Reanalysis Data

Reanalysis data is one of the most important — and most misunderstood — data types in the geospatial ecosystem. Unlike satellite imagery, which directly measures electromagnetic radiation reflected or emitted from the Earth's surface at the moment of acquisition, reanalysis data is the output of a numerical weather prediction model that has been "replayed" over historical periods using all available observations as constraints. The result is a physically consistent, spatially and temporally complete dataset of atmospheric and surface variables — temperature, precipitation, wind, humidity, pressure, soil moisture, radiation — covering the entire globe at regular time intervals, typically every hour, going back decades.

The key distinction from satellite Earth observation data is this: EO data is observed but sparse (affected by clouds, revisit gaps, and sensor limitations), while reanalysis data is modeled but complete (no gaps, no clouds, global coverage at every time step). EO data tells you exactly what the sensor measured at a specific place and time. Reanalysis tells you what the atmosphere and surface most likely looked like everywhere, based on the best available physics and observations combined.

For geospatial analysts accustomed to working with satellite imagery, reanalysis data requires a fundamentally different mindset. The spatial resolution is coarse — typically 0.25° to 1° (roughly 25–100 km at the equator), compared to 10–30 m for optical satellites. But the temporal resolution is dense — hourly data going back to 1940 in some products — and the physical consistency across variables is unmatched. You cannot get consistent, co-located temperature, wind, precipitation, and radiation from any single satellite sensor, but reanalysis provides all of them on the same grid at the same time.

Reanalysis works through a process called data assimilation, which optimally combines three ingredients: a forecast model (the prior), observations (the data), and error statistics (the weights).

The numerical weather prediction (NWP) model runs forward in time, producing a short-range forecast of the atmospheric state. This forecast is then compared against all available observations — surface weather stations, radiosondes (weather balloons), aircraft measurements, ocean buoys, and, importantly, satellite radiance measurements from microwave sounders, infrared sensors, and scatterometers. Data assimilation algorithms (typically 4D-Var or Ensemble Kalman Filter) adjust the model state to optimally fit the observations, accounting for the known error characteristics of both the model and the instruments. The adjusted state becomes the starting point for the next forecast cycle.

This process is repeated continuously, stepping forward through time, ingesting observations as they become available, and producing a best-estimate of the complete atmospheric and surface state at each time step. Because the same model version, resolution, and assimilation system are used throughout the entire period (unlike operational weather forecasts which change over time), the result is a temporally consistent dataset — trend analysis is meaningful because methodology changes don't introduce spurious shifts.

Major reanalysis products include ERA5 (ECMWF, 1940–present, 0.25°/hourly, widely considered the gold standard), MERRA-2 (NASA, 1980–present, 0.5°×0.625°), JRA-55 (JMA, 1958–present), and NCEP/NCAR Reanalysis (1948–present, 2.5°). ERA5 alone produces over 200 variables across 137 atmospheric levels.

Reanalysis is not observation — it is a model's best estimate. In data-sparse regions (oceans, polar areas, developing countries before the satellite era), the "analysis" is heavily influenced by the model's own physics rather than observations, making it less reliable. Spatial resolution (25–100 km) is far too coarse for site-specific applications like urban heat island analysis or field-level agriculture. Precipitation in reanalysis is particularly uncertain because it depends strongly on model convection schemes. Older periods (pre-satellite, before ~1979) have fewer observations to constrain the model, resulting in higher uncertainty. Different reanalysis products can disagree significantly on the same variable — ERA5 and MERRA-2 may show different temperature trends in data-sparse regions. The datasets are enormous (ERA5 is petabytes) and require specialized tools and formats (NetCDF, GRIB, xarray) that differ from the GeoTIFF/COG ecosystem familiar to EO users.

The concept of reanalysis was pioneered by the NCEP/NCAR Reanalysis project (Kalnay et al., 1996), which first applied a frozen data assimilation system to the historical observational record, producing a consistent dataset from 1948 onward. ECMWF followed with ERA-15 (1979–1993), ERA-40 (1957–2002), ERA-Interim (1979–2019), and the current ERA5 (1940–present), each dramatically improving resolution and quality. NASA's MERRA (2009) and MERRA-2 (2017) provided an independent American reanalysis with emphasis on aerosol assimilation. Japan's JRA-55 extended back to 1958. ERA5, released in stages from 2017, represents the current state of the art with its 0.25° resolution, hourly output, and back-extension to 1940. The Copernicus Climate Change Service (C3S) distributes ERA5 freely, making it the most accessible reanalysis product. Today, reanalysis is a multi-billion-euro scientific enterprise and a cornerstone of climate services, energy forecasting, and environmental monitoring.