Feature-based clustering of hydroclimatic time series

<p>Both the grouping of hydroclimatic time series (often required, e.g., for technical and operational purposes) and the identification of spatial hydroclimatic patterns can be formalized and automated through algorithmic clustering methodologies. In this presentation, we focus on a new family of such methodologies that can be applied to various types of hydroclimatic variables (e.g., temperature, precipitation and streamflow) and at various temporal scales (e.g., the daily, monthly, seasonal, annual and climatic ones) with minimal adaptations. Aiming to exploit the largest part possible of the total information encompassed in the hydroclimatic time series, this family of clustering methodologies primarily relies on massive feature extraction, a concept sourced from the data science field. Once a compilation of numerous and diverse time series features (comprising autocorrelation, long-range dependence, entropy, temporal variation, seasonality, trend, lumpiness, stability, nonlinearity, linearity, spikiness, curvature and more features) has been computed, the clustering upon them is performed using a selected machine and statistical learning algorithm, with unsupervised random forests being an appealing choice for the task. Explainable machine learning can also be applied, as part of wider methodological frameworks, for ranking the features from the most to the least informative ones in obtaining the clusters, thereby facilitating the interpretability of the clustering outcomes in a comprehensive manner. We extensively discuss the above-outlined approach to hydroclimatic time series clustering emphasizing its main similarities and differences with the current well-established approaches in hydrology (e.g., from the catchment hydrology field), as well as its strengths and current limitations. Our discussions are well-supported by global-scale and other large-scale investigations, which have been conducted for temperature, precipitation and streamflow variables at several temporal scales.</p>