Table of contents (30 pp limit, approved by nsf on 19 October 2012) – 1 pg introduction
Download 250.48 Kb.
|
- Navigate this page:
- GLACE-2 Prediction Skill from Land Initialization
- Rebound of Predictability over US Great Plains COLA has discovered a new mechanism of overlooked land-driven predictability, which is evident over
- Severe Drought Likelihood in a Changing Climate
- Large areas of North America, Europe, central Asia, South America and Africa show an increase in the return frequency of severe drought by two to four times
- Decadal Predictability and Prediction Figure 2.5.2
- Scientific Basis for Decadal Prediction COLA has established a scientific basis for decadal predictions
- 2.5.3 COLA’s Contribution to CMIP5 Decadal Predictions
- 3. PROPOSED WORK – 15 pp
Land-Climate Interactions The importance of land surface variability in modulating and contributing to climate predictability on intra-seasonal and longer time scales has long been a central theme of COLA research. In particular, the role that soil moisture anomalies play in enhancing predictability has been the focus of a great many numerical experiments that COLA has helped lead, including the multi-model Global Land-Atmosphere Circulation Experiment (GLACE; Koster et al. 2006) that evaluated the sensitivity of climate to soil moisture anomalies, and the second generation of GLACE (GLACE-2; Koster et al. 2010) that examined the impact of antecedent soil moisture anomalies on climate prediction skill.
Realistic initialization of soil moisture in ten different weather and climate models is found to increase prediction skill in hindcasts over 1986-1995. 10-member ensembles of 60-day forecasts with start dates ranging from early April to mid-August are compiled in Fig. 2.4.1 for all validation periods falling within JJA. Skill impacts over North America is defined as the change in temporal correlation with observed anomalies at each point over land between realistic and random land surface initialization. Dots denote grid boxes significant at the 95% confidence level, and each row is for forecasts with differing lead times. Skill improvements are greater for near-surface air temperature than for precipitation, but for both variables, improvements are more marked when only the cases with extreme initial soil moisture anomalies are considered.
COLA has discovered a new mechanism of overlooked land-driven predictability, which is evident over the central United States during the transition from spring to summer. Figure 2.4.2 shows 10 sets of 60-day forecasts from the GLACE2 project initialized at roughly 15-day intervals starting in April continuing until early October. Averages are shown for lead-days 1-15, 16-30… 46-60 of the forecasts. The dotted lines show predictability when initial soil moisture is randomized among the 10 ensemble members, solid lines when the initial soil moisture is identical and realistic, taken from an offline land surface model analysis. It is evident from the dotted lines that when a coupled land-atmosphere model is initialized, the atmospheric predictability decays rapidly in the first two weeks owing to chaos. However, land predictability, in the form of soil moisture anomalies, remains high owing to its long memory. Yet this high predictability in the land does not influence the atmosphere in early spring because the atmosphere and land do not communicate well during this time - surface fluxes are not controlled by the soil moisture anomalies because the region is in a radiation-limited rather than a moisture-limited regime. With the approach of summer, the region changes regimes and the high predictability in the land is "transferred" to the atmosphere, giving an apparent "rebound" in predictability. Several forecasts for both precipitation and temperature show an increase in predictability with time that cannot be attributed to sampling error (arrows). System-wide predictability cannot increase with lead-time – this rebound in predictability occurs after the land-atmosphere coupling, i.e. the positive feedback in the hydrologic cycle, becomes established, allowing the atmosphere to respond latently to soil moisture anomalies that it was disconnected from earlier in the year.  Figure 2.4.2 Average predictability (signal-to-total ratio) of hindcasts done with realistic soil moisture initial conditions (solid lines) and random soil moisture initial conditions (dotted lines). Hindcasts, each having 10 members, initialized at 5-day intervals in April (4th month) through September (9th month) are shown, with averages for lead times 1-15, 16-30, …, 46-60 shown in each curve.
A very high resolution (16km) multi-decade climate simulation with a global atmospheric model with observed SST for 1961-2007, and “time-slice” experiment with a mean annual cycle of A1B SST change superimposed over the same SST to simulate 2071-2117 conditions, were performed with the operational ECMWF weather forecast system (as part of Project Athena – see Section 3.7). The current drought threshold was defined from the first integration as the seasonal JJA precipitation from the 5th driest year of the 47 simulated. The likelihood of future severe drought was then defined as the number of years out of the 47 in the late 21st century simulation where the rainfall fell short of the current threshold (Fig. 2.4.3). Large areas of North America, Europe, central Asia, South America and Africa show an increase in the return frequency of severe drought by two to four times. Many other regions such as much of India, Indonesia, and Australia, show a smaller increase in likelihood, whereas most high-latitude regions of the Northern Hemisphere become wetter (Dirmeyer et al. 2012). The regions of increased drought occurrence  Figure 2.4.3 Number of years out of 47 in a simulation of future climate (2071-2117) for which the June-August mean rainfall was less than the 5th driest year of 47 in a simulation of current climate (1961-2007).
are concentrated in most of the productive agricultural areas, suggesting a very worrisome trend for future food production.
 Figure 2.5.2: The squared autocorrelation function of the most predictable component shown in Fig. 2.2, but for each model control run as a function of time lag (in years). The 5% significance level of the autocorrelation, for a sample size of 300, is indicated by the thick horizontal dashed line.
Scientific Basis for Decadal Prediction COLA has established a scientific basis for decadal predictions by explicitly identifying patterns in climate models that are predictable on decadal time scales and showing that these  Figure 2.5.1: (Top) The component that maximizes the average predictability time of sea surface temperature in 14 climate models run with fixed forcing (i.e., “control runs”). The left panel shows the spatial structure of the component. This component is called the Internal Multi-decadal Pattern, or IMP. Ocean points with no shading indicate regions that were omitted from the maximization (i.e., “masked out”) because of insufficient data in the corresponding observational data set. (Bottom) Time series of this component in three representative control runs.
structures also are predictable in the observed climate system. In particular, DelSole et al. (2011a) identified a pattern that maximized the average predictability time in multiple climate simulations of the CMIP3 data set. The simulations were from pre-industrial control runs with no interannual variations in climate forcing, ensuring that the multidecadal variability found in the models was generated internally by the climate system and occurs irrespective of anthropogenic or natural forcing. This component, shown in Fig. 2.5.1, has maximum loadings in the North Atlantic and North Pacific and varies on multidecadal time scales in many models. DelSole et al. (2011a) used optimal fingerprinting techniques to separate this component from the climate change signal and found the resulting amplitude to be strongly correlated with the observed Atlantic Multi-decadal Oscillation (AMO) index. This implies that unforced climate models naturally simulate a multi-decadal component very similar to the AMO. Furthermore, the squared autocorrelation of the time series of this pattern for each model, shown in Fig. 2.5.2, reveals that the predictable component has statistically significant correlations for at least ten years in some models, thereby providing a scientific basis for decadal prediction.  Figure 2.5.3 Spatial patterns of the leading predictable component of annual mean surface air temperature (top left) and precipitation (top right) over North America derived from eight CMIP3 control simulations. Also shown are the squared correlation skills of predictions of these components by a regression model trained in control simulations and verified in independent control simulations. The spatial patterns are normalized such that the absolute value at each grid point gives the standard deviation of the component at that point, in units of degree Kelvin and mm/day (after Jia and DelSole 2011)
COLA also has identified patterns over continents that are predictable beyond seasonal time scales. Jia and DelSole (2011) identified these patterns in CMIP3 control runs by generalizing the statistical optimization procedure of DelSole and Tippett (2009) to maximize the average predictability time of land variables using SST predictors. As illustration, the most predictable pattern for annual mean surface air temperature and precipitation over North America is shown in Fig. 2.5.3. The corresponding squared correlation skill, shown in the bottom panels of Fig. 2.5.4 reveal statistically significant skill as long as seven years for temperature, and one year for precipitation. Additional analysis of six different continental regions reveals predictability for 1-3 years for precipitation and 3-7 years for surface air temperature, depending on model. These results provide a scientific rationale for regional prediction on multi-year time scales. It is well established that land predictability on seasonal time scales arises from both SST and land-atmosphere feedbacks. Thus, dynamical models need to capture these relations in order to predict land variables. Figure 2.5.4 shows the regression of the AMV index (defined in the figure caption) against the SST for years 6-10 of the CFSv2 CMIP5 decadal forecasts produced by COLA and as seen in the corresponding segments of the NCEP/NCAR reanalysis. The CFSv2 results are ten 3-member ensembles of CFSv2 retrospective decadal forecasts performed by COLA using NEMOVAR ocean initial conditions (complementary to runs made by NCEP using CFS-R ocean initial  Figure 2.5.5 The skill of predicting observed annual mean SST during 1910-2004 using an estimate of the forced component derived from 35 20C simulations from eight CMIP5 models (red line), and using this estimate of the forced response plus a multivariate regression model to predict unforced variability (black curve). The black horizontal dashed line indicates zero skill.
  Figure 2.5.4 Regression of surface air temperature against the Atlantic Multidecadal Variability index (defined as area averaged surface air temperature over the ocean in the boxed region in the upper figure. Top: NCEP reanalysis 1960-2011; bottom: years 6-10 of the 1960-present core CMIP5 decadal retrospective forecasts made by COLA using CFSv2 and the NEMOvar ocean initial conditions.
conditions). Since the AMV forecasts are not more skillful than persistence at this time range, the intent of this comparison is to assess the realism of the model's internal dynamics and response to external forcing. The patterns are remarkably similar, with the familiar AMV single-signed horseshoe pattern in the North Atlantic. Interestingly, the model reproduces other details of the observed regression, including a large positive signal over North America and a weaker negative signal over the southwest US. There is also an association with positive El Niño-like SST anomalies in the tropical Pacific extending across central America and northern South America, with negative surface temperature anomalies in the North Pacific, northeast Asia, and northern Europe, and positive anomalies in northwest Africa and central Asia. The regression map also shows intriguing similarities over North America with the leading predictable component derived from CMIP3 models shown in Fig. 2.5.3. One implication of this result is that, if the AMV could be forecast on decadal time scales, it would be reasonable to expect that decadal forecasts over many land regions, and especially North America would be potentially useful. 2.5.2 Demonstration of Decadal Predictability and Prediction Skill DelSole et al. (2012) propose a new forecast system that skillfully predicts sea surface temperatures on decadal time scales. The forecast system has two parts: (1) a prediction of the response to external forcing, derived from the ensemble mean of climate simulations driven with appropriate climate forcing, and (2) a prediction of internal variability based on a multivariate regression model, trained on pre-industrial control simulations. The advantage of deriving a regression model from control simulations is that a much larger sample size is available for multivariate estimation, the entire observational record is independent of the training data and hence provides genuinely independent verification data, and all questions about the validity of cross validation in decadal prediction are avoided. The skill of the forecast system in predicting the area average North Atlantic, North Pacific, and global sea surface temperature are shown in Fig. 2.5.5. The figure shows that adding the regression model predictions to an estimate of the response to anthropogenic and natural forcing yields a prediction with higher skill than either alone, demonstrating the contribution of initial condition information to skill on multi-year time scales. 2.5.3 COLA’s Contribution to CMIP5 Decadal Predictions As part of the CMIP5 decadal prediction project, COLA has completed a preliminary series of decadal retrospective forecast and forecast ensembles (both referred to as “forecasts” below) using NCEP’s CFS_v2 CGCM starting from November initial conditions, 1960-2008, both with (WV; initialized every five years: 1960, 1965, …) and without volcanic forcing (WOV; initialized every year). These are denoted COLA-CFS in the CMIP5 database. The runs were initialized using atmospheric and land data (1980 onward) from NCEP and ocean states from the NEMOVAR ocean reanalysis. The NEMOVAR reanalysis was used both because the NCEP ocean reanalysis does not extend back before 1980, and to complement similar experiments performed at NCEP using CFSR initial conditions. The WV cases are the same as the post-1980 NCEP decadal forecasts except for the ocean initial conditions. This extensive set of forecasts provides a baseline for evaluation of skill, as well as for testing not only the impact of the strategies, such as those discussed below, to improve the forecasts, but also sensitivity to the choice of ocean initial conditions. The major results thus far are: (1) the skill of the COLA-CFS results on time scales from interannual to decadal is comparable to that from the nine other models contributing to the CMIP5 decadal prediction project and is better than persistence for some quantities, such as Atlantic Multidecadal Variability (Schneider et al. 2012); (2) ENSO exhibits predictability for at least two years; and (3) the specification of the volcanic forcing adds artificial skill.  Figure 2.6.1 North Atlantic track density as a number density per season per unit area for (a) Observed, and IFS simulations at (b) 10-km, (c) 16-km, (d) 39-km and (e) 125-km for May-Nov 1990-2008.
High Spatial Resolution A key step in meeting the societal need for accurate prediction of regional climate variability and change is to take models of the climate system to a new level of capability in which salient features and processes of weather and climate are explicitly resolved (Shukla et al., 2009, 2010; Palmer 2011). Project Athena (Kinter et al. 2013), a project led by COLA in 2009-2011, addressed this question (see Section 3.6). As part of Project Athena, a series of 13-month runs was made with the ECMWF global numerical weather prediction model (Integrated Forecast System or IFS) at several different settings of horizontal resolution from 125-km to 10-km. Manganello et al. (2011) analyzed the statistics and average structure of tropical cyclones in those runs and found that increasing resolution leads to better representation of such storms, in terms of climatological frequency, formation regions, and intensity distribution. They found that storms in the higher resolution runs have, on average, more intense warm cores of quite realistic magnitudes, higher wind maxima, smaller and also more realistic maximum wind radii, and consequently stronger radial eyewall gradients of temperature and velocity (see Section 3.2.2.5). Furthermore, although low-resolution models can simulate the spatial distributions of TC genesis locations and tracks well, there is a systematic improvement of these features with increasing resolution, particularly in the North Atlantic (Fig. 2.6.1). TC activity strengthens in the main development region and grows in the subtropical Atlantic, reaching realistic levels in the 16-km simulation (panel c) although remaining well off the coast, as in the lower resolution runs. As the resolution is refined to 10 km (panel b), a distinct change takes place in the distribution of the track density: more storms propagate closer to the east coast of North America, in better agreement with observations. The 10-km model is also the only one that produces realistic seasonal mean TC frequency in the North Atlantic (not shown). Project Athena showed that, even with the hydrostatic approximation, realistic distribution and intensities of tropical cyclones can be achieved with a good global atmospheric model, provided that sufficient spatial resolution is used. 3. PROPOSED WORK – 15 pp Further research is suggested by a number of COLA’s findings. Among these are:
The subseasonal to seasonal time scale exemplifies the efficacy of the seamless approach and the requirements for better initialization and higher spatial resolution to improve predictions. Improved subseasonal to seasonal forecasts can be of great social and economic value, because decision-makers in agriculture and food security, water resources, disaster risk reduction and human health depend on advance information at these time scales (REFERENCE). Subseasonal climate prediction has received less attention than weather and seasonal prediction due to the relative minimum of predictability at that time scale found in the previous generations of models (REFERENCE). There are, however, several sources of potential predictability at the subseasonal scale that are revealed in models with better representations of whole atmosphere and coupled ocean-, sea ice- and land-atmosphere processes, and improved initialization of the land surface, oceans and upper atmosphere. For example, the recently improved representation of the Madden-Julian Oscillation (MJO) in models (REFERENCE) may have a big impact on forecast skill through the links to other modes of variability such as ENSO and the NAO. COLA’s findings on the rebound of land-driven predictability provide another example of how the representation of coupling mechanisms can significantly alter our ability to predict climate on the subseasonal to seasonal time scales (Guo et al. 2012). More research is needed to explore all the potential sources of predictability, with specific attention to the risk of extreme weather (tropical cyclones, droughts, floods, and heat waves) and the variations of monsoon precipitation. On the seasonal to interannual time scale, the principal source of predictability remains ENSO. Model representations of ENSO frequency, spatial pattern, and magnitude are all highly sensitive to model formulation (e.g. REFERENCE). The influence ENSO exerts on climate outside the tropical Pacific, notably the Indian monsoon and weather in both North and South America, remains a challenge for both simulation and prediction. Recent work suggests that substantial progress can be made in simulating ENSO through improved representation of the clouds and convection (e.g., Neale et al., XXXX; Stan et al. 2011). It remains to be seen if similar improvements can be realized in predictions of ENSO and its remote effects. Decadal climate variability has been of interest for some time (Mehta? REFERENCE), but it has only recently been a subject of prediction research. The CMIP5 experimental protocol (CMIP5 protocol REFERENCE) includes a series of decadal prediction experiments that have been undertaken by many modeling groups, with what must be acknowledged is only modest success. COLA has used these experiments, along with long “free” integrations of coupled climate system models, to show that there is a scientific basis for decadal prediction (DelSole et al. 2011a). However, the CMIP5 forecast experiments have uncovered large biases and model drifts, and on the whole have been rather disappointing. The one feature of the climate system that has been thought to be predictable at decadal time scales, the AMOC, is very different from model to model such that its variability and predictability are essentially not well constrained to date. Common on all these time scales are the ways in which predictability is “lost” over the course of a given prediction, namely uncertainty in the initial conditions, uncertainty in model formulation, particularly with respect to the mechanisms of potential predictability that have been overlooked and under-utilized sources. For example, model mis-characterization of the stratosphere, clouds and convection, land-atmosphere coupling and noise may be masking important sources of predictability. The misrepresentation of important phenomena such as the MJO, the AMOC and coupling of predictability to the mean annual cycle can also lead to underestimates of predictability. The 2014-2018 period of COLA research will be an exciting time, building on the achievements of the past two decades of work and extending the research program in important new directions. The proposed research is guided by a single hypothesis and several overarching principles.
The overarching principles that guide this research are:
The proposed research is guided by several scientific phenomenological and process-oriented questions that are currently of high interest:
The three main areas of basic and applied core research will be:
In addition to these areas of core research, COLA will continue to actively seek to have an impact beyond the narrow academic community in which it resides by contributing to higher education, reaching out to the climate assessment, adaptation and communication communities, and providing service to the nation through improved operational climate prediction. Directory: pub -> kinter -> Projects pub -> Acm word Template for sig site pub -> The german unification, 1815-1870 pub -> Baltic Olympiads in Informatics: Challenges for Training Together pub -> Preparation of Papers for ieee transactions on medical imaging pub -> Forthcoming meetings and conferences pub -> Asilomar Conference on Circuits, Systems & Computers, Pacific Grove, ca, November 1978, pp. 55 58 pub -> Forthcoming meetings and conferences pub -> Harmonised compatibility and sharing conditions for video pmse in the 7 9 ghz frequency band, taking into account radar use Projects -> Revolutionizing Climate Modeling – Project Athena: a multi-Institutional, International Collaboration Download 250.48 Kb. Share with your friends:
|