Presentation on theme: "Studying Uncertainty in Palaeoclimate Reconstruction SUPRaNet SUPRModels SUPR software Brian Huntley, Andrew Parnell Caitlin Buck, James Sweeney and many."— Presentation transcript:
Studying Uncertainty in Palaeoclimate Reconstruction SUPRaNet SUPRModels SUPR software Brian Huntley, Andrew Parnell Caitlin Buck, James Sweeney and many others Science Foundation Ireland Leverhulm Trust
Result: one pollen core in Ireland 95% of plausible scenarios have at least one “100 year +ve change” > 5 o C Mean Temp of Coldest Month
Climate over 100,000 years Greenland Ice Core 10,000 year intervals Oxygen isotope – proxy for Greenland temp Median smooth. Past 23000 years The long summer
Past 23000 years Climate over 100,000 years Greenland Ice Core 10,000 year intervals The long summer Int Panel on Climate Change WG1 2007 “During the last glacial period, abrupt regional warmings (probably up to 16 ◦ C within decades over Greenland) occurred repeatedly over the North Atlantic region”
Climate over 15,000 years Greenland Ice Core Younger Dryas Transition Holocene Ice dynamics? Ocean dynamics? What’s the probability of abrupt climate change?
Modelling Philosophy Climate is – Latent space-time stoch process C(s,t) All measurements are – Indirect, incomplete, with error – ‘Regionalised’ relative to some ‘support’ Uncertainty – Prob (Event) – Event needs explicitly defined function of C(s,t)
Proxy Data Collection Oak treeGISP iceSedimentPollen Thanks to Vincent Garreta
Data Issues Pollen 150 slices – 28 taxa (not species); many counts zero – Calibrated with modern data 8000 locations 14 C5 dates – worst uncertainties ± 2000 years Climate `smoothness’ – GISP data 100,000 years, as published
Model Issues Climate - Sedimentation - Veg response latent processes – Climate smooth (almost everywhere) – Sedimentation non decreasing – Veg response smooth Data generating process – Pollen – superimposed pres/abs & abundance – 14 C - Bcal Priors - Algorithms …….
SUPR-ambitions Principles – All sources of uncertainty – Models and modules – Communication Scientist to scientist to others Software Bclim Future SUPR tech stuff non-linear non-Gaussian multi-proxy space-time incl rapid change dating uncertainty mechanistic system models fully Bayesian fast software
Modelling Approach Latent processes – With defined stochastic properties – Involving explicit priors Conditional on ‘values’ of process(es) – Explicit stochastic models of – Forward Data Generating Processes – Combined via conditional independence – System Model
Modelling Approach Modular Algorithms – Sample paths, ensembles – Monte Carlo – Marginalisation to well defined random vars and events
Progress in Modelling Uncertainty Statistical models – Partially observed space-time stochastic processes – Bayesian inference Monte Carlo methods – Sample paths – Thinning, integrating Communication – Supplementary materials Modelled Uncertainty Does it change? In time? In space?
SUPR Info Proxy data: typically cores – Multiple proxies, cores; multivariate counts – Known location(s) in (2D) space – Known depths – unknown dates, some 14 C data – Calibration data – modern, imperfect System theory – Uniformitarian Hyp – Climate ‘smoothness’; Sedimention Rates ≥ 0 – Proxy Data Generating Processes
Bchron Models Sedimentation a latent process – Rates ≥ 0, piecewise const – Depth vs Time - piece-wise linear – Random change points (Poisson Process) – Random variation in rates (based on Gamma dist) 14 C Calibration curve latent process – ‘Smooth’ – in sense of Gaussian Process (Bcal) 14 C Lab data generation process – Gaussian errors
Bchron Algorithm Posterior – via Monte Carlo Samples Entire depth/time histories, jointly – Generate random piece-wise linear ‘curves’ – Retain only those that are ‘consistent’ with model of data generating system Inference – Key Parameter; shape par in Gamma dist – How much COULD rates vary?
20 Bivariate Gamma Renewal Process Comp Poisson Gamma wrt x ; x incs exponential Comp Poisson Gamma wrt y ; y incs exponential
21 Compound Poisson Gamma Process We take y = 1 for access to CPG and x > 2 for continuity wrt x Slope = Exp / Gamma = Exp x InvGamma infinite var if x > 2
22 Modelling with Bivariate Gamma Renewal Process Data assumed to be subset of renewal points Implicitly not small Marginalised wrt renewal pts Indep increments process Stochastic interpolation by simulation new y unknown x
23 Stochastic Interpolation Unit Square Monotone piece-wise linear CPG Process
24 Stochastic Interpolation Monotone piece-wise linear CPG Process
25 Stochastic Interpolation Monotone piece-wise linear CPG Process
26 Stochastic Interpolation Monotone piece-wise linear CPG Process
27 Stochastic Interpolation Monotone piece-wise linear CPG Process
28 Stochastic Interpolation Monotone piece-wise linear CPG Process
29 Stochastic Interpolation Density Known Depths Known age Calendar age
Glendalough Time-Slice “Transfer-Function” via Modern Training Data Hypothesis Modern analogue Climate at Glendalough 8,000 yearsBP “like” Somewhere right now The present is a model for the past
Calibration ---------------------- ---------------------- c (t) y(t ) Modern (c, y ) pairs In space ---------------------- ---------------------- c(t) y(t ) Eg dendro Two time series Much c data missing Eg pollen One time series All c data missing Space for time substitution Over- lapping time series
Calibration Model Simple model of Pollen Data Generating Process ‘Response’ y depends smoothly on clim c Two aspects Presence/Absence Rel abundance if present Taxa not species Eg y i =0 probq(c) y i ~Poisson (λ(c)) prob1-q(c) Thus obsy i =0, y i =1very diff implications
One-slice-at-a time Slice j has count vector y j, depth d j Whence – separately - π(c j | y j ) and π(t j | d j ) ResponseChronmodule
Uncertainty one-layer-at-a-time Pollen => Uncertain Climate Depth => Uncertain depth But monotonicity Here showing 10 of 150 layers
Climate property? Non-overlapping (20 year?) averages are such that first differences are: adequately modelled as independent inadequately modelled by Normal dist adequately modelled by Normal Inv Gaussian – Closed form pdf – Infinitely divisible – Easily simulated, scale mixture of Gaussian dist
MTCO Reconstruction One layer at a time, showing temporal uncertainty Jointly, century resolution, allowing for temporal uncertainty Marginal time-slice: may not be unimodal
Rapid Change in GDD5 Identify 100 yr period with greatest change One history
Rapid Change in GDD5 One history Identify 100 yr period with greatest change
Rapid Change in GDD5 Study uncertainty in non linear functionals of past climate 1000 histories Identify 100 yr period with greatest change
Result: one pollen core in Ireland 95% of plausible scenarios have at least one 100 year +ve change > 5 o C Mean Temp of Coldest Month
Communication Scientist to scientist Exeter Workshop – Data Sets – With Uncertainty Associated with what precise support?
Modelling Approach Latent processes – With defined stochastic properties – Involving explicit priors Conditional on ‘values’ of process(es) – Explicit stochastic models of – Forward Data Generating Processes – Combined via conditional independence Modular Algorithms – Sample paths, ensembles – Monte Carlo – Marginalisation to well defined random vars and events