Research project SR/00/306 (Research action SR)
CONTEXT AND OBJECTIVES
Meteorological droughts, rainfall extremes and heatwaves are major natural disasters with diverse socio-‐economical and environmental consequences. There is a perception that these climatic events are becoming unusually abundant after recent droughts in Western United States (2011) or North-‐Eastern China (2009), unprecedented wet periods accompanied by floods in U.K. (2007) or Pakistan (2010), and unparalleled mega-‐heatwaves in Europe (2003, 2010). These events caused the failure of the agricultural and food production systems, natural biomass loss, the spread wild fires, air pollution, water scarcity, and multiple other consequences that raised the mortality tolls by tens of thousands. As we progress into the future, our climate models predict that the exacerbation and proliferation of such events will continue, following the expected rise in greenhouse gases.
There are several reasons why climate extremes are affected by the global rise in temperatures. Heatwaves will unavoidably exacerbate as average and variability of global temperatures continue to increase. Yet, higher temperatures also mean more intense terrestrial evaporation, which is expected to aggravate dry conditions in regions that are already dry, and increase the volume and rates of precipitation in regions that are wet. This 'wet-‐gets-‐wetter, dry-‐gets-‐drier' hypothesis is currently intensively investigated within the fields of climate and hydrology. In addition, climate models also predict other, more indirect, effects of global warming on climate extremes. For instance, the rise in temperatures may lead to a global reorganization of the hydrological cycle with a pole-‐wards migration of current climatic regions; the subsequent widening of the tropics implies an overall reduction in the input of rainfall to mid latitudes and an intensification of regional land-‐atmosphere feedbacks that may further intensify droughts and heatwaves. Moreover, a series of biotic feedbacks on climate can also be expected, as the intensification of climate extremes severely impacts vegetation biomass and reduces the efficiency of land as a sink of CO2.
Up to date, the expectations of a future aggravation of climatic extremes, and the impact that this aggravation may have on Earth's vegetation, remain to a large extent unstudied. This limits the societal capabilities for long-‐term adaptation. Observational evidence of trends in magnitude and variability of extreme precipitation and temperature spells is still scarce, and without observational evidence, climate model representation of extremes is condemned to remain uncertain. Just recently, the AR5 report of the Intergovernmental Panel of Climate Change (IPCC) has underlined the disagreements in the studies of past changes in these climatic extremes, particularly for the case of droughts. These discrepancies raise due to (a) the limited availability of observational datasets that can be used to evaluate past changes in these extremes at the global scale and over multidecadal periods, (b) the shortcomings of statistical and physically-‐based methods typically used to detect these changes, (c) the confronting scales at which these processes operate and the importance of disentangling the effects of multi-‐year ocean-‐atmospheric oscillations from longer-‐term trends. These inconsistencies were already noted in the IPCC AR4 report (2007) and could not be solved before the new AR5 (2013). Therefore, the understanding of recent changes in climate extremes and the effects they had on biomass is considered as a major milestone for the AR6 report, scheduled to be released by the end of 2018.
Conveniently, advances in satellite Earth observation in recent years have culminated with the development of consistent global historical records of environmental and climatic variables that are critical for the study of these extreme events. Novel continuous datasets of soil moisture, vegetation water content, fluoresce and land evaporation have been derived by merging multi-‐satellite information since the late '70s. These remote sensing datasets share the large-‐scale advantage of climate models with the observational nature of meteorological measurements, which confers them great potential as a mean to study global changes in past extreme events, but also as an observational benchmark to evaluate climate models. These new datasets can be combined with in-‐situ measurements and more conventional satellite-‐based global products of precipitation, temperature or vegetation properties, in order to: (a) unravel past global changes in frequency and severity of extreme precipitation and temperature spells, (b) uncover the spatiotemporal scales at which the processes driving these changes operate, (c) give evidence of the global impact of droughts, heatwaves and extreme rainfall events on terrestrial vegetation, (d) allow climate model selection and improvement on the basis of the model's skill to represent changes in climate extremes and their impacts on continental biomass.
Some recent studies have already used these new datasets on their own with the focus of studying past climate extremes (Dorigo et al., 2012). Others have applied them with the explicit goal of evaluating climate model representation of general average hydrological patterns (Mueller et al., 2013), and more recently to evaluate the representation of heatwaves in Europe (Stegehuis et al., 2013). Conversely, the Global Climate Model representation of climate extremes has also been evaluated by comparison to reanalysis and in-‐situ measurements in several occasions (Sillmann et al., 2013). However, these long-‐ term remote sensing datasets have not yet been applied to evaluate the global-‐scale variability of precipitation and temperature extremes in climate models, neither to identify the drivers responsible for the ongoing changes in these extremes nor their impacts on global vegetation dynamics. The objectives of SAT-‐EX strive in that direction. Several statistical techniques – that have gained popularity in recent years in other fields of research – appear optimal for analysis: the spatiotemporal clustering of extremes, the random forest machine-‐learning technique and the fingerprint analysis. These techniques will be used to analyse extreme indices of climate and vegetation fields derived from multi-‐decadal satellite records and obtained from IPCC CMIP5 Earth System Models (ESMs) archive. For more details, see the Methodology section below.
The five objectives of SAT-‐E are:
1. To provide new observational evidence of how droughts, heatwaves and extreme rain events have changed in time and space over the satellite era.
2. To identify the drivers behind these changes, e.g., intensification of the hydrological cycle, widening of the tropical belt, ocean-‐atmospheric oscillations, anthropogenic emission, etc.
3. To provide satellite-‐based observational insights into the past changes in global vegetation and the role of extreme hydrological and climatic events on these changes.
4. To show if IPCC ESMs reproduce the past changes in climatic extremes shown in the satellite records and understand the sources of discrepancies.
5. To show if IPCC ESMs reproduce the observed changes in vegetation with particular emphasis on the extremes and the vegetation response to hydrological and climatic extremes.
METHODOLOGY
The term 'extreme' is applied to those climatic events that are significantly larger than expected, considering a non-‐changing (stationary) climate as reference. While this concept still depends on the definition of stationary conditions and the setting of extreme thresholds, a wide range of statistical indices have been used in recent years aiming to minimize the subjectivity on the characterization of what is considered 'extreme'. Analysing these events is challenging given that, by definition, their frequency of occurrence is low. This is especially problematical when using remote sensing data due to the characteristic short lifespan of satellite platforms and sensors. However, recent scientific efforts have yielded global consistent records of climatic and environmental variables through the combination of data from multiple satellite sensors. In this blending of multi-‐sensor data, several cross-‐calibration techniques have been applied, validation exercises against in-‐situ measurements have been performed, and error estimates have been calculated. Now, records of 30–35 years are available for some variables (e.g., temperature, soil moisture, vegetation physical properties, evaporation). This timespan is still short, but appears long enough to start assessing some of the critical aspects of the changes in climate extremes and global vegetation dynamics in recent decades.
Assuming a sufficiently long, high-‐quality time series of a given climatic variable (e.g. precipitation, temperature), the number of extremes should remain constant over time for a stationary climate. Therefore, if a trend in the number or intensity of these extremes were detected, this would be indicative of climate change, i.e. a long-‐term change in the mean and/or the shape of the probability density function of that climatic variable. In our context, this climatic change may reflect: (a) the direct impact of greenhouse gases, aerosols or land-‐use change on the radiation budget, (b) the subsequent intensification or deceleration of the hydrological cycle, © a reorganization of the large-‐scale preferential climatic and hydrological patterns (e.g. the widening of the tropical belt), (d) the 'confounding effects' from multi-‐year and decadal climatic oscillations. All these different processes lead to impacts on the terrestrial biosphere, and especially on the variability and distribution of continental vegetation. SAT-‐EX aims to use traditional climate extreme indices derived from recent remote sensing long-‐term records, to detect and attribute changes in climatic and environmental extremes using a combination of spatiotemporal clustering methods, fingerprint analysis and random forest machine-‐ learning (see below).
Table 2 contains the workflow of the project, with the different work packages (WPs) and the partners' involvement on each of these tasks. The specific timeframes are approximate. The work is distributed in six WPs. The first two (WP1, WP2) refer to the project management, dissemination and assessment of user requirements. The WP3 and WP4 contain the core of the analysis of satellite records and extend through the first two years of the project. WP5 dedicates to the benchmarking of CMIP5 Earth System Models (ESMs) against the long-‐term satellite records based on a parallel statistical framework. Finally, a synopsis of the project is undergone in WP6.
1. Spatiotemporal clustering of extreme indices based on satellite datasets (WP3)
The climate extremes of interest for this study are three: heatwaves, droughts and extreme rainfall events. Heatwaves are considered as prolonged periods of extremely warm weather. Extreme rainfall events are simply those that significantly exceed the climatological expectation and may lead to flooding. Droughts, on the other hand, are periods of unusual rainfall scarcity (meteorological drought), that may have important repercussions for agriculture (agricultural drought) and, in the longer-‐run, serious consequences for the discharge in rivers or the volume of hydrological reservoirs (hydrological drought). These different aspects of droughts can be studied with current global remote sensing datasets of rainfall, soil moisture and evaporation. In addition, the physiological response of vegetation can also be an indicator of the intensity and persistence of drought conditions. In fact, these three climatic extremes are expected to leave a footprint on the vegetation properties that can be observed from space using datasets of fluoresce, photosynthetic active radiation, greenness, vegetation optical depth, etc.
SAT-‐EX will use the indices from the World Meteorological Organization (WMO) and World Climate Research Programme (WCRP) Expert Team on Climate Change Detection and Indices (ETCCDI); a total of 27 climate extreme indices, mostly based on the computation of absolute maxima (or minima), the number of values over (or under) certain thresholds, or different percentiles. These indices are designed for annual (or seasonal) scales and are calculated from daily time series of temperature or precipitation, but may also be calculated using time series of other variables, like soil moisture, evaporation or vegetation characteristics. Several of these indices are of particular interest for SAT-‐EX, as they combine the assessment of temporal changes in intensity, frequency and duration of extremes. The spatal and temporal distribution of these extremes can then be analysed with simple statistical techniques focused on the spatiotemporal clustering of extreme values, like the one recently described and applied by Zscheischler et al. (2014).
2. Fingerprint analysis of climate extreme indices based on satellite datasets (WP4)
The analyses will not be limited to deriving annual/seasonal climate extreme indices based on remote sensing data, neither to assessing their trends and spatiotemporal variability. A fingerprint analysis will be applied to discern whether the observed past spatiotemporal changes in those indices can be attributed to the direct impact of global warming, the subsequent intensification of the hydrological cycle, or reorganization of climatic patterns (e.g. the widening of the tropical belt). Alternatively, it may be observed that – due to the limited record-‐length of observations – the multiyear variability in those indices is simply dictated by ocean-‐atmosphere oscillations.
The proposed fingerprint analysis consists of a study of the temporal tendencies in the magnitude and spatial variability of a series of local minimum and maximum values of a given variable (for SAT-‐EX, a climate extreme index). The analysis is typically done on the basis of latitudinal averages, but can be applied using other reference domains. Applied to latitudinal averages, the technique has the potential to uncover the polewards migration of certain climate extremes, if there is an intensification of those extremes at a given latitude, or if the extremes are intensifying in regions where they are already more common (in agreement with the 'wet-‐gets-‐ wetter, dry-‐gets-‐drier' hypothesis). The analysis of the fingerprints is usually done in combination with the calculation of the leading multivariate empirical orthogonal function (EOF) of the time series of local minima and maxima to allow for a formal attribution of the observed changes (see e.g. Marvel and Bonfils, 2013).
The technique may be applied to climatic extremes of precipitation or temperature (defined using the above climate extreme indices), and also to the extremes calculated based on other environmental variables (soil moisture, evaporation, vegetation characteristics) to help assess the correspondence between the changes in climatic extremes and changes in vegetation patterns.
3. Random forest study of sensitivities of vegetation to climate extremes (WP4)
The random forest is a machine-‐learning technique based on an ensemble of trees to compute a classification. A necessary and sufficient condition for an ensemble of classification trees to be more accurate than any of its individual members, is that the members of the ensemble must perform better than random. Random forests increase diversity among the classification trees by resampling the data with replacement, and by randomly changing the predictive variable sets over the different tree induction processes. Each classification tree is grown using another bootstrap subset of the original data set, and the nodes are split using the best split predictive variable (among a subset of randomly selected predictive variables). The number of trees and the number of predictive variables used to split the nodes are user-‐defined parameters. Additionally, an unbiased estimate of the generalization error is obtained during the construction of a random forest, and this generalization error always converges as the random forest grows, so the method does not over-‐fit.
In the context of SAT-‐EX, the extreme climate indices (based on remote sensing precipitation, temperature, soil moisture, etc.) will act as predictive variables together with other satellite-‐observable climatic variables (e.g. radiation, humidity, CO2 concentrations, etc.) – see Table 1. Long-‐term records of different vegetation properties will be the variable to predict. The random forest technique will be used to estimate the importance of each predictive variable, by looking at how much the error increases when data for that variable are permuted while all other variables are left unchanged. This will therefore help reveal the importance of climatic extremes as drivers of the past vegetation variability. Random forests, in addition, will allow to cascade the uncertainty of satellite observations, even if operating with inter-‐dependent climatic variables.
4. Earth System Model evaluation (WP5)
WP3–WP4 concentrates on: (a) calculating extreme indices in climate and vegetation based on satellite data, (b) analyzing their spatiotemporal clustering and trends, © exploring those with a fingerprint approach to attribute the causes of their long-‐term variability, (d) using climate extreme indices in combination with random forest models to understand the sensitivity of vegetation to changes in climate extremes. The work package WP5 of SAT-‐EX will focus on the use of the satellite-‐based data, results and methods from WP3–WP4 to evaluate ESMs representation of (a) the changes in past climate extremes, and (b) the impact of climate extremes on vegetation dynamics.
Therefore, using the freely available output data from the range of IPCC CMIP5 ESMs, WP5 will recreate our satellite data-‐based analyses (WP3–WP4) using these ESM outputs, including the derivation of climate extreme indices for the past three decades, and the analyses of these indices using spatiotemporal clustering, fingerprints and random forest methods. A comparison to the observation-‐ based results from previous phases in the project will allow to discern which ESMs are better suited for representing the variability of climate extremes (taking the satellite-‐based results as the benchmark in these comparisons). Vegetation biomass simulations from the ESMs will be analysed in the same fashion that the satellite-‐based vegetation characteristics in earlier stages of the project.
The proposed timing of application of the different methods is detailed in Table 2. The final part of SAT-‐ EX (i.e. WP6) will comprise the synthesis of the results from the entire project, a descriptive analysis of the societal implications of our findings, and the suggested roadmap for the future.
