Multidimensional flood analysis challenges and similarities utilizing linear and non-linear approaches

Partitioning a dataset in multivariate analysis is one of the key points to better understanding the hydrological process. Different regions of a catchment may bring floods variously due to distinct types of floods or their simultaneous occurrence. Therefore, it is needed to determine the spatial extent floods brought together. In a multidimensional space, it is demanding to investigate floods. It is not clear which kind of clustering methods or dimension reduction techniques are appropriate for visualizing initial similarities among measured peaks. Two methods are applied to reduce dimensions and compare their differences in this research. Multidimensional scaling (MDS) and t-distributed Stochastic Neighbor Embedding (tSNE) are the employed models for 55 years of extreme floods in the Neckar catchment. MDS is based on Principal Component Analysis (PCA), which is a linear technique. While tSNE is a non-linear dimensionality reduction method. In theory, tSNE can handle outliers and perplexity and preserve the local structure of data. As a result, compared to the MDS, both methods react similarly in soliciting an additional algorithm to cluster data in 2D space. It is another challenge that has to be investigated in future research. Due to the fatter and heavier tails, the t-student distributions have a greater chance for extreme values than normal distributions. Therefore, tSNE can better visualize data in a high-dimensional space for assessing extreme events. However, these algorithms must be run in different climates and deal with distinct hydrometeorological variables.