Soil Moisture at Startup Shifts Landfalling Typhoon Track Forecasts

When Typhoon Hinnamnor curved toward the Korean Peninsula in September 2022, forecasters and models alike faced the notoriously difficult problem of predicting a storm that is steered by the mid-latitude atmosphere. A new modeling study now suggests that part of that difficulty may begin underground. Researchers in South Korea tested whether the initial soil moisture prescribed in a weather model changes the forecast track of poleward-moving, landfalling tropical cyclones, and found that swapping one widely used soil moisture product for another shifted simulated tracks by tens of kilometers and altered the depth and position of the mid-latitude trough that steered the storms.

The team, led by Jinyoung Park and colleagues including Eunkyo Seo, ran sensitivity experiments with the Weather Research and Forecasting (WRF) model at 12-kilometer horizontal resolution over a single fixed domain covering East Asia. They selected three recent tropical cyclones that directly affected South Korea: Maysak and Haishen in 2020 and Hinnamnor in 2022. Each storm was simulated twice with identical atmospheric initial and boundary conditions, physics schemes, and model configuration. The only deliberate difference was the soil moisture field used at initialization, drawn either from the Global Land Data Assimilation System (GLDAS) Noah v2.1 product or from the European Centre for Medium-Range Weather Forecasts Reanalysis v5 (ERA5). The two experiments are called the GL run and the ERA5 run.

The two soil moisture datasets are far from interchangeable. GLDAS Noah v2.1 is generated with the Noah land surface model version 3.6 in the Land Information System, driven by a combination of NOAA GDAS atmospheric analysis fields, Global Precipitation Climatology Project precipitation, and AGRMET radiation, with soil layers defined at 0–10, 10–40, 40–100, and 100–200 centimeters. ERA5, by contrast, is built on 4D-Var data assimilation with the Integrated Forecasting System and assimilates satellite soil moisture observations from ASCAT, TMI, AMSR-2, GMI, and SMOS alongside atmospheric variables, using the HTESSEL land scheme with soil layers at 0–7, 7–28, 28–100, and 100–289 centimeters. Because the ERA5 layer depths differed from those of the Noah land surface model used in the simulations, the researchers modified the Noah soil-layer definitions to match, and both datasets were vertically interpolated to a common depth of 0–100 centimeters using a depth-weighted method.

After initialization, the soil moisture was left to evolve freely in the land surface model; it was never nudged, re-assimilated, or replaced during the forecasts. The simulations spanned 6 days for Maysak, 5.5 days for Haishen, and 7.5 days for Hinnamnor, and the resulting tracks and intensities were verified against the Regional Specialized Meteorological Center Tokyo best-track data. Spectral nudging was applied south of 24 degrees north to constrain the early low-latitude track evolution, but crucially it was not applied over the continental region where the key soil moisture differences developed.

The headline result concerns track accuracy near landfall. In all three cases, the GL runs produced smaller mean track errors during the 12 hours preceding landfall. For Maysak, the landfall-period mean track error was 146.7 kilometers in the GL run versus 157.3 kilometers in the ERA5 run; for Haishen, 117.6 versus 133.9 kilometers; and for Hinnamnor, 113.3 versus 156.1 kilometers, the clearest separation of the three. The ERA5-run tracks were generally displaced westward relative to the GL tracks during this period. No pronounced systematic difference appeared in simulated maximum wind speed before landfall, so the analysis focused on the track response and the environmental steering differences behind it.

The soil moisture differences themselves were spatially heterogeneous across the broader land domain, with both positive and negative values, and the local extremes were not representative of the overall area. But one region stood out: a trough-development region spanning Northeast China and adjacent Mongolia, between 45 and 55 degrees north and 100 to 130 degrees east, where the ERA5 run maintained persistent positive soil moisture differences relative to the GL run. The regionally averaged 0–100 centimeter soil moisture difference there stayed positive throughout the forecast, declining only from +0.081 cubic meters per cubic meter at initialization to +0.067 at 168 hours. Similar contrasts persisted in the near-surface layer that directly drives turbulent heat and moisture exchange, and because no external soil moisture update was applied, the later differences are interpreted as persistent signatures of the initial condition.

Those surface differences propagated upward into the circulation. In the Hinnamnor case, the ERA5 run showed enhanced latent heat flux and higher moist static energy over the trough-development region, followed by lower sea-level pressure, stronger ascent at 700 hPa, and reduced 500 hPa geopotential height. The simulated mid-latitude trough in the ERA5 run became deeper and slightly more westward-displaced than in the GL run, with differences already noticeable at 48 hours and more pronounced at 144 and 168 hours. Compared against ERA5 reanalysis over the 6- and 7-day forecasts, the GL run more closely reproduced the trough position and progression: the 500 hPa geopotential height root-mean-square error was 17.7 gpm in the GL run versus 20.3 gpm in the ERA5 run, roughly 13 percent lower. A blocking index based on the meridional geopotential height gradient also indicated stronger blocking-like conditions and a trough maximum located about 3 degrees farther west in the ERA5 run.

The modified trough in turn reshaped the steering flow around the storm. Vertical wind shear in a 5-to-7-degree annulus around the simulated Hinnamnor center began increasing more strongly in the ERA5 run at 156 hours and exceeded both the GL run and the reanalysis at 168 hours, with a more pronounced northwestward shear orientation. To connect this to motion, the researchers removed the storm’s own rotational and divergent circulation and computed environmental steering winds from 850 to 200 hPa, testing radii from 1 to 8 degrees to find the layer-mean flow that best matched the simulated track. The steering-flow differences between the two runs were small at 144 hours, developed a northwestward component at 156 hours, and became coherently aligned northwestward by 168 hours, with the lower layers contributing mainly to the westward component and the upper layers to the northward component. Projecting the steering differences onto the track-deflection direction at 156 hours yielded a ratio of 1.21 for the upper layer versus 0.26 for the lower layer, indicating that the upper-level flow anomaly carried most of the dynamical link to the track displacement. The distance between the simulated storm centers stayed below about 16 kilometers through 72 hours, then grew to 23.2 kilometers at 96 hours, 41.4 at 120, 66.7 at 144, and 190.9 kilometers at 168 hours.

The authors are careful about what this chain of evidence does and does not show. The steering-flow diagnosis does not imply that inland soil moisture directly forces the tropical cyclone vortex; the proposed pathway is indirect, in which persistent soil moisture differences alter surface fluxes and lower-tropospheric moist thermodynamic conditions, which accompany modification of the mid-latitude trough, which in turn changes the steering flow. The temporal correspondence is consistent with this pathway but does not by itself establish it. Nor do the experiments establish that one soil moisture product is intrinsically more accurate than the other. Because suitable in situ observations were unavailable across the full trough-development region, the team could not attribute the track differences to a verified wet bias in ERA5, and previous regional evaluations have reported product-dependent soil moisture biases over East Asia whose sign and magnitude depend on location, season, and land cover.

The limitations are explicit. Only three deterministic cases were examined, the detailed mechanism analysis centered on Hinnamnor, and Maysak and Haishen showed weaker but qualitatively similar landfall-period responses, so the findings stand as process-oriented sensitivity evidence rather than statistically generalized conclusions. The artificial adjustment of soil layers for the ERA5 experiment may also introduce uncertainties, and results could depend on the chosen physics parameterizations, which were identical in both runs. Still, the practical stakes are considerable: the authors note that track differences of a few tens of kilometers matter in East Asia, where major cities with populations exceeding one million cluster within that range. Hinnamnor itself was initially predicted to make landfall at Ulsan but actually struck about 50 kilometers away in Pohang, causing substantial damage. The study, published open access in Geoscience Letters, argues that more accurate, model-consistent initialization of land surface states, validated with in situ observations and land data assimilation, could sharpen medium-range forecasts of the mid-latitude circulations that steer poleward-moving typhoons toward heavily populated coastlines.

Subject of Research: Sensitivity of landfalling tropical cyclone track forecasts to initial soil moisture conditions in a regional weather model

Article Title: Does initial soil moisture condition affect poleward-moving tropical cyclone forecasts?

Article References: Park, J., Kim, K., Yoon, D., Park, H., Cho, W., Cha, D.-H., Moon, J., & Seo, E. (2026). Does initial soil moisture condition affect poleward-moving tropical cyclone forecasts?. Geoscience Letters, 13(1), Article 55. https://doi.org/10.1186/s40562-026-00502-0

Image Credits: AI Generated

DOI: 10.1186/s40562-026-00502-0

Keywords: soil moisture, tropical cyclones, typhoon track forecast, landfall, WRF model, GLDAS, ERA5, mid-latitude trough, steering flow, land-atmosphere interaction, East Asia, Hinnamnor