Data-driven catchment classification: application to the PUB problem

Abstract. In many real-world flood forecasting systems, the runoff thresholds for activating warnings or mitigation measures correspond to the flow peaks with a given return period (often 2 years, which may be associated with the bankfull discharge). At locations where the historical streamflow records are absent or very limited, the threshold can be estimated with regionally derived empirical relationships between catchment descriptors and the desired flood quantile. Whatever the function form, such models are generally parameterised by minimising the mean square error, which assigns equal importance to overprediction or underprediction errors.Considering that the consequences of an overestimated warning threshold (leading to the risk of missing alarms) generally have a much lower level of acceptance than those of an underestimated threshold (leading to the issuance of false alarms), the present work proposes to parameterise the regression model through an asymmetric error function, which penalises the overpredictions more.The estimates by models (feedforward neural networks) with increasing degree of asymmetry are compared with those of a traditional, symmetrically trained network, in a rigorous cross-validation experiment referred to a database of catchments covering the country of Italy. The analysis shows that the use of the asymmetric error function can substantially reduce the number and extent of overestimation errors, if compared to the use of the traditional square errors. Of course such reduction is at the expense of increasing underestimation errors, but the overall accurateness is still acceptable and the results illustrate the potential value of choosing an asymmetric error function when the consequences of missed alarms are more severe than those of false alarms.

Section: Identification Of Balanced Cross-validation Subsets With Sommentioning

confidence: 99%

Estimation of flood warning runoff thresholds in ungauged basins with asymmetric error functions

Toth

2016

“…In the recent years, also non-supervised neural networks, and in particular of the SOM (self-organising mapping) type, were successfully applied (and sometimes compared with other methods such as K-means or Fuzzy C-means) for catchments classification purposes (Hall and Minns, 1999;Hall et al, 2002;Jingyi and Hall, 2004;Chang et al, 2008;Srinivas et al, 2008;Di Prinzio et al, 2011;Ley et al, 2011). SOM-type neural networks learn to cluster the input data by recognizing different patterns organising the data on the basis of their similarity, quantified by means of a distance measure (in the present case, like in the majority of applications, the Euclidean distance).…”

Section: Classification Of Streamflow Signatures With Som Neural Netwmentioning

confidence: 99%

Catchment classification based on characterisation of streamflow and precipitation time series

Toth

2013

Self Cite

Abstract. The formulation of objective procedures for the delineation of homogeneous groups of catchments is a fundamental issue in both operational and research hydrology. For assessing catchment similarity, a variety of hydrological information may be considered; in this paper, gauged sites are characterised by a set of streamflow signatures that include a representation, albeit simplified, of the properties of fine time-scale flow series and in particular of the dynamic components of the data, in order to keep into account the sequential order and the stochastic nature of the streamflow process.The streamflow signatures are provided in input to a clustering algorithm based on unsupervised SOM neural networks, obtaining groups of catchments with a clear hydrological distinctiveness, as highlighted by the identification of the main patterns of the input variables in the different classes and the interpretation of their interrelations. In addition, even if no geographical, morphological nor climatological information is provided in input to the SOM network, the clusters exhibit an overall consistency as far as location, altitude and precipitation regime are concerned.In order to assign ungauged sites to such groups, the catchments are represented through a parsimonious set of morphometric and pluviometric variables, including also indexes that attempt to synthesise the variability and correlation properties of the precipitation time series, thus providing information on the type of weather forcing that is specific to each basin. Following a principal components analysis, needed for synthesizing and better understanding the morpho-pluviometric catchment properties, a discriminant analysis finally assigns the ungauged catchments, through a leave-one-out cross validation, to one of the above identified hydrologic response classes. The approach delivers a quite satisfactory identification of the membership of ungauged catchments to the streamflow-based classes, since the comparison of the two cluster sets shows a misclassification rate of around 20 %.Overall results indicate that the inclusion of information on the properties of the fine time-scale streamflow and rainfall time series may be a promising way for better representing the hydrologic and climatic character of the study catchments.

“…However, some aspects of RFFA have still yet to be resolved and further research could result in substantial improvements to predictions of the design flood in ungauged catchments (Castiglioni et al, 2011). Notably, the determination and use of hydrologically homogenous groups is one such topic (see e.g., McDonnell and Woods, 2004;Di Prinzio et al, 2011;Castiglioni et al, 2011) and is the focus of this paper.…”

Section: S a Archfield Et Al: Kriging Techniques For Design Flood mentioning

confidence: 99%

“…The empirical values of the quantity of interest (e.g., empirical flood quantiles associated with a given return period T ) can be represented along the third dimension z for each gauged catchment, and can then be interpolated via kriging to estimate it at ungauged sites lying within the same portion of the physiographical space. The term canonical kriging originates from the procedure that is generally adopted to define x and y, that is canonical correlation analysis, or CCA (see e.g., Ouarda et al, 2001;Chokmani and Ouarda, 2004;Di Prinzio et al, 2011). CCA is an important multivariate statistical tool for reducing the dimensionality of an original data set.…”

Section: Canonical Krigingmentioning

confidence: 99%

Topological and canonical kriging for design flood prediction in ungauged catchments: an improvement over a traditional regional regression approach?

Archfield

Pugliese

Castellarin

et al. 2013

Self Cite

Abstract. In the United States, estimation of flood frequency quantiles at ungauged locations has been largely based on regional regression techniques that relate measurable catchment descriptors to flood quantiles. More recently, spatial interpolation techniques of point data have been shown to be effective for predicting streamflow statistics (i.e., flood flows and low-flow indices) in ungauged catchments. Literature reports successful applications of two techniques, canonical kriging, CK (or physiographical-space-based interpolation, PSBI), and topological kriging, TK (or top-kriging). CK performs the spatial interpolation of the streamflow statistic of interest in the two-dimensional space of catchment descriptors. TK predicts the streamflow statistic along river networks taking both the catchment area and nested nature of catchments into account. It is of interest to understand how these spatial interpolation methods compare with generalized least squares (GLS) regression, one of the most common approaches to estimate flood quantiles at ungauged locations. By means of a leave-one-out cross-validation procedure, the performance of CK and TK was compared to GLS regression equations developed for the prediction of 10, 50, 100 and 500 yr floods for 61 streamgauges in the southeast United States. TK substantially outperforms GLS and CK for the study area, particularly for large catchments. The performance of TK over GLS highlights an important distinction between the treatments of spatial correlation when using regression-based or spatial interpolation methods to estimate flood quantiles at ungauged locations. The analysis also shows that coupling TK with CK slightly improves the performance of TK; however, the improvement is marginal when compared to the improvement in performance over GLS.