2DkcTXceO

440 Environmental informaticsThe choice of τ is at the granularity of D t , and the choice of R depends onthe question being asked and the roughness of Earth’s surface relative to thequestion. In a classical bias-variance trade-off, one wants R(x, y) to be largeenough for g F (x, y; t) to capture the dominant variability and small enoughthat the flux in R(x, y) ishomogeneous.Carbon-cycle science has accounted for much of the dynamics of CO 2 ,but the carbon budget has consistently shown there to be a missing sink(or sinks). The OCO-2 instrument, with its almost pinpoint accuracy andhigh sensitivity near Earth’s surface, offers an unprecedented opportunity toaccurately estimate the sinks. From that point of view, the parts of Y thatare of interest are lower quantiles of g F (Y ), along with the (lon, lat)-regionswhere those quantiles occur. In Section 38.6, I argue that these queries of theprocess g F (Y ) can be formalized in terms of loss functions; Zhang et al. (2008)give an illustration of this for decadal temperature changes over the Americas.This different approach to flux estimation is centrally statistical, and itis based on a spatio-temporal model for [Y |θ]. There is another approach,one that bases [Y |θ] on an atmospheric transport model to incorporate thephysical movement of voxels in the atmosphere and, consequently, the physicalmovement of CO 2 ; see, e.g., Houweling et al. (2004), Chevallier et al. (2007),Gourdji et al. (2008), and Lauvaux et al. (2012). Motivated by articles suchas Gourdji et al. (2008), I expect that the two approaches could be combined,creating a physical-statistical model.When [Y |θ] is different, the predictive distribution given by (38.8) is different,and clearly when L in (38.9) is different, the optimal estimate givenby (38.9) is different. This opens up a whole new way of thinking about fluxestimation and quantifying its uncertainty, which is something I am activelypursuing as part of the OCO-2 Science Team.38.5.2 Regional climate change projectionsClimate is not weather, the latter being something that interests us on dailybasis. Generally speaking, climate is the empirical distribution of temperature,rainfall, air pressure, and other quantities over long time scales (30 years, say).The empirical mean (i.e., average) of the distribution is one possible summary,often used for monitoring trends, although empirical quantiles and extremamay often be more relevant summaries for natural-resource management. Regionalclimate models (RCMs) at fine scales of resolution (20–50 km) producethese empirical distributions over 30-year time periods and can allow decisionmakersto project what environmental conditions will be like 50–60 years inthe future.Output from an RCM is obtained by discretizing a series of differentialequations, coding them efficiently, and running the programs on a fast computer.From that point of view, an RCM is deterministic, and there is nothingstochastic or uncertain about it. However, uncertainties in initial and boundaryconditions, in forcing parameters, and in the approximate physics asso-