The uncertainty of hydrological simulation as an information deficit and the disentanglement of its components using information theory

Summarise with VIRAT
Authors: Evangelos Findanis, Ευάγγελος Φινδανής
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The disentanglement of the components of hydrological uncertainty and the reduction of their magnitude is considered by the hydrological community to be an unsolved problem of hydrology. In the present dissertation, this problem is examined using information theory in order to analyze hydrological uncertainty into its fundamental measurable components. This leads to the following research question, which this dissertation attempts to address: How can each component of hydrological uncertainty be mathematically equated with the lack of its corresponding information so that it becomes a measurable quantity? Thus, uncertainty is regarded as a gap in knowledge (information) within hydrological simulation, and depending on the position of this gap, different sources of uncertainty arise. The aleatory uncertainty of an observed dataset represents the difference between the available (useful) information in the set and the required information needed for runoff simulation with a selected level of accuracy. Correspondingly, the epistemic uncertainty of a simulation is defined as the difference between the available information and the information explained by the model. Therefore, computational techniques must be developed for the quantities of required, available, and explainable information, which are linear combinations of differential entropies of one, two, or multiple variables. In total, seven (7) computational routines-estimators of differential entropies were designed, whose operation is based on partitioning a domain into equiprobable one-dimensional or two-dimensional bins. In the case of multivariate differential entropies, independent component analysis is conducted so that the computational problem is reduced to estimating several one-dimensional differential entropies. Since the calculation of available information for an observed dataset displays computational instability due to the multiple independent component analyses involved, it must be stabilized. For this reason, the concept of model capacity was introduced, which corresponds to the minimum possible value of available information. At the same time, by minimizing the total mutual information shared by the generated one-dimensional signals of an independent component analysis, the quality of each analysis is assessed, and unsuitable solutions are discarded. Code implementing the above computational procedures was developed. Furthermore, a new theoretical framework was proposed, in which the epistemic uncertainty of a model is separated into its structural and parametric uncertainty. The concept of model capacity is involved here as well, since apart from constituting the lower bound of available information, it also expresses the maximum potential performance of the model. Based on this concept, the data-capacity curves are defined, enabling comparison of the structure of two models without performing calibration and validation. Concerning parametric uncertainty, it equals the reduction of information produced (explained) by a model when its parameters are set to a selected parameter set. Parametric uncertainty can thus be used to evaluate the parameter set resulting from calibration. Additionally, a new likelihood measure based on parametric uncertainty is developed for constructing confidence intervals of simulated runoff with the GLUE method. Therefore, it becomes clear that the proposed framework integrates elements of information theory with the calibration of hydrological models. Hence, fundamental principles of optimization, calibration, and genetic algorithms are developed in the body of the thesis. In particular, the basic operators that constitute a genetic algorithm are presented along with their variants. Concurrently, a simple elitist genetic algorithm is constructed, which will be used for model calibration and for searching for the parameter set that maximizes the explainable information of a model. Due to the extensive computational requirements, a computational code was developed that combines the optimization genetic algorithm with the hydrological models to be calibrated and with the algorithms for computing and diagnosing uncertainty. The proposed theoretical informational quantities are applied to synthetic and real hydrological basins, where they are compared with “conventional” evaluation measures to confirm the consistency of the former. In the synthetic basins, simple single-equation models with controlled structural error are applied. Moreover, the pseudo-observed data of the synthetic basins are generated from the normal or lognormal distribution, and controlled random error is introduced in them. It occurs that the informational quantities of available information and maximum capacity, when applied to these basins, are affected respectively only by random and structural error, in an intuitively expected manner. In contrast, traditional performance indices, such as Nash-Sutcliffe Efficiency and Kling-Gupta Efficiency, are affected by both structural and random errors, making them unsuitable for separating uncertainty into aleatory, structural, and parametric components. Next, four applications of the proposed framework are carried out in real hydrological basins in Greece and Cyprus. Specifically, applications are conducted for the Mouzaki, Pyli, and Sarakina basins of the Pinios River, for the Mesochora and Sykia basins of the Achelous River, and for the Yermasogia basin in Cyprus. In the first application, the values of available information contained in area-weighted datasets resulting from different techniques are compared. Furthermore, by maximizing the available information, for each basin, a new areal distribution of the weights of each rain gauge is derived. For each areal rainfall distribution, the NSE performance of all selected hydrological models is computed in order to examine the existence of a relationship between available information and NSE performance. The second application focuses on estimating uncertainty in monthly and daily lumped and semi-distributed models. Random normal error with a selected standard deviation is introduced into the area-weighted rainfall time series. Then, for different magnitudes of random error, aleatory, epistemic, structural, and parametric components of uncertainty are estimated in the runoff simulation using five (5) monthly lumped, three (3) daily lumped, and two (2) daily semi-distributed models. Thus, the relationship among the magnitude of introduced random errors, the informational quantities, and the conventional performance criteria of the models is investigated. At the same time, the data–capacity curves are plotted for the monthly lumped models and assessed to determine whether comparing these curves leads to correct conclusions regarding the relative performance of their structures. The third application examines the computational accuracy of the routine used to estimate the available information of observed data. Two versions of this routine have been developed. The first is simpler and faster because it takes the average of the solutions from multiple independent component analyses. In the second routine, the solution with the lowest total mutual information among the signals is selected. Thus, theoretically more accurate estimates of available information are produced, and the standard deviation of the estimated values is reduced. Application of these two routines to the Pinios and Achelous basins indeed leads to the conclusion that the second, more complex routine has higher accuracy in computing available information but requires a higher computational burden. In the fourth application, confidence intervals of simulated runoff are generated using the GLUE method, applying the proposed likelihood measure based on informational parametric uncertainty. The proposed measure is compared with three (3) other conventional likelihood measures used as benchmarks for evaluating its performance. This is achieved by counting the number of time steps for which the observed runoff falls within the confidence interval of the simulated runoff. The application is conducted for monthly and daily lumped and semi-distributed models. The conclusion is that the proposed measure is equivalent to more traditional likelihood measures while requiring fewer arbitrary assumptions. Thus, it reduces subjectivity stemming from the hydrologist’s choices during the analysis of parametric uncertainty in simulated runoff.From the aforementioned applications, we conclude that the proposed concepts succeed in evaluating the quality of information contained in observed data within a runoff basin. Models tend to perform better when more information flows into them, while it appears that a nonlinear relationship exists between informational quantities and statistical performance measures of a model. Moreover, data-capacity curves prove to be a powerful tool for model comparison: they allow identification of the simplest model that provides adequate performance for a given level of available information. In summary, the dissertation contributes to the understanding and quantification of uncertainty in hydrological simulation, offering a mathematically coherent framework based on information theory. The framework enables the disentanglement of uncertainty sources, the comparative evaluation of hydrological models, and the assessment of the quality of observed data, contributing to future developments in the field.

Fields of Study & Subjects

Hydrology and Watershed Management Studies Environmental Monitoring and Data Management Meteorological Phenomena and Simulations

Publication details

Journal / source
National Archive of Doctoral Theses (National Documentation Center (Greece))
Publication date
1 January 2026
Language
EL
Document type
Thesis