How Weather Forecasting Works: From Observations to Predictions

Observing the Atmosphere

Accurate forecasting depends on knowing the current state of the atmosphere as precisely as possible. Surface weather stations, numbering roughly 10,000 worldwide, measure temperature, humidity, pressure, wind speed and direction, precipitation, and visibility at standardized intervals. Automated stations report as frequently as every minute, while staffed stations typically report hourly. These surface observations provide the foundation for understanding conditions at ground level.

Upper-air observations are equally critical because weather systems extend through the full depth of the troposphere. Radiosondes, instrument packages carried aloft by weather balloons, measure temperature, humidity, and pressure from the surface to altitudes above 30 kilometers. Wind speed and direction are calculated by tracking the balloon's position using GPS. Approximately 900 radiosonde stations globally launch balloons at 00:00 and 12:00 UTC daily, creating a snapshot of the upper atmosphere's three-dimensional structure twice per day.

Weather radar provides continuous monitoring of precipitation patterns. Networks of Doppler radar stations scan the atmosphere every 5 to 10 minutes, detecting the location, intensity, and motion of rain, snow, and hail. Doppler capability measures the velocity of precipitation particles toward and away from the radar, enabling detection of rotation within thunderstorms and providing critical lead time for tornado warnings. Modern dual-polarization radar distinguishes between rain, snow, hail, and debris by analyzing the shape of precipitation particles.

Satellites observe the atmosphere from space, filling the vast gaps between surface and upper-air stations, particularly over oceans that cover 71 percent of Earth's surface. Geostationary satellites at 35,786 kilometers altitude provide images every 5 to 15 minutes over fixed regions, tracking cloud movement, thunderstorm development, and tropical cyclone evolution. Polar-orbiting satellites at 800 to 900 kilometers provide higher-resolution measurements of temperature profiles, moisture content, and sea surface temperatures that are critical inputs for numerical models.

Data Assimilation

Data assimilation is the process of combining millions of observations from diverse sources into a unified, physically consistent snapshot of the atmosphere called the analysis. This is one of the most mathematically sophisticated steps in the forecasting process. Observations are sparse, irregularly distributed, taken at different times, and contain measurement errors. The analysis must reconcile all of this into a coherent three-dimensional picture on the model's computational grid.

Modern data assimilation methods use variational analysis (3D-Var and 4D-Var) or ensemble Kalman filtering. These techniques compare observations with a short-range model forecast (called the background or first guess) and adjust the model state to better match the observations, weighted by the estimated accuracy of each observation and the model. The 4D-Var approach used by the European Centre for Medium-Range Weather Forecasts (ECMWF) considers observations over a 12-hour time window, making the analysis even more temporally consistent.

Satellite data now constitutes the largest volume of observations assimilated into forecast models. Infrared and microwave soundings provide temperature and moisture profiles across the globe, radiance measurements from multiple satellite channels constrain the model atmosphere, and scatterometer winds measure ocean surface winds that no other instrument can observe at comparable coverage. The dramatic improvement in forecast accuracy over the Southern Hemisphere (which has far fewer surface stations) is largely attributable to satellite data.

Numerical Weather Prediction

Numerical weather prediction (NWP) models divide the atmosphere into a three-dimensional grid and apply the fundamental equations of fluid dynamics and thermodynamics to calculate how conditions at each grid point will change over time. These equations, known as the primitive equations, express the conservation of mass, momentum, and energy along with the equation of state for air.

The model advances forward in small time steps, typically a few minutes, recalculating the state of the atmosphere at each step. At every time step, the model computes how wind moves air, heat, and moisture between grid cells (advection), how air rises and sinks (vertical motion), how the Sun heats the surface and atmosphere (radiation), how water evaporates, condenses, and precipitates (microphysics), and how the surface exchanges heat and moisture with the atmosphere (boundary layer processes).

Global models like the ECMWF Integrated Forecasting System (IFS) and the American GFS operate on grid spacings of 9 to 13 kilometers with 100 or more vertical levels. High-resolution regional models, such as the HRRR (High-Resolution Rapid Refresh) used in the United States, run on grids as fine as 3 kilometers, resolving individual thunderstorms and terrain-driven circulations that global models cannot capture. The computational cost of NWP is enormous: a single run of a global model at 9-kilometer resolution requires trillions of calculations and several hours on the world's fastest supercomputers.

The Human Forecaster

Despite the sophistication of numerical models, human forecasters remain essential to the prediction process. Models have systematic biases and known weaknesses, such as difficulty resolving terrain effects, lake-enhanced snowfall, fog formation, and the exact timing of convective initiation. Experienced forecasters learn these model tendencies and adjust model output accordingly, a skill called model output statistics when applied systematically. Forecasters at National Weather Service offices combine multiple model solutions with radar, satellite, and surface observations to produce forecasts tailored to local conditions that no single model captures perfectly. The human forecaster's role is particularly critical during high-impact events where the difference between model solutions can mean the difference between a blizzard warning and a winter weather advisory for a given area.

Ensemble Forecasting and Uncertainty

A single model run (deterministic forecast) provides one possible evolution of the atmosphere from one set of initial conditions. But because observations contain errors and models have limitations, the initial conditions are never perfectly known. Small errors in the initial state can grow over time due to the chaotic nature of the atmosphere, eventually producing significantly different outcomes. Edward Lorenz discovered this sensitivity to initial conditions in 1963, founding the field of chaos theory.

Ensemble forecasting addresses this uncertainty by running the same model many times with slightly perturbed initial conditions and, in some systems, varied model physics. The ECMWF ensemble runs 51 members; the American GEFS runs 31. The spread among ensemble members at any future time indicates forecast confidence: when members cluster tightly, confidence is high; when they diverge widely, the atmosphere is in a state where multiple outcomes are plausible.

Ensemble output is communicated as probability forecasts ("70 percent chance of rain") rather than binary yes/no predictions. This probabilistic approach is more valuable for decision-making because it conveys not just the most likely outcome but the range of possibilities. Emergency managers, farmers, and energy companies increasingly rely on ensemble-derived probabilities for operational decisions where both the forecast and the uncertainty surrounding it matter.

Forecast Accuracy and Limits

Forecast accuracy has improved steadily over the past half-century. Today's 5-day forecast for large-scale features (pressure patterns, temperature trends) is as skillful as the 3-day forecast of the early 1990s and the 1-day forecast of the 1970s. This improvement translates to gaining about one day of useful forecast range per decade. One-day temperature forecasts now achieve errors of only 1 to 2 degrees Celsius in most regions, and 3-day precipitation forecasts correctly identify rain or no-rain events about 85 to 90 percent of the time.

The theoretical limit of deterministic weather prediction is generally estimated at about 2 to 3 weeks for large-scale features, limited by the chaotic nature of atmospheric dynamics. In practice, useful skill for specific weather events rarely extends beyond 10 to 14 days. Beyond that range, forecasts can only describe statistical tendencies (warmer than average, increased probability of above-normal precipitation) rather than specific weather events. Seasonal forecasts, extending months ahead, rely on slowly varying boundary conditions like sea surface temperatures (particularly ENSO state) rather than initial atmospheric conditions.