Of data and models

Abstract

Relationship between the data, such as direct observations of nature and recorded measurements, and the models is very complicated in the ‘water domain’. It is not at all as clear and explicit as it is often presented by teachers to students, by consultants to clients, or by authors to readers of publications. A number of aspects of this relationship are discussed using examples to illustrate the author's views. Limitations of data-driven tools (correlations, Artificial Neuronal Networks, Genetic Algorithms, etc.) and data-mining, when applied without physical knowledge of the relevant phenomena, are discussed, as are those of deterministic models. The currently used ‘good practice’ paradigm in modelling (the model is to be set up, calibrated, validated and run) is rejected when deterministic models are concerned. They should not be calibrated. A new paradigm, a new ‘code of good practice’, is proposed instead. Strategic and tactical aspects of various available approaches to modelling of physical phenomena and data exploitation have practical engineering and financial consequences, most often immediate and sometimes very important: hence the significance of the subject that concerns the everyday occupations of modellers, their clients and end-users.

Flood-plain mapping: a critical discussion of deterministic and probabilistic approaches

“…channel friction much higher than flood-plain friction; Hunter et al, 2006), while, if some ground-truth high water marks were available, a different prediction would have been achieved. In this context, it must be stated that some hydrologists argue that roughness coefficients should be assessed using engineering judgement, and that physically implausible roughness values should be used as the evidence that the model does not reproduce reality (Cunge, 2003). This might be correct under the assumption that all data are error free and the model structure is perfect, and that the point roughness values derived by observation can adequately reflect the spatial variability in momentum losses on a heterogeneous flood-plain which affects effective roughness values at the model discretization scale.…”

“…This might lead to an incorrect assessment of hazard when the inundation maps are used for other purposes, such as planning decisions for future developments in the vicinity of the flood plain. Thus, conceiving inundation hazard as a probability has been encouraged more recently (Romanowicz & Beven, 1996, 2003Aronica et al, 1998Aronica et al, , 2002Bates et al, 2004;Hall et al, 2005;Pappenberger et al, 2005. In a probabilistic approach, flood-plain mapping generally consists of: construction of flood inundation models; sensitivity analysis of the model using historical flood data; and use of the multiple behavioural (acceptable) models to carry out ensemble simulation using an uncertain synthetic design event as hydrological input (e.g.…”

Point Cloud Data Fusion for Enhancing 2D Urban Flood Modelling

“…The models should be parameterised using engineering judgement informed by experience. Such issue, however, is still debated in the literature (Beven, 2000;Cunge, 2003). For example, the parameterization of bed friction is a much more important issue, because flow predictions (velocity and flood wave celerity) could be crucially depending on friction parameter values.…”

Application of artificial neural networks for modeling of biohydrogen production

“…Although ANN models may be successfully applied in biohydrogen production systems and can capture effectively the nonlinear relationships existing between variables in complex systems like fermentative biohydrogen production, one of the main limitations of ANN is the uncertainty of outputs prediction outside the data range, used in establishing the model [17,18].…”