A data-driven approach to discern the curvature spectral complexity of compound meander bends

Meandering rivers display multiform patterns along their course. The bends composing these patterns are commonly detected by inflexion-point positions along the river centreline. However, this approach prevents singling out compound and multiple-loop bends, which, despite comprising many inflexion points, represent a morphological unit in their own right. The study aims to propose a data-driven methodology for analysing compound and multiple-loop bends along meandering rivers. To this purpose, a detection technique was first developed to single out these types of meander bends by looking at their signature on the curvature power spectrum. From 11,830 rivers, almost 150,000 meander bends were extracted with single-bend and complex (i.e., compound and multiple-loop) meanders covering 56\% and 44\% of patterns extracted, respectively. Then, an autoencoder architecture was utilised to learn from the curvature power spectra of various types of Kinoshita-generated meander bends. The trained autoencoder successfully distinguished meander bends based on shape complexity (simple or compound/multiple-loop) and skewness (upstream-skewed, symmetrical, or downstream-skewed). Similarity metrics confirmed the discernment of meander bends obtained via machine learning. Meander bends along unconfined meandering rivers, such as the Ucayali and Chixoy rivers, exhibited a decrease in shape complexity, probably due to dam-related sediment blockage, whereas meander bend patterns along highly-engineered rivers, such as the Sacramento River, showed only small changes. Overall, the proposed methodology allows to identify relevant spatio-temporal morphodynamic signatures from remotely-sensed meandering rivers. The detection of these signatures can give an insight into changes in meander morphology, aiding effective river management.