290 likes | 491 Vues
Welcome to the Gap Filling Comparison Workshop September 18-20, 2006. Antje Moffat. Welcome. 12 of the 14 members from the gap filling comparison Over 30 participants - from all over Europe (Germany, Italy, Netherlands, Switzerland), US, Canada, Russia, and Australia. Goals of the Workshop.
E N D
Welcometo theGap Filling ComparisonWorkshopSeptember 18-20, 2006 Antje Moffat
Welcome • 12 of the 14 members from the gap filling comparison • Over 30 participants - from all over Europe (Germany, Italy, Netherlands, Switzerland), US, Canada, Russia, and Australia
Goals of the Workshop • Review of Gap Filling Techniques • Completion of the Gap Filling Comparison • Discussion of the Results • Review of Paper • Evaluation of the Techniques • Work Sessions and Plenary Debates to Exchange our Experiences and Expertise • Generate Ideas for Further Improvement of the Gap Filling • New Insights into the Eddy Flux Data
Outline Workshop Program GFC Analysis Performance of Techniques Error on Annual Sum
Program: This Morning (Monday) 9:00 Registration of the Participants If you need any kind of help, please contact Ulli or Silvana! 9:30 Setting the Stage of the Workshop • Antje Moffat: “Welcome” • Martin Heimann: “Biogeochemistry Research at the MPI in Jena” • Dario Papale: “The CarboeuropeIP Ecosystem Component Database: data processing and availability” • Markus Reichstein: "Gap filling: Why and how?” 11:00COFFEE BREAK (Foyer) 11:30 Review of Gap Filling Techniques: SPM and ANNs • Vanessa Stauch: “Semi-parametric models” • Dario Papale: “Gap filling of eddy fluxes with artificial neural networks” • Rob Braswell: "Gap filling by iterative regression using a regularized neural network” 12:30 LUNCH (Campus Cafeteria)
This Afternoon 14:00 Review of Gap Filling Techniques (cont.): NLRs and UKF • Ankur Desai: “Towards a robust, generalizable non-linear regression gap filling algorithm” • Asko Noormets: “NLR_AM - AQRTa-Model” (10 min recording) • Andrew Richardson: “Maximum likelihood non-linear regression model” • David Hollinger: “Data assimilation for eddy flux filling: The unscented Kalman filter” 15:30 COFFEE BREAK 16:00 Review of Gap Filling Techniques (cont.): Models and comparison • Zisheng Xing: “A gap-filling model for tower-based NEP measurements” • Jens Kattge: “Model parameter inversion against Eddy Covariance Data using a Monte Carlo Technique” • Bart Kruijt: “Comparing gap filling using neural networks and the CarboEurope tool, for Fluxes and Meteo data” 19:30 Dinner Suggestion: Restaurant Papiermuehle (Please sign up!)
Tuesday Morning 9:00 Eddy and Component Flux Measurements • Corinna Rebmann: “Eddy covariance measurements and their shortcomings for the determination of NEE” • David Hollinger: “Uncertainty in eddy flux data and its relevance to gap filling” • Eva van Gorsel: “Nocturnal Carbon Efflux: Can eddy covariance and chamber measurements be reconciled?” • Pasi Kolari: “Gapfilling submodel selection based on measured component fluxes” 10:30 COFFEE BREAK 11:00 Gap Filling of Grassland and Agricultural Sites • Christof Ammann: “Gap-Filling of CO2 Fluxes of Frequently Cut Grassland” • Mauro Colavincenzo: “A gap filling methodology used at a agricultural site in Southern Italy” • Irene Lehner: “Carbon balance of a maize canopy: comparison of different gap filling strategies” 12:00 LUNCH END of OPEN SESSIONS!
Tuesday Afternoon 13:00-15:00 Parallel Work Sessions Part 1 • Group 1: “Analysis of the partitioned GPP/ER comparison results” (Ankur Desai) • Group 2:“Gap filling of meteorological data and water and energy fluxes” (Dario Papale) 15:00 COFFEE BREAK 15:30-17:30 Parallel Work Sessions Part 2 • Group 3:"Gap filling of sites with non-steady time series, e.g. cut grassland, cropland" (Christof Ammann) • Group 4: "Using gap-filling techniques for estimating random errors in eddy covariance data" (Andrew Richardson) Please sign up for the work sessions! 19:30WORKSHOP DINNER (Restaurant: Weinbauernhaus “Im Sack”)
Wednesday Roundtable on the Gap Filling Comparison 8:30 Review of Gap Filling Comparison Paper, Antje Moffat - Interpretation of the comparison results - Derivation of key findings • Evaluation of techniques 12:30 LUNCH 13:30 Minutes from the four Work Sessions 14:30 COFFEE BREAK 15:00 Plenary Debates - Site dependency of gap filling technique performances - Filling of long gaps using previous years - Conception of a public domain code library with filling routines - Extended gap filling comparison for urban and crop - Workshop resume and outlook 17:00 End of Workshop
Handling of Presentations • 20 min presentations: 15 min talk plus 5 min discussion • Please transfer your presentation onto common laptop during coffee or lunch break (Ulli or Silvana) • Publicized on GFC webpage after workshop Questions?
Gap Filling Comparison Analysis
Basic Principle • Keyfile: Artificial Gap Flags • Golden File - fragmented: • Superimposition • Comparison of Observed NEP and Predicted NEP:
Statistical Metrics • Bias Error • Root Mean Square Error • Correlation Coefficient p - predicted NEP o - observed NEP
Daytime/Nighttime data Weighted ALL data Analysis Predicted versus Observed • Half-hourly basis • Daily sum basis for full day artificial gaps
Challenge of the Analysis 5 artificial gap length scenarios (single hh - 12 days) • 10 permutations • 3 subsets: day, night, all • 12 golden sites • 19 submissions • 15 statistical metrics: RMS, R2, Bias, Daily Sum, normalized, benchmarked, … 513,000 comparison results! (see selection on posters in foyer)
RMSE and R2:Half-hourly Basis Performance of gap filling techniques from bottom • MIM, MDV, UKF_LM, NLR & Others • 3 ANNs leading Daytime Correlation Coefficient R2 Nighttime Root Mean Square Error (gCm-2)
Daytime Nighttime Root Mean Square Error (gCm-2) RMSE and R2: Half-hourly & Daily Sums Daytime Daily sum basis Daytime • R2: 0.8 - 0.95 • RMSE: 1.0 - 1.8 gCm-2 Nighttime • R2: 0.75 - 0.9 • RMSE: 0.5 - 1.0 gCm-2 • Very good filling performance for daytime and nighttime data Techniques: • MIM, Others, ANNs leading Correlation Coefficient R2 Nighttime Root Mean Square Error (gCm-2) Half-hourly basis Daytime • R2: 0.6 - 0.8 • RMSE: 2.5 - 4.0 gCm-2 Nighttime • R2: 0.2 - 0.4 • RMSE: 1.5 - 2.5 gCm-2 • Good filling performance for daytime but not for nighttime Techniques: • MIM, MDV, UKF_LM, NLR & Others, 3 ANNs leading
DailySum Bias per Site Year:Medium Gaplength, ALL Bias Techniques Medium gap length (1.5 days): Bias of <0.07 gCm-2 per filled day
DailySum Bias per Site Year:Long Gaplength, ALL Bias Techniques Long gap length (12 days): Bias of <0.2 gCm-2 per filled day
Annual Sum Error Estimate Assumption: • representative choice of golden sites • good technique (red stars) • Error estimate on the annual sum Annual Sum Error • Small to med gaps: <0.07 gCm-2 per filled day equivalent • Periods of longer gaps: <0.2 gCm-2 per filled day equivalent • Quality of long gap filling critical
Calculation Example Example for average file with 35% gaps: • 18% small to medium gaps • 18% periods of longer gaps of 5-10 days 18% = 66 filled days Error estimation • 66 x 0.07 gCm-2: 5 gCm-2 • 66 x 0.2 gCm-2: 13 gCm-2 • Total error induced by filling of the gaps on the annual sum: ±18 gCm-2 ^
Test using Real Gap Filling Results Are 18gCm-2 an appropriate estimate of the error on the annual sum prediction? • Standard deviation between techniques of filling the real dataset with 35% gaps ≤ 16 gCm-2 1) no soil temperature 2) 30 day system failure
Let’s fill our “knowledge gaps” and have a fun and productive workshop!
Separation of Daytime and Nighttime Data Keyfile: 10% artificial gaps Fragmented Golden File: 80% daytime NEP data, 35% nighttime NEP data • Real gap filling: 20% real day gaps, 65% real night gaps 1:3 • Artificial gap filling: 8% artificial day gaps, 3.5% artificial night gaps 2:1 Important to consider daytime and nighttime data separately
Bias on daily sumsDaytime data • Distribution of bias error of the individual daily sums Daytime data Daily Bias Error: • up to 4 gCm-2 ANNs leading ANN_BR Bias Error (gCm-2)
Bias on daily sums:Nighttime data • Distribution of bias error of the individual daily sums Daytime data Daily Bias Error: • up to 2 gCm-2 ANN_PS leading Bias Error (gCm-2)