Today’s Big Picture Question Regarding Predicting and Quantifying the ‘Breathing of the Biosphere’: Can We Produce Flux Information that is ‘Everywhere All the Time’ with a Mechanistic Biophysical Model? Dennis Baldocchi Youngryel Ryu & Hideki Kobayashi University of California, Berkeley

Stomata: m Leaf: m Plant: 1-10 m Canopy: m Landscape: km Continent: 1000 km (10 6 m) Globe: 10,000 km (10 7 m) Bacteria/Chloroplast: m How Do We Transcend Flux Information from the Scales of the Stomata to the Leaf, Plant, Lanscape and Globe?

A Challenge for Leaf to Landscape Upscaling: Transform Weather Conditions from a Weather Station to that of the Leaves in a Canopy with Their Assortment of Angles and Layers Relative to the Sun and Sky And use that information to drive a variety of Non-Linear Functions (photosynthesis, energy balance, stomatal conductance)

Hierarchy of Canopy Abstractions

Cornelus T deWit (1970) ‘.. To build a model we have to consider and join two levels of knowledge. The level with the sort of relaxation times is then the level which provides the explanation or the explanatory level and the one with the long relaxation times, the level which is to be explained or the explainable level…’ The Perils of Upscaling Leaf-Scale Fluxes

Upscaling from Landscapes to the Globe ‘Space: The final frontier … To boldly go where no man has gone before’ Captain James Kirk, Starship Enterprise

Piers Sellers, Biometeorologist ( and Astronaut ), broke the ‘deWit’ barrier by attempting to incorporate Soil-Vegetation-Atmosphere Transfer models (SVATS) into Global Circulation Climate Models, but at coarse spatial resolution Global-Scale SVAT Modeling is Possible Today

Challenge for Landscape to Global Upscaling Converting Virtual ‘Cubism’ back to Virtual ‘Reality’ Realistic Spatialization of Flux Data Requires the Merging Numerous Data Layers with varying Time Stamps (hourly, daily, weekly), Spatial Resolution (1 km to 0.5 degree) and Data Sources (Satellites, Flux Networks, Climate Stations)

To Develop a Scientifically Defensible Virtual World ‘You Must get your boots dirty’, too Collecting Real Data Gives you Insights on What is Important & Data to Parameterize and Validate Models

Motivation for a High-Resolution, Space-Driven, and Mechanistic Trace Gas Exchange Model – Current Global-Scale Remote Sensing Products tend to rely on Highly-Tuned Light Use Efficiency Approach – GPP=PAR*fPAR*LUE (since Monteith 1960’s) Empirical, Data-Driven Approach (machine learning technique) Some Forcings come from Satellite Remote Sensing Snap Shots, at fine Spatial scale ( < 1 km) Other Forcings come from coarse reanalysis data (several tens to hundreds of km resolution) – Hypothesis, We can do Better by: Applying the Principles taught in Biometeorology 129 and Ecosystem Ecology 111 which Reflect Intellectual Advances in these Fields over the past Decade and Emerging Scaling Rules Merging Vast Environmental Databases at same resolution Utilizing Microsoft Cloud Computational Resources

Lessons Learned from the John Norman, Experience with the CanOak Model, and Reading the Literature We Must: Couple Carbon and Water Fluxes Assess Non-Linear Biophysical Functions with Leaf-Level Microclimate Conditions Consider Sun and Shade fractions separately Consider effects of Clumped Vegetation on Light Transfer Consider Seasonal Variations in Physiological Capacity of Leaves and Structure of the Canopy

Atmospheric radiative transfer Canopy photosynthesis, Evaporation, Radiative transfer Soil evaporation Beam PAR NIR Diffuse PAR NIR Albdeo->Nitrogen -> Vcmax, Jmax LAI, Clumping-> canopy radiative transfer dePury & Farquhar two leaf Photosynthesis model Rnet Surface conductance Penman-Monteith evaporation model Radiation at understory Soil evaporation shadesunlit BESS, Breathing-Earth Science Simulator

Necessary Attributes of Global Biophysical ET Model: Applying Lessons from the Berkeley Biomet Class and CANOAK Treat Canopy as Dual Source (Sun/Shade), Two-Layer (Vegetation/Soil) system – Treat Non-Linear Processes with Statistical Rigor (Norman, 1980s) Requires Information on Direct and Diffuse Portions of Sunlight – Monte Carlo Atmospheric Radiative Transfer model (Kobayashi + Iwabuchi,, 2008) Couple Carbon-Water Fluxes for Constrained Stomatal Conductance Simulations – Photosynthesis and Transpiration on Sun/Shade Leaf Fractions (dePury and Farquhar, 1996) – Compute Leaf Energy Balance to compute Leaf Saturation Vapor Pressure and Respiration Correctly – Photosynthesis of C 3 and C 4 vegetation Must be considered Separately Light transfer through canopies MUST consider Leaf Clumping to Compute Photosynthesis/Stomatal Conductance correctly (Baldocchi and Harley, 1995) – Apply New Global Clumping Maps of Chen et al./Pisek et al. Use Emerging Ecosystem Scaling Rules to parameterize models, based on remote sensing spatio-temporal inputs – Vcmax=f(N)=f(albedo) (Ollinger et al; Hollinger et al; Wright et al.) – Seasonality in Vcmax is considered (Wang et al., 2008) – Vcmax scales with Jmax (Wullschleger, 1993 )

Role of Proper Model Abstraction

But, We Need Big Iron to Play with the Big Guys and Gals

MOD04 MOD05 MOD06 MOD07 aerosol Precipitable water cloud Temperature, ozone MCD43albedo MOD11Skin temperature Atmospheric radiative transfer Net radiation MOD15LAI POLDER Foliage clumping Canopy radiative transfer Challenge for a Computationally-Challenged Biometeorology Lab: Extracting Data Drivers from Global Remote Sensing to Run the Model Youngryel was lonely with 1 PC

Barriers to Global Remote Sensing by the Berkeley Biometeorology Lab Data processing – Global 1-year source data: 2.4 TB (10 yr: 24 TB) – 150,000+ source files – Global 1-year calculation: 9000 CPU hours – That is, 375 days. – 1-year calculation takes 1 year!

Help from ModisAzure -Azure Service for Remote Sensing Geoscience Scientific Results Download Reduction #1 Queue Source Metadata AzureMODIS Service Web Role Portal Request Queue Analysis Reduction Stage Data Collection Stage Source Imagery Download Sites... Reprojection Queue Derivation Reduction StageReprojection Stage Reduction #2 Queue Download Queue Scientists Science results AZURE Cloud with 200 CPUs cuts 1 Year of Processing to <2 days

Photosynthetic CapacityLeaf Area Index Solar Radiation Humidity Deficits

Ryu et al (Accepted) Global Biogeochemical Cycles

118±26 PgC yr -1

BESS vs Machine Learning Upscaling Method Ryu et al (Accepted) Global Biogeochemical Cycles

Global Evaporation at 1 to 5 km scale An Independent, Bottom-Up Alternative to Residuals based on the Global Water Balance, ET = Precipitation - Runoff = 503 mm/y == m 3 /y

BESS vs Machine Learning Upscaling Method Ryu et al (Accepted) Global Biogeochemical Cycles

Big Picture Question Regarding Predicting and Quantifying the ‘Breathing of the Biosphere’: Can We Produce Flux Information with a Mechanistic Model that is ‘Everywhere, All the Time?’...Yes

Gross Photosynthesis, GPP, Across the US Lessons for Biofuel Production Indicates Less GPP in the Corn Belt, than the Adjacent Temperate Forests

Key point: 4. Temporal upscaling of fluxes from snap-shots to 8-day mean daily sum estimates Ryu et al (2011) Agricultural and Forest Meteorology Accepted R gPOT =f(latitude, longitude, time) Instantaneous LE Rg at TOA Day (1-8) 30 min Satellite overpass time

Ryu et al (Accepted) Agricultural and Forest Meteorology

Tested the scheme using 33 flux tower data from the Arctic to the Tropics Ryu et al (Accepted) Agricultural and Forest Meteorology

Conclusion Three-Dimensional Radiative Transfer models should be used to compute Mass and energy exchanges of Heterogeneous canopies – Models can be implemented with new generation of LIDAR data and powerful clusters of computers Advances in Theory, Data Availability, Data Sharing and Computational Systems Enable us to Produce the Next- Generation of Globally-Integrated Products on the ‘Breathing of the Earth’ Data-Mining these Products has Much Potential for Regional and Locale Decision making on Environmental and Agricultural Management

Data standardization MODIS Land products: standardized tiles (sinusoidal projection)

Barriers for global RS study 2. Data standardization MODIS Atmospheric products: swath => Should be gridded to overlay with the land products

Current status The Cloud includes – 10-year MODIS Terra and Aqua data over the US (1 km resolution) – 3-year MODIS Terra for the global land (5 km resolution) Quota: – 200 CPUs – 100TB storage

Help from MODIS-AZURE

Necessary Attributes of the Next-Generation Global Biophysical Model, BESS Direct and Diffuse Sunlight – Monte Carlo Atmospheric Radiative Transfer model (Kobayashi, xxxx) – Light transfer through canopies consider leaf clumping Coupled Carbon-Water for Better Stomatal Conductance Simlulations – Photosynthesis and Transpiration on Sun/Shade Leaf Fractions (dePury and Farquhar, 1996) – Photosynthesis of C3 and C4 vegetation considered Ecosystem Scaling Relations to parameterize models, based on remote sensing spatio-temporal inputs – Vcmax=f(N)=f(albedo) (Ollinger et al; Hollinger et al;Schulze et al.; Wright et al. – Seasonality in Vcmax is considered Model Predictions should Match Fluxes Measured at Ecosystem Scale hourly and seasonally.

Seasonal pattern of follows the seasonal pattern of LAI (modified version of Houborg et al 2009 AFM)

Size and Number of Candidate Data Sets is Enormous US: 15 tiles FluxTower: 32 tiles Global: 193 tiles 1.Global 1-year source data: 2.4 TB (10 yr: 24 TB) 2.How to know which source files are missed among >0.1 million files

Automation – Downloads thousands of files of MODIS data from NASA ftp Reprojection – Converts one geo-spatial representation to another. – Example: latitude-longitude swaths converted to sinusoidal cells to merge MODIS Land and Atmosphere Products Spatial resampling – Converts one spatial resolution to another. – Example is converting from 1 km to 5 km pixels. Temporal resampling – Converts one temporal resolution to another. – Converts daily observation to 8 day averages. Gap filling – Assigns values to pixels without data either due to inherent data issues such as clouds or missing pixels. Masking – Eliminates uninteresting or unneeded pixels. – Examples are eliminating pixels over the ocean when computing a land product or outside a spatial feature such as a watershed. Reprojected Data (Sinusoidal format - equal land area pixel) Tasked Performed with MODIS-AZURE h12v04h13v04h11v04h10v04h09v04h08v04 h12v05h11v05h10v05h09v05h08v05 h11v06h10v06h09v06h08v06

remote sensing of CO 2 Temporal scale Spatial scale [km] hour day week month year decade century local global forest inventory plot Countries EU plot/site tall tower obser- vatories Forest/soil inventories Eddy covariance towers Landsurface remote sensing Components of an Integrated Earth System EXIST, but are Multi-Faceted From: Markus Reichstein, MPI

Computing Carbon Dioxide and Water Vapor Fluxes Everywhere, All of the Time Dennis Baldocchi Youngryel Ryu & Hideki Kobayashi University of California, Berkeley AGU, Fall 2011