ERA5-Land Climate Reanalysis

• ERA5-Land is a land-focused climate reanalysis providing continuous, hourly data at ~9 km resolution since 1950.
• It supplies a broad set of variables—from snow depth to wind fields—that support applications in downscaling, PV simulations, and streamflow forecasting.
• ERA5-Land serves as a coarse-resolution atmospheric reference, enabling transfer learning and retrospective climatic analyses across diverse land-surface applications.

Searching arXiv for the cited ERA5-Land-related papers and closely related sources. arXiv Search Query: (Liu et al., 2024) OR (Gadal et al., 2022) OR (Camargo et al., 2020) OR (Ji et al., 2024) OR (Taccari et al., 18 Feb 2026) ERA5-Land ERA5-Land is a land-focused climate reanalysis used as a temporally continuous forcing, a regional atmospheric reference, and a historical pretraining domain across a wide range of land-surface applications. In the studies considered here, it is described as an ECMWF land reanalysis, a replay of ERA5 at finer land resolution, an “offline land-surface replay” with specific elevation corrections for thermodynamic states, and an enhanced land-focused version of ERA5 (Gadal et al., 2022, Taccari et al., 18 Feb 2026, Camargo et al., 2020). Its operational value in these works derives from a recurring combination of properties: hourly temporal resolution, spatial resolution of about $0.1^\circ$ or approximately 9 km over land, multi-decadal coverage extending from 1950 to the present in some workflows, and a variable set broad enough to support snow, wind, energy, hydrology, and compound-extreme analyses (Liu et al., 2024, Camargo et al., 2020).

1. Dataset character and core variables

ERA5-Land is treated in the snow-depth downscaling study as the main reanalysis source, specifically as ERA5-Land hourly data (version 5), covering 1950 to the present, with approximately 9 km spatial resolution, and containing over 50 variables (Liu et al., 2024). In the Namib dune-wind study, it provides hourly 10-m wind speed and direction on a $0.1^\circ \times 0.1^\circ$ grid, stated to be about $9$ km in Namibia (Gadal et al., 2022). In the photovoltaic study, its resolution is given as about 0.1° / 9 km, and it is contrasted with coarser ERA5 and with MERRA-2, which is described as about 5× coarser (Camargo et al., 2020).

The variables emphasized depend on the application. For snow-depth downscaling, the key fields are snow depth sde and monthly 10 m wind fields u10 and v10 (Liu et al., 2024). For photovoltaic power simulation, the required meteorological inputs are SSRD, t2m, u10, and v10 (Camargo et al., 2020). For global daily streamflow forecasting, the dynamic forcings are SSR, STR, SP, T2M, and TP, explicitly chosen because they are available consistently in both ERA5-Land and IFS via the CARAVAN MultiMet extension (Taccari et al., 18 Feb 2026). In DeepExtremeCubes, ERA5-Land contributes meteorological context and event-detection variables including Tmax, PE30, PE90, and PE180 (Ji et al., 2024).

Several studies also specify variable semantics or temporal handling. The snow paper states that ERA5-Land snow depth is an instantaneous average of the snow thickness on the ground, excluding snow on vegetation canopy, and that ERA5-Land snow depth is resampled at daily resolution and linearly interpolated for each ICESat-2 point $(E,N,t)$ during training (Liu et al., 2024). The photovoltaic workflow notes that radiation is provided as accumulated energy and must be transformed to hourly values by differencing consecutive accumulation values within a forecast cycle (Camargo et al., 2020). In DeepExtremeCubes, ERA5-Land is stored as a single-pixel temporal array at the minicube center and summarized as minimum, maximum, and mean values on a 5-day temporal resolution from 2016.01.01 to 2022.10.10 (Ji et al., 2024).

2. Roles in modeling workflows

Across the cited literature, ERA5-Land is not used in a single canonical way. It functions instead as a coarse-resolution backbone whose temporal continuity compensates for sparse observations or shorter-lived sensors.

In snow-depth reconstruction, ERA5-Land is the coarse-resolution climate/reanalysis backbone of the full downscaling-calibration workflow. The authors explicitly state that “ERA5 Land data primarily provides the temporal variability, and ICESat-2 the spatial variability in snow depths” (Liu et al., 2024). ICESat-2 snow depth is first retrieved from ATL08 by elevation differencing, using snow-on and snow-free classification from ATL08 attributes and flags, after which ERA5-Land snow depth is interpolated to the exact ICESat-2 point in space and time. The regression target is not snow depth alone but subgrid variability, defined operationally as ICESat-2-derived snow depth minus ERA5-Land snow depth at the same point and time. A second XGBoost model then predicts that subgrid signal from sde_era, coordinates, topography, a cumulative wind-aspect factor, and month of year; the predicted variability is then added or applied to ERA5-Land to produce monthly 10 m snow-depth maps (Liu et al., 2024).

In wind-regime analysis over the Namib Sand Sea, ERA5-Land serves as the regional atmospheric reference against which local station observations are tested (Gadal et al., 2022). The field records are averaged into 1-hour bins to match ERA5-Land’s time step, and both station winds and reanalysis winds are converted to shear velocity $u_*$ using a logarithmic wind law,

$$\frac{u(z)}{u_*} = \frac{1}{\kappa}\ln\!\left(\frac{z}{z_0}\right),$$

with $\kappa = 0.4$, while aerodynamic roughness is calibrated by minimizing a vector-mismatch metric $\delta$ (Gadal et al., 2022). Here ERA5-Land is neither training target nor truth proxy; it is the large-scale predictor that is evaluated against terrain-conditioned surface measurements.

In photovoltaic modeling, ERA5-Land is the meteorological input stream for a deterministic simulation pipeline rather than a predictive target (Camargo et al., 2020). Hourly reanalysis data for 2014–2018 are passed through a workflow that converts accumulated SSRD to hourly irradiance, computes wind speed from u10 and v10, adjusts wind from 10 m to 1 m using a logarithmic profile with roughness length 0.25, and then drives PV_LIB for Python. The pipeline further applies the Erbs model to infer DNI and DHI from GHI, assumes a fixed-tilt system with panels oriented toward North and tilt equal to site latitude, and uses a specified representative module and inverter (Camargo et al., 2020).

In AIFL, ERA5-Land is elevated to a transfer-learning domain (Taccari et al., 18 Feb 2026). The model is first pre-trained on 40 years of ERA5-Land reanalysis (1980–2019) across 18,588 curated basins using 180-day sequences and daily specific discharge targets, then fine-tuned on IFS control forecasts for 2016–2019. The paper is explicit that ERA5-Land is used to learn “robust hydrological processes” and “universal hydrological response functions,” after which fine-tuning adapts those learned representations to operational forecast biases and error structure (Taccari et al., 18 Feb 2026).

DeepExtremeCubes uses ERA5-Land in yet another way: as a compact meteorological context layer and as the source of the event logic inherited from Dheed (Ji et al., 2024). The dataset stores ERA5-Land as a central-pixel time series rather than a full raster at Sentinel-2 resolution, and uses ERA5-Land-derived Tmax, PE30, PE90, and PE180 to represent compound heatwave-drought signals. This suggests a design principle in which ERA5-Land supplies temporally structured forcing and event descriptors, while higher-resolution Earth observation data carry the spatial response signal.

Study	ERA5-Land role	Output or task
(Liu et al., 2024)	Coarse-resolution prior	Monthly 10 m snow-depth maps
(Gadal et al., 2022)	Regional atmospheric reference	Local wind-regime comparison
(Camargo et al., 2020)	Meteorological forcing	Hourly PV generation series
(Taccari et al., 18 Feb 2026)	Pretraining domain	Global daily streamflow forecasts
(Ji et al., 2024)	Meteorological backbone and event source	Minicubes for CHD impact analysis

3. Validation, skill, and scale dependence

The strongest direct validation numbers in the provided material come from the snow-downscaling and photovoltaic studies. In the Hardangervidda snow application, validation against airborne laser scanning was performed at both a mesoscale and a microscale (Liu et al., 2024). At 100 m × 500 m, the calibrated downscaling model achieved $R^2$ between 0.74 and 0.88 and RMSE between 0.64 m and 0.43 m for the six ALS flight lines in April 2008; for 2009, RMSE ranged from 0.37 m to 0.58 m. At 10 m, the reported metrics are a Spearman rank correlation of 0.77, $R^2 = 0.34$, RMSE = 1.33 m, Kolmogorov-Smirnov $0.1^\circ \times 0.1^\circ$0, and mean deviation in the difference map = 0.28 m. The semivariogram analysis further shows that the downscaled model captures spatial structure but underestimates variance relative to ALS by about 10%–13% (Liu et al., 2024).

These numbers are not direct measures of raw ERA5-Land snow-depth accuracy. They measure the quality of a derived downscaled product in which ERA5-Land acts as the coarse spatio-temporal scaffold. The paper’s phrasing makes this distinction explicit: the model learns subgrid deviation from the coarse reanalysis field rather than snow depth from scratch (Liu et al., 2024). A plausible implication is that ERA5-Land’s value in this setting lies less in pointwise microscale fidelity than in providing a stable temporal baseline that can be corrected by terrain-aware statistical learning.

In the dune-wind study, the performance assessment is more conditional than metric-centered (Gadal et al., 2022). Over flat terrain, at Etosha West and Huab, agreement between ERA5-Land and local measurements is described as quantitative. In the interdune areas, ERA5-Land also performs well during the day and during the main daytime sea-breeze season, but it misses the nocturnal dune-induced flow reorientation. The discrepancy is large: the observed wind deviation exceeds $0.1^\circ \times 0.1^\circ$1, especially when regional shear velocity is low, typically when $0.1^\circ \times 0.1^\circ$2 (Gadal et al., 2022). Once the regional shear velocity rises to about $0.1^\circ \times 0.1^\circ$3 or above, the reorientation largely disappears. The key interpretive point is that ERA5-Land reproduces the broad seasonal wind regime but not a locally generated nocturnal steering mode tied to dune geometry and boundary-layer confinement.

For photovoltaic modeling in Chile, the validation involves measured hourly output from 57 large PV plants over 2014–2018 (Camargo et al., 2020). For the full plant set, hourly correlations are mostly around 0.8, RMSE is often around 0.2, and MBE is generally around 0. On the stricter S2 subset, where plant configuration better matches the model assumptions, ERA5-Land-driven simulations achieve hourly correlation 0.88–0.97, MBE about -0.02 to 0.05, and RMSE around 0.12. When the S2 plants are aggregated spatially, hourly Pearson rises to 0.970, RMSE drops to 0.082, and MBE is 0.018 (Camargo et al., 2020). The same study concludes, however, that ERA5-Land’s advantage over renewables.ninja/MERRA-2 is only slight.

The AIFL streamflow paper does not directly benchmark ERA5-Land alone against another reanalysis in terms of forcing accuracy; instead, it measures the effect of ERA5-Land pretraining followed by IFS fine-tuning (Taccari et al., 18 Feb 2026). On the independent 2021–2024 test set, the final system reaches median $0.1^\circ \times 0.1^\circ$4 and median NSE = 0.53. Comparing the pre-trained-only and fine-tuned models over 2,003 test basins, mean $0.1^\circ \times 0.1^\circ$5 improves from 0.21 to 0.44 and mean NSE from $0.1^\circ \times 0.1^\circ$6 to $0.1^\circ \times 0.1^\circ$7, while the authors note that fine-tuning primarily contracts the lower tail of the performance distribution (Taccari et al., 18 Feb 2026). This indicates that ERA5-Land can be an effective representation-learning substrate even when the deployment forcing is different.

4. Temporal depth, reconstruction, and retrospective analysis

A recurrent reason for choosing ERA5-Land is its long temporal coverage. In the snow-depth study, the method can generate monthly snow-depth maps for the entire ERA5 time range (since 1950) because the downscaling relationship learned from the ICESat-2 overlap period can be propagated backward through the full ERA5-Land record (Liu et al., 2024). The authors state directly that the method can provide snow-depth map time series for the entire ERA5 time range (since 1950). This retrospective capability depends on the assumption that the statistical relationship between coarse ERA5-Land snow depth and local subgrid variability is stationary.

In AIFL, the temporal depth is used differently but with similar logic (Taccari et al., 18 Feb 2026). The model is pre-trained on 1 January 1980 to 31 December 2019, thereby exposing the LSTM to diverse hydroclimatic conditions and multi-decadal variability before fine-tuning on the shorter operational-forecast interval. For flood-threshold estimation, the paper further mentions 1980–2024 ERA5-Land-driven discharge simulation, although pretraining itself is limited to 1980–2019 (Taccari et al., 18 Feb 2026). Here the long historical record is not a target climatology to be reconstructed, but a data regime large enough to stabilize sequence learning across basins.

The photovoltaic paper also underscores ERA5-Land’s suitability for long-term simulation by noting that it was attractive partly because it covers a long period, ultimately intended to span 1950 to near-real-time (Camargo et al., 2020). The specific validation window is shorter, 2014–2018, but the intended use case is long hourly PV generation time series. In DeepExtremeCubes, the ERA5-Land segment spans 2016.01.01 to 2022.10.10, while the event data inherited from Dheed cover 2016-01-01 to 2021-12-31 (Ji et al., 2024). The temporal mismatch is explicitly acknowledged by the authors.

Taken together, these studies present ERA5-Land as a time-extensive forcing archive that can support both historical reconstruction and transfer learning. This suggests that one of ERA5-Land’s defining methodological roles is to supply continuity in time even when the complementary high-information data source is sparse, short, or domain-specific.

5. Limitations, assumptions, and common misconceptions

The most explicit limitations concern stationarity, scale mismatch, and local processes unresolved at ERA5-Land resolution. In the snow-downscaling study, the model assumes stationarity of the ERA5-Land relationship, meaning that relationships learned from 2018–2022 can be applied back to earlier decades (Liu et al., 2024). The same paper cautions that ERA5-Land snow depth can contain interpolation/downscaling artifacts inherited from ERA5, especially near large glaciers and large water bodies, and that monthly ERA5-Land wind fields may be too coarse to fully represent complex wind-topography interactions (Liu et al., 2024). It also states that the model underestimates extremes and therefore requires area-specific calibration, with a small number of localized control points from a one-time field survey needed to restore full snow-depth patterns.

The Namib dune study identifies an additional scale-dependent limitation: ERA5-Land is reliable for regional wind climatology over flat desert surfaces and can represent interdune winds when the boundary layer is deep, but it misses dune-aligned nocturnal deviation associated with shallow-boundary-layer confinement (Gadal et al., 2022). The paper’s theory section frames the mechanism in terms of $0.1^\circ \times 0.1^\circ$8, a Froude number $0.1^\circ \times 0.1^\circ$9, and an internal Froude number $9$0, with strong deflection expected under stronger confinement. The observed nocturnal steering exceeds the qualitative range reproduced by the linear theory, which the authors attribute to a likely nonlinear regime (Gadal et al., 2022). A plausible implication is that ERA5-Land may reproduce regional forcing regimes while failing on terrain-induced mesoscale or microscale atmospheric responses with strong diurnal modulation.

The AIFL paper defines a different limitation: domain shift between ERA5-Land and operational IFS (Taccari et al., 18 Feb 2026). The authors quantify this mismatch using a normalized Wasserstein distance on wet-day precipitation over 2016–2019, reporting a median shift of 0.045, an upper-decile threshold of $9$1, and maxima up to 0.638. Their caution is explicit: a model trained only on ERA5-Land would likely inherit ERA5-Land’s forcing characteristics and then degrade when exposed to IFS. In this usage, ERA5-Land is not criticized for internal inconsistency; rather, its physical consistency and special structure make it statistically distinct from the forecast domain (Taccari et al., 18 Feb 2026).

Two of the supplied papers are important precisely because they should not be taken as direct ERA5-Land validation. The temperature paper evaluates ERA5 2 m air temperature, not ERA5-Land, and states that any implication for ERA5-Land is only indirect (McNicholl et al., 2022). The Indian monsoon soil-moisture paper is likewise based on ERA5 reanalysis datasets, not ERA5-Land (Gautam et al., 2023). These clarifications matter because ERA5 and ERA5-Land are related but distinct products. A common misconception is to transfer metrics from ERA5 directly to ERA5-Land without qualification; the provided literature explicitly warns against that inference (McNicholl et al., 2022, Gautam et al., 2023).

6. Scientific uses and cross-domain significance

The application range represented in these studies is broad. ERA5-Land underpins snow-depth mapping in remote, above-treeline snow regions by providing the temporally continuous prior later refined to 10 m resolution with ICESat-2, topography, and machine learning (Liu et al., 2024). In aeolian geomorphology, it supplies the regional wind regime against which local dune-field flow deviations are identified and interpreted (Gadal et al., 2022). In energy systems, it drives PV generation models through variables directly relevant to irradiance, temperature, and near-surface wind, enabling long hourly simulations for utility-scale plants (Camargo et al., 2020). In hydrology, it acts as the pretraining forcing for an end-to-end global LSTM over 18,588 curated basins (Taccari et al., 18 Feb 2026). In climate-extreme impact analysis, it becomes the meteorological backbone of multicube datasets designed for compound heatwave-drought research and biosphere forecasting (Ji et al., 2024).

These roles are methodologically connected. ERA5-Land repeatedly supplies the “background field” or “forcing stream,” while other data sources provide local detail, event labels, or observed responses. In the snow application, that complementary pairing is stated directly: “ERA5 Land data primarily provides the temporal variability, and ICESat-2 the spatial variability in snow depths” (Liu et al., 2024). In DeepExtremeCubes, ERA5-Land provides “historical weather conditions” and the event reference, while Sentinel-2 provides the observed vegetation state (Ji et al., 2024). In AIFL, ERA5-Land supplies the consistent historical domain and IFS supplies the operational domain (Taccari et al., 18 Feb 2026). This suggests a general pattern in current research use: ERA5-Land is frequently positioned as the temporally coherent land-surface scaffold onto which higher-resolution observations, local calibrations, or deployment-specific forecast corrections are attached.

The literature considered here therefore presents ERA5-Land less as a universally sufficient product than as a strong intermediate representation for land applications. Its most robust uses are those that exploit its temporal continuity, standardized forcing variables, and land-oriented resolution while explicitly accounting for unresolved local physics, calibration needs, or reanalysis-to-forecast distributional mismatch (Liu et al., 2024, Gadal et al., 2022, Taccari et al., 18 Feb 2026).