|Printer friendly version of the page:|
"Project description: spatial pattern evaluation in hydrology"
Print date: Thursday, December 12 2019 - 6:29
Page last modified: June 23, 2016
© SPACE space.geus.dk
Water scarcity, overexploitation and water pollution have increased the pressure on freshwater resources worldwide, resulting in negative impacts on human welfare, deterioration of freshwater ecosystems and limitations for agricultural and bioenergy production. Water management that until now have been supported mainly by simple practical tools, therefore urgently needs to be upgraded to a truly science based water management. On the Danish scene, the Commission on Nature and Agriculture has specified demands for a targeted spatially differentiated agricultural regulation which is not possible with today’s state-of-the-art modeling frameworks.
My vision is that water managers should have access to a new generation of scientific tools that accurately can inform them about the flow and transport of water and solutes in their river basins and aquifers as well as on the impacts caused by human interventions such as changes in land use, agricultural management and climate. SPACE will make a significant contribution towards this goal.
The most advanced hydrological models which can simulate flow and transport processes with fine spatial resolution are promising tools which together with new high resolution spatial datasets from satellites and airborne geophysics can contribute to a leap forward towards such a vision. However, the scientific basis for fully exploiting the potential of these sophisticated tools and datasets is still immature. The key barrier in this respect is that hydrological models traditionally have been evaluated on their ability to simulate aggregated outputs for larger areas (river basins), while there is a lack of scientific methodologies for calibrating and evaluating models based on their ability to reproduce observed spatial patterns within the river basins.
The hypothesis behind the proposal is that:
• A fundamental shift and significant improvement in spatial model evaluation and performance can be achieved by using a new model parameter regionalization and calibration framework based on new spatial performance metrics and remote sensing based observations.
Therefore, the objectives of SPACE are to close the gaps between the current state-of-the-art in spatially distributed hydrological modeling and the requirements for a new generation of intelligent and differentiated water management. This is achieved by:
• Developing a completely new set of model performance criteria that are particularly targeted to evaluate the ability of models to describe the spatial patterns in the hydrological system and to assess the predictive skills of the models at fine spatial scales.
• Developing new satellite remote sensing based datasets of hydrological variables that are unique in their information content, spatial detail and coverage, and specifically targeted for tests of distributed hydrological model predictions at different spatial scales.
• Developing a new model calibration and evaluation framework by combining distributed model parameterization and regionalization with calibration techniques such as joint inversion and super-parameter optimization based on á priori spatial distributions. And testing the hypothesis by using the new framework together with the new performance metrics and the new remote sensing data.
Hydrological modeling encompasses the numerical description of water (and energy) fluxes as they progress through the hydrological cycle: a discipline that has expanded significantly with the increases in computational power over recent decades. Hydrological models range from simple conceptual lumped approaches to very complex physically based spatially distributed systems that explicitly account for the water exchanges between all hydrological components, including ground water, surface water and land surface processes (e.g. evapotranspiration, snow melt and overland flow) (Wood et al., 2011).
Traditionally, hydrological models have been calibrated and evaluated against point observations of stream flow, that aggregate the catchment scale hydrological processes to a single effective output, summing up all internal processes and variability (Beven and Binley, 1992; Refsgaard, 2000; Stisen et al., 2011a). State-of-the-art hydrological models provide spatially distributed outputs that simulate all relevant fluxes and model states at the grid scale. However, these complex models are still only confronted with spatially lumped observations for their calibration and validation. This is highly unsatisfying and does not honor neither the complexity of the models nor the applications they are intended for.
Previously, complex distributed models suffered severely from both computational efficiency and a lack of input data and distributed data for validating their spatial performance (Beven, 2000; Beven, 2001; Grayson and Blöschl, 2000; Refsgaard, 1997). However, with the emergence of super computers and especially satellite remote sensing and airborne geophysics, the limiting factor for bringing forward distributed hydrological modeling is a lack of methodologies for calibrating and validating the distributed models against truly distributed observations. This requires a new way of thinking with regards to the type of observations that are used and the way the models are parameterized and optimized. Aggregated catchment observations, such as stream flow are still important, but they cannot stand alone if we are to meet the requirements of modern water managers and administrators who require much more detailed information of where and when things happen within a catchment. The hydrological model codes, the data and the computational power are all available today – the methodologies for spatial calibration and validation are not. This project will develop new performance metrics and model optimizations techniques to address this model-evaluation divide.
The research within the proposed project will contribute to the development of spatial pattern evaluation of distributed hydrological models by combining three lines of research:
• WP1 will develop a set of new performance metrics that are specially designed to enable a spatial evaluation of the patterns simulated by grid-based distributed hydrological models.
• WP2 will develop, implement and validate remote sensing based estimates of hydrological states and fluxes, such as evapotranspiration, soil moisture and land surface temperature.
• WP3 will a framework for parameter regionalization and optimization that will allow for improved spatial model calibration and evaluation, while limiting the dimensionality of the parameter space.
One of the main challenges for integration of spatial data and hydrological models is the lack of suitable metrics to evaluate similarity of spatial patterns, as expressed in several reviews on the topic (Grayson et al., 2002; Wealands et al., 2005). Unfortunately, comparison of spatial patterns is not straight forward. Simple statistical metrics such as mean, bias and correlation length provide a global comparison but include little pattern information. Local comparisons, such as correlation coefficients, are better suited for assessing the spatial pattern. However, the restriction on geographical precision is often too high, especially in heterogeneous landscapes. The insufficiency of simple statistical metrics to compare spatial patterns in hydrology is caused by the complexity of the spatial field simulated by a distributed model. Positional errors in simulated patterns originate from a range of sources such as georeferencing, model parameterization, model structure, forcing data quality and interpolation (Stisen et al., 2011a). Consequently, errors in simulated patterns are a combination of static and dynamic errors. This leads to a need for flexible performance metrics. Furthermore, spatial fields in hydrology are typically continuous and heterogeneous, limiting the application of techniques generally used for assessment of classification performance (Wealands et al., 2005).
It is not expected that a single metric will describe the quality of spatial patterns exhaustively; hence the goal is to derive a set of complementary metrics. Classical geostatistical methods such as variogram analysis will be combined with investigation of new pattern matching techniques, known from fingerprint and face recognition analysis, and statistical techniques such as multiple-point statistics (Mariethoz et al., 2009). Modification of the multiple-point algorithm is required to create a dissimilarity matrix, giving scores related to the similarity in observed and simulated spatial patterns.
Several studies have demonstrated the application of satellite remote sensing data for optimization of hydrological models. These studies vary in scale and methodology, with some focusing on the temporal dynamic in single model grids (Coudert et al., 2006) while others work on larger scale with quantitative remote sensing estimates (Immerzeel and Droogers, 2008; Pan et al., 2008) or with spatial patterns obtained from remote sensing data (Franssen et al., 2008; Li et al., 2009; McCabe et al., 2005; Stisen et al., 2011a).
Remote sensing data is without doubt one of the key sources for spatially distributed observations (Krajewski et al., 2006), both for parameterizing hydrological models (Stisen et al 2008) and for data calibration and validation (Immerzeel and Droogers, 2008). Especially the observed pattern in satellite remote sensing data gives a unique spatial information that cannot be obtained other data sources. Satellite remote sensing data are however only proxies for physical hydrological variables such as soil moisture and evapotranspiration.
Today a wide range of methods exists for the retrieval of hydrological variables from remote sensing. These methods (Anderson et al., 1997; Nieto et al., 2011; Su, 2002) will be developed further and tested rigorously through comparison with data from intensely monitored research catchments (Jensen and Illangasekare, 2011) and by evaluation of the overall water balances at the catchment scale. Since remote sensing data by nature are instantaneous snapshots in time, this evaluation will require the development of methods for aggregating the estimates to monthly or annual means and ensure that hydrological consistency between various datasets is maintained (McCabe et al., 2008). Accurate knowledge of possible biases is extremely important when the remote sensing estimates are subsequently utilized as calibration targets in model optimization.
This work package will build on the two previous activities by exploiting the new performance metrics and the remotely sensed dataset for both driving variables and spatial evaluation targets. In addition, the work will be based on the vast number of distributed hydrological models that are available through previous research projects, including national scale water resources models, detailed research catchment models and large scale models in data scarce developing countries.
The challenges associated with spatial model calibration forms an active field of research which is developing in parallel with data assimilation of spatial observation data. In data assimilation (Houser et al., 1998) the simulated state variables are updated by observations in order to improve the predictive skills of the model in the next time step with the aim of hydrological forecasting (Cheng et al., 2011; Vrugt et al., 2006). The consequence is a violation of the mass balance and the approach is not likely to increase our understanding of the underlying hydrological processes. Within water resources modeling, where the models are utilized for water resources assessment and scenario modeling, it is necessary to preserve the mass balance and accept some deviations from the observations. The task at hand then becomes to minimize these deviations through evaluation of the model input data, model structure, model parameter values and not least the spatial distribution of the model parameters (Hojberg and Refsgaard, 2005).
The last element is of particular interest when integrating spatial observations with distributed models. The latest development in the field is described in a series of recent research papers on parameter regionalization and utilization of observed spatial patterns in model evaluation (Brunner et al., 2012; Gupta et al., 2009; Gupta et al., 2012; Kumar et al., 2013; Livneh and Lettenmaier, 2013; Pokhrel et al., 2008; Pokhrel and Gupta, 2010; Samaniego et al., 2010; Stisen et al., 2011a; Yilmaz et al., 2008). The proposed research will built on these findings and together with signature based approaches (Gupta et al., 2008) and pattern-oriented modelling approaches (Grimm et al., 2005) from other fields such as ecology, they will form the basis for developing new techniques which are particularly tailored to the high degree of complexity present in physically based hydrological model codes.
A fundamental idea behind a new parameter regionalization and optimization approach is to return to the “raw” measurements that usually form the basis for initial model parameter values and their spatial distribution. The raw measurements and their spatial pattern can be used as á priori spatial distributions, while the conversion from raw data to the physical properties that the hydrological model requires, is included in the parameter optimization. This is also referred to as joint inversion (Kowalsky et al., 2011), where two models with interdependent parameters are calibrated simultaneously. Examples of this could be pedo-transfer functions for estimating soil parameters based on simple soil texture data and remote sensing algorithms for estimating vegetation and land surface parameters based on measured reflectances.