ASCAT is a real aperture radar onboard the Meteorological Operational (MetOp) satellites (i.e., MetOp-A, MetOp-B, and MetOp-C), and measures radar backscatter at C-band (5.3 GHz) VV (vertical transmit vertical receive) polarization (Wagner et al., 2013). The equator crossing times of the MetOp satellites are 09:30 am/pm local solar time (LST) for the descending and ascending overpasses, respectively, with a revisit frequency of 1–3 d. The ASCAT soil moisture is retrieved based on the Vienna University of Technology (TU Wien) change detection algorithm (Wagner et al., 1999, 2010), and provides an estimate of the degree of water saturation (ranging between 0 % and 100 %) of the top 0–2 cm soil layer.

We use the MetOp-B/C ASCAT near-real time soil moisture product at 12.5 km swath grid. The soil moisture product used in this study was obtained directly from the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) for use in the Korea Meteorological Administration's operational weather prediction system while the same product can be downloaded from the EUMETSAT Earth Observation Portal (https://data.eumetsat.int/product/EO:EUM:DAT:METOP:SOMO12, last access: 23 February 2026).

The SMAP mission provides volumetric soil moisture (m³ m⁻³) for the top 0–5 cm of the soil, derived from L-band (1.4 GHz) passive microwave radiometer measurements (Entekhabi et al., 2010). It has a revisit cycle of 2–3 d, with local equator-crossing times of 6 am (descending) and 6 pm (ascending).

The present study uses the SMAP Level 2 (L2) Radiometer Half-Orbit 36 km Equal Area Scalable Earth (EASE)–Grid soil moisture data (SPL2SMP version 9, O'Neill et al., 2023), which is obtained from the National Snow and Ice Data Center (NSIDC, 10.5067/K7Y2D8QQVZ4L, O'Neill et al., 2023). The SMAP soil moisture retrievals based on the Dual Channel Algorithm (DCA, Chaubell et al., 2020) are assimilated.

Typical DA algorithms are designed to correct random errors under the assumption of unbiased state estimates between models and observations (Dee and Da Silva, 1998). However, there are generally large systematic discrepancies between modeled and satellite-retrieved soil moisture because of their different representations of soil moisture associated with the geophysical definition and horizontal/vertical scales (Koster et al., 2009; Kumar et al., 2019). Therefore, soil moisture DA systems essentially employ appropriate bias correction strategies to remove these systematic biases prior to assimilation, and thus to comply with the DA assumption of unbiased models and observations (Kolassa et al., 2017; Reichle and Koster, 2004). In this study, bias correction is implemented differently for the ASCAT and SMAP soil moisture products. That is, we apply cumulative distribution function (CDF) matching (Reichle and Koster, 2004) and anomaly correction methods (Kwon et al., 2022) to assimilate the ASCAT and SMAP soil moisture retrievals, respectively, into the Noah LSM within the KIM-LIS coupled system. The use of the anomaly correction method for SMAP follows our previous investigations (Kwon et al., 2022, 2024) aiming to minimize the loss of useful information from the original data through bias correction. In contrast, traditional CDF matching is applied to ASCAT, since the anomaly correction method is not applicable due to differences in soil moisture data types between ASCAT (soil wetness index) and the model (volumetric soil moisture in m³ m⁻³). Further details are provided in the following two paragraphs.

The CDF matching is a commonly used bias correction method in soil moisture DA. Through the CDF matching, in this study, the ASCAT soil wetness index data are transformed to volumetric soil moisture (m³ m⁻³) by correcting all the statistical moments of the original ASCAT data to those of the Noah-simulated soil moisture. Existing soil moisture DA studies use two different CDF matching approaches. One uses a lumped CDF computed using all seasons data (i.e., pixel-wise single CDF for each data type) (e.g., Draper et al., 2011, 2012; Kumar et al., 2009, 2014; Reichle et al., 2007) while the other uses monthly-stratified CDFs (i.e., pixel-wise 12 CDFs for each data type) (e.g., Jun et al., 2021; Kumar et al., 2015; Kwon et al., 2022, 2024; Santanello et al., 2016). Kumar et al. (2015) and Santanello et al. (2016) suggest using the monthly CDF rather than the lumped CDF to mitigate spurious statistical artifacts in the bias-corrected soil moisture by the CDF matching. Kwon et al. (2024) show that abnormal fluctuations are witnessed in the lumped CDF-based rescaled soil moisture, particularly in dry periods, and Kwon et al. (2022) demonstrate that soil moisture DA employing the monthly CDF achieves better soil moisture analysis than that applying the lumped CDF matching. Based on these previous findings, we implement the monthly CDF matching for the ASCAT soil moisture DA.

The anomaly correction method, proposed by Kwon et al. (2022), is a simpler alternative to traditional bias correction approaches. It aims to reduce the reliance of DA systems on rescaling methods like the CDF matching, which is known to cause significant information loss in the original soil moisture data, especially when human-induced processes (e.g., irrigation activities), poorly represented in models, are the dominant source of systematic discrepancies between observations and models (Kumar et al., 2015; Nearing et al., 2018). Instead of rescaling the SMAP soil moisture retrievals, the anomaly correction approach obtains the soil moisture temporal variability (i.e., anomaly) information by subtracting the long-term soil moisture mean from the original SMAP data. The extracted SMAP soil moisture anomaly is added to the long-term mean of the modeled soil moisture, which is then assimilated into the LSM. The anomaly correction assumes that the systematic bias between observations and models is dominated by the climatological mean difference and higher moment (e.g., standard deviation) differences are negligible. Kwon et al. (2022) and Kwon et al. (2024) demonstrate that the SMAP soil moisture data and Noah-simulated soil moisture satisfy this underlying assumption over the continental United States and global domain, respectively. In particular, Kwon et al. (2024) show that the anomaly correction-based SMAP soil moisture DA is effective in improving the global soil moisture estimates and weather forecast skill of the KIM-LIS coupled system that employs the Noah LSM.

The ASCAT and SMAP soil moisture retrievals undergo quality control before DA by removing inaccurate or uncertain soil moisture observations based on strategies employed in previous soil moisture DA studies (e.g., Blyverket et al., 2019; Draper et al., 2012; Ferguson et al., 2020; Jun et al., 2021; Kolassa et al., 2017; Kumar et al., 2014, 2019; Kwon et al., 2022, 2024; Nair and Indu, 2019). Firstly, before bias correction, the soil moisture data are discarded when the data quality flags provided with each soil moisture product indicate that the data accuracy is impacted by open water bodies, dense vegetation, urban areas, precipitation, snow cover, frozen ground, complex topography, or anthropogenic Radio Frequency Interference (RFI). Especially, the ASCAT soil moisture retrievals are assimilated only when the Estimated Soil Moisture Error (ESME) is less than 16 %, and the topographic complexity and wetland fraction are below 20 % and 15 %, respectively, as applied in Jun et al. (2021). These uncertainty thresholds are slightly higher than those used in Draper et al. (2012) and Kolassa et al. (2017) for the purpose of utilizing more data in a near-real time operational DA system.

Secondly, the model-based quality control is additionally applied to both ASCAT and SMAP after bias correction. Specifically, assimilation of ASCAT and SMAP soil moisture into the Noah LSM is not performed in the case that (1) model background estimates indicate active precipitation events or frozen/snow-covered soil conditions, that (2) the model land cover type and green vegetation fraction inputs from the Moderate resolution imaging spectroradiometer (MODIS) International Geosphere-Biosphere Programme (IGBP) data (Friedl et al., 2002) and the National Centers for Environmental Prediction (NCEP), respectively, indicate that a grid cell is classified as forests or has green vegetation fraction greater than 0.7, or that (3) bias-corrected soil moisture retrievals are close to wilting point or saturation.

Atmospheric DA in KIM is based on a hybrid four-dimensional ensemble variational (hybrid 4DEnVar) DA method as described in Kwon et al. (2018) and Song et al. (2017). In this study, we assimilate both conventional and non-conventional atmospheric data including the Advanced Microwave Sounding Unit-A (AMSU-A), Atmospheric Motion Vectors (AMVs), Microwave Humidity Sounder (MHS), Global Positioning System Radio Occultation (GPS-RO), Infrared Atmosphere Sounding Interferometer (IASI), Advanced Technology Microwave Sounder (ATMS), Cross-track Infrared Sounder (CrIS), and observations obtained from surface, aircraft, and sonde. The KIM Package of Observation Processing (KPOP, Kang et al., 2019) is employed to preprocess (e.g., quality control and bias correction) the observations before assimilation. Details of the KIM hybrid 4DEnVar DA scheme are provided in Sect. S1 of the Supplementary Material and in Kwon et al. (2018) and Song et al. (2017).

In the KIM-LIS coupled system, land DA is conducted by the LIS-DA subsystem (Fig. 1) in which various DA schemes are available. The current study applies a 1-dimensional EnKF method (Reichle et al., 2002b) to assimilate satellite soil moisture retrievals (i.e., ASCAT and SMAP) into the Noah LSM. The EnKF is one of the widely used DA schemes for nonlinear hydrological applications (e.g., Cho et al., 2023; Crow and Van den Berg, 2010; Draper and Reichle, 2019; Kim et al., 2021a; Kwon et al., 2019, 2021; Reichle et al., 2023; Renzullo et al., 2014; Xu et al., 2021) because of its relatively flexible and computationally efficient nature (Keppenne, 2000).

Within the EnKF-based DA system, model forecasts and analysis updates are performed alternately. That is, the ensemble forecasts of model prognostic state variables are propagated forward in time until observations are available, and the forecasted states are updated in the assimilation step when and where observations exist. The resulting analysis ensemble is then used as the initial condition for the next model forecast. In this study, the control vectors that are directly updated by assimilating the ASCAT and SMAP soil moisture retrievals include the Noah LSM estimates of soil moisture at four soil layers while other related hydrometeorological variables are adjusted through model physics in subsequent model integrations. As we conduct the 1-dimensional EnKF, the soil moisture analysis in a given grid is produced independently of neighboring grids.

The EnKF increments are determined depending on the relative uncertainties (error variances) of model and observation. The model uncertainty (background error covariance) is represented by the ensemble forecast spread (ensemble size of 20), which is obtained at each grid by randomly perturbing the atmospheric variables from KIM including shortwave radiation, longwave radiation, and precipitation, and by additionally (randomly) perturbing the Noah LSM-simulated soil moisture estimates. Shortwave radiation and precipitation are perturbed by applying log-normally distributed multiplicative perturbations with standard deviations of 0.3 and 0.5, respectively, while normally distributed additive perturbations are applied to longwave radiation (with a standard deviation of 50 W m⁻²) and the soil moisture (SM) estimates at four soil layers [with standard deviations of 0.01, 0.006, 0.003, and 0.0015 m³ m⁻³ for SM1 (top layer), SM2, SM3, and SM4 (bottom layer), respectively]. First-order autoregressive temporal correlations and cross-variable correlations are also considered during the perturbation (Table 1), whereas horizontal error correlations are neglected. The perturbation parameters used in this study are determined based on Kumar et al. (2017, 2019) and Reichle et al. (2008), and have also been effectively applied in Jun et al. (2021) and Kwon et al. (2024).

The spatially and temporally constant observation error standard deviations of 10 % and 0.02 m³ m⁻³ are applied for ASCAT and SMAP soil moisture retrievals, respectively, based on previous DA studies (e.g., Dorigo et al., 2010; Draper et al., 2012; Ferguson et al., 2020; Kolassa et al., 2017; Kwon et al., 2022, 2024). In the KIM-LIS coupled system, the ASCAT-derived soil wetness index data are scaled into the Noah LSM soil moisture climatology (in m³ m⁻³) through the CDF matching (see Sect. 3.3) to remove the systematic bias between the ASCAT soil moisture and Noah-simulated soil moisture. Correspondingly the 10 % ASCAT soil moisture error standard deviation is also locally scaled by the ratio of the Noah LSM and ASCAT soil moisture time series standard deviations following Draper et al. (2012) and Jun et al. (2021). Unlike ASCAT, the SMAP soil moisture data are provided in the same unit as the Noah LSM soil moisture, and only the climatological mean biases between the SMAP soil moisture and modeled soil moisture are corrected during the bias correction procedure (see Sect. 3.3). Therefore, the observation error standard deviation of SMAP is not scaled in this study.

Note that there is a mismatch in the surface soil layer depth between the soil moisture observations (i.e., 0–2 cm for ASCAT and 0–5 cm for SMAP) and Noah LSM (i.e., 0–10 cm). However, Shellito et al. (2016, 2018) and Nair and Indu (2019) have demonstrated that changing the surface layer depth in the model from 10 to 2 cm or 5 cm has only a marginal impact on the simulated soil moisture. Moreover, because we apply the systematic bias correction of the soil moisture retrievals before assimilation, the impact of the surface soil layer depth difference on the DA performance is assumed to be negligible.

A triple collocation analysis (TCA, Stoffelen, 1998; Scipal et al., 2008), a statistical random error estimation method, is applied to evaluate the global soil moisture analysis from the soil moisture DA experiments. TCA has been initially proposed by Stoffelen (1998) to quantify the error of near-surface ocean wind speed estimates, and it is now one of the most commonly used method for estimating uncertainties in satellite-based soil moisture retrievals (e.g., Dorigo et al., 2010; Gruber et al., 2016; Kim et al., 2021b, 2023; Scipal et al., 2008) or for evaluating large-scale soil moisture simulations of computational models (e.g., Kim et al., 2021a; Kwon et al., 2024; Nair and Indu, 2019; Renzullo et al., 2014) due to no requirement of reliable ground truth reference data that is hard to obtain at large scales.

The TCA approach is based on the assumption of a linear relationship between hypothetical true soil moisture and individual soil moisture estimates as expressed in Eq. (1): 1θk=αk+βkθtrue+εk where θk is independent collocated soil moisture datasets (i.e., triplet components, k∈[x,y,z]); αk and βk are additive and multiplicative systematic biases of θk, respectively, with respect to the unknown hypothetical true soil moisture signal (θtrue); and εk represents the additive random noise in each soil moisture data (θk). The random error (noise) variance (σεk2) of three collocated soil moisture triplets [Eqs. (2) to (4)] can be derived from variance and covariance equations by introducing additional assumptions, i.e., error orthogonality (independence between the random error of the soil moisture datasets and the unknown soil moisture truth) and zero error-cross correlation (independence of the random errors between the soil moisture datasets) (Gruber et al., 2016). 2σεx2=σx2-σxyσxzσyz3σεy2=σy2-σyxσyzσxz4σεz2=σz2-σzxσzyσxy where σk2 and σεk2(k∈[x,y,z]) are the variance and random error variance, respectively, of each soil moisture data; and σxy, σxz, and σyz are the covariances of two soil moisture triplet components. In this study, the fractional mean-square error (fMSEk, Draper et al., 2013), ranging from 0 (free-of-noise soil moisture data) to 1 (no meaningful soil moisture signal), is computed using Eq. (5). This metric is employed as a TCA-based global soil moisture evaluation measure, following procedures implemented by Kim et al. (2020 and 2021a) and Kwon et al. (2024). 5fMSEk=σεk2σk2 In order to meet the zero error-cross correlation assumption, we select two independent satellite-based soil moisture products, i.e., ASCAT and Advanced Microwave Scanning Radiometer 2 (AMSR2) [or Soil Moisture and Ocean Salinity (SMOS)], derived from different microwave sensors using different retrieval algorithms for the first and second soil moisture triplet components while the soil moisture simulations from the experiments (CTL, SG_AT, and SG_SP) are used for the third triplet component (Table 3). Due to the TCA assumptions, it is hard to compose the same reference frame for all experimental cases, especially for the multi-sensor soil moisture DA experiment. Therefore, for a fair comparison, we only evaluate the relative improvement in the soil moisture estimates by comparing fMSE of the single-sensor soil moisture DA with fMSE of CTL that are computed using the same first and second triplet components (i.e., satellite soil moisture retrievals that are not assimilated in the soil moisture DA experiment) as shown in Table 3. The effects of the multi-sensor soil moisture DA (MT_ATSP) are assessed only for atmospheric variables (see Sect. 6.2 and 6.3).

The ASCAT, SMOS, and AMSR2 soil moisture data used in TCA are the 12.5 km ASCAT soil moisture Climate Data Record (CDR) version 7 product (H119 and extended H120) based on the TU-Wien change detection algorithm (Wagner et al., 2013), the SMOS-INRA-CESBIO (SMOS-IC) version 2 product, and the AMSR2 X-band Land Parameter Retrieval Model (LPRM) product, respectively. The original datasets are preprocessed by conducting quality control based on quality flags provided in each data product, and spatial resampling using the nearest neighbor distance method to match the spatial resolution of the datasets with that of the LSM outputs (i.e., 25 km latitude-longitude grids). Due to the different local overpass time between the satellite soil moisture products [i.e., 09:30 am/pm LST for ASCAT, 01:30 am/pm LST for ASMR2, and 06:00 am/pm LST for SMOS], the UTC-based Noah-LSM outputs at 04:00 am/pm LST and 11 am/pm LST are extracted for TCA that uses AMSR2/SMOS/SG_AT(or CTL) and AMSR2/ASCAT/SG_SP(or CTL), respectively. We select model outputs at the approximate midpoint time (e.g., 04:00) between the overpass times of two other satellite-based soil moisture triplet components (e.g., 01:30 AMSR2 and 06:00 SMOS) for a fair comparison. While some errors may still arise due to sampling-time mismatches between the triplet components, we assume these errors are acceptable since the same sampling time was applied to both CTL and DA experimental outputs to evaluate their relative performance. Please refer to Kwon et al. (2024) and Kim et al. (2023) for more detailed procedures.

Evaluations of the specific humidity and air temperature analyses/forecasts are performed using the ECMWF Integrated Forecasting System (IFS) analysis (ECMWF, 2017) as reference data, which has been extensively used for evaluating modeling systems (e.g., Ben Bouallègue et al., 2024; Kwon et al., 2024; Lee et al., 2020; Polichtchouk et al., 2023; Reichle et al., 2023). The root mean square difference (RMSD) between the atmospheric variables (i.e., specific humidity and air temperature) from each experiment and those from the IFS analysis is computed.

Local variations in soil moisture modify boundary-layer heat and moisture fluxes, thereby altering water–energy budgets and influencing convective triggering (Findell and Eltahir, 2003; Pal and Eltahir, 2003) and subsequently influence large-scale dynamics (Cook et al., 2006; Pal and Eltahir, 2003), both of which play key roles in determining precipitation processes. A number of studies have investigated the complex interaction mechanisms between soil moisture and precipitation, referred to as the `soil moisture-precipitation feedback', using observational analyses (e.g., Catalano et al., 2016; Yang et al., 2018) and computational modeling systems (e.g., Beljaars et al., 1996; Bosilovich and Sun, 1999; Hohenegger et al., 2009; Lin et al., 2023; Pal and Eltahir, 2003). Although these studies generally agree on a predominant positive feedback, the sign and strength vary depending on modeling systems and spatiotemporal scales (Hohenegger et al., 2009; Lin et al., 2023). Differences in the sign of soil moisture-precipitation feedback can be attributed to the complexity of representing the soil moisture-evapotranspiration relationship (Yang et al., 2018) and convective development (Hohenegger et al., 2009). Considerable debate and uncertainty remain regarding the physical mechanisms determining the sign of the feedback (Hohenegger et al., 2009). Nevertheless, there is no doubt that soil moisture and precipitation are reciprocally linked, implying that better characterization of soil moisture conditions through soil moisture DA can enhance precipitation forecasts in land-atmosphere coupled systems.

We assess the impact of soil moisture DA on precipitation forecast skill using well-established metrics, such as the frequency bias (FB) and equitable threat score (ETS) based on recommendations by the World Meteorological Organization (WMO, 2008). Calculations of the FB and ETS metrics are based on a 2 × 2 contingency table (Table 4), which consists of Hits, Misses, FalseAlarms, and CorrectNegatives. Gauge-based global daily precipitation analyses from the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center (CPC) (Chen et al., 2008; Xie et al., 2007) are used as reference data for precipitation evaluation.

FB [Eq. (6)] assesses the ratio of the frequency of precipitation occurrence in the forecast to that in the observation: 6FB=Hits+FalseAlarmsHits+Misses where Hits, FalseAlarms, and Misses are the number of model grid points with correct forecasts, false alarms, and missed forecasts of precipitation occurrence, respectively. The FB metric ranges from zero to infinity where the ideal FB score is 1.0, indicating that the number of the forecasted precipitation events is the same as that of the observed events. Note that FB does not consider the timing of precipitation events.

ETS [Eq. (7)] quantifies the fraction of the forecasted or observed precipitation events that are captured correctly after excluding random hits (Hits_rnd), which is the number of correct forecasts by random chance computed using Eq. (8): 7ETS=Hits-HitsrndHits+Misses+FalseAlarms-Hitsrnd8Hitsrnd=(Hits+Misses)(Hits+FalseAlarms)N where N is the total number of events defined as N=Hits+Misses+FalseAlarms+CorrectNegatives. CorrectNegatives denotes the number of correct forecasts of no precipitation. The ETS metric ranges from -1/3 to 1, where values below 0 and 1 represent “no skill” and “perfect skill” (no Misses or FalseAlarms), respectively.

The impact of single-sensor soil moisture assimilation (i.e., SG_AT and SG_SP) on the global soil moisture estimates is evaluated using the TCA method (see Sect. 6.1). Figure 2 presents the spatial distribution of fMSE obtained from TCA during the experimental period from April to July 2022. Both ASCAT and SMAP are seen to have an overall positive impact on the surface soil moisture estimates of the Noah LSM through assimilation. Compared to the CTL experiment, which does not assimilate soil moisture data, SG_AT and SG_SP reduce the global mean fMSE by 4.0 % (Fig. 2a) and 10.5 % (Fig. 2b), respectively. In both single-sensor soil moisture DA cases, obvious improvements in soil moisture are observed in Asia while a decrease in skill is mostly found in the Australian and North American continents where CTL already exhibits a relatively good performance in estimating soil moisture.

Note that we use identical first and second triplet components for DA and CTL (Table 3), replacing only the CTL soil moisture estimates with those from the DA experiments (SG_AT and SG_SP) to assess the relative performance gain from soil moisture DA. This approach (i.e., replacing one triplet member) may alter the fMSE calculation of the other two triplet components and thus influence the comparison results between DA and CTL. However, because the soil moisture estimates from DA and CTL share the same spatial and temporal coverage and climatology, as they are generated from the identical modeling system, the impact of replacing the model-based triplet member is negligible, as shown in Fig. S1. Therefore, the fMSE comparison results (Fig. 2) can be considered reliable.

Because land surface characteristics affect the soil moisture skill of models and observations (Draper et al., 2012), the TCA results are also plotted for the MODIS-IGBP land cover types. Figure 3 shows that for all land cover types, both SG_AT and SG_SP enhance the skill of the modeled soil moisture relative to CTL in terms of the median fMSE, with SG_SP achieving greater skill gains. The soil moisture estimates are significantly improved by SG_AT for evergreen broadleaf forests and croplands, and by SG_SP for deciduous needleleaf forests, deciduous broadleaf forests, mixed forests, open shrublands, woody savannas, savannas, grasslands, croplands, and barren or sparsely vegetated land cover types. In both SG_AT and SG_SP experiments, the highest skill improvements, in terms of the median fMSE, are observed for croplands. This implies that the land DA system effectively utilizes soil moisture signals related to agricultural practices from satellite observations, especially in the case of SMAP soil moisture DA (SG_SP), which employs the anomaly-based bias correction approach.

Figures 2 and 3 indicate that SMAP DA shows higher skill than ASCAT DA for the soil moisture analysis. The superior performance of SMAP over ASCAT within a land DA system is also reported in Seo et al. (2021) where SMAP and ASCAT soil moisture DA results are evaluated against in situ measurements in the continental United States. These results can be supported by the fact that L-band brightness temperature measurements have higher sensitivity to soil moisture variations than C-band backscatter measurements (Kolassa et al., 2017), and thus the SMAP soil moisture retrievals have better accuracy (Al-Yaari et al., 2019; Kumar et al., 2018). However, note that in this study, a direct comparison of the global soil moisture analysis between SG_AT and SG_SP is not made because the model soil moisture outputs used in TCA are extracted at different LST – specifically, 04:00 am/pm for SG_AT and 11:00 am/pm for SG_SP – due to the different local overpass times of the satellite soil moisture data used for TCA-based assessment (see Sect. 6.1).

As shown in Fig. 2, soil moisture performance gains and losses by each single-sensor soil moisture DA are locally dependent. Thus, some previous studies (e.g., Draper et al., 2012; Kolassa et al., 2017) have shown that simultaneously assimilating soil moisture retrievals from both passive and active sensors achieves higher model soil moisture accuracy than assimilating a single product. However, because soil moisture triplets that fully satisfy the TCA assumptions (see Sect. 6.1) are difficult to obtain over the global domain for the multi-sensor soil moisture DA experiment, the combined effects of ASCAT and SMAP DA are discussed only in terms of atmospheric variables, which are the ultimate objective of this study, in the subsequent sections.

Domain-averaged RMSD differences (i.e., RMSD_EXP minus RMSD_CTL) in the specific humidity analysis (Figs. 4a to 4c) and air temperature analysis (Fig. 4d to f) are evaluated across atmospheric levels and over time. Compared to CTL (without soil moisture DA), ASCAT DA (i.e., SG_AT) has more beneficial impacts on the air temperature analysis (Fig. 4d) while SMAP DA (i.e., SG_SP) has more beneficial impacts on the specific humidity analysis (Fig. 4b). Figure 4c and f show that the simultaneous assimilation of ASCAT and SMAP soil moisture retrievals (i.e., MT_ATSP) improves the analysis of both atmospheric variables relative to CTL. Notably, degradations in the specific humidity and air temperature analyses by SG_AT and SG_SP, respectively, are compensated by additionally assimilating other soil moisture products.

Figure 5 more clearly demonstrates that multi-sensor soil moisture DA enhances the performance of the specific humidity analyses (Fig. 5a) and air temperature analyses (Fig. 5d) compared to the single-sensor soil moisture DA cases (i.e., SG_AT and SG_SP, respectively). Although MT_ATSP exhibits somewhat reduced performance in the air temperature (Fig. 5c) and specific humidity analyses (Fig. 5b) compared to SG_AT and SG_SP, respectively, it achieves a more balanced improvement, meaning that neither variable is degraded while both show moderate gains compared to CTL, by assimilating radar- and radiometer-based soil moisture data together.

Tables 5 and 6 summarize the domain-averaged RMSD differences between the soil moisture DA experiments and CTL for the analyses and forecasts of 2 m atmospheric variables (i.e., specific humidity and air temperature) from the 00:00 UTC cycle (Table 5) and the 12:00 UTC cycle (Table 6). In the global domain, SG_SP generally achieves better domain-averaged analysis and forecast skills for both 2 m atmospheric variables compared to SG_AT (Tables 5 and 6), except for the 2 m air temperature forecast of the 12:00 UTC cycle, where SG_AT performs slightly better (Table 6). During the experimental period, all soil moisture DA cases are more effective in improving the 2 m air temperature analysis and forecast than those of specific humidity, especially for the 00:00 UTC cycle. Overall, they perform better in the Northern Hemisphere than in the Southern Hemisphere (Tables 5 and 6), although they achieve greater 2 m air temperature forecast skill during the 00:00 UTC cycle in the Southern Hemisphere (Table 5). The lower analysis and forecast skills of soil moisture DA in the Southern Hemisphere can be attributed to the use of spatially and temporally constant observations errors (see Sect. 4.2), which do not adequately reflect the relatively higher uncertainties in winter-period soil moisture data.

For the 2 m specific humidity estimates, the assimilation of SMAP soil moisture retrievals alone (SG_SP) achieves the best domain-averaged performance (Tables 5 and 6). MT_ATSP reduces RMSD compared to SG_AT by additionally assimilating the SMAP soil moisture data, but it exhibits relatively lower skill in specific humidity than SG_SP. However, in Europe and tropical regions, MT_ATSP provides improved 2 m specific humidity forecasts compared to SG_SP, particularly when both ASCAT and SMAP have a positive impact. The synergistic impacts of the combined assimilation of ASCAT and SMAP are evident in the 2 m air temperature analysis and forecast of the 00:00 UTC cycle (Table 5).

Local performance differences in the 2 m atmospheric analysis between the single-sensor soil moisture DA cases are clearly illustrated in Fig. 6, which is plotted for the 12:00 UTC cycle from June to July 2022. SG_AT and SG_SP exhibit opposite impacts on the atmospheric analysis, especially on specific humidity, over India, Eurasia, and Brazil. ASCAT DA (SG_AT) generally leads to the improved analyses of specific humidity and air temperature over India while SMAP DA (SG_SP) performs better in Eurasia (except for India and the southern part of West Asia) and Brazil. These discrepancies in the local performance of 2 m specific humidity between ASCAT and SMAP DA may contribute to the reduced domain-averaged skill of MT_ATSP relative to SG_SP, as shown in Tables 5 and 6. In Africa, SG_AT yields better air temperature analysis than SG_SP, whereas SG_SP outperforms SG_AT in the specific humidity analysis. It can be noted from Fig. 6 that, locally, the joint assimilation of the ASCAT and SMAP soil moisture retrievals yields the best estimates of the 2 m atmospheric variables when both soil moisture products have positive impacts.

The potential added value of multi-sensor soil moisture DA for precipitation forecasts is assessed using categorical skill score metrics, including the FB and ETS, as detailed in Sect. 6.3. Daily precipitation rates (mm d⁻¹) are computed from the KIM forecasts at 0–24, 24–48, and 48–72 h lead times, and compared against reference data using seven conventional thresholds (0.5, 1.0, 5.0, 10.0, 15.0, 20.0, and 25.0 mm d⁻¹). The numbers of model grid points classified as Hits, FalseAlarms, Misses, and CorrectNegatives in the contingency table (Table 4) are then counted, and FB and ETS are calculated for six domains (i.e., global domain, Northern Hemisphere, Southern Hemisphere, Asia, Europe, and tropical area). Figure S2 presents the FB and ETS of daily precipitation forecasts from KIM, averaged over the three lead times, for CTL and the three soil moisture DA experiments (SG_AT, SG_SP, and MT_ATSP) during the 00:00 UTC cycle in July 2022. The corresponding differences (ΔFB=FBEXP-1-FBCTL-1, ΔETS=ETSEXP-ETSCTL) are shown in Fig. 7.

In the global domain, CTL (without soil moisture DA) tends to overestimate precipitation frequency (FB > 1.0), simulating excessive precipitation events, except for the precipitation threshold of 25.0 mm d⁻¹ (Fig. S2a). The model significantly overestimates precipitation at the 5.0 mm d⁻¹ threshold and exhibits an FB close to 1 for heavier precipitation events (20.0 and 25.0 mm d⁻¹) (Fig. S2a), with regional variations (Fig. S2b to f). Both the smallest FB (close to 1.0) and the largest FB (>2.5) are observed in the Southern Hemisphere for the lightest (0.5 mm d⁻¹) and heaviest (25.0 mm d⁻¹) precipitation events, respectively (Fig. S2c).

Unlike FB, ETS is higher (indicating better skill in predicting precipitation events) at lower precipitation thresholds while ETS decreases as the precipitation intensity thresholds increase in the global domain (Fig. S2g) and in the Northern Hemisphere (Fig. S2h), including Asia (Fig. S2j), Europe (Fig. S2k), and tropical areas (Fig. S2l). In the Southern Hemisphere, CTL shows the highest ETS skill at the precipitation threshold of 10.0 mm d⁻¹ and the lowest at 1.0 mm d⁻¹ (Fig. S2i). The different model performance patterns (in both FB and ETS) between the Northern and Southern Hemispheres across the range of precipitation intensity thresholds may be attributed to different weather regimes associated with cyclones and monsoons (Dare and Ebert, 2017), along with additional impacts from seasonal variations.

Overall, soil moisture DA improves the prediction of precipitation events (i.e., better ETS; Fig. 7g) while its contribution to precipitation frequency (FB) remains neutral (Fig. 7a). MT_ATSP demonstrates higher ETS skill than CTL (by up to 1.8 %) and single-sensor soil moisture DA (by up to 2.4 % and 0.6 % relative to SG_AT and SG_SP, respectively) (Fig. 7g). The impacts of MT_ATSP on the global FB are marginal, showing negligible improvements at the 0.5, 1.0, 5.0, and 25.0 mm d⁻¹ thresholds and slight overpredictions at the 10.0, 15.0, and 20.0 mm d⁻¹ thresholds (Figs. 7a, S2a). Similar soil moisture DA performance patterns are observed in the Northern Hemisphere (Fig. 7b and h) and Asia (Fig. 7d and j). In the Southern Hemisphere, MT_ATSP slightly improves FB for heavy precipitation events (precipitation thresholds ≥ 15.0 mm d⁻¹) while relatively obvious improvements in ETS by MT_ATSP are witnessed at lower thresholds (≤1.0 mm d⁻¹) (Fig. 7c and i). Compared to CTL and SG_AT, multi-sensor soil moisture DA enhances ETS across precipitation thresholds in tropical areas, although it is slightly less effective than SG_SP at thresholds ≤5.0 mm d⁻¹ (Fig. 7l). For FB, MT_ATSP shows improvements over CTL and SG_AT in tropical areas, except at thresholds of 5.0 and 10.0 mm d⁻¹, where SG_SP performs better (Fig. 7f). In contrast, MT_ATSP is generally ineffective in Europe (Fig. 7e and k), except for ETS at precipitation thresholds of 5.0 mm d⁻¹ or lower. The overprediction of precipitation in Europe (Figs. 7e, S2e), especially for heavy precipitation events (≥15.0 mm d⁻¹), may lead to a decrease in ETS (Figs. 7k, S2k).

The regional variability in precipitation performance between the CTL and soil moisture DA experiments may be attributed to seasonality. A full-year or multi-year experiment would therefore be valuable in future studies to assess the impact of seasonality on the effectiveness of multi-sensor soil moisture DA for improving precipitation forecasts.

The experimental results indicate that SMAP DA slightly outperforms ASCAT DA in enhancing near-surface atmospheric analyses and forecasts. While many factors may affect the overall performance of each experiment, one factor may be related to errors resulting from subsurface scattering, which are not accounted for in the ASCAT soil moisture data used in this study, as discussed by Wagner et al. (2024). The current TU Wien change detection algorithm (Wagner et al., 1999, 2010), used to retrieve soil moisture from ASCAT backscatter observations, is based on the assumption of a positive linear relationship between soil backscatter and wetness. However, Wagner et al. (2024) demonstrated that this assumption fails when coarse fragments (e.g., stones and rocks) or discontinuities exist in the soil profile, as they increase subsurface scattering contributions to total backscatter signals under dry soil conditions. This eventually reverses the relationship between backscatter signals and soil moisture content, deteriorating the quality of soil moisture data retrieved with the current algorithm in many arid and semiarid regions. This phenomenon partly explains why ASCAT DA exhibits better performance in croplands than in barren or sparsely vegetated areas, as shown in Fig. 3, which is consistent with results from soil moisture data evaluation studies (e.g., Dorigo et al., 2010).

Soil moisture bias correction methods used to remove systematic discrepancies between observations and models may also influence the ASCAT soil moisture DA performance. Kumar et al. (2015) and Kwon et al. (2022, 2024) have shown that different bias correction methods can substantially impact soil moisture DA performance. In SMAP DA, soil moisture temporal variability information is directly assimilated (i.e., anomaly correction method). In contrast, ASCAT DA employs the CDF matching method because it does not satisfy the underlying assumption of the anomaly correction method (see Sect. 3.3). It is a known issue that rescaling-based bias correction, such as CDF matching, causes a significant loss of information (Kumar et al., 2015). Meanwhile, Sect. S2 of the Supplementary Material and Figs. S3–S5 suggest that employing the anomaly correction method in soil moisture DA worsens the atmospheric analysis within land-atmosphere coupled systems when the underlying assumptions of the anomaly correction approach are not met. Rather than relying on CDF matching and anomaly correction methods, a more robust and appropriate bias correction approach is required for improving ASCAT soil moisture DA.

The effectiveness of the combined assimilation of multiple soil moisture products from different sources (i.e., radar backscatter and radiometer brightness temperature observations) in enhancing the soil moisture analysis skill has been demonstrated in previous studies (e.g., Blyverket et al., 2019; Draper et al., 2012; Kolassa et al., 2017; Renzullo et al., 2014). Our experimental results also show that, compared to single-sensor soil moisture DA, simultaneous assimilation of the ASCAT and SMAP soil moisture retrievals within the KIM-LIS coupled system has a synergistic impact, improving the analyses and forecasts of atmospheric variables including specific humidity, air temperature, and precipitation. However, the superiority of individual single-sensor soil moisture retrieval products and their combined performance within the DA system depend on the region and time period (Fig. 6). Specifically, locally reduced benefits of SMAP or ASCAT are observed in the multi-sensor soil moisture DA experiment when their performance impacts are opposite (Fig. 6). This may be attributed to the use of uniform observation error standard deviations for the soil moisture retrievals across space and time, which is unrealistic and does not adequately reflect the actual and relative retrieval skill of each soil moisture product. Especially, underestimated soil moisture observation errors can degrade the soil moisture and atmospheric analyses by over-relying on less reliable soil moisture data within the DA framework.

One key advantage of simultaneously assimilating individual soil moisture products, rather than a preprocessed multi-sensor-derived soil moisture product, is the flexibility to handle individual soil moisture sensors separately within a single DA system while achieving comparable performance (Kolassa et al., 2017). However, this advantage can be more effectively utilized by accurately specifying the relative errors of the soil moisture retrievals in space and time, enabling the optimal combination of models and diverse soil moisture products.

Several recent studies (e.g., Wu et al., 2021; Kim et al., 2023; Kim et al., 2025) have made efforts to estimate spatially or spatiotemporally distributed errors in satellite-based soil moisture retrievals using TCA-based methods and machine learning algorithms. A map of spatially distributed (but time-invariant) error standard deviations from Kim et al. (2025) [see Fig. S6, which is Fig. 5a in Kim et al. (2025)] exhibits that the SMAP soil moisture data have errors greater than 0.05 m³ m⁻³ in many areas, particularly in forests. Excluding forested areas where soil moisture retrievals are masked out during quality control (see Sect. 3.4) and thus not assimilated, the error standard deviations of SMAP soil moisture are still above 0.02 m³ m⁻³ (the uniformly applied error value in this study) in some savanna regions of South America and Africa, grasslands in North America, and croplands in South Asia (Fig. S6) [see Fig. S3c for the land-cover type map generated in this study using the MODIS-IGBP global land-cover classification (Friedl et al., 2002)]. As a result, this leads to degradation in the 2 m specific humidity and air temperature analyses in the regions, as noted in the SMAP soil moisture DA (SG_SP) results shown in Fig. 6.

Meanwhile, in ASCAT soil moisture DA, the spatially distributed soil moisture observation error standard deviation (m³ m⁻³) (Fig. 8) is applied after locally rescaling the uniformly specified 10 % soil wetness index observation error using the ratio of the standard deviations of modeled and observed soil moisture time series (see Sect. 4.2). However, the spatial pattern of the ASCAT observation error standard deviations in Fig. 8 does not completely match that of the TCA/ML-based ASCAT fMSE computed by Kim et al. (2023) (see their Fig. 4f), which was derived using other satellite soil moisture data as triplet components. Notably, the ASCAT errors used in the current study appear to be relatively underestimated in dry areas of Africa and Asia (Fig. 8), where the 2 m atmospheric analyses are degraded by ASCAT soil moisture DA (SG_AT in Fig. 6).

The use of pre-generated spatial or spatiotemporal observation error estimates (e.g., Kim et al., 2025) can potentially maximize the benefits of each soil moisture product in multi-sensor soil moisture DA systems. One critical issue, however, is that bias correction of soil moisture observations, an essential procedure in soil moisture DA, may substantially alter their error characteristics, especially when rescaling-based methods like CDF matching are employed. To effectively apply spatially and temporally varying observation error estimates in multi-sensor soil moisture DA, a refined approach is needed to propagate error estimates from original soil moisture retrievals to bias-corrected soil moisture values.