H42D-01 INVITED
Assimilation of hydrologic and hydrometeorological data into lumped and distributed hydrologic models for improved operational hydrologic forecasting in NWS
Despite the long history and the renewed interest in recent years, automatic data assimilation in operational hydrology is yet to find widespread acceptance and use at the National Weather Service (NWS) River Forecast Centers (RFCs). To meet the new service needs for uncertainty-quantified high-resolution soil moisture and streamflow information, however, a new paradigm is necessary for operational data assimilation that fully capitalizes on automatic techniques and fast-advancing computing power, and that recognizes, and takes full advantage of, the role of human forecasters in the forecast process. Toward that goal, the NWS Office of Hydrologic Development in collaboration with RFCs and other partners is carrying out a number of data assimilation and related projects. In this talk, we present an overview of these activities in the context of hydrologic ensemble prediction, describe in some detail research, development and research-to-operations transition activities for automatic assimilation of streamflow, soil moisture, precipitation and potential evaporation into lumped and distributed soil moisture accounting and routing models, present results, and identify challenges in assimilating hydrologic and hydrometeorological data toward improving operational hydrologic forecasting.
H42D-02
HYDROLOGICAL DATA ASSIMILATION WITH THE ENSEMBLE KALMAN FILTER: USE OF STREAMFLOW OBSERVATIONS TO UPDATE STATES IN A DISTRIBUTED HYDROLOGICAL MODEL
This paper describes an application of the ensemble Kalman filter (EnKF) in which streamflow observations are used to update states in a distributed hydrological model. We demonstrate that the standard implementation of the EnKF is inappropriate because streamflow is not Normally distributed. Transforming streamflow into log space before computing error covariances improves filter performance. We also demonstrate that model simulations improve when we use a variant of the EnKF that does not require perturbed observations. Our attempt to propagate information to neighbouring basins was unsuccessful, largely due to inadequacies in modeling the spatial variability of hydrological processes. New methods are needed to produce ensemble simulations that both reflect total model error and adequately simulate the spatial variability of hydrological states and fluxes.
H42D-03
Ensemble Streamflow Forecasting via data assimilation
Accurate streamflow forecasting is required for effective water resources management. The present research focuses on development of a streamflow forecasting methodology through a data-assimilation technique, known as Ensemble Kalman filter (EnKF). A spatially distributed hydrologic model is forced with an ensemble of meteorological inputs, thus producing an ensemble of streamflow forecasts. A digital elevation model along with the temperature and precipitation lapse rates are used to represent the spatial variability for the input meteorology within the watershed. The model structural uncertainty is handled by sampling parameters from a feasible parameter space obtained through the calibration process. The model snow states are updated using SNODAS Snow Water Equivalent (SWE) estimates, where as the soil moisture states are updated using ground-based streamflow observations, within the EnKF framework. The methodology is tested in watersheds in the Pacific Northwest. The short-term and seasonal streamflow forecasts are compared to the observations and the currently employed regression based approaches.
H42D-04
Improving Convective Precipitation Forecasting through Assimilation of Data-Forced Land Surface Parameters
Ongoing research investigates the potential of advancing our understanding on land-atmosphere interaction processes, e.g., soil state and precipitation feedbacks, through data assimilation of ground-based and remotely sensed data into atmospheric and land surface models. In this study, we examine the impact of assimilating radar rainfall data in the land surface scheme of an atmospheric mesoscale model as forcing term instead of the model-generated fields. The experimental design is based on a 2-day-long mesoscale convective system (central North America, July 2004), including multiple 84-hour runs that cover a variety of assimilation periods. The sensitivity of the model's performance in varying soil state conditions is initially verified through continuous (84-hour) data assimilation. Shorter time assimilation periods (12-, 24-, and 36-hour) prior to the 48-hour storm event are then utilized to assess the effectiveness of the technique for improving convective precipitation forecasting. In most cases, assimilation of radar rainfall data brings the simulated precipitation fields closer to the observed ones, as compared to the control simulation. The potential of identifying the most suitable time duration for implementing the assimilation technique based on the criteria of best forecasting performance and shortest assimilation period is further discussed.
H42D-05 INVITED
Ensemble Data Assimilation to Sequential Bayesian Multi-model Combination: Tackling the Hydrologic Model Uncertainty
The key step to enhance the accuracy of hydrologic prediction is the knowledge and realistic characterization of uncertainty. A method that has recently garnered the attention of researchers and practitioners is data assimilation (DA) aiming to improve the model's predictive skills and to explicitly characterize the uncertainty in water and energy balance computations. Although successful, DA methods are limited to the single model ignoring other plausible models. Analysis of predictive uncertainty in land surface fluxes and storages based on single hydrologic model are prone to systematic bias and underestimation of uncertainty. This results to overconfidence in model predictive capabilities even with the advanced data assimilation technique available to date. This motivates to employ various competitive models and use a combination technique that takes the most benefit from models for predicting the quantity of interest. Bayesian Model Averaging (BMA) has recently been used in few hydrologic prediction studies as an effective multi-model combination technique. However, BMA method is limited to the Gaussian likelihood assumption of individual model predictive distributions which may not result to optimum models combination and accurate prediction. In addition, BMA is limited to the fixed model weights ignoring the possibility that some models may behave differently at different periods of simulation/prediction owing to the merits of models for capturing the physics. To overcome these limitations two methods are proposed: (1) a sequential Bayesian multi-model combination method for prediction which is not limited to the fixed model weight as opposed to the BMA method, (2) a procedure for blending the strength of sequential data assimilation (using the particle filter) and multi-model combination.
H42D-06
Model Structure Identification and Correction Through Data Assimilation
The physical laws governing water movement at small scales have been understood for decades. What we don't understand well is how to apply these physical laws to systems that are complex and heterogeneous on all scales. To date, most ‘physically based' models of hydrologic systems are based on an implicit up-scaling premise that the behavior at the model scale can be described by the small scale governing equations by spatial averaging of the state variables and by use of ‘effective' parameters. Of course, the up-scaling assumption may be wrong, and the effective large scale governing equations for a heterogeneous system may be different in form, not just different in parameters, from the equations derived from small-scale physics. We suppose that there is a conceptual model of a hydrologic system; i.e. the major processes and their interconnections have been identified. We wish to know if it is possible to construct the mathematical relationships in question (or correct them) via data assimilation, using measurements made on the system inputs and outputs. Our approach is based upon the construction of a ‘posterior' joint probability density functions for the relationships in question, in such a way that data assimilation helps to correct ‘prior' belief about the dependences. In regions where no data are available the ‘prior' knowledge dominates. The approach permits a representation of, and discrimination between, all three sources of uncertainty: initial conditions, input and structure uncertainty, and is illustrated using case studies.
H42D-07
Distributed Soil Moisture Estimation in a Mountainous Semiarid Basin: Constraining Soil Parameter Uncertainty through Field Studies
A common practice in distributed hydrological modeling is to assign soil hydraulic properties based on coarse textural datasets. For semiarid regions with poor soil information, the performance of a model can be severely constrained due to the high model sensitivity to near-surface soil characteristics. Neglecting the uncertainty in soil hydraulic properties, their spatial variation and their naturally-occurring horizonation can potentially affect the modeled hydrological response. In this study, we investigate such effects using the TIN-based Real-time Integrated Basin Simulator (tRIBS) applied to the mid-sized (100 km2) Sierra Los Locos watershed in northern Sonora, Mexico. The Sierra Los Locos basin is characterized by complex mountainous terrain leading to topographic organization of soil characteristics and ecosystem distributions. We focus on simulations during the 2004 North American Monsoon Experiment (NAME) when intensive soil moisture measurements and aircraft- based soil moisture retrievals are available in the basin. Our experiments focus on soil moisture comparisons at the point, topographic transect and basin scales using a range of different soil characterizations. We compare the distributed soil moisture estimates obtained using (1) a deterministic simulation based on soil texture from coarse soil maps, (2) a set of ensemble simulations that capture soil parameter uncertainty and their spatial distribution, and (3) a set of simulations that conditions the ensemble on recent soil profile measurements. Uncertainties considered in near-surface soil characterization provide insights into their influence on the modeled uncertainty, into the value of soil profile observations, and into effective use of on-going field observations for constraining the soil moisture response uncertainty.
H42D-08
Improving environmental model diagnostic techniques – Development of a timestep-based performance measure for hydrological models
The idea that a number of different models may model the observed data equally well is not a new one (see discussion in Beven, 2001 and Beven and Freer, 2001). This is the concept known as equifinality. This idea was developed into the Generalised Likelihood Uncertainty Estimation (GLUE) of Beven and Binley, (1992). Beven, (2006) commented on the need to improve the equifinality technique by defining levels of acceptability for model predictions; this issue is addressed in the present study. Traditionally models have been analysed using a performance measures, such as Root Mean Squared Error (RMSE). These measures only take into account the modelled output against the observed output for the entire data range. Whilst these measures can be very useful in giving an overall picture of modelled performance, they cannot give information about model performance at individual time steps or acknowledge observational error in a meaningful way. By contrast, a performance measure based on individual time steps will give information allowing the unknown observed values to be reconstructed from an ensemble of different model estimates. In this paper a model is used to give predictions of flow at a catchment gauge using real data. The results are then analysed using the method created in this paper, an extension of the GLUE procedure. The study area, data and model are presented as is an outline of the methodology created. The paper suggests that we need to be more thoughtful about errors in our observations and be inclusive to these when we evaluate our models. K. Beven. A manifesto for the equifinality thesis. Journal of Hydrology, 320(1-2):18-36, 2006. K.J. Beven. Rainfall-Runoff Modelling: The Primer. John Wiley & Sons, Chichester, 2001 K.J. Beven and A.M. Binley. The future of distributed models - model calibration and uncertainty prediction. Hydrological Processes, 6(3):279-298, 1992. K. Beven and J. Freer. Equifinality, data assimilation, and uncertainty estimation in mechanistic modelling of complex environmental systems using the GLUE methodology. Journal of Hydrology, 249(1-4):11-29, 2001.