Comparing Satellite Soil Moisture Products Using In Situ Observations over an Instrumented Experimental Basin in Romania

Abstract

This study assessed the performance of different remotely sensed soil moisture products with in situ observations; six profile probes for the water content monitoring were selected, operating during 2016–2021 from the Voineşti Experimental Basin in the Romanian Subcarpathian region. The reliability of satellite observations has been analyzed on both single ground-based observation points and spatialized information, considering near-surface and root-zone soil moisture data. The physics-based index (HCI) and some statistical tests widely used in inter-comparison analyses have been computed. The study of HCI highlighted that the SMAP SP_L4_SM products have shown the best performances considering the near-surface and root-zone data evaluations. The comparison of SWI1km observations with in situ data produced good results for single-point and spatialized soil moisture estimations acquired at different depths over the experimental basin. The SSM1km and SMAP L2_SM_SP products exhibited the lowest performances. The results contribute to the validation of satellite products of surface and root-zone soil moisture in the Subcarpathian region, helping to provide information in an area not monitored by the International Soil Moisture Network. The findings offer valuable insights into evaluating the performance of satellite soil moisture products in the Romanian region.

1. Introduction

Knowledge of surface soil moisture (SSM) and root-zone soil moisture (RZSM) is crucial for many purposes, such as flood forecasting, the rainfall–runoff relationship, landslide prediction, drought monitoring and irrigation scheduling, etc. [1,2,3]. Therefore, in the context of ongoing climatic changes, obtaining reliable soil moisture observations is essential to understanding hydrological, ecological, and biogeochemical processes [4,5,6].

Obtaining high-resolution, continuous and spatially distributed soil moisture information is challenging due to its large variability in space and time (vegetation dynamics, soil properties, land use, topography, etc.) [7,8,9,10,11]. The in situ measurement methods provide points and profiles of the soil water content using Time-Domain Reflectometry (TDR) and Frequency-Domain Reflectivity (FDR) probes; the latter requires an accurate calibration depending on the soil type and physical–chemical characteristics of the pore fluids [12,13,14,15,16,17]. However, ground-based measurements have limited spatial coverage and are time-consuming. Due to the high variability of soil moisture, it is challenging to obtain soil moisture estimates over large areas, such as river catchments [18]. In situ measurements, even if in a limited number, remain essential for calibrating other automatic or semi-automatic methods for determining water content-related parameters.

In recent decades, space-borne remote sensing has been increasingly used to fulfill the need for quantitative measurements of climatic variables. Satellite active or passive observations, operating in the different microwave bands (L-band, C-band and X-band), provide numerous soil moisture products differing in spatial–temporal resolution and accuracy depending on sensor characteristics and retrieval algorithms [19,20]. Satellite soil moisture data are freely available in most world regions, and their use is increasing. Ma et al. [21], based on global ground-based observations, reported that different spatial–temporal resolutions of the Soil Moisture Active Passive (SMAP) and the Soil Moisture and Ocean Salinity (SMOS) products performed better than other soil moisture retrieval products.

Usually, satellite soil moisture data are compared with data from a single probe with helpful support from in situ observations of the International Soil Moisture Network—ISMN [22,23,24]. The ISMN [25] presents 2842 stations located all over the globe; ground-based observations are centralized in a few locations, and often only the surface soil moisture is measured. Al-Yaari et al. [26] evaluated the performance of microwave-based soil moisture retrieval products, comparing them to more than 1000 ground-based measurements from the ISMN network, highlighting that climate, land cover, and differences in spatial–temporal sampling affect the performance of the soil moisture products.

For the Romanian region, the Assessment of Satellite-Derived Soil Moisture Products over Romania (ASSIMO, 2015 [27]) aims to acquire surface soil moisture data (0 to 0.05 m) on twenty sites to validate satellite soil moisture-derived products. On thirteen of these, it is possible to download the data series using the ISMN data viewer [28]. Although the soil moisture network is distributed in areas with different topographical, land-cover and climate conditions, no measurement points are in the Subcarpathian region and no focused studies analyzed the spatialization of ground-based data over the satellite grid pixels or catchments. This topic has been more and more investigated in recent years in the literature to study the reliability of satellite data, which expresses the soil moisture (SM) and or soil degree of saturation over the entire grid pixel (e.g., [29,30,31]). In situ observations provide the most accurate estimates of soil moisture. Although they are limited in terms of spatial extent, they should be used preferably for calibrating and testing other products, such as satellite observations or model simulations (e.g., [8]). In this framework, the present study—thanks to soil moisture observations collected from six profile probes in an experimental catchment (Voineşti Experimental Basin, VEB)—aims to evaluate the performance of different high-resolution products of the Copernicus Global Land Service (CGLS) and NASA Soil Moisture Active Passive (SMAP). In detail, the main objective is to present a comparison analysis of different satellite products with ground-based data to investigate their reliability both in the shallow portion of the soil and in the root zone. The findings provide information for an uninvestigated area of the Romanian region, especially regarding the performance of satellite data in the root zone. The results are discussed with those obtained in previous European ecoregions and the Romanian region studies, considering the results from the Romanian Soil Moisture Network (RSMN) supported by the Romanian National Meteorological Administration (RMNA). The findings contribute to the knowledge of the reliability of satellite soil moisture data in different climatic, lithological and land-cover conditions and the evaluation of remote sensing moisture data at the catchment scale.

2. Materials and Methods

2.1. Study Area

The Voineşti Experimental Basin (VEB) is located at an altitude of about 500 m a.s.l. in the Subcarpathian curvature, Romania (45°05′05″N 25°15′09″E, Figure 1) and extends for about 0.74 km^2. This experimental site was completed in 1963 and is managed by the National Institute of Hydrology and Water Management (NIHWM) of Romania. The purposes of the VEB are (i) to evaluate the rainfall–runoff relationship, (ii) to obtain data on the effect of land use on the flow processes, and (iii) to study evapotranspiration phenomena for the soil water budget [32,33,34,35,36]. The site is located in a moderate slope area (up to 5%) and is characterized by a moderate temperate continental climate with an average annual rainfall of about 700 mm (1981–2021), mainly from April to September. The average air temperature is 9.7 °C, with July being the month of the maximum mean temperature (17.9 °C) and January the month of the minimum mean temperature (−2.2 °C). From October to March, the VEB is exposed to low temperatures, snow-covered and with frozen soil, making it difficult for staff to access. Continental fluvial Pliocene sediments belonging to the final sedimentation phase of the Dacian basin characterize the study area [37,38].

To acquire reliable data for the hydrological studies, the VEB is equipped with ten plots characterized by different areas (from 10 to 900 m²) and land cover (Figure 2). All plots are built with concrete edges and include collection channels composed of gutters and underground pipes located at different depths (10 cm, 40 cm, and 130 cm), used to collect water in calibrated tanks placed downstream of the plots [39]. In addition, the area is equipped with a rain gauge (self built and assembled by NIHWM of Romania) and eight calibrated profile probes (PR2/6 of the Delta-T Devices, Cambridge, UK) for soil volumetric water content (θ) estimations, providing data up to different depths (10, 20, 30, 40, 60 cm) acquired manually between 9 am and 12 am local time [40]. The PR2/6 profile probe works at 100 MHz (similar to FM radio), generating an electromagnetic field, the changes in which (mV) are recorded and empirically related to the square root of soil dielectric permittivity (√ε) by a six-order polynomial function. Detailed information about the operation and calibration of the PR2/6 probe can be found in [15,17].

In the VEB area, three plots are equipped with PR2/6 probes (P3, P7, and P8 in Figure 2); thus, the remaining probes are located outside the plots but within the VEB area (Figure 2). Soil moisture probes are placed in areas with different land use (Figure 2): five probes are in natural grasslands (36.5% of VEB), one probe is in forest (55.4% of VEB), and two probes are in fruit trees (6.8% of VEB). The plot area and urban fabric occupy the remaining area (1.3%). In the present study, for the comparison of satellite data with ground-based data, we used six probes (P1, P2, P3, P4, P5 and P6) that monitored the θ values for the 2016–2021 period in the VEB (see Figure 2 for the location of probes). The acquisitions of the other probes (P7 and P8) are discontinuous over time and, therefore, not valid for accurate comparative analyses. The probe number meets the minimum density of in situ sensors required for validation for both the Copernicus Global Land Service (CGLS) and NASA Soil Moisture Active Passive (SMAP) products (i.e., [41]).

The geotechnical analyses performed on samples collected along two soil profiles indicate that the shallowest layer is predominantly composed of sandy materials, with a porosity (n) of about 0.45, a dry unit weight (γ_d) of 14.32 kN/m³ and a saturated volumetric water content (θ_s) of 0.45 m³/m³ [39,42]. The sand amount increases up to the depth of about 100 cm (a maximum value of about 80% of the soil mass) with a slight increase in γ_d (14.81 kN/m³). Above 100 cm, there is a lithological transition from sands to silty–clayey soils (about 85% of the soil mass). According to Minea et al. [36], the soil characterizing the VEB area belongs to the hydrological soil group A, which represents the most porous in the Subcarpathian curvature (with over 7.6 mm/h infiltration rate).

2.2. Satellite Soil Moisture Products

2.2.1. CGLS Sentinel-1 SSM1km

The Sentinel-1 SSM1km CGLS product estimates of the degree of saturation of the first 50 mm of soil every 4–5 days with a spatial resolution of 1 km (SSM1km). The acquisitions consist of a constellation of two polar-orbiting satellites, Sentinel-1A (launched in 2015) and Sentinel-1B (launched in 2016), both performing C-band Synthetic Aperture Radar (C-SAR) imaging [43,44]. SSM1km are derived directly from the observed radar backscatter; the changes in backscatter are interpreted as changes in soil moisture, while other surface properties, such as geometry, roughness, and vegetation structure, are interpreted as static parameters. The soil moisture information is derived through the TU-Wien change detection method [45]. The dataset is freely distributed (https://land.copernicus.eu/en/products/soil-moisture/daily-surface-soil-moisture-v1.0#download, accessed on 15 June 2024). Flag values mask the areas where the retrieved SSM at time t (SSM(t)) is meaningless [40] due to surface effects such as the presence of frozen soils, snow-covered and flooded areas, slopes higher than 17°, extremely dry conditions, and dense vegetation areas. The VEB area is included in four grid pixels of the SSM1km (Figure 2). The satellite observations over the VEB area are spatialized by areally weighting the SSM1km data of each pixel.

2.2.2. CGLS Soil Water Index 1 km

The fusion of SSM data of the Sentinel-1A/B C-SAR mission and Meteorological Operational Satellites (MetOp-A/B/C Advanced Scatterometer, ASCAT) allows us to obtain the CGLS Soil Water Index 1 km (SWI1km) [46,47,48,49,50,51]. This index describes the degree of saturation of soil layers at various depths with a spatial resolution of 1 km and 1-day temporal resolution. The SWI1km product is supplied for eight T-values (2, 5, 10, 15, 20, 40, 60, 100). In practice, referring to the two-layer water balance model [52,53], characterized by a surface layer (WS) and a reservoir layer (WR), the T factor represents the ratio between the reservoir layer depth (L) and the soil diffusivity of WR (C). If C is considered constant, a high value of the factor T describes deeper soil layers [54]. The two-layer water balance model does not consider the soil texture; thus, the same T-value can describe different depths for different soils [55,56].

SWI1km values are not retrieved when the retrievals from ASCAT and Sentinel-1A/B missions are not meaningful (e.g., [57]). The Surface State Flag value (SSF) plays a key role in this process, as it is used to mask the SWI1km estimation when the detected surface state is unknown, frozen, temporally melting/water, or could not be determined. As for the SSM1km product, the satellite observations over the VEB area are spatialized by areally weighting the SWI1km data of each pixel.

2.2.3. NASA Soil Moisture Active Passive Mission

The NASA Soil Moisture Active Passive Mission (SMAP) was launched in 2015 to acquire several surface soil moisture data globally, characterized by different spatial–temporal resolutions [58]. Initially, the mission included an L-band radar and an L-band radiometer, but unfortunately, the L-band radar failed in July 2015 [59]. The L2_SM_SP (Version 9) and the SP_L4_SM (Version 7) products are used in the present study.

The L2_SM_SP data product contains geophysical retrieval based on the brightness temperature observation by the L-band radiometer and the backscattered observation by a Sentinel-1 C-SAR spatial resolution of about 20 m. The soil moisture estimation carried out by the L2_SM_SP is expressed in volumetric units (m³/m³) and refers to the first 50 mm of soil. It is provided with a spatial resolution of 3 km with two estimations in descending (am) and descending and ascending (apm) orbits [60]. A single grid pixel of L2_SM_SP data products covers the whole area of the VEB.

The SP_L4_SM product [61] provides data at a 3 h time resolution and 9 km spatial resolution. Two soil moisture products are available: the surface (SP_L4_SM ssm) and root zone (SP_L4_SM rz). The SP_L4_SM ssm investigates the first 50 mm of soil, while the SP_L4_SM rz is obtained by applying a land surface model [62] to the SMAP observations, which reflects the average soil moisture on a thickness of one meter. Due to the SP_L4_SM temporal resolution, only those observations closest in time to the in situ measurement were considered for analysis. The VEB area falls in a single grid pixel of the SP_L4_SM product.

2.3. Analysis of Soil Moisture Series

2.3.1. Soil Moisture Data Processing

The CGLS products (SSM1km and SWI1km) are available every 4–5 days and 1 day, respectively. Since the data are shared as a daily composite, in these cases, we compared the water content manually recorded by the same operator between 9 am and 12 am with the satellite daily data, paying attention to the time zone (UTC + 2). The same considerations are valid for L2_SM_SP (am and apm). Regarding the SP_L4_SM product, we have collected the water content observation at a time close to that of the manual monitoring of the profile probes. According to [63,64], the different datasets were normalized between 0 and 1 to compare in situ soil moisture data with satellite observations. This procedure is needed because the various in situ or satellite datasets acquire different soil properties ($\mathsf{\theta }$, relative soil moisture or degree of saturation). For example, the SSM1km and SWI1km products already range between 0 and 1, as they are expressed in terms of degree of saturation. On the contrary, both in situ probes and SMAP products acquire $\mathsf{\theta }$ values that need to be normalized ($\mathsf{\theta }$_n) to be comparable with the other datasets. In this way, the normalization consists of rescaling the $\mathsf{\theta }$ values of each time series over the investigation period using the maximum ($\mathsf{\theta }$_max) and minimum ($\mathsf{\theta }$_min) values recorded. For example, Equation (1) shows the procedure for normalizing $\mathsf{\theta }$ values.

$${\mathsf{\theta }}_{\mathrm{n}}={\displaystyle \frac{\mathsf{\theta }-{\mathsf{\theta }}_{\mathrm{m}\mathrm{i}\mathrm{n}}}{{\mathsf{\theta }}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathsf{\theta }}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}$$

(1)

In the remainder of this paper, the terms soil moisture or relative soil moisture are also used, even though they have been normalized between 0 and 1.

Although the comparison between satellite and ground-based data is often conducted using a single moisture probe, in recent years, attention has been given to the comparison with spatialized ground-based data. Several approaches to spatializing ground-based data are related to the number and location of the probes within a single satellite grid pixel. For instance, the ground-based data obtained from multiple moisture probes within a catchment can be spatialized using the Thiessen polygon method [1] or the arithmetic mean [65,66]; the latter is particularly useful for small catchments or plots—as for the VEB (0.74 km²)—since the arithmetic mean allows us to describe the spatial variation in the soil moisture [67,68]. Given the small size of the basin and the presence of a consistent number of probes, the water content values in this work have been spatialized using the arithmetic mean. The arithmetic mean of normalized water content (${\overline{\mathsf{\theta }}}_{\mathrm{n}}$) is computed to spatialize the $\mathsf{\theta }$_n values acquired at different depths from all the probes available ($\mathsf{\theta }$_{n_10}; $\mathsf{\theta }$_{n_20}; $\mathsf{\theta }$_{n_30}; $\mathsf{\theta }$_{n_40}; $\mathsf{\theta }$_{n_60}), obtaining five values (${\overline{\mathsf{\theta }}}_{\mathrm{n}\_10}$; ${\overline{\mathsf{\theta }}}_{\mathrm{n}\_20}$; ${\overline{\mathsf{\theta }}}_{\mathrm{n}\_30}$; ${\overline{\mathsf{\theta }}}_{\mathrm{n}\_40}$; ${\overline{\mathsf{\theta }}}_{\mathrm{n}\_60}$).

This procedure facilitates the comparison of in situ observations with SWI1km, which allows for the investigation of increasingly greater depths (i.e., Section 2.2.2) due to the variation in T-values. Since SP_L4_SM rz observations pertain to the first meter of soil, the layer-weighted method is employed to calculate the mean distributed normalized volumetric water content in the root zone (${\overline{\mathsf{\theta }}}_{\mathrm{n}\_\mathrm{r}\mathrm{z}}$).

2.3.2. Hydrological Consistency Index

A rough evaluation of the reliability of satellite datasets was carried out using the Hydrological Consistency Index—HCI [69]. HCI is physics-based, and used to verify the consistency between satellite soil moisture estimation (ΔSM) of the shallowest portion of the soil and ground-based cumulative precipitation data (P), attributing a positive (A+), negative (A−), or null (n/a) agreement to each soil moisture retrieval. According to [69], the cumulative precipitation represents the sum of rainfall in the time elapsed between two successive satellite observations. The algorithm does not consider the evapotranspiration, which affects the soil water budget. In the HCI computation, the error of the satellite datasets (ξ) is assumed to be equal to 0.04 m³/m³, corresponding to a saturation value of 4.5%. The HCI includes the possibility of assessing the agreements even in irrigated areas; the VEB has no irrigation. A synthesis of possible cases of agreements between ΔSM and P is reported in

Table 1. Possible cases of agreements between satellite soil moisture estimation (ΔSM) and ground-based cumulative precipitation data (P).

	P > 0	P < 0
ΔSM > ξ	A+	A−
ΔSM < ξ	A−	A+
−ξ < ΔSM < +ξ	n/a	n/a

, as defined by [69].

2.3.3. Cumulative Distribution Function Matching Approach

Due to substantial differences between satellite observations and in situ measurements (e.g., spatial resolution, depth of measurements, etc.), satellite data may need to be scaled to remove the systematic differences between the two datasets [26,66,70,71,72,73]. To normalize the remotely sensed data to match the distribution of ground data and remove the systematic differences between the two datasets, the Cumulative Distribution Function (CDF) matching approach was used [63]. CDF-matching transforms a time series’ mean, variance, skewness, and kurtosis to approximately harmonize the reference time series [57].

2.3.4. Statistical Tests

Measurements acquired by each probe over 2016–2021 were compared with satellite observations, performing three statistical indices widely used in the literature (e.g., [31,74]). Following are the statistical indices used for the evaluation of the soil moisture products’ accuracy:

• Pearson correlation (R)

$$\mathrm{R}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{G}}_{\mathrm{i}}-\mathrm{G}\right)\mathrm{\ast }({\mathrm{S}}_{\mathrm{i}}-\mathrm{S})}{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{G}}_{\mathrm{i}}-\mathrm{G})}^2}\mathrm{\ast }\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{S}}_{\mathrm{i}}-\mathrm{S})}^2}}}$$

(2)

• Spearman correlation (rho)

$$\mathrm{r}\mathrm{h}\mathrm{o}=1-{\displaystyle \frac{6\sum {\mathrm{D}}^2}{\mathrm{n}\left({\mathrm{n}}^2-1\right)}}$$

(3)

• Root Mean Square Error (RMSE)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{G}}_{\mathrm{i}}-{\mathrm{S}}_{\mathrm{i}})}^2}{\mathrm{n}}}}$$

(4)

where:

• G_i = soil moisture gauge observation (m³/m³)
• G = mean soil moisture gauge observation (m³/m³)
• S_i = satellite estimation (m³/m³)
• S = mean satellite estimation (m³/m³)
• n = observation number
• D = difference between ranks of G_i and S_i

The same indices have been used to evaluate the performance of satellite observations with spatialized data of in situ measurements (i.e., ${\overline{\mathsf{\theta }}}_{\mathrm{n}}$).

3. Results

Thanks to the ground-based measurements taken daily for several years, the data acquired in the VEB area are used to investigate the reliability of different satellite soil moisture products. The results presented hereafter improve the knowledge of this experimental site, which was created to establish relations between runoff and its genetic and conditional factors and to design rainfall–runoff mathematical models [36]. From the perspective of integrating satellite data into numerical modeling, the performance analysis of remote-sensing products is essential. In addition, the VEB site-specific results add information to the investigation of the performance of remote sensing-data, as they investigate a region not monitored by ISMN, focusing on the processes in the root zone. In the following subsections, we report the results obtained by analyzing the issue of satellite data performance with the different indices and statistical tests.

3.1. Analysis of HCI

The HCI was used to investigate the performance of the satellite products by considering the response of satellite observations in the shallowest portion of the soil to rainfall. Accordingly, this index was used to evaluate the performance of SSM1km, SWI1km, L2_SM_SP, and SP_L4_SM ssm satellite products. As shown in Figure 3, the satellite observations over the VEB are consistent in the long time series and respond to the trend in precipitation fluctuation.

Table 2. Results of the analysis of HCI.

	SSM1km	SWI1km T-Value = 2	L2_SM_SP am	L2_SM_SP apm	SP_L4_SM ssm
nT	842	1973	537	506	2191
% A−	58	29	43	45	25
% A+	42	71	57	55	75

summarizes the results of the analysis of HCI, considering the possible cases of agreements listed in Table 1. As expected, due to the different satellites’ overpassing time, the number of analyzable data (n_T) is different for each satellite product, from 506 observations for the L2_SM_SP apm to 2191 of the SP_L4_SM ssm. Due to the high overpassing time of SSM1km (Figure 3a) and L2_SM_SP products (Figure 3b,c), the response of data to rainfall is low, if compared to daily satellite estimations, such as SWI1km (Figure 3d) and SP_L4_SM ssm (Figure 3e). According to [69], excluding the observations of the values of which fall within the error range of satellite measurements, the highest percentage of positive agreement (A+) occurs in the SP_L4_SM ssm series (75%), followed by SWI1km (71%). Moreover, most of the data with negative agreement (A−) fall during summer periods or prolonged droughts where the role of evapotranspiration cannot be neglected.

3.2. CDF and Statistical Tests

Following the procedure described in [63], the CDF-matching approach normalized the remotely sensed data before applying statistical indices to check the reliability and performance of satellite soil moisture estimations against ground-based measurements. According to [70] and as reported in [63,71], since the satellite water content observations matched the cumulative distribution function (CDF) of the in situ data, the bias (e.g., systematical errors) are removed.

The comparison of satellite estimations with data from a single ground-based observation point was executed following the study using the RSMN data [3]. As shown in Figure 2, probes P1, P2, P3 and P4 fall within Pixel 1 (64.4% of VEB), while probes P5 and P6 fall within Pixel 3 (3.1% of VEB) and Pixel 2 (20.2% of VEB), respectively.

Table 3. Results of the statistical tests performed to evaluate the accuracy of the comparison of SSM1km and SWI1km over the grid pixels covering the VEB with in situ shallowest soil moisture data from the probes analyzed ($\mathsf{\theta }$_{n_10}).

In Situ Dataset	Statistical Test	Grid Pixel	SSM1km	SWI1km	T-Values
Probe P1 θn_10	R	1	0.444	0.566	5
	rho		0.477	0.666
	RMSE (m3/m3)		0.191	0.168
	n		471	1220
Probe P2 θn_10	R	1	0.397	0.528	2
	rho		0.410	0.583
	RMSE (m3/m3)		0.226	0.157
	Bias
	n		473	1232
Probe P3 θn_10	R	1	0.478	0.600	5
	rho		0.522	0.700
	RMSE (m3/m3)		0.148	0.126
	n		503	1275
Probe P4 θn_10	R	1	0.429	0.586	5
	rho		0.459	0.652
	RMSE (m3/m3)		0.160	0.135
	n		471	1213
Probe P5 θn_10	R	3	0.345	0.584	5
	rho		0.333	0.670
	RMSE (m3/m3)		0.149	0.112
	n		490	1164
Probe P6 θn_10	R	2	0.377	0.560	5
	rho		0.411	0.667
	RMSE (m3/m3)		0.152	0.129
	n		451	1084

shows the statistical results obtained, comparing the shallowest in situ observations ($\mathsf{\theta }$_{n_10}) with SSM1km and SWI1km datasets. All the correlations are statistically significant (i.e., p-value < 0.05) [75].

For the comparison of SWI1km data, T-values between 2 and 100 have been evaluated. According to [76], the optimal T-value for each product–probe combination is chosen by selecting the value that yields the highest Pearson correlation coefficient. Table 3 shows only the statistical results referring to the best values of T, which generally correspond to T = 5 except for Probe 2, where T = 2.

Table 4. Results of the statistical tests performed to evaluate the accuracy of the comparison of the different satellite products with the spatialized ground-based soil moisture values at different depths over the VEB (${\overline{\mathsf{\theta }}}_{\mathrm{n}}$).

Dataset	Statistical Test	SSM1km	SWI1km	T-Values	L2_SM_SP am	L2_SM_SP apm	SP_L4_SM ssm	SP_L4_SM rz
${\overline{\mathsf{\theta }}}_{\mathrm{n}\_10}$	R	0.431	0.573	5	0.418	0.421	0.651	-
	rho	0.494	0.678		0.526	0.522	0.764	-
	RMSE (m3/m3)	0.135	0.115		0.155	0.157	0.104	-
	n	529	1339		370	347	1383	-
${\overline{\mathsf{\theta }}}_{\mathrm{n}\_20}$	R	-	0.706	5	-	-	-	-
	rho	-	0.717		-	-	-	-
	RMSE (m3/m3)	-	0.145		-	-	-	-
	n	-	1314		-	-	-	-
${\overline{\mathsf{\theta }}}_{\mathrm{n}\_30}$	R	-	0.672	10	-	-	-	-
	rho	-	0.686		-	-	-	-
	RMSE (m3/m3)	-	0.105		-	-	-	-
	n	-	1339		-	-	-	-
${\overline{\mathsf{\theta }}}_{\mathrm{n}\_\_40}$	R	-	0.648	10	-	-	-	-
	rho	-	0.681		-	-	-	-
	RMSE (m3/m3)	-	0.117		-	-	-	-
	n	-	1340		-	-	-	-
${\overline{\mathsf{\theta }}}_{\mathrm{n}\_60}$	R	-	0.551	15	-	-	-	-
	rho	-	0.648		-	-	-	-
	RMSE (m3/m3)	-	0.124		-	-	-	-
	n	-	1339		-	-	-	-
${\overline{\mathsf{\theta }}}_{\mathrm{n}\_\mathrm{r}\mathrm{z}}$	R	-	-	-	-	-	-	0.703
	rho	-	-		-	-	-	0.752
	RMSE (m3/m3)	-	-		-	-	-	0.124
	n	-	-		-	-	-	1361

summarizes the results obtained by comparing satellite products and spatialized ground-based soil moisture data over the VEB (${\overline{\mathsf{\theta }}}_{\mathrm{n}}$). Figure 4 shows the comparison of ${\overline{\mathsf{\theta }}}_{{\mathrm{n}}_{\_\_10}}$, computed over the VEB area, with the five satellite products (SSM1km, SWI1km, L2_SM_SP am, L2_SM_SP apm, SP_L4_SM ssm). As expected, the L2_SM_SP am and L2_SM_SP apm have the lowest observations due to well-known technical acquisition issues (Section 2.2.3). The statistical tests confirm the results obtained by the analysis of HCI. Although the number of observations of L2_SM_SP am, L2_SM_SP apm are slightly lower than those of SSM1km, the results of the statistical tests of the three products are similar. The SP_L4_SM ssm is confirmed as the product with the best performance (i.e., the product with the highest value of R, rho and RMSE parameters). The SWI1km estimations are also accurate and perform better than SSM1km when compared with ${\overline{\mathsf{\theta }}}_{\mathrm{n}\_10}$ computed over the VEB.

Figure 5a–c compares the ${\overline{\mathsf{\theta }}}_{\mathrm{n}}$ values at three selected depths (20, 40 and 60 cm) with SWI1km estimates obtained using T-values that increase from 5 to 15, moving towards the deepest part of the soil. As shown in Table 4, at the different depths, the rho values computed on the SWI1km datasets remain high and range between 0.65 and 0.72. As expected, the T-values increase with depth.

The comparison of SP_L4_SM rz estimations over the VEB with ${\overline{\mathsf{\theta }}}_{\mathrm{n}\_\_\mathrm{r}\mathrm{z}}$ data (Figure 5d) shows the highest values of the different statistical tests used. Howerver, for about the same number of observations, the statistical tests provide similar results to the SWI1km.

4. Discussion

The accuracy of satellite-based soil moisture observations is affected by the intrinsic uncertainties typical of each product and by the local characteristics of the investigated area, such as slope, soil roughness, vegetation, etc. [77,78].

Studies carried out to validate satellite moisture products using RSMN data considered comparing satellite estimations with data from a single ground-based observation point. As encouraged by [3], our study introduced cumulative distribution function (CDF) matching in the analysis of comparing satellite and ground-based data in a selected Romanian region (VEB area).

The results of SWI1km’s performance with data recorded by the probes in the shallowest part of the soil ($\mathsf{\theta }$_{n_10}) are in line with the recent systematic study on the intercomparison between the SWI1km dataset and the data from the 20 stations of RSMN carried out by [79]. Despite the SSM1km product being less accurate than SWI1km, its performance in the VEB area is comparable with those reported in other Romanian sites by [80], who also considered the comparison with RSMN. As reported by [81,82], the SSM1km could be influenced by the strong Radio Frequency Interference (RFI) in Europe. Moreover, due to the high spatial resolution of SSM1km, the surface effects could be smoother for satellite products investigating larger areas. It should be pointed out that the Pixel 4 of SSM1km (Figure 2) does not host any soil moisture probes. Considering that most of the VEB area is occupied by Pixel 1 (representing 65% of the VEB area), the probes used to spatialize the water content over the VEB can be considered representative. It is confirmed by analyzing the agreement between SSM1km and ${\overline{\mathsf{\theta }}}_{\mathrm{n}\_10}$ (ground-based spatialized value); the results were better than the comparison of SSM1km with $\mathsf{\theta }$_{n_10} computed on a single ground-based point. In this way, the spatial representativeness of the soil moisture values matches the resolution of SSM1km, confirming the results obtained in other regions (e.g., [51,83]).

Among the satellite products analyzed, the best performance was achieved by SP_L4_SM ssm, followed by SWI1km (T-value = 5, near surface). These results agree with the findings recently presented by [84], who conducted a study over the European ecoregions regarding the accuracy of satellite SM products, including the analysis of 20 stations managed by the RSMN. Albergel et al. [53] evaluated the optimal T-values in the root zone by applying an exponential filter to the available surface soil moisture information on two networks in southern France characterized by different soil types. The soil moisture network presented by [53] is composed by several calibrated ThetaProbes (Delta-T Devices), which works at the same frequency of the PR2/6 probe (100 MHz). Albergel et al. [53], combining the information obtained from calibrating the optimal T-values by the Nash and Sutcliffe coefficient [85], presented T-values ranging from about 2 (near surface) to 20 (root zone). As reported in Table 3 and Table 4, the optimal T-values achieved by comparing ground-based data and SWI1km datasets, computed by considering the highest Pearson correlation coefficient, move from 2–5 (near surface, 10 cm) to 10–15 (root zone, 30–60 cm). This trend agrees with that proposed by [53].

Our results significantly enhance the accuracy of satellite observations in the Subcarpathian region, an area not covered by the official ISMN. This improvement is particularly important in small catchments in a transition area between the Central European and Balkan mixed forest defined by [84]. In other words, our study can help provide information in critical areas where is difficult to extract a clear picture in ecoregions where climate, soil, vegetation and the results of remote-sensing inter-comparison do not show a marked trend or a singular direction [84].

As discussed in [86], the best performance of SP_L4_SM ssm is probably due to the use of ancillary data. Surface effects (i.e., roughness, land-cover heterogeneity, etc.) are smoothed due to a larger grid of investigation of SP_L4_SM ssm compared to the other remote-sensing products analyzed. The accuracy of the SP_L4_SM product is also evident in the root-zone analysis. The SP_L4_SM rz estimations compared with ${\overline{\mathsf{\theta }}}_{\mathrm{n}\_\_\mathrm{r}\mathrm{z}}$ data over the VEB indicate that this product is promising in investigating processes occurring in the medium-to-deepest parts of the soil.

The approaches and results for the VEB area can open new research perspectives in using satellite data in other experimental sites worldwide characterized by different soil types, climates, and vegetation. Improved knowledge of satellite soil moisture in the root zone is increasingly required, considering its important role in governing the magnitude and variability of water fluxes into the soil column [87,88].

5. Conclusions

In this study, we assessed the comparison of different high-resolution satellite soil moisture products (CGLS SSM1km, CGLS SWI1km, SMAP L2_SM_SP, and SMAP SP_L4_SM) with ground-based data from six profile probes in the Voineşti Experimental Basin (Subcarpathian area, Romania). The results can be summarized as follows:

-

The SP_L4_SM products show the best performances regardless of the considered metrics, as confirmed by the physics-based index (HCI) and some statistical tests widely used in inter-comparison analysis.

-

The performance of the SWI1km observations has shown satisfactory results with single-point and spatialized soil moisture estimations over the Experimental Basin, considering both near-surface and root-zone data.

-

Among the products with the lowest performance (SSM1km and L2_SM_SP products), the C-band and L-band data fusion operating by the L2_SM_SP products shows a slight improvement over SSM1km at the expense of a lower temporal resumption.

In conclusion, the findings represent a novelty for the Romanian region as derived from a leading-edge experimental site, providing information on the reliability of satellite data not only in the shallowest portion of the soil but also in the root zone, which has not been investigated in previous studies in the Romanian territory. Therefore, the data and approach used can provide valuable insights into evaluating the performance of satellite SM products in other Romanian areas, inspiring further research in areas characterized by different soil types, climates, and vegetation, and opening up new possibilities for environmental science and remote sensing.