We used the SMOS Level-3 (SMOS-L3) TB product distributed by the Centre Aval de Traitement des Données (CATDS) for the years 2010–2024 (Table 1). SMOS-L3 TB provides multi-angle H- and V-polarized TBs recorded at the top of the atmosphere (Al Bitar et al., 2017). Despite the absence of atmospheric correction, the average atmospheric effect remains relatively mild globally, with ∼ 1 K (H-polarization) and ∼ 0.5 K (V-polarization) at 40° (De Lannoy et al., 2015). It should be noted that the SMOS L3 daily multi-angle TB data are obtained using a fixed 5° width binning method, with bin centers lie within the 2.5–62.5° interval. Previous studies have revealed that this approach may result in stronger short-term TB fluctuations at specific angles compared to alternative methods, such as two-step regression fitting. This ultimately increases the uncertainty in analyses or retrievals dependent on single-angle TB data (Li et al., 2022b; Peng et al., 2023; Schmitt and Kaleschke, 2018). This work employed the SMOS-L3 TB dataset (on a 25 km EASE-Grid 2.0), utilizing solely the ascending orbit (06:00 a.m. local time) TBs.

The ISMN in-situ SM measurements (https://ismn.geo.tuwien.ac.at/, last access: 1 October 2025) were used to evaluate the TB and satellite SM retrievals' accuracy. ISMN was considered to be the most reliable SM dataset and has been extensively utilized as a benchmark in satellite-based SM calibration and validation studies (Dorigo et al., 2021). Here, SM measurements from the 0–5 cm soil depth from 2016 to 2022 incorporating both sparse and dense in-situ networks were collected. Note that there is inherent scale mismatch between pixel-derived SM estimates and ground-based SM observation, particularly in the sparse observed networks. To maintain good data quality and minimize the issue of the spatial scale differences, only ISMN in-situ SM observations flagged as “Good” were spatially aggregated by averaging all available station observations within each respective 25 km EASE-Grid 2 cell. Ultimately, a total of 464 cells from 23 networks at a EASE-Grid 2.0 25 km scale were retained (Fig. 1 and Table A1 in the Appendix).

Given that the validation of VOD products at large scales is hindered by the lack of a well-established reference dataset, three frequently used vegetation proxies were selected to assess the performance of the VOD retrievals (Wigneron et al., 2024), including the 1 km spatial resolution Saatchi aboveground biomass (AGB) map (Saatchi et al., 2011), 0.5° canopy height derived from Global Ecosystem Dynamics Investigation Level 1B LIDAR observations collected between April to July 2019 (Simard et al., 2011), and 1 km resolution 16 d MODIS NDVI data from 2016 to 2022 (Didan, 2021). The canopy height serves as an indicator of total vegetation biomass, and NDVI reflects the greenness and photosynthetic activity within the upper layer canopy (Li et al., 2021). To preserve high-quality observations, the pixels for MODIS NDVI data flagged as “good quality” were kept following the method of Grant et al. (2016).

In addition, this study was the first to use satellite canopy water content (CWC) data from 2016 to 2022 to validate the temporal behavior of VOD retrievals, since L-band VOD has been demonstrated to have a linear relationship with vegetation water content (Wigneron et al., 2024). The CWC product was newly developed by integrating data from Sentinel-2, Landsat-8, and MODIS satellites with a spatial resolution of 0.05° to monitor canopy vegetation water variations, which has been demonstrated to have good accuracy and reliability, thus providing a robust reference for assessing VOD data (Ma et al., 2025). The dataset was obtained through personal communication but will soon be publicly available via ESA data portal. These four vegetation parameters were standardized through projection onto the EASE-Grid 2.0 and spatially aggregated to 25 km using arithmetic mean resampling to match the SMOS grid spatial resolution. This same resampling method has also been employed in several earlier VOD studies (Li et al., 2021; Fan et al., 2019).

To evaluate the performance of optimized SMOS-IB TB (see method Sect. 3.1) and the IB_HRmonoSMOSIB, IBmonoSMOSIB, IBmonoRawSMOS SM and VOD retrievals, two other L-band TB data (i.e., SMOS-L3 TB and SMAP-L3 TB) and two other L-band satellite global SM and VOD datasets (i.e., ICmultiSMOS and IBmonoSMAP) were collected (see Appendix A1).

The SMOS-L3 TB product has been detailed in Sect. 2.1. SMAP-L3 TBs were sourced from the Version 5 SMAP enhanced L3 radiometer SM product collected during the morning (06:00 a.m. local time) descending overpass for the period 2016–2022 (Chan et al., 2018). The SMAP-L3 TB observations were quality controlled based on corresponding quality flags and resampled to 25 km via weighted area averaging for consistency with the SMOS' grid resolution (Li et al., 2022b).

The ICmultiSMOS and IBmonoSMAP SM and VOD products at 25 km projected onto the EASE-Grid 2.0 from 2016 to 2022 were collected. (1) The ICmultiSMOS corresponds to the SMOS-IC dataset, originally developed by Fernandez-Moran et al. (2017a, b), and is among the most recent SMOS products available. It was retrieved using the processed multi-angle SMOS-L3 TB dataset with quality filtering provided by the CATDS using the SMOS-IC version 2 algorithm. The 25 km SMOS-IC V2 SM and VOD data retrieved from the morning ascending orbit was utilized; (2) The IBmonoSMAP was retrieved by applying the SMAP-IB algorithm to the 25 km SMAP-L3 TBs (resampled from the 9 km SMAP-L3 TB dataset) at 40° incidence angle (Li et al., 2022b). Readers refer to Wigneron et al. (2021) and Li et al. (2022a) for detailed information about the SMOS-IC and SMAP-IB algorithm.

All datasets were evaluated specifically at the 06:00 a.m. local overpass time to capitalize on optimal surface thermal equilibrium conditions characteristic of early morning periods (Entekhabi et al., 2010), following rigorous quality-controlled preprocessing that adhered to each product's specific flagging criteria. For example, the ICmultiSMOS and IBmonoSMAP unreliable retrievals were effectively removed based on two quality control thresholds: “Scene Flags” >1 and “TB-RMSE” >8 K (Wigneron et al., 2021).

The MODIS IGBP land cover classification (Friedl and Sulla-Menashe, 2022) was employed to analyze SM comparison results across different land cover types. Daily precipitation data at a resolution of 0.1°, sourced from the ERA5-Land reanalysis dataset, was collected and applied to analyze seasonal variation of the SM and VOD datasets (Muñoz-Sabater et al., 2021).

To obtain robust evaluation results, we additionally employed Triple Collocation Analysis (TCA), which provides an independent error estimate and is not affected by the representativeness errors originating from the spatial discrepancy between site points and satellite footprints (see Sect. 3.1.3). For this purpose, the active microwave Advanced Scatterometer (ASCAT) surface SM product and the model-based Global Land Data Assimilation System (GLDAS-Noah) SM product from 2016 to 2022 were obtained (Rodell et al., 2004). (1) ASCAT, onboard the Meteorological Operation-A, -B and -C satellite, acquires C-band V-polarized backscatter measurements on both ascending and descending orbits (Wagner et al., 2006). The ASCAT SM product is generated from MetOp satellite backscatter measurements using a TU Wien algorithm (Wagner et al., 2013). The ASCAT CDR(Climate Data Record) v7-H119 SM dataset at 12.5 km resolution was used, with its relative SM values converted to volumetric units (m³ m⁻³) based on soil porosity from the Harmonized World Soil Database (HWSD). (2) The GLDAS-Noah SM product, with 3-hourly temporal and 0.25° spatial resolution, is derived from the Noah Land Surface Model within the Global Land Data Assimilation System (Rodell et al., 2004). The GLDAS-Noah SM (kg m⁻²) was then converted to volumetric unit (m³ m⁻³) following the method of Cui et al. (2018) by dividing by water density and the corresponding soil layer thickness, with daily average SM computed for analysis. Both the ASCAT and GLDAS-Noah SM were aggregated to 25 km resolution by applying the arithmetic mean resampling to match the SMOS grid resolution.

To mitigate the angular-related noise and enhance the consistency of TBs, we adopted the L-MEB model, originally developed for the SMOS and shown to effectively reproduce SMOS TBs across varies land surface conditions (Wigneron et al., 2012). In our implementation, L-MEB was employed as a forward model, with multi-angular SMOS L3 TB as input. The optimal fitting results were obtained by minimizing the RMSE (root mean square error) between the L-MEB simulated and observed TB values. Figure 3 shows examples of the fitting results on 5 May, 15 June, 3 July, and 8 August 2024. It can be seen that the fitted TBs significantly reduce the irregularity and dispersion present in the raw L3 TBs, for both polarizations. The fitted TBs at 40° incidence angle, which is in line with SMAP observations, were used as the SMOS-IB product for subsequent applications. In addition, the fitted TB-RMSE for each pixel was retained in the dataset, as it has been shown to serve as a simple and effective indicator for assessing the real influence of RFI on SMOS TBs' quality (Wigneron et al., 2021; Li et al., 2021).

Note that three types of SM and VOD datasets were produced with the aim to address the key scientific questions of this study: ① IBmonoRawSMOS: implementing the SMAP-IB to the raw SMOS-L3 40° TB; ② IBmonoSMOSIB: implementing the SMAP-IB to the SMOS-IB 40° TB, and ③ IB_HRmonoSMOSIB: implementing the SMAP-IB algorithm that incorporated a refined soil roughness (Hr) parameterization scheme to the SMOS-IB 40° TB.

All three datasets used the Tau-Omega (τ-ω) radiative transfer approach to model microwave TB from soil-vegetation covered land surfaces (Mo et al., 1982). However, unlike IBmonoRawSMOS and IBmonoSMOSIB, IB_HRmonoSMOSIB used a novel globally calibrated pixel-level Hr data to represent soil roughness effects (Konkathi et al., 2025). Their approach moves beyond prior methods, which not only accounted for Hr differences between vegetation types, but also incorporating intra-type Hr differences through a methodology that synergistically combines radiative transfer modeling with machine learning.

The IB_HRmonoSMOSIB SM and VOD were jointly retrieved based on the optimized SMOS 40° incidence angle TBs using SMAP-IB algorithm, incorporating the values of a novel global calibrated Hr, to resolve the underdetermined problem of 2-Parameter retrieval from correlated SMOS TB observations, the SMAP-IB method implements an optimized least-squares iteration, which minimizes a cost function (CF) that accounts for prior knowledge of SM and VOD. 1CF=∑(TBpmes-TBp∗)2σ(TB)2+∑(SMini-SM∗)2σ(SM)2+∑(VODini-VOD∗)2σ(VOD)2 where TBpmes (TBp∗) denote the measured and simulated TBs at both polarizations, respectively; σ(⋅) is the standard deviation operator; and the second and third terms are regularization functions that involve the retrieval parameters (SM∗, VOD∗) and their initial estimates (SMini, VODini). Please refer to Li et al. (2022a) for a detail description of these initial estimations of the SMAP-IB algorithm.

Four key metrics were applied to examine the performance of the retrieved dataset: (1) Pearson's correlation coefficient (R; Eq. 2), (2) systematic bias (Eq. 3), (3) RMSD (Eq. 4), and (4) unbiased RMSD (ubRMSD; Eq. 5) (Entekhabi et al., 2010). 2R=1-(θRS-θREF)2(θRS-θREF‾)23Bias=θRS-θREF‾4RMSD=(θRS-θREF)2‾5ubRMSD=RMSD2-Bias2 where θRS is the satellite TB, SM or VOD dataset; θREF is the reference data; the overbar represents the temporal averaging operator (i.e., θREF‾). Since systematic biases between observations and satellite retrievals may distort RMSD, the ubRMSD and R typically provide more reliable metrics for validation (Xing et al., 2021).

The global IB_HRmonoSMOSIB dataset is archived in netCDF4 format and mapped to EASE-Grid 2.0, featuring a 584 × 1388 grid with a 25 km sampling resolution. The dataset contains 14 layers (Table 2), including TB, SM, and VOD, their associated uncertainty layers, expressed as the standard errors of SM and VOD, and the global soil roughness map. The RMSE values layer between the measured and modeled TB and the Scene_Flags layer are also included in the dataset. The RMSE layer, the optimal fitting results obtained by minimizing the RMSE between the L-MEB simulated and observed TB values, serves as a measure of RFI influence on the TBs and to filter out SM and VOD data substantially influenced by RFI. The Scene_Flags layer is used to filter out multiple impacts linked to specific climate or topographic conditions (Table 2). The datasets for the period 2010–2024 can be freely downloaded at website (10.5281/zenodo.17647385, Xing et al., 2025) and will be continuously maintained on the INRAE Bordeaux Remote Sensing Product website (https://ib.remote-sensing.inrae.fr/, last access: 1 June 2026). Note that these SM, VOD, and TB products are intended to support large-scale applications, including global drought monitoring, studies of vegetation water and biomass dynamics, freeze-thaw monitoring, etc. However, low-quality observations should be screened before any application or validation analysis. In particular, users should first assess potential RFI contamination, which is especially critical for SMOS. Observations under frozen soil or snow-covered conditions should also be excluded, given the limited applicability of current dielectric models in frozen and snow/ice environments (Wigneron et al., 2017). In addition, pixels with a high fraction of open water, substantial urban land cover, or strong topographic heterogeneity should be screened out or treated separately prior to use. Accordingly, filtering criteria for IB_HRmonoSMOSIB with respect to the above influencing factors were recommended and summarized in Table A2.

The global spatial pattern of high-frequency TB variations was quantified by calculating the standard deviation (SDHF) of TB after removing seasonal cycle for SMOS-IB, SMAP-L3, and SMOS-L3 (Fig. 4). It was observed that the high-frequency variability was consistently higher in H-polarization than in V-polarization across all products, particularly in water-limited regions. Critically, the spatial median SDHF for both SMOS-IB and SMAP-L3 was low and comparable (<5.33 K), whereas SMOS-L3 exhibited markedly higher variability (>7.20 K). This demonstrated that SMOS-IB and SMAP-L3 shared similarly low noise levels, while SMOS-L3 retained the strongest high-frequency fluctuations. These differences reflected their distinct processing chains: unlike the top of atmosphere SMOS-L3 TB, SMAP-L3 included atmospheric correction and dedicated RFI mitigation, and SMOS-IB benefited from noise reduction via L-MEB model optimization, bringing its variability characteristics closer to those of SMAP-L3. Spatially, SMOS-L3 showed markedly higher SDHF than the other two products over central and northeastern Africa, central and eastern Asia, and parts of eastern Europe, regions that coincide with known RFI hotspots for SMOS (Wigneron et al., 2021; Al-Yaari et al., 2019). Pixel-wise SDHF differences further reinforced these patterns (Fig. A1 in the Appendix): deviations between SMOS-IB and SMAP-L3 were minimal (within ±0.02 over most regions), whereas SMOS-L3 showed systematically higher values, particularly over the above-mentioned RFI-affected areas where differences relative to both SMOS-IB and SMAP-L3 typically exceeded 5 K. These results confirmed that SMOS-L3 preserved substantial high-frequency noise, while SMOS-IB and SMAP-L3 provided cleaner temporal TB profiles (Fig. A2).

To further investigate how the three TB products respond to SM variations, we assessed their sensitivity to ISMN in-situ SM using the coefficient of determination (R2) across 12 MODIS IGBP land cover types (Table 3). Overall, all three products showed the strongest SM sensitivity in shrublands (S), with R2 values exceeding 0.80 for both polarizations, and the weakest sensitivity in barren or sparsely vegetated areas, where R2 values fell below 0.30, reflecting the reduced radiometric sensitivity of microwave observations in regions with low SM dynamics. The land cover-specific analysis confirmed and extended the overall patterns described above: SMAP-L3 TB generally presented the highest R2 values, followed closely by SMOS-IB TB, while SMOS-L3 TB showed the lowest sensitivity to ISMN in-situ SM data. This ranking pattern (SMAP-L3 > SMOS-IB > SMOS-L3 TB) showed complete consistency across all land cover types for H-polarization and in 7 of 12 cases for V-polarization, respectively. Particularly, SMOS-IB TB achieved slightly higher R2 than SMAP-L3 TB in MF, CS, OS, WS, S and G land cover types. These findings indicated that the proposed optimization process effectively enhanced the sensitivity of SMOS TB to SM and enabled SMOS-IB to achieve performance levels comparable to SMAP-L3 TB in most cases.

Figure 8 presents the spatial density distributions of the five VOD datasets against the aboveground biomass (AGB) map. It was found that the four VOD products showed very similar R values, ranging only from 0.84 for ICmultiSMOS and 0.85 for IB_HRmonoSMOSIB, IBmonoSMOSIBand IBmonoSMAP, whereas IBmonoRawSMOS had the lowest R values of 0.80. Moreover, all five VOD products effectively captured the spatial gradients of AGB, yielding the same highest R values of 0.87 when comparing predicted and observed AGB (Fig. 8a–e). Similarly, the IB_HRmonoSMOSIB, IBmonoSMOSIB and IBmonoSMAP VOD products exhibited the same highest spatial R value (0.90) when correlated with forest canopy height, indicating a strong linear relationship – even for tall trees. Followed by IB_HRmonoSMOSIB and IBmonoRawSMOS with spatial R of 0.89 and 0.87 (Fig. A6). This aligned with previous VOD validation studies (Li et al., 2021; Rodríguez-Fernández et al., 2018; Zotta et al., 2024), as L-band VOD responded to the full vertical structure of vegetation, encompassing woody components (Frappart et al., 2020). These findings suggested that, in terms of spatial patterns, it was difficult to distinguish clear advantages among the five products. Nevertheless, given the comparable influence of VOD and Hr in the τ–ω model and their strong coupling (Eq. 1), we plotted each VOD product against the corresponding Hr used in its retrieval (Fig. 8a1–e1). It was found that IB_HRmonoSMOSIB VOD demonstrated a notably weaker spatial correlation with its Hr (R=0.39) than the other VOD products using IGBP-based Hr schemes (R>0.70). This decoupling effect was particularly evident in forested areas, where the spatial R between IB_HRmonoSMOSIB VOD and Hr (R=-0.36) was the lowest among all products, while the others showed R>0.57. These findings collectively indicate that the new Hr scheme effectively mitigated the coupling effect with VOD, leading to more physically independent VOD retrievals compared with the IGBP-based Hr schemes.

The per-pixel temporal correlation coefficient between the five VOD datasets and CWC were also calculated to examine the discrepancy of the temporal performances for the VOD datasets (Fig. 9). All VOD products exhibited consistent spatial patterns in their temporal R values with vegetation dynamics, particularly across eastern US, southern Africa, eastern Brazil, Siberia, and Australia (Fig. 9a–d), with the IBmonoRawSMOS VOD product showing particularly widespread non-significant pixels across these biomes (Fig. 9e). Figure 9f identifies the VOD dataset with the strongest per-pixel temporal R values (absolute R difference >0.1) after excluding non-significant pixels. It was found that IBmonoSMAP, IB_HRmonoSMOSIB and IBmultiSMOS demonstrated highest R values with CWC across 41 %, 24 % and 21 % of the analyzed pixels, respectively, with these pixels mainly located in mid- to low-latitude regions (e.g., Australia, South, East and West Africa, and America). Similar findings were also obtained when NDVI was used as a reference (Fig. A7). It is noteworthy that the temporal R between IB_HRmonoSMOSIB and CWC was generally higher than that derived using IBmonoSMOSIB across most regions globally, particularly in high vegetated regions (e.g., Australia, Central North America, Amazon, Central Africa, etc.). This finding underscored the advantage of incorporating optimized Hr inputs in VOD retrievals, because the key distinction between the two VOD products lied in the optimization of roughness inputs. Similarly, Konkathi et al. (2025) also showed that the improved VOD-NDVI correlation in these regions resulted from their refined Hr scheme, which mitigated SM and VOD compensations (Fig. A8).