Watershed Classification Using Isomap Technique and Hydrometeorological Attributes

Classification of watersheds into hydrologically similar groups prior to regionalization is essential for predicting streamflow in ungauged basins. The objective of this study was to improve the efficiency of classification for finding hydrologically similar watersheds by applying dimensionality reduction techniques. A new nonlinear dimensionality reduction technique called Isomap is applied for the first time along with nonlinear principal component analysis (NLPCA) and principal component analysis (PCA) to improve the efficiency of watershed classification. The dimensionality reduction techniques were applied to the selected attributes of 68 watersheds of the Arkansas-White-Red River basins and the Lower Mississippi River in the United States. The resulting reduced dimensions were used to classify the 68 watersheds into five homogenous groups, using a threshold of 90% accounted variance of the original data by $K$-means cluster analysis (KCA). In order to validate the results of classification, reference groups of watersheds identified using the runoff signatures of the 68 watersheds were compared with the classification results obtained using Isomap, NLPCA, and PCA employing a modified method of calculating the similarity index. Statistics of the sample distribution of the similarity index were used to analyze the efficiency of each technique. It was observed that the Isomap technique performed better than NLPCA or PCA for classification of watersheds, finding with the most hydrologically homogenous watersheds in each classification group. This study shows that the Isomap technique can be effectively used to extract and preserve the underlying essential structure of the watershed data that are relevant for hydrological processes to obtain more accurate watershed classification.