Colin Prentice SPEDDEXES 2014
- 1. Why big data is a game changer for terrestrial ecosystem science and what have we learned over the last 30 years I. Colin Pren,ce AXA Chair in Biosphere and Climate Impacts, Imperial College London Professor in Ecology and EvoluCon, Macquarie University Chair, ecosystem Modelling And Scaling infrasTructure (eMAST)
- 2. The significance of 30 years ago… • Orwell’s 1984 • Murakami’s 1Q84 • Shugart (1984) A Theory of Forest Dynamics – “gap models” for tree growth and compeCCon – ecosystem-‐specific, required data on every tree species – lack of integraCon of vegetaCon dynamics with ecophysiology, biogeochemistry, biogeography
- 3. Trends in ecosystem science, 1984-‐2004 • Recognizing large-‐scale drivers of ecosystem change GCTE launch (1992): promoCng experimental and modelling research on global change • From ecosystem-‐specific models to DGVMs Cramer et al. (2001) GCB: C cycle projecCons, six models • Revival of comparaCve funcConal ecology (moCvaCon to improve DGVMs) Wright et al. (2004) Nature: leaf economics spectrum
- 4. Big data for ecosystem science • Steady accumulaCon of precise atmospheric measurements (ramp up in 1980s) • Major advances in remote sensing (MODIS launch 2000; Sciamachy, GOME etc. for atmospheric consCtuents) • ‘Bodom-‐up’ syntheses of local measurements (flux, traits) => push for data sharing (N America first; big push from TERN; WIRADA) • ConCnuous exponenCal improvement in data storage and computaConal capacity • Major advances in computaConal tools (especially open-‐source languages and codes)
- 5. What can we do with big data? • Model evaluaCon and benchmarking (post facto comparison) • Data assimilaCon (model structure pre-‐defined: variables and/or parameters to be esCmated) • New model development (using data to inform model structure) 1. Process understanding flows from large-‐scale data analysis. 2. There are huge unexploited opportuniCes – hardly conceivable 30 years ago.
- 6. Role of eMAST • PredicCve models, fully informed by all relevant data • Ecosystems under pressure => requirement for predicCve power
- 7. Role of eMAST (cont.) • Without models, there is no predicCve power. • Without data, models are worthless. • We need to make it easy for models and data to talk to one another.
- 8. Example 1: CO2 seasonal cycles • Seasonal cycles at different locaCons as a benchmark for modelled NEE • Increasing high-‐laCtude seasonal cycle as a challenge for modelling NPP • Requires intervenCon of an atmospheric transport model – but this can be done ‘automaCcally’ through inversion
- 9. NH Tropics SH Kelley et al. (2013) Biogeosciences
- 10. Graven et al. (2013) Science
- 11. Graven et al. (2013) Science
- 12. Example 2: Leaf stable carbon isotopes • Global leaf δ13C data (for ci:ca raCo – coupling of water and CO2 exchanges): synthesis of > 3500 measurements led by Will Cornwell, UNSW • Leaf economics theory (PrenCce et al. 2013 Ecology LeIers) => predicts dependence on temperature, aridity, elevaCon • Requires climate data and a model to infer bioclimate variables, e.g. cumulaCve water deficit (proxy for vpd)
- 13. ParCal residual plots H. Wang et al. (unpublished results)
- 14. Global slopes: ln χ/(1 − χ) vs predictors Predicted FiIed (± 95% CI) temperature 0.055 0.050 ± 0.004 ln (dryness) − 0.250 − 0.226 ± 0.012 elevaCon − 0.082 − 0.093 ± 0.030 R2 = 0.450
- 16. Example 3: IntegraCng remotely sensed and flux measurements (ePiSaT) • OzFlux synthesis (all-‐site CO2 flux measurements) • fAPAR synthesis product (Huete et al.) ! ParCConing fluxes into respiraCon and GPP ! Analysis of monthly integrated GPP versus fAPAR x PPFD ! LUE model driven by fAPAR, PPFD, vpd… • Also requires climate data, bioclimate variables, parCConing and gap-‐filling methods… B.J. Evans et al. (2013) unpublished results
- 18. Where do we go from here? • Data-‐model comparison and evaluaCon ‘made easy’. • Data assimilaCon ‘made possible’. • IntegraCon of data sets with different properCes (e.g. spaCally versus temporally extensive) ‘made rouCne’.