Classification of Meander Bends a Data-Driven Approach

Meandering rivers are one of the most fascinating features of fluvial systems. The presence of single-thread rivers exhibiting a continuous sequence of alternating curves is widespread in alluvial floodplains. The study of meanders has attracted the scientific community; for a long time, researchers have tried to quantify the complexity of their platforms and model their morphodynamical evolution and classify the wide spectrum of meander styles. However, the existing classification methodologies are inadequate to encompass the variety of shapes observed in nature and, most often, lack physical insight into hydraulics and sediment transport processes which determine the style of the meander bend. To address the issue of classifying the various meander shapes unambiguously, we propose a data-driven approach mixing physical-based information and machine-learning algorithms. In this approach, we consider the planform of different meander bends as signals and analyse them using continuous wavelet transforms. Specifically, we compute the energy spectrum associated with each meander's spatial distribution of curvature. This physics-based information is then used to classify bend shapes by a neural-network autoencoder mixed with a clustering algorithm. As a first step towards developing a robust classification methodology, the convolutional neural network autoencoder is trained and tested using a broad series of Kinoshita-generated meander bends. The proposed classification procedure shows how the intrinsic information on a meander bend curvature's in the frequency and space domain can be captured and encoded for later classification with a physically informed basis. The results on the synthetic data open the possibility of expanding the methodology to real meander data.