Predictability of climate variability at the seasonal timescale

Predictions (predictability), projections and scenarios are different terms, although they are often interchanged. A prediction is a forecast of what will happen in future. This can be a deterministic forecast ('tomorrow it will be raining') or a probabilistic forecast ('there will be a more than average chance that tomorrow it will rain'). The predictability of a phenomenon can be defined as the degree to which its evolution can be deduced from the known initial conditions and the known evolution of factors that affect the phenomenon. It thus depends significantly upon the spatial and temporal scales of the phenomenon. Projections with, for instance, climate models can be made; but they cannot be considered as a prediction. A range of projections is often interpreted as a probabilistic forecast; but this is difficult as long as the quality of the projection (and therefore its likelihood to occur) cannot be firmly determined. A scenario is a projection following on from a set of basic 'what if' assumptions. For instance, for an assumed time evolution of greenhouse gas concentrations, the global mean temperature rise is deduced from an ensemble of climate model projections. However, within a given (concentration) scenario, the future climate can still evolve in multiple directions and, strictly speaking, cannot be predicted.

Global and regional climate predictability and the information that gives rise to predictability vary with the timescale and region considered. Predictability arises from at least two sources: initial conditions and changing external forcing. Predicting synoptic weather requires a good-quality initial condition of the atmosphere and land, and a decent meteorological model to describe the evolving dynamic weather features. Scientific and computational developments leading to improved initial conditions have extended the time range of sufficiently accurate weather predictions by approximately one day per decade since the late 1970s, up to approximately seven days at present. However, operational forecasting applications in the water sector usually rely on probabilities that extreme hydrological events occur, and the mean forecast quality (often denoted by the term 'predictive skill') is of less importance. Probabilistic weather forecasts have been used since the mid 1990s to assess the risks of, for instance, extreme river discharge, heavy precipitation events, hurricane tracks or other weather phenomena that have an impact upon society (see Figure 2.1). Applications focusing on this synoptic timescale are widely used and well known, and are not the subject of this book.

On longer timescales (such as the seasonal timescale), a likewise good initial condition of the slower components in the climate system is required: the temperature of the upper layers of the ocean and the sea surface temperature, ice cover extent, slowly varying signals in the stratosphere, and soil moisture and snow conditions on land. In addition, predictability at the seasonal timescale varies largely with seasons and across the globe since the chaotic nature of atmospheric motion destroys correlations as time proceeds. Seasonal predictions are routinely produced by a number of major weather services across the world. The El Niño Southern Oscillation (see the sub-section on 'Sources of predictability at the seasonal timescale') is an important source of predictability at seasonal timescales.

Seasonal forecasting tools

Seasonal forecasting tools have rapidly emerged in the past decade. Such tools are particularly powerful in areas and seasons where strong connections to slowly varying SST and other climate variables exist, and where the seasonal variability of the weather is substantial. In areas with small seasonal and year-to-year variability of mean seasonal

Forecast rime

Figure 2.1 Graphical display of probabilistic forecasting

Note: As forecast time proceeds, the probability distribution function (PDF) of a given event evolves and may occasionally break up in different regimes. The 'reality' line represents a retrospective check of the probabilistic forecast.

Source:Taylor and Buizza (2004)

Forecast rime

Figure 2.1 Graphical display of probabilistic forecasting

Note: As forecast time proceeds, the probability distribution function (PDF) of a given event evolves and may occasionally break up in different regimes. The 'reality' line represents a retrospective check of the probabilistic forecast.

Source:Taylor and Buizza (2004)

precipitation, such as the mid-latitudes or desert regions, less opportunities to predict anomalous climate conditions are present than in areas with strong variability (such as monsoon climates or land areas in the (sub)tropical regions). Seasonal forecasting tools do not aim to forecast a specific event at a given day, but rather the probability that the seasonal mean precipitation or temperature is higher or lower than the clima-tological mean. The existing tools can be roughly divided into two classes: statistical and numerical methods (Palmer and Anderson, 1994). In some applications, a mixture of the two is used.

Statistical methods use observed correlations between SST and regional weather patterns to make forecasts for the future. El Niño variations are an important source of predictability (see the following sub-section). Apart from giving a probability of anomalously high or low precipitation, they are often used to choose historical analogue years that serve as input to hydrological or agricultural applications (see, for example, Stone et al, 1996; Hamlet and Lettenmaier, 2000). However, they often suffer from limited observational record length needed for the calibration of the tools. And they are not able to cope with changes in statistical correlations induced by changes in the external forcings. Although future climate projections do not show strong shifts in El Niño frequency or structure (Van Oldenborgh et al, 2005 a), nor in the structure of the tele-connections (Van Oldenborgh and Burgers, 2005; Sterl et al, 2007), the statistical relations found today may be different for tomorrow's climate conditions.

Numerical methods use an ensemble of projections with coupled ocean-atmosphere general circulation models (OAGCMs) initialized with an 'observed' state of the ocean, land and ice conditions. This approach copes with the inherent uncertainty introduced by the chaotic nature of the climate system. However, the quality of the initial states is fairly poor owing to the lack of routine observations in the ocean and on land. A well-known operational system for seasonal forecasting is the multi-model EUROSIP system (EUROpean multi-model Seasonal to Inter-annual Prediction) system (see www.ecmwf. int/products/forecasts/seasonal/forecast/forecast_charts/eurosip_doc.htm, accessed 1 July 2008), where seasonal predictions from three European meteorological services are combined into a single application database. By combining multiple modelling systems, the model's uncertainty can be assessed.

Both the statistical and numerical tools for seasonal prediction rely on existing sources of predictability at the seasonal timescale. The major source is the oceanic surface temperature (of which El Niño is the strongest expression); but other sources are being investigated as well. The major sources are briefly discussed in the following sub-section.

Sources of predictability at the seasonal timescale

El Niño Southern Oscillation (ENSO)

Irregular but persistent sea surface temperature variations in the equatorial Pacific Ocean are associated with the El Niño Southern Oscillation phenomenon, which returns on average every three to seven years. During an El Niño, SSTs are warmer than normal around the Equator in the eastern half of the Pacific Basin, usually starting early in the year and peaking during November to January. This results from an interaction between the ocean and the atmosphere, where changes in the ocean surface temperatures affect tropical rainfall patterns and atmospheric winds over the Pacific Ocean, which in turn affect ocean temperatures and currents. Details about the mechanisms and the degree to which they are reproduced adequately in present-day climate models are given by Neelin et al (1998). Popular documentation is given on many websites, including the National Center for Environmental Predictions (NCEP)/Climate Prediction Center (CPC) website (see products/precip/CWlink/MJO/enso.shtml, accessed 1 July 2008).

The strength of an ENSO event is usually expressed as a SST anomaly in a particular Pacific section at the Equator. Different regions are used, leading to different (related) indices. For instance, one index (NINO1.2) concentrates on SST anomalies near the coast of Ecuador and Peru and is indicative of coastal precipitation variability. NINO3.4 is located in the centre of the Pacific and is related to weather phenomena around the world.

The ENSO phenomenon has a clear impact on (hydro-)climate in many regions of the world (see Plates 3 and 4, centre pages). The strongest relationships are found in the Pacific equatorial zone and coastal areas bordering the Pacific Ocean. Apart from a weak positive impact of ENSO upon precipitation in South-Western Europe during spring (Van Oldenborgh et al, 2000) a teleconnection between ENSO and European climate variability is not detectable.

Other variability modes

Apart from ENSO, a number of other modes of atmospheric and oceanic variability exist that bear some seasonal predictability in some regions of the world: the Pacific Decadal Oscillation (PDO), the North Atlantic Oscillation (NAO) and the Indian Ocean Dipole (IOD). For instance, rainfall in Eastern Africa is correlated to SST anomalies in the Western Indian Ocean. The NAO index (usually expressed as an anomaly in the pressure difference between Iceland and Lisbon) is positively correlated to precipitation in the northern part of Europe and reversed in Southern Europe (mainly during the winter season; see Plate 5, centre pages). Although these signals are weaker than ENSO teleconnections, they are used in some statistical seasonal forecasting tools.

Land-atmosphere interactions

An active field of research is the possible predictive skill (or forecast quality) present in the slowly varying terrestrial components of the climate system, such as snow and soil moisture. Statistical analyses have demonstrated a detectable positive correlation between springtime snow amounts and temperature up to one month later in NorthWestern Europe (Shongwe et al, 2007). Extreme hydrological events in Europe and the US (like the European 2003 summer heat wave) have incited a number of studies demonstrating increased likelihood of anomalous heat-wave intensities during summer when the winter/spring soil moisture content is relatively low (see, for example, Ferranti and Viterbo, 2006). These studies justify investments in widespread observation and data assimilation of these quantities in order to increase the quality of the terrestrial initial conditions of numerical seasonal prediction tools.

Regional differences in the predictability of the climate

The persistent ENSO feature is a powerful source of climate predictability at the seasonal timescale. If the initial condition related to the anomalous ENSO state is captured well, the relatively high correlations to weather phenomena elsewhere in the world enhance the quality of the forecast owing to the large persistence of the phenomenon.

Nevertheless, regional differences in the predictability of the climate exist. For instance, due to the high heat capacity of ocean water, the predictability of the temperature is higher over oceans than over land. Systematic changes in air circulation related to ENSO are also a source of high predictability. For example, the predictability of precipitation in high-rainfall regions in the tropics is strong due to a clear effect of ENSO.

Van Oldenborgh et al (2005b) compared the skill of seasonal predictions from a statistical forecast model and a number of European Centre for Medium-Range Weather Forecasts (ECMWF) dynamic coupled modelling systems performing a three-month forecast (Plates 6 and 7, centre pages). For December, January and February (DJF), positive skill in terms of 2m temperature is seen in areas where strong teleconnections with ENSO are present. The comparison shows that the dynamic model is better in areas where other factors than ENSO play a role, such as the Indian Ocean, or over many land areas outside the tropics. In general, forecasting precipitation is much more complex than temperature. Consequently, the precipitation forecast skill is generally lower (not shown).

Availability and formats of seasonal forecasts

Many (climate) institutes around the globe offer seasonal forecasts or outlooks for different periods and different parts of the globe. It is beyond the scope of this book to give an exhaustive list of products, but a few examples are provided here.

An example of statistical seasonal forecast products is provided by the Australian Bureau of Meteorology, which issues probability maps of above or below median rainfall for a three-month period (see, accessed 1 July 2008). In addition, guidance is given on the results and the interpretation. The Canadian Weather Office (see saisons/index_e.html, accessed 1 July 2008) provides similar maps for lead times longer than three months, but uses numerical forecasts for shorter lead times. The US Climate Prediction Center (see, accessed 1 July 2008) issues tercile maps (probabilities above, at or below normal) based on a mix of statistical and numerical methods. The statistical methods use various observed correlations, including ENSO and a soil moisture index. The skill of each method is assessed separately. Outlooks are given for the US only.

Products from comprehensive multi-model systems are given by the International Research Institute for Climate and Society (IRI; see, accessed 1 July 2008) and the EuroSIP consortium (see forecasts/seasonal/forecast/forecast_charts/eurosip_doc.htm, accessed 1 July 2008). The EuroSIP continues an earlier successful seasonal prediction project DEMETER (Palmer et al, 2004), which has generated a widely used data set of seasonal forecasts for evaluating model quality, climate variability at the seasonal timescales, maximum predictability across the world, and optimal interfaces between the meteorological products and end users in hydrology, agriculture, safety management and other applications. The most common products emerging from EuroSIP include an ENSO forecast plume and a range of probability maps for anomalously high or low precipitation, SST or temperature (see Plates 8 and 9, centre pages).

All seasonal forecast products are issued with a skill assessment. However, the skill parameters and criteria vary among the groups. A simple quality (or skill) measure is the correlation between forecasted and observed seasonal mean temperature, precipitation or surface pressure. More advanced quantities include the relative operating characteristics (ROC) score, which expresses whether a forecast is actually successful in predicting a certain event to happen. A useful data portal to compare skill and seasonal forecasts from a range of forecast centres is the Climate Explorer (see, accessed 1 July 2008). Here the seasonal forecasts are collected near real time and can be processed or verified using a common metric and verification database.

Was this article helpful?

0 0

Post a comment