Error and Uncertainty in Modeling George H. Leavesley, Research Hydrologist, USGS, Denver, CO Sources of Error and Uncertainty • Model Structure • Parameters • Data • Forecasts of future conditions Conceptualization of Reality Sacramento Effective Parameters and States real world heterog. input output measurement identical Identical ? homog. input output (xeff,eff) model After Grayson and Blöschl, 2000, Cambridge Univ. Press Precipitation Measurement Sparse precip gauge distribution Mountain blockage of radar 3000M 2000M 1000M Source: Maddox, et. al. Weather and Forecasting, 2002. SAHRA – NSF STC Streamflow Measurement Accuracy (USGS) • Excellent – 95% of daily discharges are within 5% of true value • Good – 95% of daily discharges are within 10% of true value • Fair – 95% of daily discharges are within 15% of true value • Poor – Do not meet Fair criteria Different accuracies may be attributed to different parts of a given record Dimensions of Model Evaluation From Wagener 2003, Hydrological Processes Dimensions of Model Evaluation From Wagener 2003, Hydrological Processes Performance Measures Mean (observed vs simulated) Standard deviation (observed vs simulated) Root Mean Square Error (RMSE) Mean Absolute Error (MAE) Performance Measures Coefficient of Determination R2 Not a good measure -High correlations can be achieved for mediocre or poor models Performance Measures Coefficient of Efficiency E Widely used in hydrology Range – infinity to +1.0 Overly sensitive to extreme values Nash and Sutcliffe, 1970, J. of Hydrology Seasonal Variability of Nash-Sutcliffe Efficiency Seasonal Analysis of Daily Nash-Sucliffe Efficiency 01608500 0.9 0.8 0.7 Efficiency 0.6 0.5 SAC 0.4 GR4J 0.3 PRMS 0.2 0.1 0 -0.1 0 2 4 6 8 Month 10 12 14 Performance Measures Index of Agreement d Range 0.0 – 1.0 Overly sensitive to extreme values Willmott, 1981, Physical Geography Analysis of Residuals Performance Measure Issues • Different measures have different sensitivities for different parts of the data. • Are assumptions correct regarding the nature of the error structures (i.e. zero mean, constant variance, normality, independence, …)? • Difficulty in defining what constitutes an acceptable level of performance for different types of data. Dimensions of Model Evaluation From Wagener 2003, Hydrological Processes Definitions • Uncertainty analysis: investigation of the effects of lack of knowledge or potential errors on model components and output. • Sensitivity analysis: the computation of the effect of changes in input values or assumptions on model output. EPA, CREM, 2003 Parameter Sensitivity The single, “best-fit model” assumption Error Propagation Magnitude of Parameter Error 5% %> VAR %> SE soil_moist_max 10% 20% 50% 0.23963 0.95852 3.83408 23.96303 0.11974 0.47812 1.89901 11.33868 0.15243 0.60973 2.43891 15.24316 %> VAR %> SE hamon_coef 0.16210 0.08102 0.10311 0.64840 0.32367 0.41245 2.59359 1.28849 1.64981 16.20993 7.80071 10.31133 %> VAR %> SE ssrcoef_sq 0.07889 0.03944 0.05018 0.31556 0.15766 0.20073 1.26224 0.62914 0.80293 7.88900 3.86963 5.01829 joint 0.32477 1.29908 5.19632 32.47698 Objective Function Selection et baseflow soil moisture routing gw rech gw rech soil moisture et routing Relative Sensitivity SR = ( QPRED / PI) * (PI / QPRED) Relative Sensitivity Analysis Soil available water holding capacity Evapotranspiration coefficient Relative Sensitivity Analysis Snow/rain threshold temperature Snowfall adjustment Parameter Sensitivity The “parameter equifinality” assumption • Consider a population of models • Define the likelihood that they are consistent with the available data Regional Sensitivity Analysis • Apply a random sampling procedure to the parameter space to create parameter sets •Classify the resulting model realizations as “behavioural” (acceptable) or “non-behavioural” •Significant difference between the set of “behavioural” and “non-behavioural” parameters identifies the parameter as sensitive Spear and Hornberger, 1980, WRR Regional Sensitivity Analysis 1 F ( i | B ) F ( i ) 0 F ( i | B ) i B = behavioural Not sensitive cumulative distribution cumulative distribution Sensitive 1 F( j | B ) F( j ) F( j | B ) 0 j B = non-behavioural Generalized Likelihood Uncertainty Analysis (GLUE) • Monte Carlo generated simulations are classified as behavioural or non-behavioural, and the latter are rejected. • The likelihood measures of the behavioural set are scaled and used to weight the predictions associated with individual behavioural parameter sets. • The modeling uncertainty is then propagated into the simulation results as confidence limits of any required percentile. Dotty Plots and Identifiability Analysis behavioural GLUE computed 95%confidence limits Rockies Sierras Objective Function obsQ – predQ Animas Cascades Cle Elum Carson 300000.0 300000.0 300000.0 250000.0 250000.0 250000.0 200000.0 200000.0 200000.0 150000.0 150000.0 150000.0 100000.0 100000.0 100000.0 50000.0 0.0 0.2 0.4 0.6 0.8 50000.0 0.0 0.2 0.4 0.6 0.8 Parameter Equifinality 50000.0 0.0 0.2 0.4 0.6 0.8 rad_trncf 300000.0 300000.0 250000.0 250000.0 200000.0 200000.0 150000.0 150000.0 100000.0 100000.0 50000.0 0.0 5.0 10.0 15.0 300000.0 250000.0 Regional Variability 200000.0 150000.0 100000.0 50000.0 0.0 5.0 10.0 15.0 50000.0 0.0 5.0 10.0 15.0 soil_moist_max (inches) 300000.0 300000.0 300000.0 250000.0 250000.0 250000.0 200000.0 200000.0 200000.0 150000.0 150000.0 150000.0 100000.0 100000.0 100000.0 50000.0 25.0 30.0 35.0 40.0 50000.0 25.0 30.0 35.0 40.0 tmax_allsnow (deg F) 50000.0 25.0 Uncalibrated Estimate 30.0 35.0 40.0 Multi-criteria Analysis Increasing the information content of the data A single objective function: • cannot capture the many performance attributes that an experienced hydrologist might look for • uses only a limited part of the total information content of a hydrograph • when used in calibration it will tend to bias model performance to match a particular aspect of the hydrograph A multi-criteria approach overcomes these problems (Wheater et al., 1986, Gupta et al., 1998, Boyle et al., 2001, Wagener et al., 2001). Identifying Characteristic Behavior Developing Objective Measures FD 1 nD n Q obs Q Q 2 2 2 peaks/timing com D FQ n Q n 1 Q obs com quick recession Q FS 1 ns n Q obs s Q com baseflow Pareto Optimality F D FS FQ FS 500 Pareto Solutions F D F u n c t i o n S p a c e N o r m a l i z e d P a r a m e t e r S p a c e F Q SAC-SMA Hydrograph Range Overall Performance Measures RMSE min BIAS min Parameter Sensitivity by Objective Function Dimensions of Model Evaluation From Wagener 2003, Hydrological Processes Evaluation of Model Component Processes Solar Radiation Observed SCE Final SCE J F M A M J J A Month PET S O N D J Annual Runoff 1991 1993 F M A M J J A Month 1995 Year 1997 1991 1993 1995 Year Nash-Sutcliffe 1993 N D Percent Groundwater Daily Q 1991 S O 1995 Year 1997 Day 1997 Coupling SCA remote sensing products with point measures and modeled SWE to evaluate snow component process East River Integrating Remotely Sensed Data Percent Basin Snowcover 1 0.8 MODEL SATELLITE 0.6 0.4 0.2 0 1/0 1/20 2/9 2/29 3/20 4/9 1996 4/29 5/19 6/8 6/28 7/18 Identifiability Analysis Identification of the model structure and a corresponding parameter set that are most representative of the catchment under investigation, while considering aspects such as modeling objectives and available data. Wagener et al., 2001, Hydrology Earth System Sciences Dynamic Identifiability Analysis - DYNIA Information content by parameter Dynamic Identifiability Analysis - DYNIA Identifiability measure and 90% confidence limits A Wavelet Analysis Strategy • Daily time series Precipitation • Seasonally varying daily variance (row sums) 12 100 11 90 10 80 9 70 8 60 • Seasonally varying variance frequency decomposition (column sums) 7 50 6 40 5 30 4 3 20 2 10 1 1 2 3 4 5 6 7 8 • Annual average variance frequency decomposition John Schaake, NWS 0 Variance Decomposition Precipitation Precipitation OBSQ 12 100 12 11 90 11 10 80 10 70 9 9 8 6 9 0.45 10 0.4 6 7 6 5 6 4 5 4 20 2 10 1 2 3 4 5 6 7 8 3 3 1 1 2 3 8-day Average Precipitation 20 12 11 18 11 16 10 10 9 4 5 Component 6 7 8 1 8 9 7 5 1 6 4 4 5 6 7 8 0 0.5 0.45 10 0.4 0.35 8 6 5 6 0.3 0.25 4 3 2 2 2 1 0 1 0.2 5 0.15 4 4 3 8 3 11 7 4 7 2 12 6 5 6 0.05 1 7 8 5 2 9 8 Month 10 6 4 0.1 8 12 7 3 3 8-day Avg Variance Transfer Function - Observed 10 9 14 2 0.15 4 OBSQ2 12 1 0.2 2 2 0 0.3 0.25 5 3 0.35 8 7 30 4 0.5 11 7 40 5 12 9 8 Month 50 Variance Transfer Function - Observed 10 8 60 7 1 Variance Transfer Functions Streamflow 3 0.1 2 2 0.05 1 1 3 1 2 3 4 5 Component 6 7 8 1 2 3 4 5 6 7 8 0 Streamflow OBSQ Estimated Streamflow - Linear 12 10 12 10 11 9 11 9 10 8 10 8 9 7 8 Month Obs 9 7 8 6 7 6 7 5 5 6 6 4 5 3 4 4 5 3 4 3 2 3 2 2 1 2 1 0 1 10 12 10 9 11 9 8 10 8 1 1 2 3 4 5 Component 6 7 8 1 2 3 ESTQ H 11 10 9 5 6 7 8 9 7 8 0 6 7 5 6 4 5 3 4 7 8 Month Month 4 ESTQ B 12 PRMS 6 7 5 6 4 5 3 4 3 2 3 2 2 1 2 1 0 1 1 1 2 3 4 5 Component Linear 6 7 8 1 2 3 4 5 Component 6 7 8 0 SAC Variance Transfer Functions OBSH Variance Transfer Function - Linear 12 0.5 12 0.5 11 0.45 11 0.45 10 0.4 10 0.4 9 0.35 8 Month Obs 9 0.35 8 0.3 7 0.3 7 0.25 0.25 6 6 0.2 5 0.2 5 0.15 4 0.15 4 3 0.1 3 0.1 2 0.05 2 0.05 0 1 1 1 2 3 4 5 Component 6 7 8 1 2 3 ESTH H 5 6 7 8 0 ESTH B 0.5 12 0.5 11 0.45 11 0.45 0.4 10 0.4 9 9 0.35 8 0.3 7 0.25 6 0.2 5 0.15 4 0.35 8 Month Month 4 12 10 PRMS 0.3 7 0.25 6 0.2 5 0.15 4 3 0.1 3 0.1 2 0.05 2 0.05 0 1 1 1 2 3 4 5 Component Linear 6 7 8 1 2 3 4 5 Component 6 7 8 0 SAC Forecast Uncertainty Ensemble Streamflow Prediction Using history as an analog for the future NOAA USGS Probability of exceedence BOR Simulate to today Predict future using historic data 2005 ESP Forecast ESP Animas River @ Durango 1981 1982 1983 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 12000 8000 6000 4000 2000 9/18/2005 All historic years ESP - Animas River @ Durango 9/18/2005 9/4/2005 8/21/2005 8/7/2005 7/24/2005 7/10/2005 6/26/2005 6/12/2005 5/29/2005 5/15/2005 5/1/2005 9/4/2005 Forecast Period 4/3 – 9/30 Made 4/2/2005 10000 9000 8000 7000 6000 5000 4000 3000 2000 1000 0 4/3/2005 Observed 2005 Streamflow (cfsd) Only el nino years 8/21/2005 8/7/2005 7/24/2005 7/10/2005 6/26/2005 6/12/2005 5/29/2005 5/15/2005 5/1/2005 4/17/2005 4/3/2005 0 4/17/2005 Streamflow (cfsd) 10000 1982 1983 1987 1991 1992 1993 1994 1995 1997 1998 2002 2003 2004 2005 Ranked Probability Skill Score (RPSS) for each forecast day and month using measured runoff and simulated runoff (Animas River, CO) produced using: (1) SDS output and (2) ESP technique RPSS 0.1 0.3 0.5 0.7 0.9 Perfect Forecast: RPSS=1 8 8 6 6 4 4 2 2 0 0 JFMAMJJASOND JFMAMJJASOND Month Month SDS ESP Given current uncertainty in long-term atmosphericmodel forecasts, seasonal to annual forecasts may be better with ESP Summary • This presentation has been a selected review of uncertainty and error analysis techniques. • No single approach provides all the information needed to assess a model. The appropriate mix is a function of model structure, problem objectives, data constraints, and spatial and temporal scales of application. • Still searching for the unified theory of uncertainty analysis. Necessary Conditions for a Model to be Considered Properly Calibrated • Input-output behaviour of the model is consistent with measured behaviour performance • Model predictions are accurate (negligible bias) and precise (prediction uncertainty relatively small) • Model structure and behaviour are consistent with the understanding of reality Gupta, H.V., et al, in review National and international groups are collaborating to assess existing methods and tools for uncertainty analysis and to explore potential avenues for improvement in this area. http://www.es.lancs.ac.uk/hfdg/uncertainty_workshop/uncert_intro.htm A Federal Interagency Working Group is developing a Calibration, Optimization, and Sensitivity and Uncertainty Analysis Toolbox International Workshop Proceedings describes this effort: available at http://www.iscmem.org Future Model Development and Application Coupled Hydrological Modelling Systems Geochemical Flowpaths INPUTS Model Complexity GIS Landuse Aquatic, Riparian & Terrestrial Hydrological Modelling Scale PREDICTIONS Uncertainty Analysis Increased Model Complexity More Parameters More Spatial Interactions More Complex Responses but still data limited …. MORE MODELLING UNCERTAINTY Visual Uncertainty Analysis Framework Modular Modelling System - USGS Deeper Valley Zone Bedrock Depth (m) 5 4 3 2 1 0 Conceptualised Bedrock Model Structures Field Measurements Visualisations of Models & Uncertainty Visualisations of Measurements Improved Representations of Hydrological Processors and Predictions Freer, et al., Lancaster Univ., UK