Weather and climate glossary

The terms used in the forecasting and El Niño notes, defined

A companion to How weather forecasting works, the El Niño note and the ENSO explainer. Terms are grouped by topic rather than alphabetically, so that related ideas sit together; use the site search box if you know the word you want. Assumes senior high school science and nothing more.

The basic idea

Numerical weather prediction (NWP) — Forecasting the weather by solving the equations of fluid motion and thermodynamics on a computer, rather than by pattern-matching against past weather. Almost all modern forecasting is NWP.

Initial conditions — The state of the atmosphere at the moment a forecast starts: temperature, humidity, wind and pressure everywhere. Weather forecasting is an initial-value problem, meaning the answer depends almost entirely on getting this starting state right.

Boundary conditions — Things imposed on the model from outside rather than predicted by it. For a regional model, the weather flowing in from the edges of its domain. For a climate model, greenhouse gas concentrations. The distinction matters: weather forecasting is controlled by initial conditions, climate projection by boundary conditions.

Chaos — The property of some physical systems whereby tiny differences in the starting state grow into large differences in outcome. It is a property of the system, not a flaw in the model or the computer. The atmosphere is chaotic, which is why forecasts have a hard limit.

Butterfly effect — The popular name for chaos in the atmosphere, from Edward Lorenz’s 1963 work: the idea that a disturbance as small as a butterfly’s wingbeat could, weeks later, change whether a storm forms.

Predictability — How far ahead a forecast can say anything useful. For specific days it runs out at roughly 10–15 days. Beyond that you can still forecast the odds of a wet or hot season, which is a different kind of statement.

Lead time — How far ahead the forecast is looking. A “three-day forecast” has a lead time of three days.

Deterministic forecast — A single forecast giving one value: 24 °C on Thursday. Simple to present, but it hides how confident the system is.

Probabilistic forecast — A forecast expressed as odds: a 70% chance of rain. Derived from an ensemble (below). More informative than a deterministic forecast, because the uncertainty is part of the answer.

Skill — How much better a forecast is than some cheap alternative, usually the long-term average for that date, or simply “tomorrow will be like today”. A forecast with no skill is not necessarily wrong; it just adds nothing.

Observing the atmosphere

Observation — Any measurement fed into a forecast: a temperature from a weather station, a wind from an aircraft, a radiance from a satellite. A modern global model ingests millions per run.

Radiosonde — The instrument package carried aloft by a weather balloon, transmitting temperature, humidity, pressure and wind as it rises to 25–30 km. Balloons give the most trustworthy vertical profile of the atmosphere, but there are only about 1,000 launch sites worldwide.

Reflectivity — What weather radar measures: how strongly a pulse of microwave energy bounces back off raindrops, hail or snow. Stronger return means heavier precipitation.

Doppler radar — A radar that also measures the frequency shift of the returned pulse, which gives the speed of the raindrops towards or away from the radar — and so the wind, along that direction only.

Dual-polarisation radar — A radar transmitting pulses polarised both horizontally and vertically. Comparing the two returns reveals the shape of the particles, which distinguishes rain from hail from melting snow.

Geostationary satellite — A satellite orbiting at about 36,000 km directly above the equator, where its orbital period matches the Earth’s rotation, so it stays over the same spot. It images the same hemisphere every ten minutes or so. Japan’s Himawari covers Australia.

Polar-orbiting satellite — A satellite at about 800 km circling roughly pole to pole every 100 minutes while the Earth turns beneath it, so it eventually sees everywhere, twice a day. These carry the most valuable instruments for forecasting.

Microwave sounder — A satellite instrument measuring microwave energy emitted by oxygen and water vapour in the atmosphere. Because microwaves pass through cloud, sounders give temperature and humidity profiles where visible and infrared instruments cannot see. The photograph in the forecasting note is one of these.

Radiance — The raw quantity a satellite actually measures: energy arriving at the sensor at particular wavelengths. It is not a temperature. Converting between radiances and atmospheric temperature is part of what data assimilation does.

Argo float — An autonomous instrument that drifts at depth in the ocean, rises to the surface every ten days to transmit a temperature and salinity profile, then sinks again. Roughly 4,000 are in the water at any time. They are the reason seasonal forecasting works.

World Meteorological Organization (WMO) — The UN agency under whose framework countries exchange weather observations without charge. For a country downwind of empty ocean, this arrangement is what makes forecasting possible.

Turning observations into a starting state

Data assimilation — The process of combining millions of scattered, irregular, imperfect observations with a previous short forecast to produce a single best estimate of the atmosphere on a complete grid. It is the step that made modern forecasting possible, and often consumes as much computer time as the forecast itself.

Background (or first guess) — The short forecast, usually six hours old, that assimilation starts from. You never build the starting state from observations alone; you correct a prior estimate.

Analysis — The output of data assimilation: the best available estimate of the atmosphere right now, on the model’s grid. The forecast starts from the analysis, and the analysis becomes the background for the next cycle six hours later.

4D-Var (four-dimensional variational assimilation) — The leading assimilation method. It searches for the atmospheric state which, when run forward through the model, best fits all the observations taken across a time window while staying close to the background. The “four-dimensional” part means each observation is used at the time it was actually taken.

Ensemble-variational assimilation (En4DEnVar and relatives) — A refinement that uses an ensemble of forecasts to estimate how the model’s errors are distributed on that particular day, rather than assuming a fixed error structure. Australia’s global system uses this approach.

Reanalysis — What you get by re-running the whole assimilation process over past decades using one fixed modern system. The result is a physically consistent, gridded history of the atmosphere, hourly, for 40+ years. Wind and solar resource assessment depends on it, because scattered historical observations alone cannot give you consistent weather at an arbitrary site.

Inside the model

Grid — The three-dimensional mesh of boxes the atmosphere is divided into. The model carries one value of each variable per box, and cannot represent anything smaller.

Horizontal resolution — The width of a grid box, in kilometres. Smaller is better and much more expensive: halving the spacing costs roughly 8–16 times more computing, because you gain boxes in two directions and must also shorten the timestep.

Vertical levels — The number of layers the atmosphere is sliced into from the surface upwards, typically 70–140. They are not evenly spaced; layers are packed close together near the ground and near the tropopause.

Timestep — The increment of simulated time the model advances in one calculation, typically seconds to a few minutes. Too long a step and the computation becomes numerically unstable.

Parameterisation — A sub-model that estimates the aggregate effect of processes too small to sit on the grid — individual clouds, turbulence, drag over hills, radiation, transpiration from plants — using the large-scale variables the model does resolve. This is where most model error lives.

Convection — Vertical overturning driven by warm air rising, ranging from a fair-weather cumulus to a thunderstorm. Because a thunderstorm is only a few kilometres across, a global model with 12 km boxes cannot represent one and must parameterise it. Convective rainfall is the largest single source of forecast error in the tropics and in Australian summer storms.

Convection-permitting model (also convection-allowing) — A model with boxes of roughly 1–4 km, fine enough that storms form on the grid itself, so the convection parameterisation can be switched off. Too expensive to run globally, so it runs over a limited region. It gives far more realistic storm structure, but not necessarily the storm in the right place.

Nesting — Running a fine-resolution regional model inside a coarser global one, which supplies the weather flowing in at the regional model’s edges. It is why a regional model adds little skill beyond about three days: errors arrive from the boundaries.

Coupled model — A model in which the atmosphere runs together with an ocean model, a wave model, a sea-ice model and a land-surface model, exchanging heat, moisture and momentum as they go. A refinement for tomorrow’s forecast; essential for next season’s.

Uncertainty and how forecasts are checked

Ensemble — A set of forecasts run from slightly different starting conditions, and in modern systems with slightly different model physics, to sample the range of possible outcomes. Typically 20–100 members.

Member — One individual forecast within an ensemble.

Perturbation — The small deliberate change applied to the starting state to create each member, sized to match how uncertain the observations actually are.

Spread — How much the members disagree. Spread is not noise: it is the forecast of the forecast’s own reliability, and it changes from day to day. Tight spread means confidence; rapid divergence means little confidence.

Post-processing — Correcting raw model output before it reaches the public. A grid box is an area average at a nominal height; a weather station is a point in specific terrain. Post-processing learns the systematic difference from years of past forecasts and fixes it.

Bias — A systematic error that repeats: a model always too warm overnight in valleys, or always too windy in sheltered places. Distinct from random error, and correctable.

Calibration — Adjusting forecast probabilities so they mean what they say. If the raw ensemble says 70% and it rains only half the time on those occasions, the system is not calibrated. Well-calibrated 70% forecasts verify at 70%.

Verification — Scoring forecasts against what actually happened. The only way to know whether a model upgrade helped.

RMSE (root-mean-square error) — A standard error measure for continuous variables like temperature: square the errors, average them, take the square root. Squaring means large errors dominate.

Brier score — The equivalent for probability forecasts of yes/no events, such as rain or no rain. The mean squared difference between the forecast probability and the outcome (1 or 0). Lower is better.

CRPS (continuous ranked probability score) — The generalisation of the Brier score to continuous variables, scoring the whole forecast distribution rather than a single threshold. The standard measure for ensembles.

Hindcast — Forecasts re-run retrospectively over past years, from the initial conditions that applied at the time. Seasonal systems need 30+ years of hindcasts, because that is the only way to know what their probabilities mean. Distinct from a reanalysis, which estimates past states rather than past forecasts.

Timescales

Nowcasting (0–6 hours) — Too short for the model chain to help, since assimilating and running a global model takes hours. Nowcasting works directly from observations, chiefly by tracking radar echoes and projecting them forward. The domain of flash flood and thunderstorm warnings.

Short range (1–3 days) — Where convection-permitting regional models earn their keep, and where forecasts are most accurate.

Medium range (3–10 or 15 days) — Global models and their ensembles. Skill falls steadily with lead time and the useful output shifts from specific values to probabilities.

Sub-seasonal (2–6 weeks) — Long called the “predictability desert”: too far out for the starting state to carry information, too near for slow ocean signals to have taken over. Skill here comes from the MJO, SAM and the stratosphere.

Seasonal (1–6 months) — Predictability comes almost entirely from the ocean, whose heat anomalies persist for months. Requires a coupled model and a large ensemble; the output is always probabilistic.

Climate projection — Not a forecast. Given an assumed emissions pathway, what does the statistical distribution of weather look like in the 2050s? Initial conditions are irrelevant at this range; boundary conditions are everything.

Emissions scenario — An assumed future pathway for greenhouse gases, aerosols and land use, fed into climate models as a boundary condition. Projections are conditional on the scenario, which is why they come in sets.

CMIP (Coupled Model Intercomparison Project) — The international framework under which modelling centres run the same scenarios so their results can be compared. The basis of the IPCC assessments.

Downscaling — Refining coarse global climate projections to regional detail, either statistically or by running a finer regional model over the area. Necessary because a 100 km global model cannot see the Great Dividing Range.

The Pacific: ENSO and its relatives

ENSO (El Niño–Southern Oscillation) — The single coupled ocean–atmosphere see-saw in the tropical Pacific whose warm and cool phases are El Niño and La Niña. “Southern Oscillation” is the atmospheric half — the pressure see-saw between Darwin and Tahiti — recognised decades before anyone realised it was the same phenomenon as the ocean warming.

El Niño — The phase in which the trade winds slacken, warm water sloshes back east, and the eastern Pacific ends up warm where it is normally cool. Rainfall follows the warm water eastward, away from Australia.

La Niña — The opposite phase: stronger trades, a colder eastern Pacific, convection concentrated further west, and generally wetter conditions over eastern Australia.

Trade winds — The persistent easterly winds along the equator, a consequence of the Earth’s rotation and the general circulation. They drag surface water westward, which is what sets the whole Pacific arrangement up.

Western Pacific Warm Pool — The mound of warm water piled up near Indonesia by the trade winds, the largest expanse of ocean above 28 °C on Earth. The sea surface there sits roughly 40–50 cm higher than off Ecuador.

Thermocline — The layer in the ocean where temperature drops sharply with depth, separating the warm surface water from the cold deep water. Its depth is the key variable in ENSO: deep in the western Pacific, shallow in the east, and it flattens during El Niño.

Upwelling — The rising of deeper water to replace surface water driven away by wind. It draws from a roughly fixed depth of 50–100 m, so whether the water arrives warm or cold depends on where the thermocline is sitting relative to that depth.

Cold tongue — The band of cool surface water stretching west from South America along the equator, produced by upwelling through a shallow thermocline. Worth being clear that the cold tongue is what normal looks like in the Pacific, not an anomaly.

Walker circulation — The closed loop of air that results: rising over the warm pool (deep cloud, heavy rain over Indonesia and northern Australia), sinking over the cold tongue (which is why coastal Peru is a desert), flowing west at the surface and east aloft. The winds create the temperature pattern and the temperature pattern drives the winds, which is what makes the system a see-saw rather than a steady state.

Sea surface temperature (SST) — The temperature of the top layer of the ocean. Because water holds far more heat than air and turns over far more slowly, SST anomalies persist for months, which is what makes seasonal forecasting possible at all.

Anomaly — A departure from the long-term average for that place and time of year, rather than an absolute value. Almost all climate monitoring is expressed in anomalies, because it separates the signal from the ordinary seasonal cycle.

ONI (Oceanic Niño Index) — The standard measure of ENSO state: the three-month running mean of sea surface temperature anomalies in the Niño 3.4 region of the central Pacific. NOAA counts an El Niño when the ONI reaches +0.5 °C or more for five consecutive overlapping three-month periods. This is the index plotted in the El Niño note.

Indian Ocean Dipole (IOD) — A comparable see-saw in the Indian Ocean, measured as the SST anomaly difference between its western and eastern tropical halves. A positive IOD means a cooler eastern Indian Ocean off Indonesia, and typically a drier winter and spring in south-eastern Australia. It matters more for Australia than for most countries, and a positive IOD arriving alongside an El Niño amplifies the drying signal.

Southern Annular Mode (SAM) — The north–south shift of the westerly wind belt that circles Antarctica. A major control on southern Australian rainfall and wind. Unlike ENSO it carries no slow ocean memory, so it can only be forecast a week or two ahead.

Madden–Julian Oscillation (MJO) — A pulse of tropical convection that travels eastward around the globe every 30–60 days. It shifts the odds of rain in northern Australia and helps trigger monsoon bursts, without determining any particular day. One of the few sources of skill at two to six weeks.

Composite — An average built from many past events of the same kind — for instance, the average rainfall pattern across 23 El Niño years since 1951. A composite shows the typical signal; it does not promise that any individual event will look like it.

Median — The middle value of a distribution, so “above-median rainfall” means the wetter half of past years. Seasonal outlooks use the median rather than the mean because rainfall is heavily skewed: a few very wet years drag the mean above what a typical year looks like.

Machine learning in forecasting

Machine-learning weather model — A neural network trained on decades of reanalysis to learn the mapping from the atmospheric state now to the state six hours ahead, then applied repeatedly to build out a forecast. It solves no physics equations. GraphCast, Pangu-Weather and ECMWF’s AIFS are the leading examples.

Training data — The historical record a model learns from — for weather models, reanalysis. This creates a dependency worth understanding: reanalysis is itself produced by physics models and data assimilation, so AI forecasts currently rely on the conventional system continuing to exist.

Blurring — The tendency of models trained to minimise average error to produce smooth, washed-out fields, understating extremes. A blurred forecast scores well on average and fails at exactly the cases warnings exist for.

Systems and acronyms

ACCESS — Australian Community Climate and Earth-System Simulator, the Bureau of Meteorology’s model family, built on the UK Met Office Unified Model. ACCESS-G is the global model (~12 km); ACCESS-GE its ensemble (~33 km); ACCESS-C the convection-permitting city models (~1.5 km); ACCESS-CE the city ensemble (~2 km); ACCESS-A a newer pan-Australian convective-scale model; ACCESS-S2 the coupled seasonal system (~60 km atmosphere, 99 members).

APS (Australian Parallel Suite) — The version numbering for operational ACCESS upgrades. APS4 is current.

BARRA — The Bureau’s regional reanalysis for Australia.

BARPA — The Bureau’s regional climate projection system, which downscales coarse global climate models over Australia. BARPA-C is a convection-permitting version.

STEPS — The Bureau’s nowcasting system, blending radar extrapolation with NWP as lead time increases.

IMPROVER — The Bureau’s statistical post-processing and calibration system, developed jointly with the UK Met Office.

OceanMAPS / AUSWAVE — The Bureau’s operational ocean model and wave model. AUSWAVE is built on WAVEWATCH III.

AGCD — Australian Gridded Climate Data, the Bureau’s gridded record of observed rainfall and temperature. The basis of the El Niño composite maps.

ECMWF — European Centre for Medium-Range Weather Forecasts, generally regarded as the leading global forecast centre. Its model runs at about 9 km with 137 vertical levels.

AIFS — ECMWF’s Artificial Intelligence Forecasting System. The deterministic version became operational in February 2025 and the ensemble in July 2025.

NOAA — The US National Oceanic and Atmospheric Administration. NESDIS is its satellite arm; CPC the Climate Prediction Center, source of the ENSO monitoring; NCEI the National Centers for Environmental Information, source of the temperature records.

Copernicus / C3S — The European Union’s Earth observation programme and its Climate Change Service, which publishes the monthly global temperature record referenced in the records chart.

Himawari — Japan’s geostationary satellite series, the one that images the Australian region.

JPSS — The US Joint Polar Satellite System: Suomi NPP, NOAA-20 and NOAA-21, which carry the microwave sounders.