Informed by the complementary strengths of active and passive microwave sensing at divergent frequencies, this study assimilates soil-moisture retrievals from three mainstream sensors: the U.S. Soil Moisture Active Passive (SMAP), the European Advanced Scatterometer (ASCAT), and China's Microwave Radiation Imager (MWRI). The data period covers June to August 2022, and the observations were aggregated to a daily timescale before assimilation.

This study utilized the Soil Moisture Active Passive (SMAP) Level-2 Enhanced Passive Soil Moisture product (L2_SM_P_E) provided by the U.S. National Snow and Ice Data Center (NSIDC) (O'Neill et al., 2023). Launched in January 2015, the SMAP satellite has been continuously providing global radiometer observations since April 2015. It operates in a sun-synchronous orbit with an L-band (1.41 GHz) radiometer. The descending and ascending overpasses occur at approximately 06:00 a.m. and 06:00 p.m. Local Solar Time, respectively, yielding a global revisit time of 2–3 d. Based on the well-validated τ-ω radiative transfer model, this product implements both the Single Channel Algorithm using vertical polarization (SCA-V) and the Dual Channel Algorithm (DCA) for soil moisture retrieval (Chaubell et al., 2020; O'Neill et al., 2021). The characteristics of the L-band provide it with canopy penetration capabilities. Although the native spatial resolution of the radiometer is approximately 36 km, this enhanced product utilizes the Backus-Gilbert optimal interpolation technique to be distributed at a spatial sampling of 9 km on the EASE-Grid 2.0 projection.

The ASCAT retrieval data are derived from the Advanced Scatterometer onboard the MetOp-C satellite operated by the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT). Following its launch in November 2018, MetOp-C has continuously provided high-quality global data, extending the long-term ASCAT data record. The satellite has equatorial crossing times of approximately 09:30 a.m. (descending) and 09:30 p.m. (ascending), providing a global revisit time of about 3 d. As an active microwave radar operating at the C-band (5.255 GHz, VV polarization), ASCAT provides global soil moisture estimates by measuring the surface radar backscatter coefficient (σ0). This study utilizes the ASCAT Level 2 Surface Soil Moisture product (EO:EUM:DAT:METOP:SOMO12) at 12.5 km spatial sampling on the swath grid. The temporal resolution of this product relies on daily morning and evening discrete observations. Regarding the retrieval algorithm, this product is based on the TU Wien change detection algorithm, obtaining relative soil moisture, which is expressed as a degree of saturation percentage from 0 % to 100 %, by linearly scaling the backscatter coefficient between historical dry and wet reference lines (Bartalis et al., 2007). Subsequently, this study utilizes soil porosity data provided by the European Space Agency Climate Change Initiative (ESA-CCI) to convert this saturation into volumetric soil water content.

The MWRI soil moisture dataset was developed by the National Satellite Meteorological Center (NSMC) of the China Meteorological Administration (CMA). The FY-3D afternoon satellite was successfully launched in November 2017, with its MWRI data record officially available since early 2018. The dataset is derived from global passive microwave brightness temperature observations acquired by the MWRI instrument onboard FY-3D, which has equatorial crossing times of approximately 02:00 a.m. (descending orbit) and 02:00 p.m. (ascending orbit), providing a global revisit time of 1–2 d. Surface soil moisture is primarily retrieved from the X-band (10.65 GHz) dual-polarized brightness temperatures using a radiative transfer equation-based retrieval algorithm, in which vegetation effects are corrected and soil surface roughness effects are addressed using the Qp model (Kang et al., 2020). This daily product combines ascending and descending overpasses and is provided on a 25 km Equal-Area Scalable Earth (EASE-Grid) projection. We selected X-band retrieval products because X-band applications remain relatively limited, and we aimed to demonstrate their value through direct assimilation. Direct assimilation of radiances is certainly an important direction for land surface data assimilation. However, it requires a reasonable estimation of the errors in simulations from the radiative transfer model. This aspect warrants further investigation in future studies. The key characteristics of the three satellite soil moisture products used in this study are summarized in Table 1.

Given that the spatial resolution of satellite products exceeds that of the LSM (about 1.4°×1.4°), all satellite soil moisture retrievals within each model grid cell were spatially averaged to prevent introducing excessive small-scale variability. The resulting spatially averaged soil moisture distribution is shown in Fig. 1.

Before assimilation, quality control was conducted for each satellite dataset, primarily based on their native quality flags. For SMAP, we kept retrievals with quality flags of 0 or 8. For ASCAT, we retained observations with an aggregated quality flag below 50 % (Chen et al., 2018). For FY3D, only observations with a quality flag of 0 were used. We then performed bias correction using cumulative distribution function (CDF) matching. Specifically, satellite-derived soil moisture was matched to the CoLM background simulations for summer 2022 and then mapped to the model's second soil layer, consistent with Albergel et al. (2017). This layer was selected because it better matches the effective sensing depth of the satellite products. In this CoLM setup, the top layer is 0–1.75 cm thick and responds quickly to rainfall, evaporation, and surface thermal forcing, so it mainly captures short-lived surface fluctuations. The second layer, which spans 1.75–4.51 cm, is closer to the main sensing depth of SMAP and ASCAT (about 0–5 cm) and provides a more stable representation of near-surface soil moisture. It was therefore used for observation mapping and bias correction.

In-situ data were selected to validate the assimilation results. Soil moisture and precipitation observations were obtained from the ISMN and the CMA station network. The ISMN comprises over 1400 stations across 35 international networks (Dorigo et al., 2021). The CMA network maintains an extensive infrastructure of surface meteorological stations across China. To ensure data reliability, quality control was performed based on the original quality flags provided by each network prior to validation, retaining only valid measurements. For the ISMN dataset, which includes the USCRN, COSMOS-UK, SOILSCAPE, ARM, TAHMO, SMOSMANIA, and FR_Aqui networks, we utilized the station-specific quality flags and only kept observations marked as “G”. Similarly, only valid observations with a quality flag of 0 were retained for the CMA network. To maximize the spatial coverage for evaluation, this study did not impose a strict threshold on time-series completeness during the quality control process, ensuring that all valid observations passing the physical screening were included in the analysis.

A total of 3057 stations remained after these quality control procedures. Figure 2 illustrates the spatial distribution of the CMA and ISMN in-situ stations to provide a clear visualization of the ground-based data coverage. Among these, 2792 CMA stations are densely distributed across most of China as shown in Fig. 2a. Additionally, a total of 265 ISMN stations were utilized, with the majority located in the United States and others distributed across Europe and Africa. Given that the highest station density occurs in the U.S. region, Fig. 2b primarily highlights the distribution of ISMN stations in that area.

In this study, the 10 cm soil depth was used for primary validation to bridge the gap between satellite sensing and in-situ monitoring. Since L-, C-, and X-band microwave observations mainly reflect near-surface soil moisture, and 10 cm is the shallowest standardized measurement depth also has the fewest missing values across the CMA station network. To better align with the research objectives and ensure a sufficient number of validation data points, we focus on evaluating the impact of land data assimilation on 10 cm soil moisture.

To improve the global applicability of validation, this study incorporates the ERA5-Land dataset developed by the European Centre for Medium-Range Weather Forecasts (ECMWF) under the Copernicus Climate Change Service (C3S) (Muñoz-Sabater, 2019). ERA5-Land is a high-resolution enhancement of the ERA5 reanalysis, focusing on land surface variables. Its soil moisture product includes four layers (0–7, 7–28, 28–100, and 100–289 cm) providing volumetric water content.

This study uses the CoLM to simulate land surface processes. Developed primarily by Chinese research institutions, CoLM originated from integrating the IAP94 and CLM frameworks through collaboration between the Institute of Atmospheric Physics, Chinese Academy of Sciences, and NCAR. The model represents coupled soil hydrothermal and vegetation ecophysiological processes, and is widely applied in Earth system studies. We use the 2014 version, which incorporates major advances in soil physics and vegetation parameterizations, including an updated soil hydraulic scheme and a two-big-leaf canopy model. The model consists of 10 vertically stratified soil layers from the surface to bedrock. Specifically, the bottom depths of these 10 layers are 1.75, 4.51, 9.06, 16.55, 28.91, 49.29, 82.89, 138.28, 229.61, and 380.19 cm, respectively. Soil moisture is simulated by numerically solving the Richards equation, accounting for root uptake, infiltration, and surface and subsurface runoff. Vegetation processes follow a plant functional type (PFT) approach, with a two-big-leaf scheme separating sunlit and shaded leaves to better represent canopy radiative and energy transfer (Dai et al., 2003; Li et al., 2017). In this study, the Leaf Area Index (LAI) is prescribed from the monthly reprocessed MODIS LAI dataset, which was generated following the approach of Yuan et al. (2011). This prescribed LAI drives the two-big-leaf canopy scheme in CoLM to partition the canopy into sunlit and shaded leaves and to compute canopy-scale radiation absorption, photosynthesis, stomatal conductance, and transpiration. Hydrological processes include precipitation interception, infiltration, evapotranspiration, and runoff generation, with subgrid-scale heterogeneity in topography, soil, and vegetation types explicitly considered. For the simulations conducted in this study, the model is configured with a horizontal spatial resolution of 1.4°×1.4° and an output temporal resolution of 1 h.

The atmospheric forcing data are derived from the ERA5 near-surface atmospheric variables, including downward shortwave and longwave radiation, precipitation, air pressure, specific humidity, air temperature, and wind speed (Hersbach et al., 2020). A bilinear interpolation method was used to interpolate the ERA5 data from its original 0.25° horizontal resolution to a 1.4° grid. The interpolated forcing data have a temporal resolution of 3 h and cover the period from 1 January 1979, to 31 December 2022. For the model integration, the CoLM employs an explicit finite-difference method with a time step of 30 min to ensure simulation stability and accuracy. The model outputs, including soil temperature, moisture, surface fluxes, evapotranspiration, and runoff, provide a foundation for subsequent data assimilation and evaluation.

To accommodate the strong nonlinearities characteristic of land-surface processes, the Simplified Extended Kalman Filter (SEKF) – an assimilation scheme that dispenses with the adjoint model – is employed herein for soil-moisture data assimilation. SEKF was proposed by Hess (2001) and Balsamo et al. (2004) based on the Extended Kalman Filter (EKF). It approximates the derivative of the nonlinear observation operator by adding appropriate perturbations, thereby avoiding the development of a cumbersome adjoint model. Compared to the EKF, SEKF does not require local linearization of model equations and updates the observation operator gradient in real time, thereby retaining some flow-dependent error characteristics. The method is computationally efficient and has been implemented in land data assimilation systems such as that of ECMWF (Drusch et al., 2009; de Rosnay et al., 2013; Herbert et al., 2024).

The analysis update step of the SEKF is consistent with the Kalman filter. First, the state is updated by weighting the difference between the observation and the simulated observation (i.e., Observation-Minus-Background, OMB) with the Kalman gain matrix K: 1xa=xb+Ky-Hxb

Here, xa is the analysis state, xb is the background state, y is the observation, and H is the nonlinear observation operator. The Kalman gain matrix K is computed as follows: 2K=BHTHBHT+R-1

B is the background error covariance, R is the observation error covariance, and H is the linearized observation operator. To account for the nonlinear nature of the observation operator in LSM, the SEKF computes H using a finite-difference perturbation approach. Specifically, a small perturbation δxj is applied to each component xi of the state vector, and the resulting change in the simulated observation is used to approximate the first-order derivative of the observation operator with respect to that state component: 3Hij=Hxi+δxi-Hxiδxj

The observation operator is estimated using the finite-difference approach described above. Each perturbation must be small enough to satisfy the linear approximation yet large enough to avoid numerical noise. According to the limit definition of the derivative, the derivative of a function can be approximated via the finite difference method. For any given model grid point, the first-order derivative of the observation operator can thus be calculated by introducing a small perturbation to the independent variable. As the perturbation amplitude becomes sufficiently small, the difference between the perturbed and unperturbed function values of the observation operator approximates its first-order derivative. To identify a sufficiently small perturbation, the right derivative H+ and left derivative H- are evaluated and compared. When the discrepancy between H+ and H- approaches zero, their mean value is adopted as the first-order derivative of the observation operator at the corresponding grid point (Balsamo et al., 2004). To minimize the influence of meteorological variability, we calculated globally averaged derivatives over all grid points under different perturbation magnitudes (Fig. 3). As an example, for the second soil layer (Fig. 3b), a perturbation of 10⁻³ m³ m⁻³ produced a stable derivative of approximately 0.35, while the difference between H+ and H- is close to zero. Based on this criterion, the optimal perturbation values for the first seven soil layers are selected as 0.001, 0.0001, 0.0001, 0.0001, 0.0005, 0.001, and 0.0005 m³ m⁻³, respectively.

The configuration of background and observation errors in this study follows the methodology established by Bai et al. (2025). For areas without station observations, errors are characterized by the root-mean-square error (RMSE) against the ERA5-Land product. In station-covered regions, observation and background errors are estimated via TC using CoLM simulations, in-situ data, and satellite retrievals as independent inputs. These TC estimates were scaled to match RMSE magnitudes to ensure consistency. Spatial patterns of these errors are shown in Fig. 4. Based on these estimated errors, and following de Rosnay et al. (2013), we specified both the background and observation error covariance matrices as diagonal, with their diagonal entries derived from the errors. In the assimilation process, only land pixels are considered according to CoLM's underlying surface data, while inland water and frozen soil grids are masked. Additionally, a 3σ innovation check is applied, rejecting any observations that deviate from the background by more than three standard deviations.

The primary objective of this study is to quantify the extent to which assimilating diverse soil-moisture retrievals can enhance the skill of the CoLM in simulating the land surface variables. To ensure the deep-layer soil variables reached a statistical equilibrium, the model was first spun up for 342 years from a cold start. This spin-up was run entirely in open loop, with all assimilation experiments starting from the same open-loop initial conditions rather than from a separate assimilation spin-up. Subsequently, a series of sensitivity experiments were designed to quantify the assimilation efficacy of different soil moisture products.

Three groups of numerical experiments were conducted. The first is a control experiment (CTL), serving as a reference benchmark. In CTL, the CoLM was integrated from a hot start for two months (2 June–1 August 2022) without data assimilation. In the second experiment suite, the SEKF assimilation module is activated to sequentially ingest SMAP-, ASCAT-, and MWRI-derived soil-moisture products, designated as the DA_L, DA_C, and DA_X experiments, respectively. Each experiment consists of one month of daily assimilation cycles, followed by a one-month free forecast. Observations were assimilated once daily at 00:00 UTC. The third group of experiments was conducted to investigate the effects of continuous data assimilation on model performance in simulating the soil moisture. This set involved continuous assimilation over the entire two-month study period, encompassing the assimilation of single datasets (DA_L, DA_C, and DA_X), their pairwise combinations (DA_L+C, DA_L+X, DA_C+X), and all three datasets simultaneously (DA_ALL). A summary of all experiment configurations is provided in Table 2.

A once-daily assimilation cycle was used in this study. Soil moisture has clear temporal persistence, and its day-to-day variability can therefore be adequately represented at a daily update interval. This choice is also consistent with the temporal characteristics of the input products, since ASCAT and SMAP are instantaneous satellite retrievals, whereas the FY-3D MWRI product used here is a fused daily product without a specific observation time. Using a 24 h assimilation window thus provides a common framework for integrating the three datasets. In addition, a higher update frequency would markedly increase the computational burden of the SEKF, which requires perturbation integrations to estimate the Jacobian of the observation operator at each analysis step. Furthermore, this daily approximation has negligible impact on the results. As shown in Fig. S1 in the Supplement, the global mean diurnal variation in soil moisture is about 0.0006 m³ m⁻³, much smaller than the typical satellite retrieval uncertainty of about 0.04 m³ m⁻³. This approximation is therefore adequate for the temporal scale of this study and has little effect on the global soil moisture estimates.

To quantitatively evaluate the differences among various assimilation experiments, the correlation coefficient, RMSE, and unbiased RMSE (ubRMSE) were selected as the primary evaluation metrics in this study. The specific formulas are defined as follows: 4R=∑i=1Nxi-x‾yi-y‾∑i=1Nxi-x‾2∑i=1Nyi-y‾25RMSE=1N∑i=1Nxi-yi26ubRMSE=1N∑i=1Nxi-x‾-yi-y‾2 where xi and yi denote the evaluated assimilation data and the reference data at the ith sample, respectively; x‾ and y‾ represent their respective sample means; and N is the total number of samples involved in the statistics. The correlation coefficient is used to characterize the consistency between the two datasets in spatial distribution or temporal variation, RMSE is used to measure the overall error, and ubRMSE is used to represent the error after removing the systematic bias.

Comparing Group 1 (CTL) and Group 2 (assimilation experiments) clarifies the forecast improvements from land data assimilation. ERA5-Land soil moisture is used to quantify the effect of assimilating L-, C-, and X-band satellite retrievals on the simulations. Figure 5 shows spatial correlation coefficients (R) and root-mean-square error (RMSE) relative ERA5-Land. For each soil layer and each time step, R and RMSE were calculated over all available global land grid cells, with ERA5-Land as the reference. Assimilation from all three bands improves soil moisture accuracy across 0–1 m, with band- and depth-dependent magnitudes. C-band yields the largest gains, R increases by 0.15, 0.12, and 0.09 and RMSE decreases by 0.020, 0.015, and 0.012 m³ m⁻³ for the 0–7, 7–28, and 30–100 cm layers, respectively. L-band ranks second in performance and exhibits clear depth-dependent variations. At the surface layer, its ΔR reaches 0.13, which exceeds that of X-band (0.11). In the second layer, the improvements are comparable between the two, while in the third layer, L-band performs slightly worse than X-band. In the forecast phase, the results differ from those in the assimilation phase. L-band shows the smallest improvement in both R and RMSE, suggesting a more rapid loss of its assimilation impact. In contrast, the effects of C- and X-band assimilation remain more persistent throughout the forecast period. It should be noted that ERA5-Land is an offline simulation driven by ERA5 atmospheric forcing. Because ERA5 itself has already assimilated ASCAT soil moisture data, this indirect coupling implies that the validation process is not strictly independent. This relationship might lead to an overestimation of the skill scores for ASCAT to some degree.

To further explore the causes of performance differences among the frequency bands, the mean first-layer soil moisture is analyzed. Since microwave signals at different frequencies are affected differently by vegetation, the vegetation distribution is also considered in the analysis. The vegetation type data used in this study are derived from the built-in dataset of the CoLM. Considering that each model grid cell typically contains multiple vegetation types, we select the vegetation type with the largest area fraction within each grid cell as the representative vegetation type. As shown in Fig. 6, the analysis fields exhibit notable spatial variations and a clear latitudinal dependence. Over southeastern China between 10–30° N, which is predominantly covered by forest vegetation, all three experiments identify the region as having high soil moisture, yet the DA_L experiment is comparatively wetter and in closer agreement with ERA5-Land. At higher latitudes, such as northern Canada and Siberia, the dry–wet contrast is more pronounced in the DA_X and DA_C experiments. However, theDA_L experiment exhibits an overly dry bias that deviates from ERA5-Land. In the wetter regions of southeastern Australia, which are mainly dominated by grass vegetation, the X-band results most closely match ERA5-Land, while the DA_C experiment displays a distinct dry bias. These results indicate that the assimilation performance of different frequency bands exhibits strong regional dependence.

To further clarify the regional dependence of assimilation performance, the global land surface was divided into three latitudinal bands: 90–30° S, 30° S–30° N, and 30–90° N. Spatial correlation coefficients and RMSE were computed for each zone, as shown in Fig. 7. The calculation follows the same method as the global analysis, but is applied separately within each latitudinal band, and the resulting time-step metrics are averaged over the corresponding period. In the CTL, the best model performance can be found in the tropics (30° S–30° N), where spatial correlation reached 0.8 and RMSE was approximately 0.115 m³ m⁻³. In contrast, performance was poorest in the Northern Hemisphere (30–90° N), with correlation around 0.4 and RMSE exceeding 0.15 m³ m⁻³.

Assimilation experiments show that in the Southern Hemisphere and tropical regions, the L-band provides the most notable improvement during both the assimilation and forecast phases. For example, in the 90–30° S band, the second-layer soil moisture correlation increased by 0.10 and RMSE decreased by 0.015 m³ m⁻³. However, in the Northern Hemisphere (30–90° N), the DA_C and DA_X experiments outperformed DA_L. In particular, C-band assimilation improved the second-layer correlation by 0.16 and reduced RMSE by 0.024 m³ m⁻³ relative to the CTL, indicating the best performance in this region.

Given the differing penetration capabilities of microwave frequencies and the strong correlation between vegetation density and latitude, variations in vegetation type likely contribute to the observed latitudinal differences in assimilation performance. Figure 8 shows the evaluation results of the assimilation experiments across different vegetation types, validated against ERA5-Land to ensure sufficient sample sizes. The vegetation types are sorted in ascending order based on the average precipitation of each region during the assimilation period. For clarity, a comprehensive list of vegetation abbreviations has been provided in Table 3. The calculation follows the same method as the latitudinal-band analysis, but is applied separately to land grid cells belonging to each representative vegetation type. Assimilation of MWRI X-band data performs best in sparsely vegetated regions, with consistent improvements across all soil depths. In contrast, in densely vegetated areas, the deeper penetration of lower-frequency sensors becomes evident. SMAP (L-band) demonstrates the most stable performance, particularly in deeper soil layers. For example, in temperate broadleaf forests (BET Temperate), the improvement in correlation for deep-layer soil moisture exceeds that of the surface layer, while the DA_X experiment exhibits a negative impact. These results align with findings from Mousa and Shu (2020), who used the TC method and reported that SMAP significantly outperformed ASCAT in forested regions of Africa.

Recognizing that reanalysis data are not absolute truth, we further evaluated the assimilation performance using in situ observations from ISMN and CMA. Figure 9 presents correlation and RMSE statistics for each experiment. The vegetation types follow the same order as in Fig. 8. For each station, the correlation coefficient and RMSE are calculated from the time series of simulated and observed soil moisture during the assimilation period, and the results are then grouped by vegetation type. Some vegetation types are omitted due to a lack of corresponding station data. The correlation results show that both L-band and C-band assimilation consistently improve model performance across all available vegetation types, except in the Arctic Grass region where RMSE increases slightly. Notably, in densely vegetated areas, assimilation of the MWRI product leads to a decline in correlation, highlighting its limited effectiveness under dense canopy conditions. These findings are consistent with the reanalysis-based evaluation and reinforce the robustness of the earlier conclusions.

As defined in the experimental design, the primary difference among the experiments lies in the observation data used. Given that all satellite products achieve near-global coverage during the one-month continuous assimilation period, differences in assimilation performance are largely attributed to observation errors. To explore the relationship between assimilation performance and vegetation type, we evaluated the retrieval accuracy of each satellite product using in situ soil moisture observations under different vegetation types. Figure 10 presents the correlation and RMSE of SMAP, ASCAT, and MWRI retrievals against in situ measurements, with vegetation types ordered as in Fig. 8.

The retrieval accuracy of all three products exhibits a clear dependence on vegetation density. Owing to the stronger canopy penetration capability of L-band microwave radiometry, SMAP maintains relatively higher correlation and lower RMSE in densely vegetated areas compared to the other products. ASCAT exhibits competitive overall performance relative to SMAP, owing partly to its robust time-series change detection algorithm. It tends to slightly outperform SMAP in low- and sparse-vegetation regions but performs marginally worse in moderate-to-high vegetation zones due to greater C-band signal attenuation. In contrast, the X-band MWRI product is notably inferior to both SMAP and ASCAT in moderately and densely vegetated areas owing to its shorter wavelength and limited penetration. Over sparsely vegetated and bare-soil regions, however, MWRI achieves accuracy that is comparable, or occasionally marginally superior, to the other two products. These results suggest that spatial variations in assimilation performance are primarily driven by the intrinsic differences in retrieval accuracy among the products. In summary, the relatively longer wavelengths of the L-band and C-band, combined with robust retrieval algorithms, enable SMAP and ASCAT to maintain high accuracy in vegetated areas, while the shorter X-band (MWRI) is more severely limited by vegetation-induced scattering and attenuation.

The vegetation-dependent differences in assimilation performance may help explain the pattern observed in Fig. 5, where the impact of L-band assimilation decays more rapidly during the forecast phase. Figure 11 illustrates the impact of precipitation on the persistence of assimilation increments. Figure 11a uses a typical BET Tropical grid point (6.38° N, 59.06° W) as an example to present the soil moisture time series. During the assimilation phase, continuous observational constraints result in a distinct divergence between the soil moisture sequences of the Data Assimilation (DA) experiment and the Control (CTL) experiment. However, in the forecast phase, external forcings such as heavy precipitation dominate the soil moisture state. This causes the DA and CTL simulation results to rapidly converge, which makes the information introduced via assimilation more susceptible to being masked. Figure 11b further demonstrates that densely vegetated regions, including BET Tropical and NET Temperate, typically experience higher average precipitation. Because L-band assimilation performs relatively well in these regions, frequent and intense precipitation may partly explain the more rapid decay of its impact during the forecast phase. In contrast, based on the statistical results, the assimilation of C band and X band products led to improvements in assimilation performance in regions with low to moderate vegetation cover. In these areas, less frequent precipitation and slower soil moisture dynamics allow the assimilated signal to persist longer, resulting in better forecast retention.