A probabilistic approach to river network detection in digital elevation models

Abstract

Abstract River networks are often derived from digital terrain models and are affected by uncertainty and errors of the corresponding elevation data. The analysis of the spatial distribution of the errors provides information on the confidence level of the derived networks. However indications on the most probable river network as a whole are missing. This study proposes a method to indicate which is the river network maximising the sum of the probability values along the network, given a map reporting the likelihood that a cell belongs to the network itself. The method is considering the inverse of the channel probability map as pseudo-DEM from which drainage networks are derived. A reference network is used to assess the spatial match of the extracted river networks using the Euclidean distance as simple comparison parameter. The network extracted from the inverse of the channel probability map is the closest to the reference. The use of a probabilistic approach to error modelling significantly increases the values of channel probability for extracted river networks and the spatial match with a ground reference dataset.