies in producing data to support a range of applications, encompassing soil moisture, ocean wind, sea ice, and wetlands and flood inundation.Detailed specifications for these two satellite batches are presented in Table 1.
The Spire GNSS-R mission tracks multi-constellation GNSS signals, including GPS and Quasi-Zenith Satellite System (QZSS) L1 C/A, Galileo E1-CA, and BeiDou Navigation Satellite System (BDS) B1CP, in order to enhance the coverage per receiver.In this study, level 1B (L1B) data, which have undergone correction for the power ratio between reflected and direct paths, are used.The average coherent reflection coefficient, also known as surface reflectivity, Γ rl , is computed in the L1B product as follows: where P R and P D respectively correspond to the received powers of the reflected and direct GNSS signals, N R denotes the noise power of the reflected GNSS signal, and C D and C R respectively represent the calibration signals transmitted through the direct and reflection path's radios.R I , R R , and R D are the incident, reflected, and direct path ranges, respectively.Furthermore, G T;D and G T;I respectively stand for the transmitter antenna gains pointed at the receiver and the specular point, and G Rx;D and G Rx;R respectively represent the receiver antenna gains pointed towards the transmitter and the specular point.
The power of the reflected signal, P R , is determined by the peak power of the downselected delay-Doppler map (DDM), which is coherently integrated over 1 ms, followed by incoherent averaging over 500 ms interval for land surfaces.The onboard DDM has a nominal size of 65 � 16 (Doppler � delay) pixels.A relatively large number of Doppler pixels are used to minimize the amplitude roll-off that could emerge along the edges of the DDM frequency band.Due to the constraints associated with data volume, a subset of DDM is downlinked.For Batch-1, this entails 5 Doppler and 5 delay pixels, while for Batch-2, the configuration features 5 Doppler and 8 delay pixels.The calibration parameters in Equation 1 are derived from a reference GPS-like calibration signal that is continuously injected into each front-end.This calibration signal serves as the reference for comparison with the received signals.
The specular point position is estimated onboard using the Earth ellipsoid as the reference, and then an offset is applied in the direction of the local ellipsoid normal by the geoid and digital elevation model (DEM) (Gleason et al. 2020).The Earth Gravitational Model 2008 (EGM2008) (Pavlis et al. 2012) gridded across a 1 arcminute global grid serves as the geoid model.Additionally, the Earth2014 (Hirt and Rexer 2015) in 1 arcminute spatial resolution is used as the DEM.The specular point location is then recalculated on the ground to generate the L1 product.

Methodology
The Spire L1B data used in this study was obtained through the Earthnet programmes of the European Space Agency (ESA).The dataset available for analysis covers a region in southeast Australia, delineated by the red polygon in Figure 1a, and includes several tracks spanning the entirety of east Australia.The time frame for this coverage spans from April 20th to 10 December 2021.While the primarily focus of the numerical analysis is within this specific region, it is important to note that figures will have semi-transparent areas indicating results from other tracks across east Australia.Encompassing a total of 6978 tracks, the Spire dataset consists of observations from multiple constellations: GPS (51.5%),Galileo (24.8%),BDS (18.5%), and QZSS (5.2%).Based on the International Geosphere-Biosphere Programme (IGBP) land cover classification, the selected area predominantly features grasslands, croplands, and open shrublands.Figure 1a presents the SMAP surface roughness parameter for eastern Australia, where the relatively rougher areas (depicted in light grey) primarily correlate with the local topography, notably concentrated in the eastern highlands.Figure 1b provides insight into the mean vegetation water content (VWC) during the study period, derived from SMAP.The central region exhibits sparse vegetation, whereas areas of dense vegetation are evident near the coastline within evergreen broadleaf and needleleaf forests.
For comparison, the CYGNSS v3.0 level 1 data (CYGNSS 2020a), covering the same region and time period as the Spire dataset is used.We computed CYGNSS surface reflectivity by applying the bistatic radar equation for coherent reflections (Garrison et al. 2020;Setti, Tabibi, and Van Dam 2022): where P t r is the transmitted power and λ denotes the carrier wavelength.It is noteworthy that the peak power of the DDMs is considered as P R .
The Spire and CYGNSS surface reflectivity observations are gridded to 9 km � 9 km cells on the Equal-Area Scalable Earth Grid (EASE-Grid) 2.0 (Brodzik et al. 2012).Figure 2 illustrates the number of available observations per grid cell during the period from April to December 2021, encompassing both Spire GPS and multi-constellation observations, as well as CYGNSS observations.Once again, the red polygon highlights the main study area.Observations from Spire's multi-constellation approach, which uses four Low Earth Multi-Use Receiver (LEMUR) nanosatellites, are slightly fewer than those from CYGNSS, a constellation of eight satellites in orbit.Within the scope of our study region, Spire outnumbers CYGNSS observations in 24% of the cells.Nevertheless, the presence of two polar-orbit satellites allows Spire to extend its coverage to more regions in south Australia, setting it apart from CYGNSS.
The assessment of Spire GNSS-R data proceeds in two steps: initially, an analysis of surface reflectivity properties is conducted, followed by the use of variations in surface reflectivity to retrieve large-scale near-surface soil moisture.The surface reflectivity (Γ rl in Equation 1 and Equation 2) is affected by the dielectric constant of the surface (Fuks 2001), which relies on both soil moisture and soil characteristics.Additionally, parameters such as bare soil roughness and vegetation contribute to this relationship.This relationship is described by Equation 3, as explained in De Roo and Ulaby (1994).It considers factors like the Fresnel reflection coefficient (R rl ), the standard deviation of the surface (σ), the incidence angle (θ), and the vegetation optical depth (τ), which is proportional to the VWC: The soil moisture retrieval algorithm applied in this study assumes a linear relationship between variations in surface reflectivity and variations in the near-surface soil moisture content.This approach does not account for the impact of changing surface roughness and vegetation water content over time on the estimates of surface reflectivity.Given the constraints of our limited dataset, we were unable to employ more sophisticated models.It is worth noting that Hodges et al. (2023) have demonstrated that such approximate models, if they do not outperform models using auxiliary data, still exhibit strong performance in estimating soil moisture.To calibrate our model, we used SMAP level 3 radiometer global daily soil moisture products.These products, available at both 9 km and 36 km spatial resolutions (O'Neill et al. 2018), are derived from L-band microwave radiometer observations of brightness temperature, which are then converted to soil moisture estimations at a 5-cm depth.The primary product is generated at a 36-km spatial resolution, with an enhanced 9-km resolution product produced using Backus-Gilbert optimal interpolation methods (Chan et al. 2018).
We gridded the GNSS-R observations to 9 and 36-km grids.Within each grid cell, soil moisture can be estimated through a linear regression model as follows: where SM est represents the estimated soil moisture, b is the slope of the linear regression, Γ rl denotes the surface reflectivity of each observation in the cell, Γ rl stands for the mean surface reflectivity for the cell, and SM ref is the SMAP mean reference soil moisture.The Global Surface Water Explorer (GSWE) seasonal water body product (Pekel et al. 2016) is used as reference to remove reflections over water.For our analysis, only GNSS-R and SMAP observations corresponding to the the same calendar day are considered.The Mironov model is employed to correct the observations for the influence of incidence angle.A minimum of three matching observations per grid cell is required for model calibration.
Using the outlined model, we generate daily soil moisture products at spatial resolutions of 9 and 36-km, incorporating data from Spire GPS, Spire multi-constellation, and CYGNSS.Our assessment relies on SMAP products as a reference.For validation purposes, we included in-situ soil moisture observations from 21 stations, located within or near the study area (marked as black dots in Figure 2c).These stations are part of the OzNet network (Smith et al. 2012;Young et al. 2008), and they measure soil moisture at a depth of 5 cm across grasslands and croplands in southeast Australia.By comparing the daily averages of these in-situ observations with those retrieved from GNSS-R, we validate our approach.Furthermore, we assess the GNSS-R soil moisture retrievals by comparing them with the UCAR-CU CYGNSS level 3 soil moisture products (CYGNSS 2020b).The UCAR products offer daily (and sub-daily) soil moisture retrievals at a 36-km spatial resolution.

Results and discussion
This section is divided into two parts.Initially, we compare Spire surface reflectivity with that of CYGNSS and verify their sensitivity to variations in soil moisture.In the second part, we assess the produced soil moisture maps using a set of reference products.

Spire surface reflectivity
As outlined in Section 3, the estimation of surface reflectivity is computed based on the DDMs. Figure 3 showcases examples of both Spire and CYGNSS DDMs over a 36-km grid cell within our study region.It is worth noting that the Spire downselected DDMs over land have a reduced size, consisting of only a few pixels around the peak power, while the CYGNSS DDMs have a size of 17 delay and 11 Doppler pixels.Despite this size disparity, both sets of DDMs show the presence of both coherent and incoherent reflections.In coherent DDMs, the majority of reflected power is centred at the specular point, but weaker signals are also evident with longer delays, as exemplified in the CYGNSS DDMs that represent a larger subset of the full DDM.Incoherent DDMs exhibit greater noise, capturing signals further from the specular point that reach the receiver.
Figure 4a,b present the surface reflectivity derived from Spire multi-constellation and CYGNSS observations, respectively.The results from both missions reveal a similar pattern, highlighting the main variations in land features.As expected, the strongest reflections originate from still water bodies (depicted in yellow) under calm wind conditions, showcasing the effect of specular scattering on flat surfaces.Conversely, regions with complex topography, leading to higher surface roughness (Figure 1a), exhibit weaker reflections due to diffuse scattering, causing coherent signal attenuation or even extinction (Loria et al. 2020).Despite the resemblances in the observed behaviours between both missions, a small bias of a few decibels is evident.This bias, attributed to distinct calibration protocols (Freeman et al. 2021), becomes more apparent in Figure 4c, where the histogram of Spire and CYGNSS surface reflectivity is presented.
In order to assess the sensitivity of Spire and CYGNSS surface reflectivity to soil moisture, we conducted a comparative analysis.Correspondent soil moisture retrievals from SMAP's 9 km-product are matched with observations from our study area throughout the analysis period.We divided the observations into 10 � incidence angle bins and selected only those with a VWC below 1 kg.m −2 in order to minimize the impact of vegetation canopy on surface reflectivity estimates.In alignment with previous studies (Chew and Small 2020;Freeman et al. 2021), surface roughness variations are not considered in this analysis.The results are illustrated in Figure 5, where we also showcase the expected surface reflectivity curves computed using the Mironov semiempirical model (Mironov, Kosolapova, and Fomin 2009).A mineralogy-based soil dielectric model, it estimates the soil dielectric constant based on soil moisture and texture.The results indicate an excellent agreement between GNSS-R observations and Mironov model within the 0-0.35 m 3 .m−3 range.Beyond this range, an accelerated rise in surface reflectivity relative to soil moisture is observed, possibly due to soil saturation and the subsequent increase in coherent reflections.Comparing Spire and CYGNSS observations, similarities emerge in their responses.Both missions exhibit the ability to sense variations in soil moisture.Notably, Spire observations show increasing noise as soil moisture increases.
The sensitivity of Spire's surface reflectivity to soil moisture is further assessed by examining the correlation coefficient between these variables, as shown in Figure 6 for Spire GPS, Spire Galileo, and CYGNSS.Overall, a good agreement is observed for both Spire and CYGNSS reflections, displaying similar spatial trends.The highest correlations are observed over moderately vegetated areas.However, consistent with observations by Chew and Small (2020), lower correlations are found over arid areas in central Australia.This can be attributed to the low amplitude of the GNSS-R (and to some extent SMAP) time series in these regions, as soil moisture variation is minimal throughout the analysis period, amplifying the impact of noise.Additionally, regions with excessive vegetation or mountainous terrain (Figure 1) exhibit low correlation due to diffuse scattering, particularly for Spire data due to their lower surface reflectivity relative to CYGNSS.In the southeastern study area of the country, median correlation coefficients of 0.57 for Spire GPS, 0.50 for Spire Galileo, 0.53 for Spire BDS, 0.71 for Spire QZSS, and 0.66 for CYGNSS data are found.

Retrieved soil moisture
The sensitivity of GNSS-R observations to soil moisture is quantified by the slope between surface reflectivity and soil moisture (b in Equation 4).The estimated values for Spire GPS and Galileo and CYGNSS are shown in Figure 7. Spatial variability in this slope is evident, mainly driven by differences in topography and land cover.Lower slopes generally indicate higher sensitivity; however, this trend can also be found in regions with minimal soil moisture variation over the calibration period, such as the arid central Australia.Conversely, higher slopes indicate a lower sensitivity of GNSS-R to soil moisture and are found in the densely vegetated and topographically complex southern regions of Australia.
In Figure 7, it is evident that the slopes for Spire GPS and Galileo and CYGNSS demonstrate similar spatial patterns but with varying magnitudes.This relationship is further illustrated in Figure 8 and summarized in Table 2, which show Spire and CYGNSS observations and their respective linear regressions for two selected 9-km cells within the study area.Notably, CYGNSS reflections exhibit slightly stronger slopes compared to those of Spire (Figure 4), stemming from the distinct calibration procedures mentioned earlier.Although the number of observations used for each linear regression can influence slope estimation, slight differences are evident between each system.To determine if a single slope would effectively capture the reflectivity variations caused by soil moisture changes for all four Spire-tracked constellations, Figure 9 illustrates the correlation between Spire GPS and Spire Galileo slopes across the study area, taking into account varying observations counts for each system.The figure demonstrates that when a substantial number of observations are used for slope computation, the maximum correlation is approximately 60%, indicating that Spire GPS and Spire Galileo should indeed be treated separately to achieve optimal performance.These differences are likely attributed to differences in signal modulation, bandwidth, and autocorrelation functions.
We then calibrated a different slope for each GNSS in the Spire dataset, and retrieved daily soil moisture content at both 9 and 36-km resolutions.Figure 10 illustrates the  average soil moisture during our study period for Spire multi-GNSS and CYGNSS.SMAP estimates are also included for reference.Notably, these three maps present very similar spatial patterns, depicting wetter soils areas near the coastline and drier regions within the interior lowlands of Australia.
The unbiased root-mean-square difference (ubRMSD) between GNSS-R soil moisture retrievals and the SMAP 9 and 36-km products is then computed.The results, covering the 9-km grid and the analysis period spanning from April to December 2021, are presented in Figure 11.It is clear that the three different products exhibit similar spatial trends.Lower ubRMSD values are observed in the central region, characterized by lower soil moisture variability.Conversely, higher ubRMSD values are found in densely vegetated areas and regions with complex topography (Figure 1), where the slope is higher and the sensitivity of surface reflectivity observations to soil moisture is reduced.It is important to note that the results presented in regions outside of the selected study area (illustrated using semi-transparent colours in the figure) should be treated with caution, as the correlation coefficient and ubRMSD computations are influenced by the number of observations.In those regions, only a limited number of Spire observations were available (Figure 2).
The results from the study area in southeast Australia, where the number of observations from Spire multi-constellation and CYGNSS is comparable, are presented in Table 3.The table provides details on the ubRMSD and standard deviation of the different products, along with the median revisit time, which represents the median temporal resolution of the products.It can be seen that upscaling the product from 9 to 36-km grids has led to a degradation in data quality, but it has significantly improved the temporal resolution.This deterioration could be attributed to the spatial averaging inherent in upscaling, which may not accurately capture the average soil moisture within a 36-km cell due to the limited number of 9-km retrievals.Comparing the ubRMSD of the Spire GPS and CYGNSS retrievals, they appear to be quite similar at both 9 and 36-km resolutions.While the Spire multiconstellation retrievals exhibit slightly lower performance than those derived using only  GPS data, there is an improvement in temporal resolution from 13 days to 7 days for the 9-km products.This temporal resolution is comparable to the median revisit time of CYGNSS products, which uses eight satellites in orbit.The comparison highlights that our CYGNSS 36-km ubRMSD of 0.065 m 3 .m−3 is consistent with UCAR's reported value of 0.076 m 3 .m−3 .It is worth noting that while the UCAR products use CYGNSS v2.1 data, our study uses v3.0, incorporating improvements in the estimation of GPS-related calibration parameters necessary for surface reflectivity estimation (Wang et al. 2021).
Our final assessment is focused on the soil moisture measurements from the in-situ stations, as described in Section 3. The location information and statistics, including the number of observations, correlation, and unbiased root-mean-square error (ubRMSE) for this comparison, as well as those of the SMAP 9-km product retrievals are outlined in Table 4.
Additionally, examples of time series for the in-situ observations, along with Spire, CYGNSS, and SMAP retrievals are depicted in Figure 12.Upon a thorough examination of Table 4 and Figure 12, it becomes evident that Spire and CYGNSS retrievals effectively capture variations in soil moisture at most of the in-situ stations.However, it should be noted that the performance of our soil moisture estimates is intricately tied to the performance of SMAP in the region, as it serves as the calibration reference for surface reflectivity observations.Furthermore, it is worth considering the spatial resolution of the measurements.In-situ observations represent point measurements and may not holistically represent the broader GNSS-R footprint.Taking into account all 21 stations, the median ubRMSE for Spire multi-GNSS retrievals stands at 0.057 m 3 .m−3 , which is comparable to the ubRMSE of SMAP.For CYGNSS, the median ubRMSE is estimated as 0.049 m 3 .m−3 .Although direct comparisons are complex due to the varying number of observations for each product, it is worth noting that Spire retrievals outperformed CYGNSS in 35% of the stations.YB5d.The locations and full statistics for these stations can be found in Table 4.

Conclusions
In this study, we conducted an extensive analysis of Spire multi-constellation GNSS-R data's potential for estimating soil moisture in southeast Australia.Through a thorough comparison of Spire and CYGNSS surface reflectivity, we demonstrated their capacity to detect soil moisture variations.Notably, our findings underscored the significance of treating observations from distinct navigation systems separately in order to achieve optimal performance in the context of multi-constellation data.Using SMAP soil moisture retrievals as the reference, we used a linear regression to estimate soil moisture from Spire GPS and multi-constellation and CYGNSS data.Furthermore, we validated our retrievals by comparing them to in-situ measurements.
It is important to emphasize that the results presented herein pertain to the specific attributes of the study area and the time frame from April to December 2021.Further investigations are needed to extend these findings over longer time spans and different land cover types, with a particular focus on using BDS and QZSS observations, which were limited in availability during this study.
The results of this study are particularly relevant in light of the upcoming ESA's Scout Hydrology using Global Navigation Satellite System reflections (HydroGNSS) mission (Unwin et al. 2021).With the mission slated for launch before the end of 2024, it will measure GPS L1 and Galileo E1 reflected signals.Our study showcases the comparable performance of Spire multi-constellation GNSS-R and CYGNSS in estimating soil moisture.These datasets could potentially be combined to improve spatiotemporal resolution of soil moisture retrieval.The fusion of these datasets with SMAP products remains a topic for future research.

Figure 1 .
Figure 1.SMAP (a) surface roughness and (b) mean VWC for the study area and period.The red polygon outlines the main coverage of Spire data.

Figure 2 .
Figure 2. Distribution of observations: (a) Spire GPS, (b) Spire multi-constellation, and (c) CYGNSS, for each 9 km � 9 km EASE-Grid 2.0 cell during the period from April to December 2021.Black dots in panel (c) indicate the location of in-situ soil moisture stations used for validation.

Figure 3 .
Figure 3. Examples of Spire and CYGNSS DDMs within a 36-km EASE-Grid cell in our study area, along with their corresponding location indicated on a Landsat 8 true-color image acquired during the same period.

Figure 4 .
Figure 4. Surface reflectivity comparison for east Australia using (a) Spire and (b) CYGNSS observations.(c) histogram of surface reflectivity for the same region.

Figure 5 .
Figure 5. Theoretical surface reflectivity (solid lines) from Mironov's dielectric constant model with a clay fraction of 0.25 is superimposed on (a) Spire and (b) CYGNSS surface reflectivity observations, plotted against corresponding SMAP soil moisture measurements (points), and categorized by seven incidence angle intervals.Only observations with VWC < 1 kg.m −2 are represented.A 5-dB offset is applied between different incidence angle bins for improved visibility.

Figure 6 .
Figure 6.Correlation coefficient between SMAP soil moisture and (a) gridded Spire GPS, (b) Spire Galileo, and (c) CYGNSS surface reflectivity observations.It is important to consider the number of observations in each cell when interpreting the correlation coefficient results, and this should be compared to observations shown in Figure 2.

Figure 7 .
Figure 7. Slope of the linear regression between SMAP soil moisture and (a) Spire GPS, (b) Spire Galileo, and (c) CYGNSS surface reflectivity.

Figure 8 .
Figure8.Relationship between SMAP soil moisture and CYGNSS and Spire (GPS, Galileo, BDS, and QZSS) surface reflectivity observations for two 9-km grid cells, encompassing all observations within the analysis period.The slope, number of observations, and correlation coefficient for each system are detailed in Table2.The primary land cover type in the two cells is (a) grassland and (b) cropland, respectively.

Figure 9 .
Figure 9. Slope correlation between Spire GPS and Galileo for all cells within the study region, from April to December 2021.The correlation coefficient is computed for each cell, with the minimum number of GPS and Galileo observations represented on the horizontal axis.

Figure 10 .
Figure 10.Average soil moisture content during the study period for (a) Spire multi-GNSS and (b) CYGNSS.SMAP values are shown in (c) for comparison.

Figure 12 .
Figure 12.Time series examples of daily-averaged in-situ observations, as well as 9-km Spire multiconstellation, CYGNSS, and SMAP soil moisture retrievals for OzNet stations (a) Y10, (b) YA3, and (c) YB5d.The locations and full statistics for these stations can be found in Table4.

Table 2 .
Number of observations, correlation coefficient, and slope of the different systems and constellations for the two cells illustrated in Figure8.

Table 3 .
ubRMSD, standard deviation, and median revisit time of the different products.

Table 4 .
Location information and statistics (correlation, ubRMSE, and number of observations) for the 21 OzNet in-situ stations used for validation, comparing CYGNSS and Spire multi-constellation retrievals.SMAP statistics are also included for context.The best results for each station in terms of ubRMSE are highlighted in grey.The locations of these stations are also presented in Figure2.