Insight into hydrologic system complexity through catchment classification framework

The need for simplification in hydrological modelling is present in the literature since some time ago and has lead in the last years to the development of catchment classification studies, with the objective of identifying similarities between catchments as a basis to transfer information and improve knowledge of hydrologic systems. In my research activity, I analysed 46 basins of the Upper Po territory by calculating a lot of characteristics indices; these basins have been classified from two different point of view: from one hand, for each catchment six hydrological signatures (runoff coefficient, baseflow index, streamflow elasticity, median flow, ten percentile flow and slope of the flow duration curve) has been calculated, so that we can have a description of the hydrological behaviour of the catchments; from the other hand, all the basins have been classified also with fifteen climatic, physiographic, and soil usage-type characteristics. The main question is: how long the second classification scheme is useful to identify group of hydrologically similar catchments and so convenient to characterize also ungauged basins? The methodology used for classification purpose is the cluster analysis, in particular a two level approach has been used: first clustering with Self Organizing Maps, that is a type of artificial neural network and unsupervised learning algorithm, and then with hierarchical clustering method that allows to build a clustering tree, dendrogram, useful to interpret data structure and to determine the number of cluster. The number of cluster identified in this case is three. The results obtained from the two different classification scheme were compared through a contingency table and the degree of overlap was 68.5%, this means that in 68.5% cases the second classification well explains the results of the hydrological one. This work would not be a propose for a universal classification system, but tries to discover the potential of catchment classification method in a relatively restricted area, in order also to improve knowledge about the possibility of extrapolation of information up – or downstream along a river network, useful for distributed hydrological models calibration. Starting from the results gained at this point, my research activity focused on some applications that use operatively classification results: improve estimation of regionalization models for flow duration curves, analyse different formulation for the computation of concentration time, improve and make efficient the calibration process of a distributed hydrological model. The first application examines the estimation of regional model for flow duration curves, FDCs, that were estimated using log-normal distribution. The two parameters describing the distribution were estimated through stepwise multiple linear regression. The estimation of parameters done by taking advantage of classification, so by building a different regionalization model for each hydrological class, allowed to improve considerably the results in term of estimation of flow duration curves, maintaining the mean absolute percentage error of estimation lower than 7% in more than a half cases. Concerning the estimation of time of concentration, sixteen equations present in literature and widely used in Italy and in USA, were compared with the observed time of concentration, computed from the data. In this case, the subdivision in hydrologically similar catchments doesn’t improve the estimate of time of concentration, but the main conclusion remains that time of concentration is a an extremely variable and difficult to evaluate parameter. Furthermore, a new equation to estimate time of concentration, was implemented using stepwise multiple linear regression and fitted on the observed data. The advantage of this equation is that produces good results and all the parameters are easily obtainable from a DTM. The last investigated point concern the distributed hydrological model. To calibrate a distributed model, several are the steps to do before reaching good results. First of all, after doing an accurate sensitivity analysis of the model, is crucial to define a rigorous parameterization process, so to reduce over-parameterization: we proceeded by assigning specific values to all parameters, obtained from soil usage and attributable to a physical range of variation. Each cell of the basin has its own value, but the process of calibration is implemented in such a manner that each parameter is corrected by an equal multiplicative correction factor, so to reduce substantially the numbers of parameters to be calibrated. The calibration was done with a trial and error approach. Two different components of the hydrological models were calibrated: the snow model, calibrated by using MODIS images and the soil processes, calibrated by comparison of observed and simulated discharge. In the analysed data set, there are several nested basins; in this case the classification scheme retrieved before can help the user in the calibration process, because it was verified that if basins are nested, so geographically close and member of the same class, then using only one series of streamflow to calibrate the model allows to obtain good results in all basins considered; on the contrary, it was verified that if basins are nested, but member of different classes, then it is fundamental using all the streamflow series available to obtain good results. This approach proved to be useful in finding information about the differences or similarity present in nested, spatially close, catchments and the need of taking or not into account more information than the only ones about the biggest catchment, especially for distributed hydrological models calibration.

Questa tesi affronta la tematica del “catchment classification” e della possibilità di utilizzarne i risultati in modo operativo. L’area di studio consiste in 46 bacini nell’area dell’Alpo Po, questo data-set è stato classificato attraverso tecniche di cluster analysis e sono state individuate tre classi idrologiche e tre classi fisiche, che si corrispondono con una percentuale del 68.5%. I risultati della classificazione sono stati applicati ed utilizzati in due differenti campi: la regionalizzazione di grandezze idrologico/idrauliche e la calibrazione di modelli idrologici distribuiti. Per quanto riguarda il primo dei due argomenti, la classificazione si è rivelata essere un valido supporto per migliorare notevolmente la stima delle curve di durata tramite regionalizzazione, mentre le stessa cosa non può essere affermata per quanto riguarda la stima dei tempi di corrivazione. Per quanto concerne la seconda applicazione, la classificazione ottenuta è risultata uno strumento utile per perseguire un’efficiente procedura di calibrazione e validazione del modello idrologico distribuito attraverso numerose misure di portata.