This chapter will
- Outline important concepts in the consideration of predictability and its relation to practical aspects of prediction and prediction skill;
- Identify important sources of subseasonal to seasonal (S2S) predictability and highlight recent progress in understanding and modeling these sources; and
- Recommend research in these areas that will further our understanding of sources of predictability and allow us to better exploit them to extend and improve S2S forecast skill.
During the early developments of seasonal to interannual climate prediction, weather prediction was often described as an “initial-value” problem and seasonal or longer-term climate prediction as a “boundary-value” problem (Box 4.1). However, as our aspirations and capabilities for providing skillful predictions across timescales and Earth system components has increased, the value of such distinctions has become limited.
The schematics in Figure 4.1 are meant to further illustrate the above considerations and the complexities in estimating predictability across a range of timescales and phenomena (e.g., S2S) and implementing them in a forecasting system. The top schematic shows time series depicting variations in an arbitrary quantity (e.g., temperature, precipitation) that are typical of weather, subseasonal, and seasonal variability over a roughly 6-month period, with an indication of what processes might be associated with the given variation and timescale (e.g., green colors are associated with the Madden-Julian Oscillation [MJO], soil moisture). The bottom schematic is similar but for longer timescale variations. For weather forecasts, the forecast proceeds from an observed initial state (solid blue circle) out to lead times of a few days. For subseasonal or seasonal forecasts, the same is true out to a few weeks or months, respectively. Along the lines of the discussion above, the influence on weather or S2S forecasts
from the sorts of (long timescale) phenomena indicated in the lower schematic could feasibly be provided as a fixed boundary value. However, this practical separation between processes and timescales is not always possible or straightforward, particularly as the different timescales become closer. In fact, filling the forecast capability gap between the weather (initial value) and the seasonal (previously referred to as boundary value) problems with subseasonal capabilities has served to strongly blur the perception that they are separate sorts of forecast “problems” and helped to instigate the desire for “seamless” forecasting systems (e.g., Palmer et al., 2008, see Box 5.2).
Even when considering only the timescales associated with the S2S forecast range, a spectrum of phenomena and processes contribute to the observed variations. This is schematically indicated in Figure 4.1 by the rainbow of colors equated from high (purple) to low (red) frequency phenomena. Within this spectrum of phenomena/processes are some, highlighted by the individual colored lines, that dominate the variability and provide valued sources of predictability. The study of predictability is to answer the question of whether a prediction of a phenomenon is possible given all antecedent observations.
In this report, predictability refers to a phenomenon’s potential, that is, its upper limit, for being predicted. Theoretically, this is inherent to the phenomenon itself and its limit results from inevitable errors in initial conditions, which are amplified through nonlinear processes in a perfect model. Practically, predictability can only be estimated through various means using empirical or numerical models that are not perfect (NRC, 2010b).
Such models can lead to over- or underestimates of predictability. For example, by not adequately accounting for “noise” relative to the specific phenomenon (i.e., the “signal”), models will often overestimate predictability. Similarly, if the specific phenomenon is only weakly represented, then models will often underestimate its predictability. A firm estimate of the lower bound of predictability is the upper limit on the observed forecast skill of operational systems. Achieving the upper limit of predictability with a prediction system is hindered by practical factors. In addition to a lack of accurate specification of the initial conditions (typically due to inadequate observation sampling in space, time, and physical quantity), there are shortcomings in forecast systems (typically limited by models with too coarse spatial resolution and incomplete or inaccurate physical process representations) and shortcomings in the data assimilation systems (see Chapter 5). For the purposes of this report and the associated research agenda, it is critical to explore and quantify the predictability of various components of the Earth system, especially weather and climate. The estimated upper bounds of predictability for the various phenomena and processes discussed in this chapter are key to identifying unexploited or underexploited prediction capabilities and providing a quantified means to measure our progress in practical forecast skill against our predictability (i.e., upper limit) estimates (e.g., Figure 4.2). Together these help to prioritize areas of research and model development across the range of sources of predictability to pursue.
Research on predictability and its sources is a central part of carving a path to new and improved forecast capabilities (e.g., S2S). Advancements in this research critically hinge on observations, a variety of models, forecast system analogs, and ensemble retrospective forecast data sets. Such research typically begins with theoretical considerations or empirical analysis based on observations (Figure 4.3, Facet I; e.g., a lagged correlation analysis between two or more variables) that point to a process or phenomenon that exhibits predictability. From this perspective, it is essential to have long records of multivariate observations for both the predicted and the potential predictor(s). Often predictability of a particular observed phenomenon is investigated through process-oriented studies using a hierarchy of models (Figure 4.3, Facet II). Models used for this purpose may be reduced order or idealized. In other cases researchers may create a series of sensitivity experiments in complete Earth system models or make intercomparisons across models.
Further advancement (Figure 4.3, Facet III) is made by the development of robust models that incorporate the physical relationships underlying the phenomena or coupled interactions that yield the predictability, as demonstrated by simulation or
retrospective forecast experiments that are evaluated against observations. Such model development generally requires additional targeted process observations, research and analysis in order to properly understand the underlying processes, and significant effort to encapsulate and validate them in the models. Because S2S forecasts are inherently probabilistic, an ensemble of simulations is usually needed to tests whether relationships that give rise to predictability are accurately simulated. The number of ensemble members needed to capture the probability is a research question that can be explored either theoretically or practically.
With the dynamical/coupled models developed in Facets II and III, carefully designed experiments can be performed to estimate the predictability and assess sources of skill (Figure 4.3, Facet IV). To identify the contribution of a process to forecast skill, numerical retrospective forecasts are usually performed and compared using model configurations with and without the process. In an example with such an approach,
it was found that seasonal forecast skill is enhanced following stratospheric sudden warmings (Sigmond et al., 2013). Roff et al. (2011) assessed the effect of stratospheric resolution on extended-range forecast skill.
Another approach involving relaxation retrospective forecast experiments—where, for example, atmospheric fields in one region are relaxed toward analysis fields—has been used to assess the impact of teleconnections and to identify the origin of skill. With this approach, Ferranti et al. (1990) found an important contribution of tropical subseasonal variability on the forecast skill in the North Hemisphere middle latitudes. Vitart and Jung (2010) assessed the influence of the Northern Hemisphere extratropics on the skill in predicting the MJO.
In summary, research on predictability and its sources is critical for helping to identify and prioritize advancements between the various phenomena and Earth system components that might impact/offer S2S predictability. Such research is also critical to guiding model and forecast system development and helping to identify observing systems for sustained observations.
A useful technique to estimate predictability is the analysis of variance (ANOVA). An atmospheric variable, such as air temperature or precipitation, may be decomposed into a predictable component (signal) and an unpredictable component (noise), Z = Zs + Zn. The predictable signal comes from sources of predictability such as those discussed in the next section, and the unpredictable noise comes from chaotic processes with respect to S2S, such as high-frequency eddies. The total variance is then the sum of the signal variance and the noise variance, that is, Var(Z)=Var(Zs)+Var(Zn). The extent to which the signal variance, Var(Zs), exceeds the noise variance, Var(Zn), determines the potential predictability. Thus potential predictability can be defined as the ratio of signal variance to the total variance, Var(Zs)/Var(Z), or the ratio of signal to noise, Var(Zs)/Var(Zn). In S2S, forecasts averaged over a period of a week, a month, or a season, are usually produced. Such time averaging can increase the signal-to-noise ratio, because it reduces high-frequency noise variance while keeping most of the slow-varying signal variance, and thus improve predictability. Estimates of predictability with ANOVA can be performed using observational data (Facet I of Figure 4.3) or with ensemble model integrations (or retrospective forecasts) (Facet II of Figure 4.3). For example, Madden (1976) estimated weather noise variance of seasonal means by extrapolating the power spectrum derived from observed daily time series in a season. In the case of a dynamical ensemble retrospective forecast, the ensemble mean represents the “predictable signal component,” because it is independent of the uncertainties (in initial condition or model parameter). On the other hand, the difference among the members of the ensemble retrospective forecast (spread) represents the “unpredictable noise
Estimates of predictability of a given process or phenomenon when accounting for (unavoidable) uncertainties in the initial conditions and model configurations can also be done from “twin experiments” or an ensemble of experiments where one of the ensemble members is considered truth (or the “observed” state) and the other member(s)—which only differ by some small perturbation in the initial conditions and/or model parameters—are used to predict it. These predictability estimates can then be put into the same context as retrospective forecast experiments that instead compare the same predictions to observations in order to quantify forecast skill. This is the type of prediction skill and predictability experimentation that is shown in Figure 4.2. One of the main messages from this figure is that the practical forecast capabilities are still far from what might be achieved given the associated estimates of predictability (i.e., at least 2-3 weeks of additional lead time might be possible). An additional message is that predictability estimates are model dependent, as stated earlier. Further research and exploration can be performed with this type of system of experimentation through categorizing results by season or the conditions of other portions of the climate system (e.g., warm or cold El Niño-Southern Oscillation [ENSO] state). In summary, predictability research is critical for helping to identify and prioritize advancements between the various phenomena and Earth system components that might impact/offer S2S predictability, as well as guiding model and forecast system development and helping identify observing systems for sustained observations.
Finding 4.1: Predictability research is critical for identifying predictable phenomena and providing lead-time-dependent upper limits on prediction skill. These serve to guide model and forecast system developments with practical targets for forecast skill. Further research on predictability is needed to more completely identify sources and quantify levels of predictability, including interactions across scales and phenomena and how these impact predictability of extreme events.
As illustrated in Figure 4.1, predictability derives from a number of processes and phenomena that exhibit a wide range of timescales. For the purpose of this discussion, these sources will be generalized into three types.1 The first occurs in the form of
1 Using an analog from basic physics, these three types can very loosely be equated to a harmonic oscillator, a strongly damped harmonic oscillator, and a forced harmonic oscillator.
recurring and/or quasi-oscillatory patterns of variability—often referred to as “modes” of variability—that vary with S2S timescales. When a space-time pattern of variability tends to reoccur in the observed record, particularly when it includes positive and negative phases and/or space-time propagation of the given pattern, it is often referred to as a natural“mode” of variability. Attempts are made to understand the physics behind the pattern(s) and the evolution of a typical event life cycle for such modes. Examples include ENSO, MJO, Quasi-Biennial Oscillation (QBO), and Indian Ocean Dipole (IOD). Referring to Figure 4.1, this type of S2S predictability would be exhibited as a quasi-oscillatory phenomena with a period ranging somewhere between about 2 weeks to 1 year (based on this report’s definition of S2S).
The second source of predictability occurs from an anomaly in the initial state of one of the components of the Earth system whose typical timescale of evolution (i.e., persistence time) is similar to the target forecast. For the S2S timescale, this might be large-scale anomalies in upper ocean heat content, sea ice, snowpack, soil moisture, etc. Given their relatively slow variation compared to weather, such anomalies are said to retain “memory” of the initial state and impart “inertia” to the system’s subsequent evolution. They typically have a systematic or recurring manner of evolving on timescales much longer than the forecast. For the purposes of this discussion, we refer to these as “slowly varying processes.”
The third type of predictability stems from anomalous external forcing that is extensive or strong enough to have an impact globally or regionally for weeks to months (such as cyclic or anomalous solar output, anthropogenic factors, and events such as volcanoes). In this case, its predictability in relation to S2S is derived from a combination of being able to specify the anomalous external forcing and the forcing evolving relatively slowly or in a well-defined way over the forecast lead times (e.g., the annual cycle of solar radiation).
Understanding and being able to model the dynamics of each of these three types of predictability sources, as well as their interactions and teleconnections, is essential to generating S2S forecasts. Although much progress has been made in recent years in furthering understanding of how some of these sources of predictability, such as the MJO or soil moisture, influence environmental conditions or events that forecasters would like to be able to predict (e.g., precipitation anomalies, heat waves, or tropical cyclones), more work is needed and continued progress in this area remains fundamental to advancing S2S predictions (e.g., NRC, 2010b; Vitart et al., 2014). Indeed large gaps remain in our understanding of the sources of predictability and how they may interact, and discoveries of new sources of predictability for forecasts of different phenomena remain likely. Each of the subsections below describes progress and gaps
in understanding the three types of predictability important for S2S prediction—natural modes of variability, slowly varying processes, and external forcing.
Predictability from Natural Modes of Variability
Natural modes of variability display distinct and organized patterns that are typically oscillatory or cyclic in some fashion, or at least bimodal with the given “mode” having a tendency to occur with an anomaly pattern of one sign or its opposite (e.g., see Figure 4.4). The modes are typically identified in a given field (e.g., sea surface temperature [SST], 500 hPa heights, 200 hPa zonal winds) but are correlated to impacts on, or interactions with, other features in the Earth system, such as temperature, precipitation, drought, bio-productivity, and ozone. These modes of natural variability are characterized by dynamical interactions within or across Earth system components. A canonical example of a coupled mode of variability is ENSO. When these modes have life cycle lengths similar to S2S forecasts (e.g., 2 weeks to 12 months), their characteristic evolution offers a source of S2S predictability (Figure 4.1). For such cases, it is imperative that the forecast system be able to accurately represent the mode of variability and its life cycle. If the life cycle is much longer than the S2S timescale, then practically speaking for the purpose of the S2S forecast, the mode’s variation would likely be considered a “slowly varying process.” Natural modes of variability are often associated with teleconnection properties that relate variability at one location to conditions in another. For example, the mechanisms that produce ENSO occur and evolve in the tropical Pacific Ocean, yet influence midlatitude variability through atmospheric dynamics. As a result, the sign, strength, and frequency of occurrence of known patterns of extratropical atmospheric circulation (such as the Pacific North American [PNA] pattern) partly depend on ENSO (e.g., Zhang et al., 1997). Atmospheric patterns are in turn important drivers of winter weather and climate over North America. Some of the more well recognized natural modes of variability already found to be or expected to be important for S2S predictability are discussed in more detail below, with specific attention to areas that are ripe for or in need of further research.
A large part of the signal for S2S weather and climate predictions have tropical origins (NRC, 2010b). Through relatively long-lived SST anomalies (e.g., ENSO, IOD) and/or systematic dynamic flows (e.g., wavelike motions, MJO, Kelvin waves), large-scale storm systems become highly organized and produce systematic variations in atmospheric heating. This excites circulation anomalies that have local impacts on rainfall and temperature in the tropics but that also “propagate” to the extratropics via sequences of circulation anomalies of alternate sign, often referred to as “waves” (Horel and Wallace,
El Niño-Southern Oscillation (ENSO)
ENSO, treated in detail in the NRC (2010b) report Assessment of Intraseasonal to Interannual Climate Prediction and Predictability, is a coupled atmosphere-ocean mode of variability that involves slow equatorial waves in the ocean that impact SST, particularly in the central and eastern Pacific, and associated changes in surface pressure and wind variations in the atmosphere that extend over most of the tropical regions. The SST variations in the Pacific and associated circulation changes result in strong modulations of organized convection and precipitation in this region, which in turn influence the extra-tropical circulation via teleconnections as described above. For seasonal prediction, ENSO’s coupled dynamics provides a major source of skill (e.g., NRC, 2010b; Shukla et al., 2000), while for subseasonal predictions, the relatively slowly varying SST anomalies provide a relatively persistent tropically forced atmospheric circulation anomaly. Because the evolution of ENSO is anchored to the seasonal cycle, it is often described as an event, with terms El Niño for warm events and La Niña for cold events (e.g., Figure 4.4). The signature of an ENSO event first emerges during boreal spring or summer, and the associated SST anomalies peak the following fall or winter and then typically decay in spring.
The normal progression of ENSO and its impacts on the tropics and elsewhere through teleconnections are relatively well studied (e.g., Hoerling and Kumar, 2000; Latif et al., 1998; Rasmusson and Mo, 1993). Recent work has advanced our understanding of ENSO’s impact on predictability, the impacts of a number of distinct types of ENSO events, and ENSO’s decadal and longer timescale change. For example, there are lagged impacts of ENSO on predictability of the Indian Ocean in summer. In the summer after a positive ENSO first emerges, the tropical Pacific SST returns to normal, but the Indian Ocean SST is anomalously high, with a strong suppression of tropical cyclones and impacts on temperature and precipitation across Southeast Asia and Japan (Chowdary et al., 2011; Kosaka et al., 2013). Despite the large body of existing work on ENSO, there are important gaps related to understanding both ENSO and its influence on S2S predictability (McPhaden, 2015). For example, there is neither consensus on a theory nor agreement on the predictability limit of ENSO. Recent work shows that variations in the structures and seasonal timing of ENSO strongly affect the persistence and predictability (Lee et al., 2014), as well as ENSO teleconnections (Capotondi et al., 2014). In recent decades, the SST anomalies in El Niño events have
often peaked in the central Pacific rather than the more typical location of the past in the eastern Pacific. Whether such ENSO diversity is a consequence of greenhouse warming, and hence the recent shift can be expected in the future, is also unclear (Cai et al., 2015; Capotondi et al., 2014).
Madden-Julian Oscillation (MJO)
The MJO, also discussed in detail in the NRC (2010b) ISI report, is the dominant mode of intraseasonal variability in tropical convection, precipitation, and circulation. Through its local influences in the tropics and its teleconnections to higher latitudes, it represents a primary source of predictability at the subseasonal timescale (e.g., Waliser, 2011). The MJO is mainly an atmospheric phenomenon, but it also exhibits some modest interactions with the upper ocean—both in forcing and responding to coupled SST anomalies and exciting ocean currents and waves. It exhibits planetary-scale structures along the equator in pressure, winds, clouds, rainfall, and many other variables, with its strongest anomalies in precipitation propagating from the Indian to central Pacific Oceans over a period of about 30-50 days. An eight-phase index for MJO, referred to as the Real-time Multivariate MJO (RMM1 and RMM2) indices of Wheeler and Hendon, are usually used to describe the east-west location and amplitude of the MJO (Wheeler and Hendon, 2004). See NRC (2010b), Lau and Waliser (2011), and Zhang (2005) for further description.
The MJO has been shown to have significant connections to a number of important global weather and climate phenomena, including high-impact events (e.g., see reviews in Lau and Waliser, 2011; Zhang, 2005, 2013). This includes a strong influence on the onset and breaks of the Asian and Australian summer monsoons and on the modulation of synoptic variability—including tropical cyclones—and even the triggering of ENSO variations. Improving representation of the MJO in global models has led to better prediction on S2S timescales at high latitudes as well as in the tropics (e.g., Ferranti et al., 1990; Vitart, 2014). For example, North American wintertime surface temperatures are found to be anomalously warm 10-20 days after MJO-related convection occurs in the Indian Ocean (Lin and Brunet, 2009) (Figure 4.5). Such a lagged relationship implies predictability of North American temperature anomalies up to about 3 weeks given knowledge of the initial state of the MJO. Forecasts using statistical models have demonstrated that it may be possible to extend the forecast range of North American temperature anomalies beyond 20 days, especially for strong MJO cases (Johnson et al., 2014; Rodney et al., 2013; Yao et al., 2011).
The stratosphere is a potential source of S2S predictability because of its persistent and slowly varying circulation anomalies (NRC, 2010b). In boreal winter, such persistent circulation anomalies in the lowest part of the stratosphere interact vigorously with the upper troposphere and influence prediction of tropospheric circulation (Baldwin et al., 2003; Gerber et al., 2012). The QBO is an easterly-to-westerly reversal of tropical stratospheric winds driven by stratospheric waves originating from the troposphere. The QBO has a mean period of about 28 months, and its phase is predictable a few years ahead. The QBO influences the strength of the mid-to-high latitude westerly winds in the stratosphere, or the polar vortex. Prominent strengthening or weakening of the wintertime stratospheric polar vortex tends to be followed, with a lag of about 1 month, by similar variations in the large-scale tropospheric circulation patterns known as the annular modes (Baldwin and Dunkerton, 2001; Thompson and Wallace,
2000). Variability in the annular modes has in turn been associated with episodes of extratropical surface air temperature anomalies (warm spells or cold surges) and sea ice anomalies (Rigor et al., 2002; Thompson et al., 2002) (see below).
Extratropical weather is frequently dominated by recurring circulation patterns, often referred to as weather regimes or extratropical modes of variability. Because of their large-scale and low-frequency nature, these circulation patterns can contribute to atmospheric S2S predictability. For example, it has long been recognized that the PNA pattern (PNA), has a significant impact on the North American surface air temperature and precipitation (e.g., Wallace and Gutzler, 1981). Although the state of the PNA and its predictability on S2S timescales is influenced by ENSO (Zhang et al., 1997) and MJO (Mori and Watanabe, 2008) variability, it is unclear how interactions between these coupled modes and/or additional drivers may influence PNA variability and its associated weather patterns.
The North Atlantic Oscillation (NAO) is another major circulation pattern that influences weather from eastern North America to Europe, and it is highly correlated with the Northern Annular Mode (NAM). The NAO/NAM exhibits predictability on S2S timescales because its variability is linked to other components of the Earth system that are more predictable, such as the stratospheric polar vortex. Observational studies also show a robust lagged connection between the MJO and NAO (Cassou, 2008; Lin et al., 2009), and indeed a higher level of skill in predicting the NAO on a subseasonal timescale can be achieved when a strong MJO signal occurs in the initial condition (Lin et al., 2010). Similarly, skillful seasonal NAO predictions have been made by improving the initialization procedure to more realistically capture the initial state of the QBO and ocean and sea ice conditions (Scaife et al., 2014b). It follows that forecasts of the NAM have also been found to be skillful on a seasonal timescale and that this skill was improved through more realistic initialization (Riddle et al., 2013).
Understanding and correctly representing phenomena such as the NAO/NAM, the PNA and the Southern Annual Mode (SAM) in the Southern Hemisphere are additionally important for S2S predictions because their state can influence the development of strong and persistent anomalies in midlatitude atmospheric circulations that are sometimes caused by blocking events. Blocking can be exploited as a source of predictability (Hoskins and Woollings, 2015) and has been linked to high-impact weather such as severe cold spells in winter and droughts in summer. Variability in the NAO has been related to blocking episodes (Woollings et al., 2008) and, as the
NAO has proved to be more predictable than previously thought, so has blocking (Athanasiadis et al., 2014).
Future Directions on How Modes of Variability Influence Predictability
Natural modes of variability represent key sources of S2S predictability. Although much progress has been made in understanding these modes, in particular ENSO and MJO, less is known about how the interactions between coupled modes and slowly varying processes influence the development of specific environmental conditions. Continued research into variability in coupled modes, and their interactions across timescales, is necessary to fully exploit their predictability for S2S forecasting. Important questions that need to be addressed include How does the MJO influence rainfall over southeast Asia during El Niño vs. La Niña, or in different phases of the IOD? How do tropical Kelvin and other atmospheric waves influence the initiation, amplitude, or decay of the MJO? Under what conditions can the various modes of tropical variability ensure the high or low occurrence probability of tropical cyclones in a given region? These sorts of investigations fall under Facet I in Figure 4.3. Moreover, as correlations between these modes and impactful weather/climate are discovered, it is essential that our models can re-create such variability and its impact in simulations and retrospective forecast experiments (i.e., Facet II and III in Figure 4.3). For example, some models do quite well at representing intraseasonal variability in the eastern Pacific—which has a strong impact on tropical cyclones in that area—while others perform relatively poorly (Jiang et al., 2012a, 2012b). Many such examples are evident from the literature where empirical analysis has indicated potential relationships that can be exploited for S2S predictability, yet models still struggle to represent the variability and relationships correctly. These include IOD and boreal summer monsoon interactions (Ajayamohan et al., 2009), Kelvin wave and MJO interactions (Guo et al., 2015), and many others. Of particular challenge are those modes of variability that stem from coupled processes, including ENSO, but that could also include land-atmosphere or cryosphere-atmosphere coupling.
Finally, as our models become more capable of representing these processes, it is critical to carry out the predictability experimentation described above and highlighted as Facet IV in Figure 4.3. Such experimentation can point to forecasting system research and development avenues that would yield the greatest benefits and help to identify and/or characterize forecasts of opportunity based on specific modes of variability being in a particular phase (e.g., when the MJO is in phase 4-5, there is a strong enhancement of tropical cyclones to the west of the maritime continent and a suppression of them to the east).
Finding 4.2: Natural modes of variability represent key sources of S2S predictability, and it is essential that S2S models accurately represent them. Further research is needed especially to understand the interaction of natural modes across timescales, associated impacts on teleconnection patterns, and the formation of extreme environmental conditions. Long and sustained observational records are essential for such research.
Predictability from Slowly Varying Processes
As discussed briefly above, S2S predictability can stem from persistence in the initial state of various components of the Earth system. For example, anomalous conditions in the stratosphere or ocean can persist for several months owing to their vertical stability and slowly overturning circulation. In addition, persistence in anomalous environmental conditions often stems from storage of anomalous energy, typically in the form of heat or water in a given phase, such as in snow, sea ice, soil moisture, or ocean heat content. For example, the heat capacity of the entire atmosphere column is about the same as just the top 2.5 m of the ocean, and the melting of a global 25 cm shell of ice would take as much energy as warming the entire atmosphere by 10°C. When these anomalous stores of heat occur on large spatial scales (e.g., greater than ~1,000s of kilometers), their evolution/dissipation typically occurs on timescales of several days, weeks, or months and thus provides predictability to the Earth system. Smaller anomalies may also provide predictability for important ocean and coastal properties that are of interest to predict in their own right. Similarly, anomalies in momentum (e.g., ocean currents or atmospheric circulation patterns), aerosols and chemical species, and phytoplankton can also instill slow and anomalous variations on the coupled Earth system, impacting the ability to make skillful S2S forecasts.
Sometimes persistence is used as a threshold for predictive skill, which does not preclude considering persistence as a source of predictability. By the committee’s definition, a phenomenon can exhibit predictability even if it can be predicted with an idealized model or theoretical means. Furthermore, the threshold for predictive skill must itself have a source of predictability. Even so, the probability of a phenomenon occurring due to persistence in a system with many interacting processes may not be possible to predict with an idealized model or theoretical means, and may require a predictive system, even though the mechanism for predictability at some level appears basic and might not be considered “dynamical.” Additional details of a number of slowly varying processes within the coupled Earth system that provide predictability on S2S timescales are provided below.
Given the ocean’s relatively larger heat capacity compared to other components of the Earth system and the persistence of its temperature, salinity, and currents, the ocean represents a key source of predictability on S2S timescales in a number of ways (NRC, 2010b). Here we focus on mechanisms involving the ocean surface conditions owing to their relevance for humans (e.g., fisheries, harmful algal blooms, controls on the atmosphere and sea ice), rather than those that primarily affect the deep ocean. These mechanisms include large- and small-scale ocean dynamics in the tropics (e.g., Alexander, 1992) and the extratropics (e.g., Hartmann, 2015), as well as ocean interactions with the atmosphere and sea ice through surface exchange of energy, moisture, and momentum, yielding both one-way influences and coupled feedbacks.
The persistence of surface anomalies depends primarily on the depth of the upper ocean mixed layer. Other secondary factors include the net surface energy and freshwater fluxes, upwelling rates (via Ekman pumping and entrainment), and the properties of upwelling subsurface waters. Anomalous upwelling driven by persistent wind regimes associated with atmospheric modes of variability can lead to predictable anomalous surface conditions because subsurface waters generally also have longer-lived properties, including concentrations of nutrients that can drive biological productivity (e.g., Waliser et al., 2005). Subsurface anomalies may even lie “dormant” (unrelated to the mixed-layer properties) until one or more storms with high winds mixes the upper ocean, transporting the anomaly vertically to the surface (Alexander et al., 1999).
Small-scale (10s-100s of kilometers) surface ocean features, such as circular motions known as eddies and regions of strong gradients known as fronts, can also exhibit persistence for months to years (Chelton and Xie, 2010; Chelton et al., 2004. These small-scale variations in sea surface temperature (SST) cause divergence and convergence in the surface wind and vertical motions that link the small-scale ocean features to cloud properties and other atmospheric features (e.g., Chelton and Xie, 2010). Ocean eddies also have an association with ocean biogeochemistry through their influence on upwelling or downwelling, horizontal advection, and isolation of nutrients and ecosystems (Gaube et al., 2014). Because of their persistence and coupling with the atmosphere (20 percent of the heat flux between the atmosphere and ocean is related to the ocean eddy field [Boas et al., 2015]), these eddies represent a potential source of S2S predictability for the ocean and even the entire Earth system if feedbacks to the atmosphere are prominent.
Soil Moisture and Vegetation
Soils have the capacity to hold substantial amounts of water relatively close to the surface (e.g., centimeters to meters), depending on soil texture, structure, and vegetation. This water-holding capacity lends predictability to the atmosphere for several weeks or months by influencing surface energy budgets (e.g., heat and moisture fluxes to the atmosphere) (Koster et al., 2010; NRC, 2010a). For example, given soil moisture’s influence on heat flux, the number of hot days over land in many regions has been found to correlate highly with precipitation summed over a preceding period (Figure 4.6 and Mueller and Seneviratne, 2012). Soil moisture’s influence on surface temperature is coupled with a direct impact on the surface moisture flux to the lower atmosphere, which together influence subsequent precipitation anomalies (e.g., Guo et al., 2012; Koster et al., 2011; Roundy et al., 2014). It follows that soil mois-
ture is also strongly associated with drought predictability (e.g., Kumar et al., 2014; Roundy and Wood, 2015; Thomas et al., 2015). This predictability may be especially pronounced during boreal spring and summer, when coupled Earth system models often exhibit lower predictive skill due to weaker links between midlatitude climate systems and the oceans and an increase in land-atmosphere interactions. Along with its coupling to atmospheric conditions, the slow variations of soil moisture are also important for predicting quantities such as runoff to rivers, lakes, and the ocean, as well as plant growth—and thus land cover, albedo, and flood potential. For example, there is increasing evidence that vegetation states and anomalies can be sources of weather and climate predictability on S2S timescales (Koster and Walker, 2015).
Snow also contributes to predictability of atmospheric and land conditions due to its storage of surface water and its influence on surface energy budgets. The latter occurs because of its high albedo relative to snow-free areas; it acts as a significant surface heat sink via the latent heat required to melt the snow, and in changing the interface conditions it influences the fluxes of heat and moisture between the land and atmosphere. Knowledge of anomalous snow conditions, particularly the snow water equivalent as opposed to just snow cover, can improve forecasts of air temperature and humidity, runoff, and soil moisture during the winter and spring seasons (Jeong et al., 2013; NRC, 2010b; Peings et al., 2011; Thomas et al., 2015). For large-scale anomalies in snow conditions, there is also some evidence that snow can influence remote atmospheric conditions by altering large-scale atmospheric circulation features (e.g., Rossby waves) (also see section below on Sea Ice and Polar Land Surface). For example, correlations have been documented between autumn anomalies in Eurasian snow and the large-scale Northern Hemisphere atmospheric circulation a few weeks to months later through the influence of snow cover on the vertical propagation of wave energy into the stratosphere and the NAO (Brands et al., 2012; Fletcher et al., 2007; Orsolini et al., 2013, 2015). Snow cover and snow water can have a profound influence on the evolution of the local, regional, and even large-scale weather patterns as well as a number of Earth system components. This influence places a high priority on ensuring observations of snow are available for process understanding and forecast initialization (e.g., Orsolini et al., 2013) and that our terrestrial hydrology and atmospheric models properly represent snow and related processes (see Chapter 5).
Sea Ice and Polar Land Surface
Sea ice lends predictability to the Earth system because its presence strongly reduces heat and moisture fluxes from the ocean to the atmosphere, it serves as a significant reservoir of freshwater within the upper ocean, and it is an excellent reflector of solar radiation. The persistence of sea ice anomalies has several important timescales (Figure 4.7). There is an initial persistence of anomalies in the sea ice cover that varies from 2-4 months (Lemke et al., 1980), depending on the season (Blanchard-Wrigglesworth et al., 2011a; Day et al., 2014) and location (Bushuk and Giannakis, 2015). After this initial period of persistence, there is a reemergence that occurs in some seasons owing to sea ice internal dynamic and coupled interactions between sea ice and SST. Modeling studies suggest anomalies of sea ice thickness are far more persistent and about as important as SST in controlling the persistence characteristics of the sea ice cover (Bitz et al., 1996; Blanchard-Wrigglesworth and Bitz, 2014; Blanchard-Wrigglesworth et al., 2011b; Chevallier and Salas-Melia, 2012; Holland et al., 2013; Lindsay et al., 2008). The lack of long-term sea ice thickness measurements forces researchers and forecasters to turn to models to estimate these quantities. When models factor in transport, sea ice thickness anomalies can persist for almost two years and exhibit typical length scales of about 500-1,000 km (Blanchard-Wrigglesworth and Bitz, 2014).
Through coupling to the atmosphere, the presence and persistence of sea ice affects the trajectories of atmospheric storms and ocean circulation (Balmaseda et al., 2010; Bitz et al., 2006; Screen et al., 2011) and has considerable impacts on coastal erosion (Barnhart et al., 2014), marine and terrestrial biology (Post et al., 2013), and shipping (Khon et al., 2010). Researchers are actively exploring the extent to which sea ice anomalies and polar conditions in general can influence the lower latitudes, with longer lasting cold air outbreaks in years with an anomalously warm Arctic surface as one possibility (Francis and Vavrus, 2012). A proposed mechanism stems from polar controls on atmospheric meridional temperature gradients and the subsequent coupled interactions among temperature gradients, the midlatitude jet stream, and storms. However, the multitude of interactions involving the midlatitude jets has made it difficult to find conclusive evidence of Arctic-midlatitude weather linkages (Figure 4.8; Cohen et al., 2014). Although the mechanisms remain obscure, when global forecast models include more realistic Arctic sea ice and other Arctic variables, forecasts improve in lower latitudes (Jung et al., 2014; Scaife et al., 2014a). Because of the persistence of sea ice and arctic snow cover, it is important to improve our understanding of sea ice and related processes and the mechanisms linking Arctic and midlatitude conditions, as well as to incorporate these processes and mechanisms into models used for S2S predictions.
Sudden Stratospheric Warmings
Occasions of rapid slowdown of the stratospheric Arctic vortex are usually accompanied by sudden stratospheric warmings (SSWs) and a subsequent negative phase of the NAM. However, experimentation with models that have adequate resolution in the stratosphere to capture the relevant dynamics is a relatively new endeavor. Recent studies show that SSWs can be predicted only 1 or 2 weeks in advance (Gerber et al., 2009; Marshall and Scaife, 2010). Yet for several months following an SSW, enhanced forecast skill has been found in extratropical surface temperatures and sea level pressure (Sigmond et al., 2013). More recent work has found multiscale/mode interactions between the MJO, SSWs, and the QBO (Liu et al., 2014).
Finding 4.3: A number of slowly varying processes impart predictability to the Earth system in the S2S time range, including processes and interactions related to sea ice, the thermal and dynamic evolution of the upper ocean, and soil moisture, surface water, snow and vegetation on the land surface.
Finding 4.4: It is essential to maintain and increase observations of the slowly varying components of the Earth system relevant to S2S (e.g., snow, soil moisture, sea ice, and near-surface ocean) in order to improve process understanding, advance model development, and improve the initial conditions of the forecast system. Further studies are needed to understand the relative importance of these components as sources of S2S predictability.
Predictability from External Forcing
Variability on S2S timescales may also be driven by external forcing, such as from anomalies in solar forcing, anthropogenic emissions of pollution or aerosols, or the episodic input of aerosols from volcanoes. Advanced knowledge of changes in the radiative, thermal, biogeochemical, or hydrological forcing of some part of the Earth system can lead to skillful predictions of other quantities of interest such as surface temperature or precipitation. Two leading cycles in the Earth system, the diurnal cycle and seasonal cycle, are very predictable precisely because they are driven by highly repeating patterns in the incoming shortwave radiation at the top of the atmosphere. Although these timescales are very relevant to S2S variations, particularly when considering interactions across scales, it has not been a trivial matter to represent their impacts in global weather/climate forecast models.
Over the past decade or so, numerical weather and climate models have started to be able to better reproduce credible seasonal variability through careful representation of the relevant processes. However, significant shortcomings remain in representing the effects from these very well defined external forcings that are highly relevant to S2S prediction, such as the diurnal cycle. For example, the diurnal cycle over the maritime continent could influence the MJO as it propagates eastward from the Indian Ocean into the western Pacific Ocean. Observations exhibit a relative minimum in the MJO-driven subseasonal variability over the maritime continent, possibly because of the relatively stronger diurnal cycle in this region relative to the open ocean to the east and west. Representing this scale interaction in models has been challenging and has represented a barrier to producing accurate forecasts of MJO amplitude and propagation in this region (e.g., Weaver et al., 2011).
Aerosol variability in a number of forms holds the potential to influence variability on S2S timescales and represents an important source of S2S predictability in some cases.2 A volcanic eruption has the potential to loft significant ash and dust into the troposphere and stratosphere, which can result in substantial anomalies in incoming solar radiation and outgoing infrared radiation. Depending on the magnitude of the mass injection and its altitude, the anomalous aerosol forcing can last for days to a couple of weeks in the troposphere and months to a year in the slow, stable circulation of the stratosphere. Accurate representation of the aerosol content, types, and interactions with clouds and radiation provides a potential source of predictability. Demand for realizing this forecast potential stems from the needs to better represent and forecast its influence on weather and short-term climate as well as to better understand and predict the lifetime of the aerosol anomaly itself and its societal impacts (e.g., how long will the ash plume last, and will it affect air traffic?).
In some cases, ash, dust, and other aerosols can influence the Earth system even after they are removed from the atmosphere, most notably when they are deposited on ice or snowpack. In this case, they can have a substantial influence on the subsequent evolution of the surface, producing considerably faster melting than would otherwise be the case. This has both hydrological implications (Qian et al., 2009) via the change in the runoff and implications for the evolution of the snow pack and the manner it influences weather and short-term climate (see section above on Slowly Varying Processes). Aerosols can also impart predictability on near-surface ocean biology by providing input of key nutrients, namely iron, that can facilitate the development of widespread phytoplankton blooms (Langmann et al., 2010), which have life cycles of days to weeks. Such blooms influence the vertical profile of solar absorption in the upper ocean, typically leading to greater warming of SST and a more stable surface mixed-layer than would otherwise occur (Siegel et al., 1995). The latter can have considerable implications for large-scale variations and spatial structure of SST anomalies, which in turn can influence weather and short-term climate.
Although the lifetimes of other atmospheric constituents can be much longer, it is still critical that they be accounted for in S2S forecast systems. Notable examples of these are the concentration of anthropogenic greenhouse gases (GHGs, e.g., carbon dioxide, methane). The typical lifetime of anthropogenic GHG anomalies is on the order of a decade to centuries, and fluctuations and trends in the emissions of GHGs also tend to occur on timescales that are long relative to the S2S forecast. These long timescales
2 Aerosols also play a key role in cloud formation and the development of precipitation. Accurate understanding and modeling of this process are critical to producing high-fidelity models of the atmosphere for nearly all forecast timescales. Given its place as a key physical process, rather than a source of predictability, aerosol-cloud interaction is treated in Chapter 5.
imply that, for a given forecast, the GHG concentration can be specified to be a constant. However, because multi-decade retrospective forecast data sets are a crucial component of an S2S forecast system for bias correction (see Chapter 5), it is imperative that the values of impactful constituents be specified to the forecast system as varying boundary conditions over the time period of the retrospective forecasts. This type of slowly varying forced signal can lead to systematic shifts in the probability distributions of variables (e.g., temperature and precipitation) that can be predicted given the known value of the forcing. Furthermore, such external forcing has caused the seasonal minimum of Arctic sea ice extent to decline by greater than 40 percent, radically changing the probability of where the sea ice edge lies at the end of summer in recent years compared to the beginning of the satellite record in 1979. As S2S forecast systems encompass more Earth system components and coupled processes that are influenced by such external forcing, it is important to have an accurate representation of GHG forcing and other slowly varying external forcing (e.g., solar constant, surface albedo).
Finding 4.5: Given the requirement that S2S forecasts have multi-decade retrospective forecast data sets for the purposes of bias correction, is it imperative that the model forecast system account for all slowly varying external forcings that influence the frequency, spatial distributions, and temporal distributions of S2S forecast quantities (e.g., temperature, precipitation). Such external forcing includes the influences from natural and anthropogenic aerosol emissions, GHG concentrations, variations in the solar constant, and surface albedo, where the latter may derive from snow/ice cover or land use/land cover changes.
Prediction of Disruptive and Extreme Events and “Forecasts of Opportunity”
A strong motivation for developing and improving S2S forecasts is to provide guidance on the likelihood, magnitude, and impacts of disruptive events (see also Chapter 3), which could be severe rain or wind storms (e.g., tropical or extratropical cyclones, large mesoscale convective systems, tornado outbreaks), Santa Ana or Chinook wind conditions, severe rain or snow events, drought, prolonged cold surge or heat wave conditions, and other events. These types of events may exhibit predictability, but usually only when they are associated with other phenomena that are predictable, such as the MJO and ENSO (e.g., Pepler et al., 2015). For example, Figure 4.9 illustrates the impacts, typically felt through one or a series of extreme weather events, from El Niño conditions (or “warm episodes”). Similarly, Figure 4.10 illustrates the impact of the MJO on the frequency and spatial variability of tropical cyclones.
Indeed ENSO and the MJO have remarkable impacts on the modulation of the frequency, spatial distribution, and types of extreme environmental events that occur in a number of regions around the globe. Other sources of variability discussed above can similarly impact the occurrence of extreme or disruptive events. Particularly important for disruptive events is the additive effect of the various sources of predictability. As a simple consideration of this effect, both the cool phase of ENSO (i.e., La Niña) and phases 4-6 of the MJO increase the likelihood of precipitating conditions over the Maritime continent region. These two phenomena then work in concert to facilitate
the development of more frequent, longer, and/or more severe precipitating events, one operating on a timescale of months (i.e., ENSO) and the other weeks (i.e., MJO; also see Figure 4.1). Similarly, an El Niño condition along with MJO phases 8 and 1 will produce subsidence and thus dry conditions over the same region. On top of modes of variability, other processes lending predictability can act on the given anomaly to further exacerbate the condition. For example, wet conditions will produce positive soil moisture anomalies, which can in turn positively influence the further development of precipitation in the given region. Such timescale and process interactions, in terms of their additive or in some cases counteracting influences, can occur in a number of places around the globe depending on the phenomena, region, and season. These multiscale interactions of an inherent, albeit intermittent, source of S2S predictability, represent “forecasts of opportunity”—a foundational consideration in S2S forecasting. Better understanding these interactions will make it possible to develop more forecasts of opportunity, for example, forecasts that take advantage of windows of time in which higher predictability is possible. This will be particularly important for the prediction of events that are of interest to decision-makers.
Finding 4.6: The nature of sources of S2S predictability, namely intermittent natural modes of variability, wide and often disaggregated variations in anomalous conditions in a number of slowly varying processes/quantities, and varied natural and anthropogenic external forcings, liken the S2S prediction challenge to the identification and successful prediction of a series of“forecasts of opportunity.” Identifying such windows of predictability will be particularly important for forecasts of extreme and disruptive events.
The relative value of predictability sources is dependent on location of the forecast and time of the year. Although some processes have a stronger local impact, others influence the climate through teleconnections and have a far-reaching effect. For example, initial anomalies in soil moisture can influence the local forecast precipitation and surface air temperature through changes in surface energy budget associated with evaporation. Anomalies of tropical convection associated with ENSO and the MJO influence the midlatitude climate through teleconnections related to Rossby wave energy propagation, and thus a large impact is usually observed along the path of Rossby wave train. In the Northern Hemisphere extratropics, the wintertime westerlies provide a more favorable background for Rossby wave propagation than in summer, thus the teleconnection contribution is stronger in winter. On the other
hand, the influence of soil moisture becomes relatively more important in summer than in winter.
Our understanding of the source of S2S predictability is still lacking. The relative value of predictability sources has not yet been established. The approaches that have been used in predictability study may not be appropriate to separate the contributions of different sources. For example, the specification of soil moisture in the initial condition in the retrospective forecast experiment that is designed to identify the contribution of soil moisture may contain information of ENSO or other sources of predictability. In the relaxation experiment that is designed to identify the origin of skill source, the analysis fields that are relaxed in a given region may already contain variability propagating from other regions. The combined effect of several different sources of predictability may not be a simple sum of individual processes. More studies are needed to understand how different sources of predictability interact.
Climate models that are used for retrospective forecasts are imperfect, and different models have different model errors, leading to inaccurate, incomplete, and model-dependent estimates of signal and noise variability in ANOVA analysis, as well as false representation of the “truth” in the twin experiments used to estimate the upper predictability limit. The assessment of impact of a particular process on forecast skill is also model-dependent. Encouraging future studies to use a multi-model framework would help to reduce the uncertainty related to model configurations. In addition, innovative methodologies to estimate predictability need to be explored in order to better understand the nature of S2S forecast.
It is essential to maintain and increase observational records for different components of the Earth system. These observations can be used to explore new sources of predictability and to better initialize S2S models. It is important for S2S models to capture the natural modes of variability, slow processes, and externally forced trend and variability.
Recommendation C: Identify and characterize sources of S2S predictability, including natural modes of variability (e.g., ENSO, MJO, QBO), slowly varying processes (e.g., sea ice, soil moisture, and ocean eddies), and external forcing (e.g., aerosols), and correctly represent these sources of predictability, including their interactions, in S2S forecast systems.
- Use long-record and process-level observations and a hierarchy of models (e.g., theory, idealized models, high-resolution models, global Earth system models) to explore and characterize the physical nature of sources of predict-
ability and their interdependencies and dependencies on the background environment and external forcing.
- Conduct comparable predictability and skill estimation studies and assess the relative importance of different sources of predictability and their interactions, using long-term observations and multi-model approaches (such as the World Meteorological Organization-led S2S Project’s database of retrospective forecast data).
Decision-makers are particularly interested in guidance on the likelihood, magnitude, and impacts of disruptive events (see also Chapter 3). Prediction of these types of events will rely on identifying multiscale interactions of inherent, albeit intermittent sources of S2S predictability. Thus prediction of such features will require developing “forecasts of opportunity”—a foundational consideration in S2S prediction. Although any given extreme event (e.g., storm) is typically not predictable more than a few days in advance, understanding interactions between sources of S2S predictability offer the means to infer changes in the likelihoods of extreme events—including their spatial distribution, occurrence frequency, magnitude, and type. More specifically, this means that an accurate S2S forecast system will provide quantitative forecast information—likelihoods and uncertainties—on extreme events with lead times from weeks to months. However, it is critical that all of the important and impactful phenomena be represented faithfully in order to yield accurate forecasts. For example, based on the discussion above, if the ENSO modulation is accurately depicted by the forecast system but there are temporal, spatial, and/or amplitude biases in the representation of the MJO, then the forecast accuracy of precipitation amounts and extreme events will be heavily compromised.
In summary, accurate prediction of extreme weather/environmental events hinges critically on the accurate representation of all of the dominant modes of variability and slowly varying processes that operate and yield predictability on S2S timescales. Forecast models must represent these processes individually as well as collectively, with specific attention to their multiscale interactions and influences on the development of extreme events. Thus all four facets of predictability research highlighted in Figure 4.3 need to be undertaken to improve the prediction of disruptive, high-impact, or extreme events.
Recommendation D: Focus predictability studies, process exploration, model development, and forecast skill advancements on high-impact S2S“forecasts of opportunity” that in particular target disruptive and extreme events.
- Determine how predictability sources (e.g., natural modes of variability, slowly varying processes, external forcing) and their multiscale interactions can influence the occurrence, evolution, and amplitude of extreme and disruptive events using long-record and process-level observations.
- Ensure the relationships between disruptive and extreme weather/environmental events—or their proxies—and sources of S2S predictability (e.g., modes of natural variability and slowly varying processes) are represented in S2S forecast systems.
- Investigate and estimate the predictability and prediction skill of disruptive and extreme events through utilization and further development of forecast and retrospective forecast databases, such as those from the S2S Project and the North American Multi-Model Ensemble (NMME).
This page intentionally left blank.