Spaceborne GNSS-Reflectometry for Surface Water Mapping in the Amazon Basin

To limit the abstract to 250 words, please change it to “The Amazon basin, one of the world's largest and most vital ecosystems, presents a formidable challenge for accurately mapping its extensive surface water extent and dynamic seasonal and long-term variations. This challenge arises from the region's dense vegetation and persistent cloud coverage, limiting the applicability of conventional remote sensing technologies. Spaceborne global navigation satellite system reflectometry (GNSS-R) offers a promising avenue to complement existing techniques, due to its distinctive sensitivity to surface water, coupled with its vegetation and cloud penetration and high spatio-temporal resolution. A trackwise method is developed for mapping surface water using data from cyclone GNSS (CYGNSS). Over a period of more than six years, from April 2017 to May 2023, monthly 3-km binary water/land maps covering the entire basin are produced. Comparative analyses are conducted using a comprehensive set of classification metrics with datasets derived from optical, passive microwave, and radar missions. The findings revealed that these products tend to underestimate the full extent of inundation across the basin. In contrast, CYGNSS, while exhibiting a slight tendency to overestimate surface water due to its heightened sensitivity to water and wet soils, provides valuable insights into inundation dynamics even in densely vegetated areas. Furthermore, monthly inundated area estimates are compared with water levels measured by four gauging stations across the basin. The results demonstrated a strong agreement between them and a maximum correlation coefficient of 0.74. Overall, the monthly water maps produced in this study hold significant promise for a wide range of hydrological investigations and practical applications.


I. INTRODUCTION
T HE Amazon River basin, spanning across nine South American countries (see Fig. 1(a)), is renowned as the world's largest drainage basin, encompassing approximately 6 million km 2 [1].It also boasts the distinction of being the top contributor to global freshwater and sediment transport to the oceans [2], with a river discharge exceeding 200 000 m 3 /s-an ecological heavyweight of unparalleled significance [3].Beyond its sheer scale, the Amazon's aquatic ecosystem harbors the richest freshwater fauna of any basin globally, profoundly impacting regional and global biogeochemical cycles.This impact manifests in the alteration of carbon and nutrient biogeochemistry and the release of substantial quantities of carbon dioxide and methane into the atmosphere.Moreover, the basin serves as a vital lifeline for local communities, supporting food supply, transportation, as well as crop and livestock cultivation [4].Additionally, the basin's energy needs are intrinsically tied to the utilization of existing and future hydropower reservoirs [5].Hence, the accurate mapping and assessment of inundation extent and its fluctuations across different spatial and temporal scales are imperative for comprehending and effectively managing the basin's resources.
However, the task of detecting, mapping, and measuring inundation in this region is formidable due to the substantial spatial and temporal variability of surface water.In situ measurements of flooding extent are prohibitively expensive and scarce across the expansive Amazon basin.Satellite-based optical imaging has traditionally been a primary source for large-scale surface water mapping [6], [7].While these sensors offer high spatial resolutions, their applicability is limited by cloud cover and absence of sunlight.The study of [8] underscored the challenge of persistent cloud coverage, making monthly observations of the basin unfeasible using optical sensors, with annual observations feasible for most but not all parts of the basin.Additionally, the presence of above-ground vegetation poses difficulties in accurately estimating inundation extent, often leading to underestimation.Passive microwave radiometers have also been used for this purpose, but their typical spatial resolution in the tens of kilometers range, due to a broad angular beam, is often insufficient for detecting minor floods.On the other hand, radar systems can achieve spatial resolutions in the order of tens of meters, albeit with revisit times spanning several days.However, in densely vegetated environments, radars have shown limitations in accurately mapping surface water [9], [10], [11].
These variations in sensor properties and capabilities have resulted in notable inconsistencies among inundation estimates derived from these different technologies.In a comprehensive study, Fleischmann et al. [4] conducted an intercomparison of 29 inundation datasets for the Amazon basin, encompassing estimates derived from optical sensors, radar systems, radiometers, and hydrological models.Their findings revealed a significant standard deviation of 204 800 km 2 in the maximum inundated area among these products, highlighting substantial knowledge gaps in Amazon inundation mapping with far-reaching implications across multiple applications.
An alternative approach for estimating inundation extent is spaceborne global navigation satellite system reflectometry (GNSS-R), a bistatic remote sensing technique relying on reflected GNSS signals [12].GNSS-R stands out for its high sensitivity to surface water, as demonstrated in Fig. 1(c).Other advantages include its operational capability under all weather conditions, day and night, and the ability to penetrate aboveground vegetation to some extent [13].NASA's cyclone GNSS (CYGNSS) mission, initially designed to measure ocean wind speeds for hurricane prediction [14], has extended its applications to land parameters including soil moisture [15], [16], [17], [18], above-ground biomass [19], and freeze/thaw state [20].Recent studies have also demonstrated the potential of GNSS-R for detecting surface water in different regions globally [21], [22], [23], [24], [25], [26], [27], [28], [29].These studies have generated surface water maps, representing binary water/no-water or water fraction, at different spatial and temporal resolutions, employing diverse assumptions.Evaluation through multiple approaches and reference products has revealed both strengths and weaknesses in estimating surface water using GNSS-R.
Nevertheless, the literature lacks an in-depth and comprehensive exploration of CYGNSS's capabilities and unique characteristics in detecting and mapping surface water within the Amazon basin, an area of paramount hydrological importance.In this study, we propose a trackwise method for mapping surface water using CYGNSS data, focusing on the Amazon basin and its distinctive features.Through this method, each classified track holds the potential for independent use.Consequently, the need for aggregating CYGNSS observations in both spatial and temporal dimensions to establish thresholds and generate water maps is alleviated.Additionally, the impact of GNSS transmitted power and the angle of incidence on our observations is minimized.
The rest of this article is organized as follows.Section II provides an in-depth introduction to the Amazon River basin, elucidating its key geographical and hydrological attributes.Section III offers a detailed overview of the CYGNSS mission and the wide range of datasets used in our comparative analysis.In Section IV, we delineate the methodology developed to detect and classify inland water bodies in the Amazon basin.Section V presents the results based on over six years of CYGNSS data, followed by an in-depth discussion in Section VI.Finally, Section VII concludes this article.

II. AMAZON RIVER BASIN
The Amazon River basin encompasses diverse geographic features, including the Andes, the Guyanese and Brazilian shields, and the Amazon plain [1].In Fig. 1(b), the land cover of this region is depicted according to the ESA Climate Change Initiative (CCI) 2020 map.Predominantly, the tree cover is composed of broadleaved evergreens.This basin features an extensive array of aquatic systems, which include temporally and spatially dynamic habitats such as savannas, grasslands, floodable forests, large and small rivers, and lakes [30].These ecosystems give rise to intricate flow patterns both in open areas and under dense vegetation [31].The majority of Amazon wetlands are classified as floodplain due to their vulnerability to seasonal or periodic inundation resulting from river overflow.Additionally, the region hosts substantial interfluvial wetlands, experiencing periodic inundation driven by local rainfall and runoff, with the distinguishing feature of being shallow water bodies [4].
Fig. 2 illustrates the precipitation anomaly over the Amazon basin, derived from Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) v2.0 data [32].CHIRPS v2.0 provides a quasi-global rainfall dataset by incorporating satellite imagery at a 0.05 • resolution with in situ station data.On average, the basin receives an annual rainfall of approximately 2200 mm [4], with the majority of this precipitation occurring between September and May.However, noteworthy spatial variations exist, with certain regions in the northeast, southeast, near the Amazon delta, and portions of the Andes receiving annual rainfall exceeding 3000 mm [33], [34].Due to its equatorial location, the basin displays distinct rainfall patterns in its northern and southern segments.The northern region experiences its rainy season from June to August, while the southern region's rainy season falls between December and February [1], [33].

III. DATASETS
This section provides an introduction to the CYGNSS mission and outlines the additional datasets used for comparative analysis in this study.The term "reference datasets" is avoided due to the unique inherent limitations in each dataset, a topic that will be explored in more detail in subsequent sections.

A. Cyclone Global Navigation Satellite System
CYGNSS is a GNSS-R constellation consisting of eight microsatellites orbiting at approximately 510-km altitude and 35 • orbital inclination [35].Each satellite is equipped with a multichannel GNSS-R receiver featuring a zenith antenna for receiving the direct signal, along with high-gain nadir antennas to capture surface-reflected and scattered signals.Each instrument simultaneously selects the four global positioning system (GPS) L1 signal (1575.42MHz) specular reflections located in the highest sensitivity region of its nadir antenna pattern.Initially, the signal is integrated for 1 s, later reduced to 0.5 s after mid-2019.Consequently, the full constellation records 32 simultaneous reflections every 0.5 s, effective since mid-2019.As of November 2022, the CYGNSS constellation has been operating with seven satellites.
Due to the continuous movement of both transmitting and receiving satellites, CYGNSS observations follow ground tracks that appear pseudorandom, as illustrated in Fig. 3(a) showing observed CYGNSS tracks over the Amazon basin during a one-day period.The spatial resolution of each measurement depends significantly on the scattering mechanism.The theoretical footprint of a fully coherent reflection recorded by CYGNSS is related to the first Fresnel zone and can be estimated to be approximately 0.5 × 0.5 km [36].Due to the signal integration time, the footprint elongates to an ellipse of approximately 7.0 km in the along-track direction until mid-2019, and 3.5 km afterward.This is the typical scenario for nondisturbed inland water bodies.With increased surface roughness, the glistening zone expands due to incoherent scattering, resulting in a reflection footprint that can extend beyond 20 km across-track.
The pseudorandom sampling pattern of CYGNSS also results in an imprecisely defined temporal resolution.To address this, observations are typically aggregated in grids of regular size based on the specular point's position.Fig. 3(b) and (c) depicts the one-month coverage of CYGNSS over a segment of the Amazon, with reflections aggregated to 1 and 3-km grids, respectively.As spatial resolution decreases, the revisit rate increases, albeit at the cost of potentially overlooking smaller ground features.
In this study, CYGNSS level 1 data are used, containing geolocated delay Doppler maps (DDMs).The DDMs depict the distribution of reflected and scattered signal power reaching the receiver, considering the time delay and the frequency offset caused by the relative motions of the transmitter, reflecting surface, and the receiver concerning the direct signal [12].We use the most recent publicly available version, v3.1 [37], starting from August 2018.For the period spanning from April 2017 to July 2018, we use v2.1 [38].

B. Global Surface Water
The global surface water (GSW) product [7] offers GSW maps at a 30-m resolution.By applying expert systems, visual analytics, evidential reasoning, and the multispectral and multitemporal attributes of Landsat images, it classifies pixels as open water, land, or nonvalid observation (snow, ice, cloud, or sensor-related issues).The GSW seasonality product characterizes the intraannual water distribution, revealing both perennial and seasonal water occurrences by indicating the duration in months when water consistently appeared.In this study, available seasonality maps from 2018 to 2021 are used.

C. Surface Water Fraction High-Resolution
The passive microwave surface water fraction (SWAF) product [39] was developed to monitor inundation extent over the Amazon basin.It uses L-band frequency brightness temperature data from the Soil Moisture and Ocean Salinity (SMOS) mission to map inland waters at a temporal resolution of three days and a spatial resolution of 25 × 25 km.In SWAF high-resolution (SWAF-HR) [31], coarse SWAF data were downscaled to 1 km by using high spatial resolution products: the water occurrence maps of [7], derived from Landsat optical images, and the multierror-removed-improved-terrain (MERIT) digital elevation model (DEM) [40].We use the SWAF-HR surface water occurrence frequency product developed by [4] with data from 2010 to 2016.

D. Chapman
Chapman et al. [10] generated regional inundation maps using advanced land observing satellite (ALOS) phased array L-band synthetic aperture radar (PALSAR) bimonthly data.ALOS PAL-SAR was an orbiting L-band SAR launched by the Japan Space Exploration Agency (JAXA) in a mission that spanned from 2005 to 2011.We use the "class" product, indicating for each 100-m pixel whether it has never, occasionally, or always been observed as open water or inundated vegetation.The product covers the Amazon basin and uses data acquired between 2006 and 2011.

E. Surface water microwave product series
The surface water microwave product series (SWAMPS) provides daily fractional surface water at a 25-km resolution globally from 1992 to 2020 [41].It integrates intercalibrated brightness temperature records and active microwave radar backscatter from missions such as the Advanced Scatterometer (ASCAT), Quick Scatterometer (QuikSCAT), Special Sensor Microwave/Imager (SSM/I), Special Sensor Microwave Imager/Sounder (SSMIS), and the European Remote Sensing Satellite (ERS) [42].Despite daily maps availability, the data's temporal resolution averages about 3 days.

F. ANA Gauges
A substantial portion of the Amazon basin is monitored by the Brazilian National Water and Sanitation Agency (ANA -Agência Nacional de Águas e Saneamento Básico1 ).Distributed gauging stations positioned across the basin continuously provide water level and discharge information at 15-min intervals.Table I provides descriptions of the four stations used in our study, which are also visually represented in Fig. 1(c).

IV. METHOD
This section comprises two parts: the first part introduces the proposed method for converting CYGNSS L1 observations into surface water maps, while, the second part outlines the approach for achieving consistent spatio-temporal resolution across different products and elaborates on the chosen classification metrics.

A. CYGNSS Surface Water Maps
The method proposed for classifying surface water on a trackby-track basis as outlined and demonstrated for a specific region within the Amazon basin in Fig. 4, comprises five primary steps.Initially, permanent water bodies within the designated region of interest are identified.These bodies serve as reference for calibrating subsequent tracks and observations.CYGNSS L1 v3.1 DDMs are used for this purpose, with observations converted to surface reflectivity, typically expressed in decibels (dB), under the assumption that the coherent component predominates and can be modeled effectively using the bistatic radar equation for coherent reflections [43] Γ rl [dB] = 10 log where P c represents the reflected power, P t G t corresponds to the transmitter's effective isotropic radiated power, with P t denoting the transmitted signal power and G t representing the gain pattern of the transmitting antenna.G r represents the gain of the receiving antenna, R ts signifies the distance between the transmitter and the specular reflection point, R sr represents the distance between the specular reflection point and the receiver, and λ is the carrier wavelength (approximately 19 cm for GPS L1).Surface reflectivity primarily depends on surface water, soil moisture, vegetation water content, and surface roughness over land, as well as the reflection incidence angle [17].In our method, we represent P c by the peak value of each DDM.
To define permanent water bodies, we generate monthly surface reflectivity maps over a three-year period (August 2019-July 2022), considering CYGNSS observations with an incidence angle below 40 • and a signal-to-noise ratio (SNR) above 1.5 dB.These surface reflectivity estimates are projected onto a 1-km equal-area scalable Earth (EASE) grid (version 2.0).We then compare the monthly reflectivity values of each grid cell and produce a new map showing the minimum surface reflectivity observed over this three-year period.The map, shown in Fig. 1(c), highlights areas with consistently high reflectivity, indicating permanent water bodies.Based on this, a 1-km binary map of permanent water bodies is created across the Amazon basin, classifying cells as water if the minimum surface reflectivity is higher than −23.5 dB (box i in Fig. 4).To account for our assumption that permanent waters exhibit consistently high reflectivity, we add a constraint that the three-year reflectivity standard deviation should not exceed 4 dB.
After identifying the permanent water bodies, the next step involves the classification of observations into either water (inundated) or land (noninundated) areas.In this phase, we remove the previous restriction of a 40 • incidence angle and lower the SNR threshold to 0 dB.Two possible scenarios emerge in our trackwise classification: observations within a track reflect from a permanent water body compared to our reference water mask (illustrated by track A in Fig. 4), or all observations reflect from points designated as land in the water mask (illustrated by tracks B and C in Fig. 4).In the first scenario, points marked as water establish a surface reflectivity threshold for water-reflected observations (horizontal orange line in box ii, Fig. 4).This threshold takes into account the GPS transmitted power and the signal incidence angle for that specific track.Any observations within the track exceeding this threshold are labeled as inundated, while the rest are classified as noninundated.Once all tracks, akin to track A in our example, have undergone classification, the resultant output produces a map resembling the one illustrated in Fig. 4, box iii.The specular points are then classified as land or seasonal water, in addition to the cells representing permanent water bodies.This step generates 1-km binary maps for a one-month period, classifying cells as water (permanent + seasonal) or land.
The subsequent step replicates the second step for tracks unclassified due to the absence of reflections from permanent water bodies.During this step, we replace the permanent water mask (box i) with the newly generated 1-km map for that specific month (box iii).This substitution increases the likelihood of classifying new tracks, leveraging insights from seasonally inundated regions identified in step two.For instance, track B (box iv) is illustrative in this context.It now exhibits reflections over seasonal waters, courtesy of previously classified tracks.Consequently, its classification follows the same procedure as track A.
The next step is dedicated to tracks with no reflections from water but with observations identified as land in steps two and three, exemplified by track C (box v in Fig. 4).A new threshold is established, categorizing unclassified specular points with reflectivity below the defined threshold as noninundated.Any observations with surface reflectivity higher than the threshold (typically less than 5% of the dataset) are discarded as unreliable.The permanent water bodies are then overlaid onto each monthly map.
With the 1-km grid and one month of CYGNSS data, we cover approximately 25%-30% of the Amazon basin area.In the final step, we perform nearest neighbor interpolation to upscale the monthly maps to a new 3-km spatial resolution.This coarser grid enables coverage of more than 90% of the study area with one-month temporal resolution.Each 3 × 3-km cell is classified as water if the majority of the 1-km cells within it were flagged as water in steps one to four.
By applying this methodology, we generated monthly 3-km binary water/land maps for the Amazon basin from April 2017 to May 2023, switching from CYGNSS v2.1 to v3.1 in August 2018.

B. Comparison
To perform comparative analysis among difference products, we initially regridded the finer spatial resolution GSW (30-m), SWAF-HR (1-km), and Chapman (100-m spatial resolution) products to our 3-km grid.As proposed in [29], we found the fractional water in each cell, which is defined as the number of high-resolution cells with water divided by the total number of high-resolution cells in each of our 3-km cells.To convert fractional water information into water or land information, we defined an individual threshold for each product by minimizing the difference in the total detected inundated area between the original and the 3-km resolution products.Consequently, we obtained thresholds of 0.30, 0.40, and 0.30 for GSW, SWAF-HR, and Chapman, respectively.
However, considering variations in the temporal coverage of each product, we established a comparative framework based on the long-term maximum and minimum inundation estimates, encompassing the entirety of available data.The maximum inundated area refers to all the pixels that were detected as water at least once in each product.In contrast, the minimum inundated area comprises grid cells that were consistently flagged as water.
Given that none of the products offers a flawless representation of surface water in the Amazon basin, our approach initially used SWAF-HR's minimum and maximum inundation maps as reference.We evaluated the results from CYGNSS, GSW, and Chapman against these references.Subsequently, we employed the Chapman product as another reference and assessed the performance of CYGNSS, GSW, and SWAF-HR against it.
For the assessment, we used the metrics and guidelines for remote sensing classification as outlined in [44].These metrics are defined with respect to the confusion matrix, which includes true positives (TP, representing cells classified as water in both the reference and the test products), true negatives (TN, denoting cells classified as land in both the reference and the test products), false positives (FP, indicating cells classified as water in the test product but as land in the reference), and false negatives (FN, signifying cells classified as land in the test product but as water in the reference).The binary assessment metrics include overall accuracy (OA), recall, specificity, precision, negative predictive value (NPV), F1 score, and intersection-over-union (IoU), which are, respectively, given by [44] OA = TP+TN TP+TN+FP+FN (2) The OA is anticipated to yield favorable values because there are significantly more land grid cells than water cells in the dataset.Recall represents the proportion of true water cells that were detected by the test product, while specificity represents the proportion of true land pixels that were detected by the test dataset.Precision indicates the probability of the test dataset to correctly detect surface water compared to the reference data, and the NPV quantifies the likelihood of the test dataset to correctly detect land compared to the reference data.The F1 score serves as the harmonic mean between precision and recall, providing a consolidated indicator that encapsulates both precision and recall in a single measure.The IoU quantifies the degree of overlap or agreement between the predicted and reference inundated regions.Additionally, we included the Matthews correlation coefficient (MCC), a correlation coefficient that ranges from −1 to +1 and is calculated as follows [45]: Given the overlap between the SWAMPS product and our maps from April 2017 to December 2020, we performed a numerical comparison of the total inundated area on a monthly basis.Furthermore, we conducted a comparative analysis between our monthly inundated area estimates and the water levels of four rivers in the Brazilian portion of the Amazon.Since ANA gauge data are sampled every 15 min, our comparison was based on the monthly mean water levels.

V. RESULTS
The results are presented in three main parts.First, we conduct a comprehensive comparative analysis of minimum and maximum surface water maps derived from four different products.Following that, we proceed to compare our monthly estimates of inundated areas.In this comparison, we first compare them against the SWAMPS product.Finally, we evaluate our estimates against data from four gauge stations situated in the Amazon basin.

A. Minimum and Maximum Inundation
Fig. 5 depicts the maximum and minimum inundation derived from the CYGNSS, GSW, SWAF-HR, and Chapman products.A visual similarity in the maximum inundation extent is evident among the CYGNSS, SWAF-HR, and Chapman products.However, the GSW product, due to its inability to sense water beneath vegetative cover, frequently omits most of the wetlands and smaller rivers from its mapping.Additionally, during the rainy season, characterized by prevalent cloud cover, is when the region typically experiences its maximum surface water levels.
In terms of minimum inundation extent, the products were effective in detecting larger rivers, but most of the smaller ones remained undetected, except for CYGNSS, which outperformed the other products by detecting a greater number of smaller water bodies.CYGNSS, SWAF-HR, and Chapman products effectively distinguished between the wet and dry periods of seasonally inundated wetlands.
Our CYGNSS-derived method tends to classify the majority of the Branco floodplains in the northern part of the basin (see Fig. 1(c)) as consistently inundated, which differs from the results obtained by the other three products.The presence of extensive low vegetation cover in the region, as opposed to the densely forested surroundings (see Fig. 1(b)), contributes to the strong CYGNSS minimum surface reflectivity, as seen in Fig. 1(c).It is important to note that our method may potentially misclassify certain grid cells as water when they actually represent wet soil or inflate the size of small water bodies due to the CYGNSS strong sensitivity to surface water.Nevertheless, Fleischmann et al. [4] reported that large discrepancies between different products are observed in this region.They emphasize the need for further investigations to gain a comprehensive understanding of the actual extent and dynamics of inundation in the Branco floodplains.Thus, it remains uncertain which product Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE II SURFACE WATER DETECTION METRICS FOR THE DIFFERENT PRODUCTS AND THEIR MAXIMUM AND MINIMUM INUNDATION EXTENTS
offers the most accurate representation of inundation extent in the area.
Disagreement is also noted over the Bolivian floodplains (see Fig. 1(c)).As noted in [4], flooding in the region is predominantly shallow and occurs in vegetated areas, primarily savannas and grasslands.Furthermore, it exhibits significant year-to-year variability.In this particular region, they reported a standard deviation of 52 400 km 2 in inundation extent across the datasets they examined.
In comparing the selected products, Table II presents a summary of performance metrics (see ( 2)-( 9)) for both maximum and minimum inundated areas.The difference in inundated area, relative to the passive microwave-based SWAF-HR and the synthetic aperture radar (SAR)-based Chapman products as references, is assessed.In every scenario, the GSW product consistently demonstrated better agreement with the reference datasets in terms of precision and specificity.GSW's precision of 0.96 for the maximum inundation extent, when compared to SWAF-HR, implies that GSW's identification of water grid cells carries a 96% probability of being accurate; the 1.00 specificity in all cases indicates that GSW perfectly identified all land cells in the reference products.These results primarily stem from the product's limitation of mapping only open waters, resulting in an extremely low number of FP and justifying the excellent precision and specificity results, which do not incorporate the high number of FN of the product.However, it identifies 80% less water in maximum inundation and 30% less water in minimum inundation when compared to both SWAF-HR and Chapman products.
When evaluating maximum inundation extent with SWAF-HR and Chapman products as references, CYGNSS excels in recall, NPV, F1 Score, IoU, and MCC.It is important to note the inherent challenges posed by the absence of a universally accepted ground truth for cross-product comparison, and this evaluation is based on the chosen reference datasets.The recall values are notably higher than those of the other products, signifying CYGNSS's ability to accurately identify a greater number of water grid cells.With an NPV value of 0.97, CYGNSS demonstrates a 97% probability of accurately recognizing land grid cells during maximum inundation periods.The IoU value of 0.52, when compared to SWAF-HR, signifies a 52% overlap in the water cells flagged as such in either CYGNSS or SWAF-HR.The satisfactory performance can be attributed to the spaceborne GNSS-R's capability to detect water in complex landscapes featuring open water, inundated vegetation, and dry land, leading to a substantial number of TP.Compared to SWAF-HR, CYGNSS exhibits an increment of approximate 27% in water detection.Contrasted with Chapman, it manifests an augmented inundated area, demonstrating an increase of approximately 56% at the maximum extent and around 19% at the minimum extent.
In summarizing this initial analysis, Fig. 6 delineates the agreement and disagreement zones between our CYGNSS estimates and the SWAF-HR and Chapman products, particularly concerning the maximum inundation area.Noteworthy is the robust agreement among the three products in larger water bodies, showcasing CYGNSS's enhanced capability in detecting smaller water bodies.As elucidated and examined earlier, the Branco and Bolivian floodplains emerge as regions marked by significant disparities.The question of which product holds greater accuracy in these areas remains uncertain.

B. Monthly Inundation Extent
In this section, we delve into an analysis of the monthly inundation estimates.Fig. 7 presents the monthly data spanning from April 2017 to May 2023, accompanied by corresponding precipitation rates.We observe a strong correlation between inundated area and precipitation, consistently showcasing an expansion of surface water toward the end of each rainy season.Conversely, the minimum extent occurs just before the onset of a rainy season, at the end of each dry period.The total surface water area fluctuates between approximately 404 100 to 841 200 km 2 , representing 6.9%-14.4% of the entire basin area, respectively.These findings align with those of [4], which reported a maximum inundation extent on the order of 490 300 ± 204 800 km 2 .It is worth noting that nearly all datasets examined in that study exhibited a consistent tendency to underestimate inundation.Importantly, our transition from CYGNSS v2.1 to v3.1 in August 2018 (red vertical line in Fig. 7) did not appear to impact our trackwise inundation estimates.Fig. 7 also highlights an increased inundated area from 2017 to 2019, coinciding with the adoption of CYGNSS 1-Hz data.With the implementation of a 2-Hz integration rate, it is likely that the impact of smaller water bodies on surface reflectivity is limited to a reduced set of observations, thus, enhancing our method's ability to differentiate between water and land reflections and reducing estimated inundation extent.Notably, a similar trend is observed in the SWAMPS estimates during this timeframe.Given SWAMPS' daily inundation extent reporting, our approach involved aggregating the maximum inundation values for each month.The scatterplot in Fig. 8 presents the monthly inundation area reported by both our product and SWAMPS from April 2017 to December 2020, with each year represented by a different color.In both products, a significant surge in inundated area is evident during the years 2017 and 2018.Furthermore, there exists a robust seasonal consistency between both products, marked by a high correlation coefficient of 0.82.However, it is worth noting that CYGNSS estimates magnitude surpass SWAMPS values by a factor of over two.Previous studies have demonstrated SWAMPS has a relatively small annual amplitude and limited sensitivity to detect small fractional flooded areas, flooding under dense vegetation, and short-duration flooding [4], [41].

C. River Water Level
The final phase of our analysis involves a comparison between our monthly inundation estimates and data from four gauge stations distributed across the Amazon basin, as previously described in Table I and Fig. 1(c).Initially, we computed CYGNSS monthly inundated areas within square regions approximately 200 km in size, centered around each gauge.We then downsampled river water levels from 15-minute intervals to monthly data, calculating the mean level for each respective period.
Fig. 9 illustrates the averaged water levels measured by the gauges at these four sites, represented in purple, alongside the CYGNSS inundated area estimates shown in continuous blue lines.In terms of water levels, we did not observe substantial amplitude variations over this period, except for the Boa Vista station (see Fig. 9(b)), which exhibited a clear trend of water levels rising at a rate of 32 cm/yr.The CYGNSS inundation estimates surrounding these four sites were generally higher in 2017 and 2018, most likely due to the factors discussed in Section V-B.Upon examining the CYGNSS time series, we notice subtle spikes in the inundated area, deviations typically unexpected in a realistic scenario.These fluctuations may be attributed to uneven CYGNSS sampling from month to month, leading to variability in the inundation maps.To mitigate this effect, we applied a 3-month low-pass filter to smooth the CYGNSS time series, represented by the dashed blue lines in Fig. 9.
Remarkably, we observe a good overall agreement between the water level measurements and our inundation extent estimates at the four selected sites.The Rio Branco gauge (see Fig. 9(a)), located in the central part of the Amazon basin, exhibits peak water levels between January and March, which aligns with our maximum inundation estimates for the region.Similarly, the other three sites (see Fig.  Before smoothing the time series, we found correlations between water level and inundation extent of 0.68, 0.74, 0.69, and, 0.58 for the Rio Branco, Boa Vista, Manaus, and Itacoatiara sites, respectively.Following the application of a low-pass filter, the correlation increased to 0.74, 0.70, and 0.59 for the Rio Branco, Manaus, and Itacoatiara sites, respectively.For the Boa Vista gauge, characterized by more intricate water level fluctuations throughout the year, the application of the low-pass filter resulted in a slight reduction of the correlation to 0.73.This suggests that our unfiltered estimates provide a more accurate representation of the water variations in the area.The weakest correlation, standing at 0.59, pertains to the data derived from the gauge placed along the Amazon River.Consistently mapping surface water in a river of such considerable width presents significant challenges that will be later.It is also important to note that the extent of inundation is influenced by a complex interplay of factors, including river water levels, precipitation, topography, and soil moisture.Therefore, while variations in water levels observed by the gauges do not always directly correspond to changes in the total inundation extent, these results are indeed promising, demonstrating strong correlation in most cases.

VI. DISCUSSION
The monthly surface water maps generated through the described methodology allow for synthesizing the results into a single water occurrence map, aggregating all estimates from April 2017 to May 2023 (see Fig. 10).This product effectively communicates the frequency with which each 3 × 3-km grid cell is inundated, ranging from 0 (always dry) to 1 (or 100%, permanently underwater).Dark blue cells surrounding permanent water bodies tend to remain inundated for almost the entire period, while lighter shades of blue signify seasonal or sporadic inundation.A detailed analysis of Fig. 10 reveals that, in a few instances, areas known to be consistently submerged, such as the central parts of rivers, are represented with a flood frequency slightly below 100%.Considering the Amazon River's considerable width, which can span over several kilometers [46], we hypothesize that certain grid cells may have been erroneously classified as land in a few of our maps.This misclassification could be attributed to wind-disturbed waters, causing signal scattering and subsequently reduce surface reflectivity.
In our comparative analysis, we identified both agreements and disparities when examining the water maps generated from CYGNSS data in contrast to those derived from other sensors.Notably, CYGNSS tended to overestimate surface water when compared to the other products.We hypothesize that some of the excess-detected water by CYGNSS may be obstructed by dense vegetation and is not sensed by any of the other products; another part of this excess surface water, however, may be the result of the coarser spatial resolution of our CYGNSS product.Enhancing CYGNSS's spatial resolution could be achieved by reducing the current 3-km grid size.However, such an adjustment would require a tradeoff, leading to a reduction in the product's temporal resolution, potentially causing it to overlook small, short-term inundations.Considering the original 1-km grid and the 2-Hz CYGNSS data available since July 2019, we find that approximately 52% of the entire Amazon basin is encompassed within a three-month data window.Near the equator, where the Amazon basin is located, CYGNSS reflections occur less frequently due to the arrangement of CYGNSS and GPS satellites.A better revisit time and spatial coverage are expected in higher latitudes within the CYGNSS operating range, which could lead to an improved spatio-temporal resolution in water mapping in those regions.
Additionally, the pseudorandom sampling pattern of CYGNSS is not uniformly distributed over the one-month period during which we aggregate the data.This uneven distribution could lead to CYGNSS missing smaller flooded areas or inaccurately interpreting short-term inundated regions as being submerged for the entire month.It is also important to acknowledge the variations in product timeframes, as described in Section III, which may contribute to long-term variations in the amount and distribution of surface water.Nevertheless, it is evident that CYGNSS brings new information to the existing inundation mapping remote sensing techniques.

VII. CONCLUSION
The Amazon basin, as the world's largest, holds immense environmental and societal significance.Accurate mapping of the surface water extent and its seasonal variations is critical.Leveraging CYGNSS's high sensitivity to surface water, this study proposed a trackwise method for mapping inundations across the Amazon River basin.Our approach minimizes the influence of signal incidence angle and GPS transmitted power on surface reflectivity estimates by independently categorizing each track with respect to permanent water bodies.Unlike traditional remote sensing with predefined swaths, CYGNSS follows a pseudorandom pattern, resulting in variable grid sizes and a tradeoff between spatial and temporal resolution.Employing a 3-km grid, we generated monthly surface water maps spanning the entire Amazon basin over six years, from April 2017 to May 2023.These water maps offer substantial potential for various hydrological studies and practical applications.While our observations were gridded for consistent water maps, we can enhance spatial resolution by classifying each track individually.
We conducted comparisons between the maximum and minimum long-term inundation extents using SAR, passive microwave, and optically derived products, highlighting the differences between them.Furthermore, we compared our monthly inundated areas estimates with the SWAMPS product and river water level measurements from four gauging stations.We identified and discussed both areas of agreement and limitations within each product.CYGNSS demonstrates enhanced capabilities in sensing inundated vegetation, small water bodies, and shallow Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
waters.However, it tends to overestimate inundation extent due its footprint size and its strong sensitivity to water.our study, we applied a range of products in a comparative analysis.Yet, the prospect of combining them into a more robust product, mitigating individual weaknesses while capitalizing on their strengths, remains a subject for future research.Recently, we expanded our method's application to produce surface water maps in various global regions, including flatter and/or less vegetated landscapes.Preliminary assessments suggest the potential applicability of this proposed method across CYGNSS's entire coverage area, especially when integrated with permanent water bodies masks from external sources.

Fig. 1 .
Fig. 1.Amazon River basin.(a) Geographical location within South American countries.(b) Land cover representation based on the 2020 ESA CCI map, illustrating the distribution of water bodies, floodplains, forested, and nonforested areas across the basin.(c) CYGNSS 1-km gridded minimum surface reflectivity spanning from August 2019 to July 2022.The Branco and Bolivian floodplains are highlighted in green, while the white dots indicate the location of four river level gauges.

Fig. 2 .
Fig. 2. Precipitation anomaly over the Amazon basin spanning from January 2006 to May 2023, using historical data as a reference from 1981 to 2023.The data are derived from the CHIRPS v2.0 precipitation product.

Fig. 3 .
Fig. 3. Observed CYGNSS tracks over the Amazon basin for one day on January 1, 2022 (a), Panels (b), and (c) display the coverage within the region outlined by the red rectangle shown in (a) over one month in January 2022, with the data aggregated to 1 and 3-km grids, respectively.

Fig. 4 .
Fig.4.Surface water mapping using CYGNSS L1 data algorithm concept.For a detailed explanation, please refer to Section IV-A.

Fig. 7 .
Fig. 7. Monthly GNSS-R inundated area retrievals of the Amazon basin and precipitation rate from April 2017 to May 2023.The red vertical line indicates the transition from v2.1 to v3.1, and the green vertical line marks the commencement of CYGNSS operations with a 2-Hz integration rate.

Fig. 8 .
Fig. 8. Scatterplot illustrating the total inundated area of Amazon basin derived from CYGNSS and SWAMPS monthly data spanning from April 2017 to December 2020.Each color represents a different year, and the black line represents the linear best fit.
9(b)-(d)), situated in the northern part of the basin, reach maximum water levels from May to July, corresponding to our maximum surface water estimates in the same region.

Fig. 9 .
Fig. 9. Inundated area and relative water level for gauges (a) Rio Branco, (b) Boa Vista, (c) Manaus, and (d) Itacoatiara.The location of these sites are described in Table I.The dashed blue lines represent the inundated area, smoothed with a 3-month low pass filter.Please note the difference in the vertical axes ranges.

Fig. 10 .
Fig. 10.Amazon basin flood frequency computed on the CYGNSS 3-km monthly surface water maps spanning from April 2017 to May 2023.