Vegetation-aware satellite assimilation improves soil moisture modeling

Satellite soil-moisture data can improve land-surface simulations, but adding every available microwave observation does not always make the system better. A preliminary technical study found that the most useful observations depend strongly on vegetation cover: longer-wavelength L-band data performed well through many canopies, C-band data were especially effective in low-to-moderate vegetation, and high-frequency X-band data supplied useful extra information mainly over sparse vegetation. A vegetation-aware strategy that combined these products selectively raised the domain-mean correlation with ground observations in central and western China from about 0.25 to 0.5 and reduced unbiased root-mean-square error from 0.068 to 0.058 cubic metres per cubic metre.

The experiments, published in Geoscientific Model Development on 21 October 2026, used the Common Land Model, or CoLM, to examine how soil-moisture retrievals from three satellites affect simulated land conditions. The study assimilated retrievals from NASA’s Soil Moisture Active Passive mission, or SMAP; the European Advanced Scatterometer, or ASCAT; and the Microwave Radiation Imager aboard China’s FY-3D satellite, or MWRI. The three instruments observe at different microwave frequencies and therefore sample the land surface in different ways. SMAP uses L-band radiation at 1.41 GHz, ASCAT is a C-band scatterometer operating at 5.255 GHz, and MWRI retrieves soil moisture primarily from X-band brightness temperatures at 10.65 GHz.

All three products provide broad coverage, but their sensitivity to vegetation differs. L-band wavelengths penetrate vegetation more effectively than shorter wavelengths, making SMAP comparatively useful in densely vegetated areas, although retrieval quality still deteriorates when canopy water content is high. C-band observations are more strongly affected by vegetation, but ASCAT benefits from its change-detection algorithm and multi-angle viewing geometry. X-band radiation is especially sensitive to vegetation scattering and attenuation. MWRI can therefore be unreliable beneath moderate or dense canopies, yet it can provide complementary information over bare soil and sparsely vegetated land. The central question was whether these differences could be used to build a genuinely synergistic assimilation system rather than simply adding more observations.

To do so, the researchers used a Simplified Extended Kalman Filter, or SEKF, to update CoLM’s soil-moisture state with daily satellite retrievals. The method approximates the observation operator through finite-difference perturbations rather than requiring an adjoint model, making it computationally practical for nonlinear land-surface processes. Model background and observation errors were estimated using ERA5-Land where stations were absent and triple collocation in station-covered regions. The researchers also applied cumulative-distribution-function bias correction, quality controls based on each product’s native flags, and a three-standard-deviation innovation check. Satellite values within each approximately 1.4-degree model grid cell were averaged, and observations were mapped to CoLM’s second soil layer, from 1.75 to 4.51 cm, because that layer best represented the effective sensing depth of the microwave products.

The main summer experiment covered June through August 2022. CoLM was first spun up for 342 years in open-loop mode, after which a control simulation and several assimilation configurations were run from identical initial conditions. Single-product experiments ingested SMAP, ASCAT, or MWRI alone; pairwise experiments combined L-band with C-band, L-band with X-band, or C-band with X-band; and a further experiment assimilated all three concurrently. A final “NEW” configuration used a targeted rule: SMAP and ASCAT were assimilated in densely vegetated regions, while all three products were retained in sparsely vegetated areas, including bare soil and several grass and crop classes. The strategy was therefore designed to screen observations according to their expected retrieval quality rather than treating every observation as equally useful.

Evaluations against ERA5-Land showed that each single-band product improved soil-moisture estimates over the upper metre, although the magnitude depended on frequency and depth. C-band assimilation produced the largest gains, increasing correlation by 0.15, 0.12, and 0.09 in the 0–7, 7–28, and 30–100 cm layers while reducing RMSE by 0.020, 0.015, and 0.012 cubic metres per cubic metre. L-band generally ranked second, whereas X-band performed best in sparsely vegetated regions but could have a negative effect beneath dense vegetation. The benefits also decayed at different rates during the month-long free forecast: L-band improvements faded more quickly, apparently because frequent precipitation in the humid, densely vegetated regions where it performed best rapidly overwrote the assimilated signal. C- and X-band effects were more persistent in areas with less rainfall and slower soil-moisture change.

Ground-based results from 3,057 retained stations—2,792 from the China Meteorological Administration and 265 from the International Soil Moisture Network—supported the same vegetation-dependent pattern. L- and C-band assimilation generally improved both correlation and RMSE across the available vegetation classes, while X-band MWRI retrievals performed poorly under dense canopies. Dual-sensor assimilation usually outperformed single-sensor assimilation, and the combination of SMAP and ASCAT was the strongest of the tested pairs. At a temperate needleleaf-forest station, for example, the three-product experiment had a time-series correlation of 0.53, below the 0.63 obtained by SMAP and ASCAT alone, because unusually wet MWRI retrievals pulled the analysis away from the station observation. In a corn field, however, adding MWRI raised correlation with station data from 0.79 to 0.83, helping the model reproduce a rainfall-driven wetting and drying sequence. At non-Arctic grassland, the corresponding correlation rose from 0.23 to 0.43 when X-band data were included alongside the other two products.

The targeted NEW experiment produced a mean correlation improvement of 0.17 over the control when evaluated with station observations. Its distribution of improvements was more consistently positive than that of indiscriminate three-product assimilation, with less probability of degradation. The benefit was especially clear in central and western China, where low-stature vegetation and bare ground dominate. Including X-band data in these sparsely vegetated regions increased correlation by 0.014, because the shorter wavelengths experienced relatively little canopy attenuation and could supplement L- and C-band observations. Under dense vegetation, where the new scheme excluded MWRI, the targeted approach improved correlation by 0.1 relative to the all-product experiment and reduced unbiased RMSE by approximately 0.001 cubic metres per cubic metre across most of the region. In southeastern China, where dense canopies dominate, correlation rose from 0.52 in the control to 0.58 with the targeted strategy, while unbiased RMSE declined from 0.055 to 0.052 cubic metres per cubic metre.

Seasonal tests exposed a limitation of using broad vegetation classes as a permanent selection rule. X-band observations were more likely to help in summer and winter, when grass and crop cover was relatively stable, but could reduce performance during spring and autumn, when rapid vegetation growth or senescence changed canopy structure faster than the seasonal classification could capture. Even with the targeted selection rule, the NEW experiment achieved higher mean seasonal correlations than both SMAP-plus-ASCAT assimilation and unfiltered three-product assimilation in every season examined. The authors caution that terrain, rainfall, and temperature can also alter retrieval errors and should be represented more fully in future systems. The study used offline land-surface simulations without coupled land–atmosphere interactions, and its deeper soil layers adjusted more slowly because no independent assimilation spin-up was available. The results are therefore preliminary evidence for vegetation-aware data selection, not a demonstrated improvement in operational weather or climate forecasts.

Subject of Research: Vegetation-aware assimilation of multi-band satellite soil-moisture retrievals for improved land-surface modeling

Article Title: A preliminary study on a synergistic assimilation scheme for multi-band satellite soil moisture data

Article References: Bai, X., Lin, Z., Qin, Z., & Li, J. (2026). A preliminary study on a synergistic assimilation scheme for multi-band satellite soil moisture data. Geoscientific Model Development, 19(20), 10051-10076. https://doi.org/10.5194/gmd-19-10051-2026

Image Credits: AI Generated

DOI: 10.5194/gmd-19-10051-2026

Keywords: soil moisture, data assimilation, satellite remote sensing, vegetation effects, land-surface modeling, SMAP, ASCAT, MWRI, Kalman filtering, China, preliminary, synergistic