Partial recovery of meter-scale surface weather

Abstract

Near-surface atmospheric conditions can differ sharply over tens to hundreds of meters due to land cover and topography, yet this variability is absent from current weather analyses and forecasts. It is unclear whether such meter-scale variability reflects irreducibly chaotic dynamics or contains a component predictable from surface characteristics and large-scale atmospheric forcing. Here we show that a substantial, physically coherent component of meter-scale near-surface weather is statistically recoverable from existing observations. By conditioning coarse atmospheric state on sparse surface station measurements and high-resolution Earth observation data, we infer spatially continuous fields of near-surface wind, temperature, and humidity at 10 m resolution across the contiguous United States. Relative to ERA5, the inferred fields reduce wind error by 29% and temperature and dewpoint error by 6%, while explaining substantially more spatial variance at fixed time steps. They also exhibit physically interpretable structure, including urban heat islands, evapotranspiration-driven humidity contrasts, and wind speed differences across land cover types. Our findings expand the frontier of weather modeling by demonstrating a computationally feasible approach to continental-scale meter-resolution inference. More broadly, they illustrate how conditioning coarse dynamical models on static fine-scale features can reveal previously unresolved components of the Earth system.

Figures (16)

• Figure 1: Inference framework combining reanalysis, surface stations, and Earth observation to recover sub-kilometer weather variability.(a) (i) Example over the Los Angeles metropolitan area illustrating that kilometer-scale reanalysis (ERA5) does not resolve meter-scale weather variability. (ii) 2 m air temperature from ERA5 (grid) and as measured by surface weather stations (points). (iii) Histogram of the spatial standard deviation of 2 m air temperature (from 2020--2023) across the weather stations versus across the nine ERA5 grid cells. (b) Overview of the inference framework. The model combines coarse atmospheric dynamics from ERA5, observations from sparse surface stations, and high-resolution Earth observation data to generate spatially continuous 10 m resolution fields of 2 m air temperature, 2 m dewpoint temperature, and 10 m wind.
• Figure 2: Quantitative evaluation of inferred micro-weather fields relative to baseline methods.(a) Mean absolute error (MAE) or vector error and spatial coefficient of determination ($R_{\text{spatial}}^2$, de-meaned per timestep to isolate spatial variability) across all held-out test stations for temperature, dewpoint, and wind. Baselines include spatial interpolation of weather stations, ERA5 reanalysis, and a model using ERA5 and stations without Earth observation inputs. (b) Histograms of spatial $R^2$ values across hourly timesteps from 2020--2023, where each $R_{\text{spatial}}^2$ is computed for one timestep across all held-out test stations. (c) Map of station-level error reduction relative to ERA5, defined as the difference between ERA5 error and full inference model error, averaged over 2020--2023. Positive values indicate improvement over ERA5.
• Figure 3: Physically coherent meter-scale structure in inferred near-surface weather fields. Case studies comparing inferred 10m resolution fields with ERA5 reanalysis. Satellite imageryis shown for spatial context. For temperature panels, land surface temperature from ECOSTRESS is included for qualitative comparison. (a) Urban surfaces are hotter than vegetated areas (Washington, D.C.). (b) Temperature decreases with elevation (Mount Lemmon, Arizona). (c) Dewpoint likewise decreases with elevation (Sawtooth Range, Idaho). (d) Crop transpiration increases humidity (Calexico, California). (e) Wind speed is higher over open areas and roads and lower over built-up and forested areas (Long Island, New York). (f) Higher winds occur over ridges; a southwesterly wind corresponds to windward-leeward contrasts in wind speed (Yokuts Valley, California).
• Figure 4: Example of Earth observation inputs for a given location. Panels (a)--(m) include satellite layers used for the end-to-end learned satellite embedding. (a--d) DSM, DEM, and derived indices. (e--h) Sentinel-2 for spring, summer, fall, and winter seasons visualized in RGB. (i--l) Sentinel-2 for the four seasons visualized as a False Color Composite (FCC; NIR, R, G). (m) Land cover map with seven classes selected (of which only four are present at this location). (n) Geospatial foundation model embeddings produced by AlphaEarth brown2025. (o) Aerial image of the same location, not used in the model, for context only.
• Figure 5: Spatial partitioning of contiguous U.S. weather stations for training and evaluation. Surface weather stations from NOAA MADIS across the continguous U.S. are split into four disjoint sets: backbone, training, validation, and test. Only test set stations are used for evaluation.