Assessing the performance of GNSS-R observations in drought monitoring: a case study in Jiangxi and Hunan, China

Abstract

Drought is a disaster that seriously constrains economic development and endangers human life. This paper explores the potential of Global Navigation Satellite System Reflectometry (GNSS-R) for drought monitoring, using Cyclone Global Navigation Satellite System (CYGNSS) data to monitor drought in Jiangxi and Hunan Provinces, China, in 2022. This study applies the Random Under-sampling Boosting (RUSBoost) algorithm to detect waterbodies and linear regression to retrieve soil moisture (SM). Result shows that drought in September was heaviest, with the area of Poyang Lake in Jiangxi and Dongting Lake in Hunan decreasing by 70.2% and 76.9%, respectively, compared to that in June. The variation in retrieved SM shows that the Poyang Lake Plain and Jitai Basin in Jiangxi and the Dongting Lake, Yuanjiang River, and Xiangjiang River basins in Hunan suffered from the most serious drought. The variation in retrievals shows high consistency with various reference datasets, including Soil Moisture Active Passive (SMAP) SM data and vegetation condition index (VCI). The correlation coefficient between retrieved SM and VCI is 0.93 in Jiangxi and 0.94 in Hunan.

Keywords:

• Cyclone global navigation satellite system (CYGNSS)
• drought monitoring
• global navigation satellite system reflectometry (GNSS-R)
• water detection
• soil moisture (SM)

1. Introduction

The impact of hydrological extremes, including droughts and floods, has not been effectively controlled (Kreibich et al. 2022). Among them, drought has a great impact on agricultural and ecological development. To reduce the losses of lives and economy due to drought, there is an urgent need for real-time monitoring methods. The American Meteorological Society has classified drought into four categories: meteorological, agricultural, hydrological, and socioeconomic. Traditional drought monitoring methods rely on ground station observations, which are limited by station coverage. With the development of space technology, remote sensing provides an unprecedented opportunity for drought monitoring with high spatial and temporal resolution and global coverage. However, traditional remote sensing satellites, such as Soil Moisture and Ocean Salinity (SMOS) (Kerr et al., 2016), Sentinel-1 (Ramon et al. 2012), etc., are limited due to a long revisit time, low penetrating ability, and enormous cost (Pekel et al. 2016; Shen et al. 2019). Spaceborne Global Navigation Satellite System Reflectometry (GNSS-R) is a burgeoning and promising approach for remote sensing. It applies reflected L-band Global Navigation Satellite System (GNSS) signals that can penetrate through clouds and vegetation, with low-cost and low-power receivers.

The concept of using Global Positioning System (GPS) reflected signals was first proposed by Martin-Neira (Martin-Neira 1993). Since then, GNSS-R technique has been rapidly developed with many GNSS-R satellites launched, such as UK-Disaster Monitoring Constellation (Gleason et al. 2005), TechDemoSat-1 (Unwin et al. 2016), Cyclone Global Navigation Satellite System (CYGNSS) (Ruf et al. 2016), Bufeng-1 A/B (Jing et al. 2019), Spire GNSS-R (Freeman et al. 2020) and FY-3E (Xia et al. 2021), etc. CYGNSS provides billions of measurements from the ocean and land, although it was originally designed to detect tropical cyclones (Valencia et al. 2014; Clarizia and Ruf 2016). CYGNSS data has been widely used for land remote sensing, such as water body detection (Chew et al. 2018; Gerlein Safdi and Ruf 2019; Wan et al.2019; Ghasemigoudarzi et al. 2020a; Zhang et al. 2021; Wei et al. 2023), vegetation estimation (Chen et al. 2016) and soil moisture (SM) retrieval (Maria et al. 2019; Chew and Small 2020a; Zhu et al. 2022), etc. Inspired by the success of CYGNSS, HydroGNSS which is dedicated to land applications will be launched by ESA in 2024 (Pierdicca 2021). Considering the sensitivity of CYGNSS measurements to inland water bodies and SM, this paper focuses on monitoring agricultural and hydrological droughts.

Spatiotemporal changes in inland water bodies are important for assessing hydrological drought conditions. Recently, researchers have used GNSS-R measurements to study water body monitoring under extreme weather conditions. Chew et al. (2018) found that surface reflectivity (SR) observations from CYGNSS were sensitive to changes in inundation extent. The flood areas were detected using a SR map divided by a given threshold. Compared with Soil Moisture Active Passive (SMAP) bright temperature data, the retrieved flood distribution map had a higher spatial and temporal resolution. The correlation coefficient between SR observations and brightness temperature data was −0.80 in the target area. Gerlein Safdi and Ruf (2019) proposed a water detection algorithm based on the random walker segmentation algorithm. Although it can detect inland water bodies, it mislabeled small land areas within water bodies as water. Wan et al. (2019), Zhang et al. (2021), and Wei et al. (2023) used CYGNSS data and the threshold method to monitor typhoons and extreme precipitation events in different regions. However, the optimal threshold varied due to different factors, such as topography, land surface roughness, and vegetation. Determining an optimal threshold remains a challenge for GNSS-R waterbody detection in large-scale drought studies. To solve these problems, data-driven machine learning methods are developed. Ghasemigoudarzi et al. (2020a) performed feature engineering in SR delay-Doppler maps (DDM) and proposed a model based on the Random Under-sampling Boosting (RUSBoost) algorithm. It achieved over 90% accuracy for water and land classification. The water detection accuracy on the validation set increased by 14.2% compared to that of the study of Gerlein Safdi and Ruf (2019). In this method, multiple statistical features extracted from the DDM can provide more information about the land surface and improve the robustness of the classifier. Moreover, under-sampling algorithms and ensemble learning models improve the model availability for a larger area. Hence, this study applies the RUSBoost algorithm to detect waterbodies for hydrological drought monitoring.

SM is an important indicator for reflecting the degree of agricultural drought (Dai 2013). Current spaceborne GNSS-R SM retrieval methods mainly fall into two categories: empirical models and machine learning models. A unique function between GNSS-R observations and SM is determined in the traditional empirical model. Chew and Small (2020a) used a one-order polynomial to build a univariate linear regression model for CYGNSS SM retrievals. This model serves as the official CYGNSS SM retrieval algorithm with a spatial resolution of 36 km. Good agreement was achieved with the in-situ SM measurements, with an unbiased root-mean-square error (ubRMSE) of $\text{0.049 c}{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3.$ Maria et al. (2019) considered the effects of vegetation and roughness on signals and proposed a reflectivity-vegetation-roughness (RVR) ternary linear regression model. The results showed good consistency with SMAP data, with a root-mean-square difference (RMSD) of $0.07 c{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3.$ Zhu et al. (2022) first considered the influence of surface temperature in the linear regression method, introducing an improvement of SM retrieval accuracy, especially in desert areas.

In conclusion, the GNSS-R technique is now widely used in water body detection and SM retrieval. Recent studies have explored the feasibility of flood monitoring using CYGNSS data. These studies have been conducted in different regions impacted by hurricanes and extreme precipitation events (Chew et al. 2018; Wan et al. 2019; Ghasemigoudarzi et al. 2020b; Zhang et al. 2021; Wei et al. 2023). However, the potential of GNSS-R technology for drought monitoring has yet to be investigated. Traditional drought monitoring relies on temperature and precipitation data because ground-based observations cannot provide global-scale SM information (Song 2022). With the development of remote sensing, it is possible to measure multiple surface characteristics globally, including SM and waterbody distribution, which are crucial for drought detection. Thus, benefiting from the capabilities to detect SM and water bodies, spaceborne GNSS-R can provide a method with a highly dynamic monitoring capability for global drought monitoring. Hence, the main objectives of this study are to (1) validate the feasibility of the GNSS-R technique in drought monitoring and (2) propose a drought assessment method combining water bodies and SM based on GNSS-R measurements. For this purpose, the study was carried out in Jiangxi and Hunan Provinces in 2022. This study further validates the possibility of GNSS-R measurements to detect droughts and shows how CYGNSS data can be applied to climatology and hydrology studies.

2. Methods and materials

2.1. Research data

In this study, CYGNSS data are used for water body detection and SM retrieval. The reduction in water area can characterize the intensity and range of hydrological drought. As an important indicator of crop loss, changes in SM can reflect the degree of agricultural drought. The decrease in SM can indicate the deterioration of vegetation status, indicating the exacerbation of drought. To evaluate the monitoring results and analyze the factors influencing drought, SMAP SM data, Global Precipitation Measurement (GPM) precipitation data, ERA5-Land temperature data, and the normalized differential vegetation index (NDVI) of Moderate-resolution Imaging Spectroradiometer (MODIS) products are used.

2.1.1. Study area

Jiangxi Province is located in the middle and lower reaches of the Yangtze River, China, within $[113.5°E,118.5°E]$ and $[24.5°N,30°N]$ with rich agricultural resources. The Jitai Basin in the central part and Poyang Lake Plain in the northern part serve as important commercial grain production bases. Hunan Province is located in the middle reaches of the Yangtze River. Its main water resources include Dongting Lake, the Xiangjiang River, and the Yuanjiang River. In 2022, the Yangtze River basin suffered from low precipitation since June, and drought developed rapidly since July. Continuous high temperature, low precipitation, and reduction in water flow in the Yangtze River resulted in severe drought disasters in these two provinces, causing extensive damage to agriculture and substantial economic losses.

The land-use and land-cover change products in these two provinces are derived from the Resource and Environment Science Data Registration and Publication System (Xu et al. 2023) (Figure 1). Considering the temporal resolution of CYGNSS data and the duration of this drought, this study focuses on monthly analyses of water bodies and SM changes in Jiangxi and Hunan Provinces from June to November 2022.

Figure 1. Land use and land cover of the study area. (a) Jiangxi Province. (b) Hunan Province.

2.1.2. CYGNSS data

CYGNSS is a constellation of eight small satellites launched by NASA in 2016. It collects GNSS-R measurements within approximately 38°S-38°N. The CYGNSS data used for SM retrieval in this study are level 1 version 2.1 (L1 v2.1) science data products (CYGNSS 2017), ranging from 1 January to 31 December in 2022. Due to data gaps in August and September in SMAP data, CYGNSS L1 v2.1 data from January 1, 2022, to December 31, 2022, except August and September, are used as the training set. The data in August and September are used as the test set. The data used for water detection are level 1 version 3.1 (L1 v3.1) science data products (CYGNSS 2021), covering the period from 1 June to 30 November 2022. The CYGNSS L1 v3.1 data in June are used as the training set, while the data from the remaining months are used as the test set. Compared with the previous version, v3.1 adjusts the CYGNSS science antenna gain patterns and applies a correction for the quantization loss in the noise floor correction, which improves the stability of the observables and reduces the effects of radio frequency interference. In addition, CYGNSS v3.1 product has a larger spatiotemporal coverage than the v2.1 product (Ruf et al. 2019; Wang et al. 2022). However, due to the correction for quantization loss, the sensitivity of v3.1 observations has decreased. It may cause errors in SM retrieval, which is based on a continuous geophysical model function. Therefore, CYGNSS v2.1 product is selected to retrieve SM for higher accuracy in this study. The latency of the dataset is approximately 6 days (or better) from the last recorded measurement time.

DDM is the scattered power map in the delay and Doppler dimensions. It is generated by cross-correlating the reflected signal with the local replica code at different delays and Doppler shifts (Ruf et al. 2016). The scattering bins on the surface are divided by iso-delay ellipses and iso-Doppler lines. The power in each DDM bin at a given delay and Doppler shift is the total power of the scattered signals from its corresponding physical scattering areas. The reflected signal is a combination of coherent and incoherent scattered components depending on the surface roughness. As the surface roughness decreases, the reflected signal will become more coherent. As seen in Figure 2, both the coherent and incoherent CYGNSS level 1 A (L1A) DDMs consist of 17 delay bins and 11 Doppler bins with resolutions of 0.25 chips and 500 Hz respectively (Ruf et al. 2016). In the dominantly coherent case, the power is concentrated around the specular point. In the other case, the DDM shows a clear horseshoe shape, and the power distribution is more diffused.

Figure 2. Examples of power DDM. (a) Dominantly coherent. (b) Dominantly incoherent.

To ensure the quality of the data, the CYGNSS data need to be filtered. The CYGNSS L1 v3.1 data are filtered using the following criteria:

1. Samples with incidence angle less than 15° or greater than 60° are removed (Carreno-Luengo et al. 2018);

2. The CYGNSS quality flags mentioned in Table 1 are applied to the data (Jensen et al. 2018).

Table 1. CYGNSS data quality flags in this study.

Quality Flag	Filtered out
Poor Overall Quality	No
S Band Powered Up	Yes
Small Spacecraft Attitude Error	No
Large Spacecraft Attitude Error	Yes
Blackbody DDM	Yes
DDMI Reconfigured	Yes
Spacewire CRC Invalid	Yes
DDM is Test Patten	Yes
Channel Idle	Yes
Low Confidence DDM Noise Floor	Yes
SP Over Land	No
SP Very Near Land	No
SP Near Land	No
Large Step Noise Floor	Yes
Direct Signal in DDM	Yes
Low Confidence GPS EIRP Estimate	Yes
Radio-frequency Interference Detected	Yes
BRCS DDM SP Bin Delay Error	No
BRCS DDM SP Bin Doppler Error	Yes
Negative BRCS Value Used for NRBCS	Yes
GPS Position, Velocity and Time SP3 Error	No
SP Non-Existent Error	Yes
BRCS Look-up Table Range Error	No
Antenna Data Look-up Table Range Error	No
Blackbody Framing Error	Yes

For CYGNSS L1 v2.1 data, the quality control criteria are consistent with the study of Chew and Small (2020a) except for one. We keep the data with the peak value of the analog DDMs occurring in a delay bin outside 7–10 pixels, as removing them will reduce samples by more than 40%. This result can be attributed to the large area of cities and regions with elevations greater than 600 m in these two provinces. Since CYGNSS was originally designed to detect tropical cyclones, the calculation of specular points with the geoid assumption works well for the ocean situation. However, it produces a large offset in the specular point position over land as the elevation and topographic relief increase. For example, in urban and mountain areas, the calculated position of the specular point is not accurate and the analog DDMs peak delay rows may fall outside the range of [7, 10].

2.1.3. Reference data

The Landsat global surface water (GSW) product is one of the global watermask products generated by optical sensors with a high spatial resolution of $30\times 30 \mathrm{m}$ (Pekel et al. 2016). In this paper, GSW seasonality data in 2020 were selected as the reference data for water detection and open water removal in SM retrieval. Each grid in the GSW seasonality maps represents the number of months that water was present. For our purpose, the seasonality map needs to be transformed to a binary image, where 0 represents the land and 1 represents the water bodies. The seasonality maps were processed differently in water detection and SM retrieval. For water detection, grids whose value in the seasonality map is not less than 1 will be flagged as water (1), with the remaining grids being land (0). Additionally, the chosen spatial resolution ($0.01°\times 0.01°$) is lower than that of GSW seasonality data. Therefore, to match the reference value for each CYGNSS grid, the values of $30\times 30 \mathrm{m}$ grids located in one $0.01°\times 0.01°$ cell in the binary image are averaged. If the average value of the cell is above/below the threshold, the cell will be labeled as water (1)/land (0). The threshold is set as 0.6 according to the study of Ghasemigoudarzi et al. (2020a). In SM retrieval, grids whose value in the seasonality map is greater than 6 will be flagged as water (1), and the others will be flagged as land (0).

The SMAP satellite mission, launched by NASA in 2015, is designed to monitor the global SM with a revisit time of 2-3 days. The SMAP mission consists of active radar and passive radiometer sensors operating in the L-band. The radar ceased operation after launch due to the failure of power supply and the SMAP SM retrievals rely solely on the passive radiometer. Since CYGNSS and SMAP operate in the same band, their observations have similar sensitivities to surface features, and there is a linear relationship between SMAP SM and CYGNSS measurements (Chew and Small 2018). Therefore, in this study, the SMAP Enhanced L3 Radiometer Global and Polar Grid Daily SM data (O'Neill et al. 2021) with a spatial resolution of $9\times 9 \text{km}$ are used as the reference data for SM retrieval.

The satellite precipitation product obtained by GPM based on the Integrated Multi-satellite Retrievals (IMERG) algorithm is a widely used global precipitation product. To study the temporal changes in precipitation during drought, the GPM IMERG L3 V06 product (Huffman et al. 2019) is used to obtain daily precipitation data with a $0.1°\times 0.1°$ spatial resolution over the target area.

The ERA5-Land dataset (Muñoz Sabater 2019) is a reanalysis dataset providing a consistent view of the evolution of land variables over several decades at an enhanced resolution compared to ERA5. It is utilized to obtain monthly temperature data with a spatial resolution of $0.1°\times 0.1°$ for the target area.

MODIS is a series of Earth observations instruments on board NASA’s Earth observing system satellites. It is designed to monitor global climate changes in the ocean, land, and lower atmosphere. Unlike meteorological drought, agricultural drought is not characterized by meteorological indicators, such as precipitation, but is characterized by SM and vegetation growth status, which are reflected in the vegetation condition index (VCI). NDVI data from the MODIS L3 V061 product (MOD13Q1) are used to calculate VCI (Didan 2021). The MOD13Q1 product is produced every 16 days with a spatial resolution of 250 meters. The algorithm selects the optimal pixel value from all the acquisitions within the 16-day period based on criteria such as low cloud cover, low view angle, and the highest NDVI value.

2.2. Methods

To further explore the potential of GNSS-R measurements in drought monitoring, this study adopts a high-accuracy drought monitoring method. Based on the RUSBoost algorithm (Ghasemigoudarzi et al. 2020a), we generate a monthly high-resolution ($0.01°\times 0.01°$) CYGNSS watermask map to analyze the development of drought. Based on the SM retrieval algorithm described in the study of Chew and Small (2020a), we choose SMAP SM data with 9 km resolution as the reference data for SM retrieval with a high spatial resolution ($9\times 9 \text{km}$).

2.2.1. CYGNSS observables

The received reflected signals are mainly influenced by the wavelength of the GNSS signal, angle of incidence, surface dielectric constant, and roughness (Liu et al. 2007). At L-band, the surface dielectric constant is primarily a function of the moisture content of the surface (Chew et al. 2018). Thus, the dielectric constant of the water body is quite different from that of land. In addition, the roughness of the water and land is also different, and reflections from the smooth water surface are mostly coherent. Generally, the water surface has higher dielectric constants and lower roughness, which will result in higher SR than the land surface (Chew and Small 2020b). Here, the parameters related to SR are selected as features for water detection and SM retrieval.

GNSS-R works like bistatic radar, satisfying the conditions of the bistatic radar equation. In the bistatic configuration, the processed power DDM can be expressed as follows: (1) $$P_r=\frac{P^t{G}^t{G}^r{\lambda }^2}{{\left(4\pi \right)}^3{\left(R_{\mathit{\text{ts}}}\right)}^2{\left(R_{\mathit{\text{sr}}}\right)}^2}\sigma$$(1) where $P_r$ is the received power, $P^t$ is the transmitted power, $\lambda$ is the wavelength of GPS, $G^t$ and $G^r$ are the transmitter and receiver gains, respectively, $R_{\mathit{\text{ts}}}$ is the distance between the transmitter and the specular point, $R_{\mathit{\text{sr}}}$ is the distance between the specular point and receiver, and $\sigma$ is the bistatic radar cross section (BRCS).

The received power described in Equation (1)(1) $$P_r=\frac{P^t{G}^t{G}^r{\lambda }^2}{{\left(4\pi \right)}^3{\left(R_{\mathit{\text{ts}}}\right)}^2{\left(R_{\mathit{\text{sr}}}\right)}^2}\sigma$$(1) is the sum of coherent and incoherent components (Voronovich and Zavorotny 2018) as (2) $$P_r=P_r^{\mathit{\text{coh}}}+P_r^{\mathit{\text{inc}}}$$(2) where $P_r^{\mathit{\text{coh}}}$ is the coherent power component and $P_r^{\mathit{\text{inc}}}$ is the incoherent power component. The coherent power component derived from the Friis radar equation is given as (De Roo and Ulaby 1994) (3) $$P_r^{\mathit{\text{coh}}}=\frac{P^t{G}^t{G}^r{\lambda }^2}{{\left(4\pi \right)}^2{\left(R_{\mathit{\text{ts}}}+R_{\mathit{\text{sr}}}\right)}^2}\mathrm{\Gamma }$$(3) where $\Gamma$ is SR.

The reflection from the smooth water surface is mostly coherent, and the incoherent component is approximately equal to zero, which means $P_r=P_r^{\mathit{\text{coh}}}.$ By substituting Equation (3)(3) $$P_r^{\mathit{\text{coh}}}=\frac{P^t{G}^t{G}^r{\lambda }^2}{{\left(4\pi \right)}^2{\left(R_{\mathit{\text{ts}}}+R_{\mathit{\text{sr}}}\right)}^2}\mathrm{\Gamma }$$(3) into Equation (1)(1) $$P_r=\frac{P^t{G}^t{G}^r{\lambda }^2}{{\left(4\pi \right)}^3{\left(R_{\mathit{\text{ts}}}\right)}^2{\left(R_{\mathit{\text{sr}}}\right)}^2}\sigma$$(1) , $\Gamma$ can also be found as (4) $$\mathrm{\Gamma }=\frac{{\left(R_{\mathit{\text{ts}}}+R_{\mathit{\text{sr}}}\right)}^2}{4\pi {\left(R_{\mathit{\text{ts}}}\right)}^2{\left(R_{\mathit{\text{sr}}}\right)}^2}\sigma .$$(4)

In this study, four different features, including the maximum (${\Gamma }_{\mathrm{\text{max}}}$), variance (${\sigma }_{\Gamma }^2$), kurtosis ($K_{\Gamma }$) of each SR DDM, and effective surface reflectivity ($P_{r,\mathit{\text{eff}}}$), are selected. The ${\Gamma }_{\mathrm{\text{max}}}$ of water is significantly higher than that of land, which can effectively distinguish the water area from the land. On the other hand, the surface roughness can change the statistical characteristics of the DDM. Statistical parameters, including the mean, variance, and kurtosis, are commonly used to describe the characteristics of a dataset. The variance shows how the data are distributed around their mean and can measure the dispersion of the data. Kurtosis, which characterizes the height of the probability density curve at the mean, can assess the values outside the standardized region. Studies have shown that ${\sigma }_{\Gamma }^2$ and $K_{\Gamma }$ are highly correlated with water bodies and watersheds (Ghasemigoudarzi et al. 2020a). Therefore, these three features are selected as inputs of the classifier for water detection.

Variance and kurtosis are defined as follows: (5) $${\sigma }_{\mathrm{\Gamma }}^2=\frac{1}{187}{\sum }_{i=1}^{17}{\sum }_{j=1}^{11}{\left({\mathrm{\Gamma }}_{\mathrm{\text{ij}}}-\overline{\mathrm{\Gamma }}\right)}^2$$(5) (6) $$K_{\mathrm{\Gamma }}=\frac{{\sum }_{i=1}^{17}{\sum }_{j=1}^{11}({\mathrm{\Gamma }}_{\mathrm{\text{ij}}}-\overline{\mathrm{\Gamma }}{)}^4}{187{\sigma }_{\mathrm{\Gamma }}^4}$$(6) where $\overline{\Gamma }$ is the mean of the SR DDM (7) $$\overline{\mathrm{\Gamma }}=\frac{1}{187}{\sum }_{i=1}^{17}{\sum }_{j=1}^{11}{\mathrm{\Gamma }}_{\mathrm{\text{ij}}}$$(7) ${\Gamma }_{\mathit{\text{ij}}}$ represents the SR value of the $i$th row and $j$th column in the SR DDM. ${\sigma }_{\Gamma }^2$ and ${\Gamma }_{\mathrm{\text{max}}}$ will be converted into dB.

$P_{r,\mathit{\text{eff}}}$ is the peak value of each DDM corrected for antenna gain, bistatic range, and transmit power. Studies have found a strong linear relationship between changes in $P_{r,\mathit{\text{eff}}}$ and changes in SMAP SM (Chew and Small 2018). By converting all terms to dB, $P_{r,\mathit{\text{eff}}}$ can be rewritten as (8) $$P_{r,\mathit{\text{eff}}}=10{\mathit{\text{logP}}}_r^{\mathit{\text{coh}}}-10\mathit{\text{log}}\hspace{0.17em}{P}^t-10\mathit{\text{log}}\hspace{0.17em}{G}^t-10\mathit{\text{log}}\hspace{0.17em}{G}^r-20\mathit{\text{log}}\hspace{0.17em}\lambda +20\mathit{\text{log}}\hspace{0.17em}\left(R_{\mathit{\text{ts}}}+R_{\mathit{\text{sr}}}\right)+20\mathit{\text{log}}\hspace{0.17em}\left(4\pi \right)$$(8)

As in the study of Chew and Small (2020a), the peak value of the analog DDMs is used in this paper to calculate $P_{r,\mathit{\text{eff}}}$. $P_r^{\mathit{\text{coh}}}$ in Equation (8)(8) $$P_{r,\mathit{\text{eff}}}=10{\mathit{\text{logP}}}_r^{\mathit{\text{coh}}}-10\mathit{\text{log}}\hspace{0.17em}{P}^t-10\mathit{\text{log}}\hspace{0.17em}{G}^t-10\mathit{\text{log}}\hspace{0.17em}{G}^r-20\mathit{\text{log}}\hspace{0.17em}\lambda +20\mathit{\text{log}}\hspace{0.17em}\left(R_{\mathit{\text{ts}}}+R_{\mathit{\text{sr}}}\right)+20\mathit{\text{log}}\hspace{0.17em}\left(4\pi \right)$$(8) corresponds to the maximum value of the CYGNSS dataset variable ‘power_analog’.

Among these four observables, ${\Gamma }_{\text{max}}$, ${\sigma }_{\Gamma }^2$, and $K_{\Gamma }$ are used in the water detection method. $P_{r,\mathit{\text{eff}}}$ is used for SM retrievals. To ensure a consistent spatial resolution for analysis and comparison, the features of each specular point are gridded into a certain cell based on their latitude and longitude. In water detection, considering the temporal resolution of CYGNSS data and the size of the selected area, we choose a grid resolution of $0.01°\times 0.01°$ for the feature maps and the water body distribution maps. In SM retrieval, we choose a subgrid resolution of $3\times 3 \text{km}$ and a final grid resolution of $9\times 9 \text{km}$. When multiple data points fall into the same grid, their average value is used to represent the grid.

2.2.2. Water detection method based on RUSBoost

To improve water detection accuracy, machine learning methods can be applied to data classification. Machine learning is a branch of artificial intelligence and computer science. It can achieve high accuracy classification benefits from billions of samples. Ensemble learning is a widely used branch of machine learning. It works by combining multiple weak classifiers according to certain rules to form a more comprehensive strong classifier. Boosting, bagging, blending, and stacking are commonly used algorithms for constructing ensemble classifiers.

One of the problems of inland water classification is that the land area is much larger than that of water bodies. If the goal is to maximize the overall accuracy, the water detection accuracy trained from the dataset with mostly land samples cannot be guaranteed. This kind of dataset is called an imbalanced dataset. To tackle the imbalanced dataset, oversampling and under-sampling are two commonly used data processing methods (Fernández et al. 2018). The oversampling method creates new samples in the minor class to achieve the same size as the major class. The under-sampling method eliminates samples from the major class to create a balanced dataset. Random Under-Sampling (RUS) is one of the under-sampling methods. After getting features, the next step is to classify each sample and map the water body distribution (Figure 3).

Figure 3. Flowchart of water detection method.

In this paper, the RUSBoost algorithm consisting of RUS and boosting is chosen to construct the classifier due to its efficient computational time, high accuracy, and widely available resources (Galar et al. 2012).

From Figure 4, at the beginning of the first iteration, a balanced subset is created using the RUS method. The principle of RUS is to randomly eliminate samples from the majority class to balance the dataset (Fernández et al. 2018). All samples obtain the same initial weight of $1/m$ ($m$ is the total number of samples).

Figure 4. Flowchart of RUSBoost classification.

In this study, the decision stump is used as the weak classifier because it is fast and easy to compute. The decision stump is the simplest decision tree model, which makes decisions based on only one feature (Kotsiantis et al. 2006). When the feature value is greater than the decision threshold, it is labeled as a water body (1); otherwise, it is labeled as land (0). Therefore, the result of each weak classifier depends entirely on the selection of features and decision threshold.

To determine the optimal threshold, it is necessary to judge the classification results of each threshold with certain criteria. Hence, the Gini impurity is chosen to determine the best thresholds. It is a measurement used to represent the purity of the data. Assuming that the dataset $D$ contains $k$ classes of samples, the proportion of the $i$th class samples in $D$ is $p_i,$ and the Gini impurity of $D$ is defined as (9) $$\mathit{\text{Gini}}\left(D\right)=1-{\sum }_{i=1}^k{p}_i^2.$$(9)

For the binary classification, when the proportion of one certain category is closer to 1, the Gini impurity factor is smaller, indicating higher sample purity. Conversely, when the proportion is closer to 0.5, the Gini impurity reaches its maximum. The final weak classifier is determined by iteratively calculating the Gini impurity of each feature and selecting the feature and threshold with the minimum Gini impurity.

At iteration $t$, the weak classifier is constructed as (10) $$h_t({\mathbf{x}}_i,y)=\{\begin{array}{c}{\pi }_r\left(y\right)\text{if}{x}_{i,k}>c_t^k,\\ {\pi }_l\left(y\right)\text{otherwise}\end{array}$$(10) where $k$ is the selected number of feature and $c_t^k$ is the selected threshold of iteration $t$. ${\mathit{x}}_i=[x_{i,1},x_{i,2},x_{i,3}]$ is the sample vector in the 3-dimensional feature space, including ${\Gamma }_{\text{max}}$, ${\sigma }_{\Gamma }^2$, and $K_{\Gamma }$. $y\in \left\{0,1\right\}$ is the class label. When the sample is passed to the weak classifier $t$, it will be divided into right or left splits based on whether its feature value is higher or lower than the threshold. ${\pi }_r(y)$ and ${\pi }_l(y)$ are the label proportions of the right and left splits, respectively. For example, ${\pi }_r(y)=N_r(y)/N_r$ is the label proportion of the right split, which is the ratio between the number of $y=\left\{0,1\right\}$ labeled samples in the right split ($N_r(y)$) and the total number of the right split ($N_r$). Here, $h_t({\mathit{x}}_i,y)$ represents the probability of the sample being classified as $y.$

After constructing the weak classifier, its classification error $\epsilon$ needs to be calculated, and a weight updating factor, $\alpha ,$ is calculated as (11) $$\alpha =\frac{\epsilon }{1-\epsilon }.$$(11)

Then, a new set of weights for iteration $t+1$ is computed as (12) $$D_{\text{t + 1}}\left(i\right)=D_t\left(i\right){{\alpha }_t}^{\frac{1}{2}(1+h_t({\mathbf{x}}_i,y_i)-h_t({\mathbf{x}}_i,1-y_i\left)\right)}$$(12) where $h_t({\mathit{x}}_i,y_i)$ represents the probability that the sample is correctly classified and $h_t({\mathit{x}}_i,1-y_i)$ represents the probability of misclassification. As $0<\alpha <1$, the misclassified samples will receive higher weights in the next iteration. Consequently, the weak classifier will tend to focus on classifying the samples that were misclassified by the previous classifier.

The iteration stops when the epoch reaches 150, and the final strong classifier is created by finding the label that maximizes the summation of these weak classifiers with respect to $\alpha .$ The final strong classifier can be found as follows: (13) $$H\left(\mathbf{x}\right)=\underset{y\in \left\{0,1\right\}}{\mathit{\text{argmax}}}{\sum }_{t=1}^T{h}_t(\mathbf{x},y)\mathit{\text{log}}\hspace{0.17em}\frac{1}{{\alpha }_t}$$(13)

where $\mathbf{x}$ is a feature vector of the training set. Notably, compared with the algorithm proposed in the study of Gerlein Safdi and Ruf (2019), the RUSBoost classifier determines the label of each grid only by its corresponding feature values. Therefore, when a small land area is located within water bodies, the water detection method in the study of Gerlein Safdi and Ruf will mislabel it as water concerning the surrounding pixels, while the RUSBoost method can classify it correctly.

Using Landsat-based product data as ‘ground truth’ has several drawbacks. For example, GSW data are based on optical remote sensing and are not able to detect water bodies under heavy vegetation and clouds. In addition, the distribution of water bodies is constantly changing and the use of yearly data as ground truth can also have an impact on monthly water detection. Despite the drawbacks associated with using Landsat-based product as reference for the training, GSW data are one of the main water extent data sets with high precision and high spatial resolution (Gerlein Safdi and Ruf 2019).

2.2.3. SM retrieval methods

From Figure 5, after calculating $P_{r,\mathit{\text{eff}}}$, the second step is to remove the observations in the open water area. As the presence of open water will greatly reduce the sensitivity of $P_{r,\mathit{\text{eff}}}$ to SMAP SM, the removal of specular points affected by the open water is critical. The Landsat-based GSW product is used to remove these samples. For each point, the amount of water within a $3.5\times 3.5 \text{km}$ region surrounding it is calculated, and if the water area exceeds 1%, the observation will be removed.

Figure 5. Flowchart of the SM retrieval.

For each SMAP observation, there may be several CYGNSS observations within the grid cell. If we collect all the observations of $P_{r,\mathit{\text{eff}}}$ in the grid cell and observe how $P_{r,\mathit{\text{eff}}}$ varies within that grid cell, we will find that the change in $P_{r,\mathit{\text{eff}}}$ is caused by factors, such as land cover type and topography, rather than changes in SM (Chew and Small 2020a). This problem can be effectively addressed by dividing the grid cell into smaller subgrids in which $P_{r,\mathit{\text{eff}}}$ and SM are more closely related. Therefore, samples are first divided into approximately $3\times 3 \text{km}$ subgrids and matched with the SMAP SM.

Within each subgrid, the mean values of $P_{r,\mathit{\text{eff}}}$ ($\overline{P_{r,\mathit{\text{eff}}}}$) and SMAP SM ($\overline{S{M}_{\mathit{\text{SMAP}}}}$) are calculated. After removing $\overline{P_{r,\mathit{\text{eff}}}}$ and $\overline{S{M}_{\mathit{\text{SMAP}}}}$ in that subgrid, a linear regression between their match-ups is performed. The slope of the linear regression is defined as β.

$\overline{P_{r,\mathit{\text{eff}}}}$, $\overline{S{M}_{\mathit{\text{SMAP}}}}$, and β serve as the reference values according to the training set. When retrieving the SM, we can use the following equation: (14) $$S{M}_{\mathit{\text{CYGNSS}}}=\beta \times (P_{r,\mathit{\text{eff}}}-\overline{P_{r,\mathit{\text{eff}}}})+\overline{S{M}_{\mathit{\text{SMAP}}}}.$$(14)

The mean value of the retrievals for all subgrids within each SMAP grid cell is used as the final SM retrieval.

It should be noted that in Jiangxi and Hunan Provinces, the SMAP data we selected for modeling contained a large amount of ‘not recommended for retrieval’ data. To ensure the quality of data modeling for each subgrid, there must be sufficient data for each grid. Therefore, we used all SMAP data to allow for the CYGNSS SM retrieval in the entire target area. Future works should consider carefully when utilizing SMAP data for CYGNSS SM retrieval in regions where it is known that SMAP performs poorly.

2.2.4. Drought monitoring metric

Among those remote sensing drought monitoring parameters, the remote sensing drought index is an effective indicator for drought monitoring. VCI is a widely used index due to its simplicity in calculation and ability to eliminate the effects of topographic and soil factors (Gebrehiwot et al. 2011). It was proposed by Kogan in 1995 (Kogan 1995). Given that vegetation growth is closely related to meteorological factors, Kogan assumed that the maximum NDVI occurs in good weather conditions, while the minimum value occurs in poor weather (such as drought). The difference between the NDVI and its maximum and minimum values can effectively reflect vegetation growth and drought conditions. The monthly VCI is given as follows: (15) $${\mathit{\text{VCI}}}_i=\frac{{\mathit{\text{NDVI}}}_{\mathrm{i}}-{\mathit{\text{NDVI}}}_{\mathrm{\text{min}}}}{{\mathit{\text{NDVI}}}_{\mathrm{\text{max}}}-{\mathit{\text{NDVI}}}_{\mathrm{\text{min}}}}$$(15) where ${\mathit{\text{NDVI}}}_i$ is the NDVI in month $i$ and ${\mathit{\text{NDVI}}}_{\mathrm{\text{max}}}$ and ${\mathit{\text{NDVI}}}_{\mathrm{\text{min}}}$ are the maximum and minimum values of the NDVI between January 1, 2021, and December 31, 2022, respectively. A smaller value of ${\mathit{\text{VCI}}}_i$ indicates a more severe drought condition.

3. Results

To evaluate the capability of GNSS-R in drought monitoring, the drought changes in Jiangxi and Hunan Provinces are analyzed by comparing the changes in features, water coverage, and SM distribution at a monthly timescale. The variation in ${\Gamma }_{\text{max}}$, water body area, and SM serve as indicators of the extent and severity of agricultural and hydrological droughts.

3.1. Temporal changes in waterbody distribution maps

As shown in Figure 6, in general, the ${\Gamma }_{\text{max}}$, ${\sigma }_{\Gamma }^2$, and $K_{\Gamma }$ maps show similar spatial and temporal patterns at Poyang Lake, indicating that the three features can effectively detect water bodies. Additionally, high reflectivity also exists in some areas outside Poyang Lake. As the land use map in Figure 1(a) shows, most of these areas are paddy fields and small water bodies, which also show high reflectivity. On the other hand, as $K_{\Gamma }$ can identify outliers, grid cells with high $K_{\Gamma }$ represent interconnected low-altitude land areas that could potentially contain surface water bodies.

Figure 6. Monthly ${\mathcal{\Gamma }}_{\mathcal{\text{max}}}$, ${\sigma }_{\Gamma }^2$, and $K_{\Gamma }$ maps ($0.01°\times 0.01°$) in Jiangxi Province from June to November, 2022. (a) to (f): ${\mathcal{\Gamma }}_{\mathcal{\text{max}}}$ map from June to November, 2022. (g) to (1): ${\sigma }_{\Gamma }^2$ map from June to November, 2022. (m) to (r): $K_{\Gamma }$ map from June to November, 2022.

By observing the variations between the subplots in Figure 6, changes in these features indicate changes in drought conditions. Due to the similar variation trend of these three features, we choose ${\Gamma }_{\mathrm{\text{max}}}$ maps for discussion. As depicted in Figure 6(a) and (b), the difference between June and July is not obvious, with the mean ${\Gamma }_{\mathrm{\text{max}}}$ of Jiangxi Province decreasing by only 0.34 dB, showing that the drought is not severe at that time. In August, the decrease in ${\Gamma }_{\mathrm{\text{max}}}$ by 1.23 dB indicates a reduction in Poyang Lake. Moreover, due to continuous high temperature and low rainfall, drought begin to affect the paddy fields in southern Poyang Lake (black box in Figure 6(c)). In September, the ${\Gamma }_{\mathrm{\text{max}}}$ at the confluence of Poyang Lake and the Yangtze River in northern Jiangxi Province (blue box in Figure 6(d)) also decreases, and the average ${\Gamma }_{\mathrm{\text{max}}}$ of Jiangxi Province decreases by 1.98 dB compared with June. In November, the cold air turns to South China, and the rainfall along the Yangtze River basin increases (Lei et al. 2023). The overall ${\Gamma }_{\mathrm{\text{max}}}$ of Jiangxi Province slightly increases, and the drought in Poyang Lake is relieved.

A similar analysis is also performed in Hunan Province (Figure 7). Overall, ${\Gamma }_{\text{max}}$, ${\sigma }_{\Gamma }^2$, and $K_{\Gamma }$ effectively detect the geographical location of Dongting Lake and the Xiangjiang River. It can also be seen that a large area outside Dongting Lake, particularly its northern part, shows larger feature values. Considering Figure 1(b), the large values in that area may still be caused by the presence of a large paddy field around Dongting Lake.

Figure 7. Monthly ${\mathcal{\Gamma }}_{\mathcal{\text{max}}},$ ${\sigma }_{\Gamma }^2,$ and $K_{\Gamma }$ maps ($0.01°\times 0.01°$) in Hunan Province from June to November, 2022. (a) to (f): ${\mathcal{\Gamma }}_{\mathcal{\text{max}}}$ map from June to November, 2022. (g) to (1): ${\sigma }_{\Gamma }^2$ map from June to November, 2022. (m) to (r): $K_{\Gamma }$ map from June to November, 2022.

Further analysis shows that the temporal changes in ${\Gamma }_{\text{max}}$, ${\sigma }_{\Gamma }^2$, and $K_{\Gamma }$ are closely related to drought in Hunan Province. Similarly, we only compare the changes in ${\Gamma }_{\mathrm{\text{max}}}$ maps. As shown in Figure 7(b), the Xiangjiang River region (red box in Figure 7(b)) is affected first by drought, with ${\Gamma }_{\mathrm{\text{max}}}$ decreasing by 1.17 dB compared with June. In August, the ${\Gamma }_{\mathrm{\text{max}}}$ of Dongting Lake and its surrounding paddy fields starts to decrease rapidly. In September, the area of the water body decreases sharply, particularly in the eastern part of Dongting Lake (black box in Figure 7(d)), where the ${\Gamma }_{\mathrm{\text{max}}}$ decreases by 3.5 dB compared with June. The mean ${\Gamma }_{\mathrm{\text{max}}}$ of Hunan Province decreases by 2.03 dB in September, and it is the most severe drought period. In November, the overall drought condition in Hunan Province is slightly relieved, and the mean ${\Gamma }_{\mathrm{\text{max}}}$ increases by 0.97 dB compared with September.

As one of the main water sources of Jiangxi and Hunan Provinces, Poyang Lake and Dongting Lake are selected as the target areas for water distribution maps. The temporal changes in the water body distribution map are plotted with a resolution of $0.01°\times 0.01°$ (Figure 8). (Yellow box in Figure 8(a) is Dongting Lake region and green box in Figure 8(a) is Poyang Lake region.)

Figure 8. Monthly water body distribution maps of Poyang Lake (green box in (a)) and Dongting Lake (yellow box in (a)) from June to November, 2022 ($0.01°\times 0.01°$). (a) June, 2022 (b) July, 2022. (c) August, 2022. (d) September, 2022 (black box shows the eastern Dongting Lake region and red box shows the main body of Poyang Lake). (e) October, 2022. (f) November, 2022.

In Figure 8, the blue pixels are the water body, and the rest are land. The black line is the provincial boundary, with the left gray region being Hunan Province and the right being Jiangxi Province. In this paper, water maps of both lakes in each month are obtained by quantifying the number of blue pixels in the water body distribution map. From July to August, the Dongting Lake area is severely impacted by drought, with an area reduction by 42.3% compared with June. Similarly, the area of Poyang Lake also decreases, with some small lake areas around it disappearing. From August to September, the area of both lakes is reduced sharply. The sustained high temperature results in significantly greater evaporation of the water body than precipitation. The eastern Dongting Lake area (black box in Figure 8(d)) and the main body of Poyang Lake (red box in Figure 8(d)) are nearly dry. In September, the area of Poyang Lake decreases by 70.2% compared with June, and it decreases by 76.9% for Dongting Lake. In November, the area of both lakes is slightly recovered, and the drought is alleviated.

3.2. Temporal changes in SM distribution maps

From Figure 9, the regions most severely impacted by drought in Jiangxi are the Poyang Lake Plain in the northern part of Jiangxi (black box in Figure 9(b)) and the Jitai Basin in the central region (red box in Figure 9(c)). Agriculture in the hilly areas of southern Jiangxi Province is less affected by drought. From June to July, there is a noticeable reduction in SM by approximately $\text{0.018 c}{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3$ over a large area of the Poyang Lake Plain. In August, the agricultural drought becomes more severe, and the Jitai Basin begins to be affected as well. By September, the drought further develops, and the SM decreases by $\text{0.043 c}{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3$ compared with June. The SM retrievals in these months are validated using SMAP SM data, with a mean unbiased RMSD (ubRMSD) of $\text{0.085 c}{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3.$

Figure 9. Monthly SM maps in Jiangxi Province from June to November. (a) June, 2022. (b) July, 2022. (c) August, 2022. (d) September, 2022. (e) October, 2022. (f) November, 2022.

Combined with the locations of rivers and lakes in Figure 1(b), agricultural land in Hunan Province is mostly built around water; thus, according to Figure 10, the most severe areas of this agricultural drought in Hunan are concentrated around rivers and lakes. SM in the areas around the Yuanjiang River (black box in Figure 10(b)) and the Xiangjiang River (red box in Figure 10(b)) begins to decrease in July with a reduction by $\text{0.011 c}{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3$ and $\text{0.025 c}{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3$ compared with June, respectively. From June to September, the Dongting Lake Plain (pink box in Figure 10(d)) underwent a severe loss of $\text{0.042 c}{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3$ in SM. Retrievals of these months are also validated using SMAP SM data, with a mean ubRMSD of $\text{0.077 c}{\mathrm{m}}^3/\mathrm{c}{\mathrm{m}}^3.$

Figure 10. Monthly SM maps in Hunan Province from June to November, 2022. (a) June, 2022. (b) July, 2022. (c) August, 2022. (d) September, 2022. (e) October, 2022. (f) November, 2022.

The main reason for the low accuracy in the SM retrieval is that the SMAP data we selected for modeling contained a large amount of ‘not recommended for retrieval’ data in the target area. Nevertheless, what we concern in our study is not the absolute value but the pattern of variation. As Figures 11 and 12 demonstrate, a good agreement is observed between SMAP SM and CYGNSS SM variations (the correlation coefficient is 0.91 in Jiangxi and 0.97 in Hunan). Therefore, we believe that this result can effectively detect the changes of SM in the target area.

Figure 11. Changes in the monthly mean measurements of precipitation (gray bars), temperature (pink bars), SMAP SM (green triangles), CYGNSS SM (brown stars), VCI (blue squares), and water area (yellow dots) in Jiangxi Province from June to November 2022. The correlation coefficient (r) between CYGNSS SM and VCI is 0.93.

Figure 12. Changes in the monthly mean observations of precipitation (gray bars), temperature (pink bars), SMAP SM (green triangles), CYGNSS SM (brown stars), VCI (blue squares), and water area (yellow dots) in Hunan Province from June to November 2022. The correlation coefficient (r) between CYGNSS SM and VCI is 0.94.

4. Discussions

To compare the accuracy of the RUSBoost algorithm with other algorithms, the methods based on the random walker segmentation algorithm (Gerlein Safdi and Ruf 2019) and the threshold segmentation algorithm (Chew et al. 2018) are selected as the comparison method. As in Gerlein Safdi and Ruf (2019) and Chew et al. (2018), SR is calculated from Equation (8)(8) $$P_{r,\mathit{\text{eff}}}=10{\mathit{\text{logP}}}_r^{\mathit{\text{coh}}}-10\mathit{\text{log}}\hspace{0.17em}{P}^t-10\mathit{\text{log}}\hspace{0.17em}{G}^t-10\mathit{\text{log}}\hspace{0.17em}{G}^r-20\mathit{\text{log}}\hspace{0.17em}\lambda +20\mathit{\text{log}}\hspace{0.17em}\left(R_{\mathit{\text{ts}}}+R_{\mathit{\text{sr}}}\right)+20\mathit{\text{log}}\hspace{0.17em}\left(4\pi \right)$$(8) and $P_r^{\mathit{\text{coh}}}$ in Equation (8)(8) $$P_{r,\mathit{\text{eff}}}=10{\mathit{\text{logP}}}_r^{\mathit{\text{coh}}}-10\mathit{\text{log}}\hspace{0.17em}{P}^t-10\mathit{\text{log}}\hspace{0.17em}{G}^t-10\mathit{\text{log}}\hspace{0.17em}{G}^r-20\mathit{\text{log}}\hspace{0.17em}\lambda +20\mathit{\text{log}}\hspace{0.17em}\left(R_{\mathit{\text{ts}}}+R_{\mathit{\text{sr}}}\right)+20\mathit{\text{log}}\hspace{0.17em}\left(4\pi \right)$$(8) corresponds to the CYGNSS dataset variable ‘ddm_snr’. The mean of the bottom 5% of SR values is removed to produce values in a more visually meaningful range. For the random walker segmentation classification, the method and parameters are consistent with the study of Gerlein Safdi and Ruf (2019). For the threshold segmentation classification, we did a series of experiments with thresholds ranging from 10 to 18 dB with a step of 0.5 dB. The results demonstrated that the detection accuracy of land and water bodies varied continuously with the threshold. The threshold where the overall accuracy from the experiments was closest to the RUSBoost algorithm was chosen, which was 12 dB. The overall accuracy of the RUSBoost classifier is 80.43%, whereas that of the threshold segmentation classifier is 80.22%. A comparison of accuracies of different methods is performed in Dongting Lake and Poyang Lake region in June 2022 (Table 2).

Table 2. Accuracies of inland water detection methods.

Algorithm	Class	Accuracy
RUSBoost classification (Ghasemigoudarzi et al. 2020a)	Water	62.96%
	Land	81.76%
Random walker segmentation classification (Gerlein Safdi and Ruf 2019)	Water	57.24%
	Land	89.05%
Threshold segmentation classification (Chew et al. 2018)	Water	43.02%
	Land	83.13%

As shown in Table 2, the accuracy of RUSBoost classifier in waterbody detection is improved by 5.72% and 19.94% compared to the random walker segmentation classifier and threshold segmentation classifier, respectively. In terms of land detection, the random walker segmentation classifier and threshold segmentation classifier are 7.29% and 1.37% more accurate than RUSBoost classifier, respectively. Therefore, the RUSBoost algorithm has a better performance regarding water detection and can better detect changes in water bodies during drought. However, the water detection accuracy of each method is relatively low. As mentioned at the end of Section 2.2.2, using Landsat-based product data as ‘ground truth’ has its drawbacks, which may lead to low accuracy in these two provinces. Another possible explanation for this might be that the use of static watermask data as the truth value for dynamic monitoring of water bodies is inherently problematic.

To demonstrate the ability of our method, the consistency and relevance of our result with other reference data are analyzed. The precipitation, temperature, SMAP SM, and VCI are selected as the reference data. The comparison between retrieved data and reference data is conducted in Jiangxi and Hunan from June to November 2022 (Figures 11 and 12). In Jiangxi Province, from June to July, as precipitation decreases and temperature increases, the water area, CYGNSS SM, SMAP SM, and VCI all begin to decrease (Figure 11). These six datasets except temperature reach the lowest values in September, indicating that this is the most severe period of drought. After September, all of them begin to rebound. With the increase in precipitation and decrease in temperature, the fastest rebound occurs in the water area, showing that the waterbody is more sensitive to the change in precipitation than SM and VCI. The increases in CYGNSS SM and SMAP SM are much lower. This finding implies that SM changes with a lag compared with water bodies. Water bodies are usually on the Earth’s surface and respond quickly to precipitation or evaporation changes. On the other hand, SM is distributed within the soil profile beneath the surface. The time it takes for precipitation to infiltrate the soil and reach deeper layers can cause a delay in the SM response compared with water bodies. In addition, there is a high degree of consistency between the CYGNSS SM and SMAP SM data, while the SM results are different from the SMAP SM in November. This result means that the assumption that the sensitivity of CYGNSS reflectivity to SMAP SM does not change over time may not be accurate, particularly in agricultural regions (Chew and Small 2020a). The linear relationship between CYGNSS reflectivity and SMAP SM may change in November due to the influence of drought on vegetation. Notably, although both GNSS-R measurements and SMAP SM can reflect the changes in drought, the revisit period of GNSS-R satellites is shorter than that of SMAP mission, which can provide a quick response of drought in a shorter period. Moreover, spaceborne GNSS-R measurements have higher spatial resolution than SMAP for more detailed mapping and quantification of drought areas.

As seen in Figure 12, the temporal evolution of water area and CYGNSS SM in Hunan Province is similar to that of SMAP SM, precipitation, and VCI (the correlation coefficient between CYGNSS SM and VCI is 0.94). On the one hand, the water area, CYGNSS SM, SMAP SM, and VCI all achieve the highest values in June and the lowest in September. However, the precipitation is lowest in October. A possible explanation for this might be that precipitation is not the only influential factor leading to drought. Temperature, human activities, and air humidity can also affect drought conditions. On the other hand, the VCI in August rebounds compared with July, which is not consistent with the changes in the SM, water area, and precipitation. This shows that SM and precipitation are not the only factors that affect the vegetation growth status. Considering the large area of crops in Hunan Province, this may be related to the effect of the growth period of crops with different drought tolerances on the VCI.

From the above analyses, the correlations between CYGNSS retrieved data and precipitation, temperature, SMAP SM, and VCI are significant. This finding illustrates that GNSS-R can offer a different perspective and potentially enhance the understanding of drought conditions. In addition, compared with other remote sensing methods for drought monitoring, this method is not affected by cloud and weather conditions, and the results have a higher resolution, which can provide more information for drought monitoring.

It is worth mentioning that because of the long duration of this drought, the temporal resolution chosen in this study is one month. However, future works can choose one week or smaller timescales for drought monitoring. Due to the pseudo-random sampling, CYGNSS satellites have a broader daily coverage compared with traditional remote sensing satellites. On the other hand, the pseudo-random sampling has many spatial gaps and the revisit period varies in the study area. In a 3 km grid, the revisit period is approximately 8 days near the edge of the CYGNSS latitude coverage and around two weeks near the equator (Chew 2021). Hence, a finer spatial grid is inherently associated with a coarser temporal resolution. To visualize the drought distribution more frequently, e.g. on a daily basis, a coarser spatial grid of 25 km or larger is necessary (Zhang et al. 2021). Otherwise, we can use different interpolation methods such as the previously-observed behavior interpolation (POBI) method to achieve a 3-days CYGNSS watermask in 3 km (Chew et al. 2023).

In addition, GNSS-R drought monitoring is also affected by vegetation, elevation, and surface roughness (Chew et al. 2023). For example, the presence of wind may increase the water roughness, especially in large inland lakes. As the surface roughness increases, its reflection becomes more incoherent. Consequently, distinguishing between water and land under such conditions becomes more challenging and requires further study. To obtain more accurate waterbody and SM retrievals, these factors need to be considered and corrected in future models.

5. Conclusion

In this paper, we explore the use of CYGNSS data to monitor drought in Jiangxi and Hunan Provinces in China in 2022. By using statistical characteristics, hydrological drought areas are detected, and time series maps of water body distribution are generated using the RUSBoost algorithm. Then, effective surface reflectivity is converted into SM based on the linear relationship with SMAP SM. The variation in CYGNSS SM is used to assess the damage to agriculture due to drought. The results show that the regions most severely impacted by agricultural drought are the Poyang Lake Plain in the northern part of Jiangxi and the Jitai Basin in the central region. The most severely affected areas in Hunan Province are the Dongting Lake, Yuanjiang River, and Xiangjiang River basins. The most severe drought occurred in September, with Poyang Lake decreasing by 70.2% and Dongting Lake decreasing by 76.9% compared with June. Our findings are highly consistent with SMAP SM data, GPM precipitation data, ERA5-Land temperature data, and remote sensing drought index, indicating the feasibility of CYGNSS data for drought monitoring.

In this paper, we only discussed the applicability in Jiangxi and Hunan provinces. However, considering the previous studies, the water body detection method performed well in the Congo and Amazon basins (Ghasemigoudarzi et al. 2020a) and the SM retrieval method was utilized to obtain SM retrievals for +/− 38 degrees latitude (Chew and Small 2020a). However, since CYGNSS observations are sensitive to surface roughness and vegetation, it is more accurate to assess the performance of our method before applying it to waterbodies with rough water surface or regions under dense biomass.

Based on our findings, future works may concentrate on better classification algorithms, such as deep neural networks (DNNs), to further improve the accuracy of water detection and SM retrieval. Additionally, auxiliary data, such as city information modeling, can help to reduce the impact of buildings on water detection and SM retrieval. Given the high temporal resolution and all-weather monitoring capabilities of the GNSS-R technique, future studies can also explore the potential of assimilating GNSS-R data with other remote sensing data to enhance real-time monitoring methods and predisaster warning systems for drought monitoring.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Access to this data for approved use can be obtained via the corresponding author.

Additional information

Funding

This work was supported in part by the National Key Research and Development Program of China under Grant 2016YFB0501804, in part by the National Natural Science Foundation of China under Grant 41604021 and Grant 41974031.