Regional flood frequency analysis using complex networks

Proper regionalisation (identification of homogeneous regions) is key to reliable regional flood frequency analysis. Several methods have been proposed in the literature for regionalisation, including the method of residuals, L-moment method, and fuzzy c-means clustering algorithm. The present study explores the suitability of the theory of complex networks for regionalisation of watersheds, with an aim to perform regional flood frequency analysis. The west-flowing rivers of Kerala in India are considered for this study. Two complex networks-based methods, namely degree centrality and clustering coefficient, are applied for regionalisation, and daily streamflow data are analysed. To identify possible links between streamflow gauging stations, different correlation threshold values (i.e. linear correlations in streamflow between stations) are used. Two approaches are adopted in the use of correlation threshold values: the first (Method I) with threshold as mean, median, and mode of correlation values; and the second (Method II) with arbitrary threshold values. The regionalisation results suggest that Method II yields better results, both with degree centrality and clustering coefficient. Based on Method II, the use of degree centrality results in seven regions (five homogeneous and two heterogeneous) and clustering coefficient results in eight regions (seven homogeneous and one heterogeneous). Comparison of predicted and observed flood quantiles indicates that the degree centrality-based regionalisation yields R2 values in the range 0.94–0.86 for return periods 2, 5, 20, 50, and 100 years, whereas the clustering coefficient-based regionalisation yields R2 values in the range 0.98–0.91. The results from this present study suggest that complex network theory is a suitable alternative for identifying homogeneous regions for regional flood frequency analysis.