Complex network based catchment classification using Canberra distance metric obtained for lumped catchment characteristics

Modelling of hydrological processes is a challenging task due to two main reasons: i) hydrological processes are complex and ii) our inability to collect data for understanding hydrological processes is very limited. In particular, it is difficult to predict streamflow in an ungauged region, i. e. a region which has no past streamflow data. As a result, prediction in ungauged catchments often involves complex hydrological model with multiple parameters. The traditional solution is to transfer model parameters from gauged catchments to ungauged catchments, which is known ‘regionalization’. However, the success of a regionalization method depends mainly the degree of similarity between the gauged catchments and the ungauged catchments. That means parameter transfer is meaningful if the gauged and the ungauged catchments are hydrological similar. Therefore, many methods have been proposed in the recent past to classify catchments into hydrologically similar groups by considering several factors like mean rainfall, mean potential evapotranspiration and mean elevation of the catchments. Since no prior information is available with us in general, most of the catchment classification practices are unsupervised. \nA major limitation of the traditional unsupervised classification methods like K-means clustering algorithm is that they are not vey suitable when the classes are not well separated from each other. Furthermore, the traditional methods cannot determine the number of catchments in a dataset automatically. To overcome these limitations, many studies recently have attempted to use more efficient catchment classification methods. This study in particular concerns about catchment classification with the help of multilevel modularity optimization (MMO), which is a community structure complex networks algorithm. It operates by first identifying links between catchments and their strengths (link weights). Although this algorithm is not few in catchment classification exercise, so far it has been used with time series data, e.g., discharge time series data, particularly by computing coefficient of determination values between time series data from different catchments. In otherwise, the existing MMO \nvii \nbased method is not applicable for prediction in ungauged catchments as it cannot handle lumped data like mean rainfall. \nThis study attempts to overcome the above limitation by adopting ‘Canberra distance’ metric, which is used to determine link strengths between catchments with lumped data. To demonstrate the proposed method, we have used several lumped characteristics for 500 catchments situated in United States. Furthermore, the results obtained from the proposed catchment classification method is compared with those obtained from K-means clustering algorithm by forcing the K-means clustering algorithm to have the same number of classes given automatically by the proposed method. Our observation suggest that two methods can give very different results. Overall, by introducing a new method that enables the use of lumped characteristics for catchment classification, our study opens up an alternative avenue for hydrologic prediction in ungauged catchments.