An Unsupervised Classification of Natural Rivers

Abstract:

While no two rivers are exactly the same, in many cases it is desirable to group rivers into similar categories that generalize their morphologies and behaviour. Many researchers (e.g. Rosgen 1994) have developed classification systems for river morphologies based on channel characteristics such as entrenchment, gradient, width/depth ratio, and sinuosity. However, as with any classification system developed through the human inspection of data, classification systems may be subject to biases imparted by the researchers, and boundaries between classes may not maximize discriminability.

The National Hydrographic Network is a publicly available database that provides geospatial data on Canada’s inland waterways including lakes and rivers. River polylines for a single watershed from within this dataset were segmented and sampled to produce two-dimensional image arrays which were fed into pre-trained convolutional-neural-networks to perform dimensionality reduction. A clustering algorithm was then used to generate unbiased classes of rivers based exclusively on channel planform geometry.

The results of the experiments showed that the model appeared to produce classes that could be described using human-interpretable channel characteristics. The use of such models to produce channel classification systems and to identify similar river reaches has wide-reaching potential applications in the fields of fluvial geomorphology and river restoration.