Before weather satellites, a forecast map was a hand-drawn field of isobars surrounded by blank ocean, and forecasts built from those maps were usually not accurate beyond two days. NASA’s history of the GOES program says forecasters depended on manually plotted maps, cloud observations and barometric pressure readings because so much of the atmosphere remained unobserved.
The limitation was even sharper in the Southern Hemisphere. Weather systems could form over enormous stretches of ocean and travel toward inhabited coasts before forecasters had enough observations to describe them confidently.
On 5 June 1944, Allied commanders postponed the invasion of Normandy after a team led by Group Captain James Stagg warned that the weather would be too poor. The Met Office archive records how the team then identified a narrow window of marginal but workable conditions for 6 June using synoptic charts, reconnaissance reports and observations gathered across Europe and the Atlantic.
How forecasters built a map from fragments
Picture a national weather office in the middle of the twentieth century. Teleprinters delivered coded observations from ships, airfields, lighthouses and weather stations while forecasters drew pressure lines across paper charts large enough to cover a table.
Each observation represented one place and one moment. Between those points, forecasters had to infer the positions of fronts, pressure systems and cloud bands using experience, earlier charts and whatever reports had arrived over the wire.
The equations needed for numerical weather prediction were known, but solving them quickly enough to produce a useful forecast required electronic computers. Early numerical systems could simulate the atmosphere, yet their results were only as reliable as the description of the atmosphere supplied at the beginning of each run.
That starting description was the central problem. A model cannot reconstruct a storm over an unobserved ocean simply because its equations are correct, and errors in the initial state grow as the forecast moves farther into the future.
The Southern Hemisphere was mostly blank
Around 1960, the weather-balloon network covered only about 10 per cent of the troposphere and did not meaningfully cover the Southern Hemisphere, the tropics or the oceans. The global importance of closing those gaps is why the World Meteorological Organization describes satellites as a fundamental part of the observing system used for weather and climate services.
The Northern Hemisphere had dense clusters of land stations across North America and Europe. South of the equator, large sections of the Pacific, Indian and Southern oceans contained few regular upper-air observations, even though those waters generated and steered weather toward Australia, New Zealand, southern Africa, South America and Pacific islands.
A vessel crossing the Indian Ocean might transmit wind and pressure readings, but a single ship supplied only one dot on an immense map. If no ship crossed the developing system, the first useful evidence might arrive only when falling pressure, growing swell or cloud changes were detected closer to land.
This was why Southern Hemisphere forecasts generally lost accuracy sooner. Sparse observations created a less reliable initial state, and the errors spread as the model or human forecast tried to project the atmosphere forward.
TIROS showed the clouds, and GOES watched them move
TIROS-1 launched on 1 April 1960 carrying two television cameras and video recorders. During 78 days of operation, the first successful weather satellite returned more than 19,000 usable pictures and demonstrated that cloud systems could be monitored from orbit, according to the National Weather Service history of the mission.
Those images did not provide continuous observation. Early polar-orbiting satellites could generally see the same location only twice a day, leaving long intervals during which thunderstorms could develop, cyclones could intensify and cloud structures could change.
Geostationary satellites altered that geometry. A spacecraft placed about 35,800 kilometres above the equator and moving with Earth’s rotation can remain over the same region, allowing forecasters to watch weather develop instead of comparing isolated snapshots.
GOES-1 launched in October 1975 with the Visible and Infrared Spin-Scan Radiometer inherited from the earlier Synchronous Meteorological Satellite missions. The instrument supplied day-and-night observations of cloud and surface temperatures, cloud heights and wind fields, and the first GOES generation helped track Tropical Storm Claudette and Hurricane David in 1979, as documented in NOAA’s GOES history.
Modern GOES instruments can produce full-disk images every ten minutes, regional images every five minutes and localized views as often as every 30 seconds. That push toward faster orbital observation continues in new satellite projects that combine Earth-observation sensors with onboard artificial intelligence.
The ceiling rose through satellites, models and computing
Forecast skill does not have one universal horizon because it varies by location, atmospheric variable and spatial scale. An ECMWF assessment of the forecast skill horizon found that improved observations, data assimilation, models and ensemble methods can preserve useful signals beyond ten days for some large-scale patterns, even though precise local conditions become uncertain sooner.
Tropical cyclone tracks show one of the clearest long-term gains. The National Hurricane Center’s verification record documents a substantial decline in Atlantic track errors across standard forecast periods, while intensity changes remain more difficult to predict.
Satellites are a major contributor, but they are not the only reason forecasts improved. Faster computers, better atmospheric models, aircraft observations, radar, ocean measurements and increasingly sophisticated data-assimilation systems all help turn the observations into a forecast.
Different forecast products also answer different questions. A satellite can reveal a cyclone’s evolving cloud structure within minutes, while seasonal systems examine broad patterns such as the developing Pacific conditions discussed in Space Daily’s coverage of El Niño forecasting.
AI still depends on the observing system
Machine learning is changing how quickly the atmospheric picture can be projected forward. GraphCast can generate a ten-day forecast in under one minute on a single Google TPU v4 system, and its developers reported in their 2023 Science paper that it outperformed ECMWF’s deterministic system across most of the variables and lead times tested.
Aardvark Weather goes further by learning a path from observations to global and local forecasts without requiring a conventional numerical weather-prediction product during deployment. The system was described in a 2025 paper in Nature, not Nature Communications, and its results included skilful station forecasts at lead times of up to ten days for the variables examined.
Neither system makes orbital observations unnecessary. GraphCast was trained on reanalysis that combines satellites, radar, stations and conventional models, while Aardvark directly uses remote-sensing and surface observations and loses substantial skill when satellite inputs are removed.
What better warning can and cannot do
The 1970 Bhola cyclone shows the stakes without reducing disaster history to a single technology. The storm struck what was then East Pakistan on 12 and 13 November and killed an estimated 300,000 to 500,000 people, mainly when its storm surge overwhelmed low-lying islands and tidal flats, according to the World Meteorological Organization.
Bhola helped drive the creation of stronger international cyclone-warning programs, but satellites alone do not move people to safety. Survival also depends on forecasting institutions, communications, public trust, evacuation routes, shelters and whether warnings reach vulnerable communities in time.
The modern observing system has removed much of the blank space that once covered the Southern Hemisphere, but it has not removed uncertainty. Satellites can reveal a storm over an empty ocean, models can project where it may travel and emergency agencies can issue warnings, yet rainfall, intensity and local impacts can still change rapidly.
The paper charts from June 1944 remain in the Met Office archive, their observations and pressure systems fixed in ink and pencil. Above the oceans those forecasters could barely sample, satellites now return new measurements every few minutes, repeatedly filling a map that once became blank at the coastline.