River Delineation from Remotely Sensed Imagery Using a Multi-Scale Classification Approach

River delineation is an initial yet critical step in river studies. Although the analysis of satellite images shows great potential in river delineation, only a few approaches have been developed. These methods usually focus on rivers at mono-scale and may ignore the large variations in river size. In particular, they may fail to capture the small rivers in the imagery. This paper presents a novel automated multi-scale procedure for delineating complete river networks. This method classifies the large and small rivers separately and combines the two classified results to generate the final delineated river networks. First, a modified normalized difference water index (MNDWI) is adapted to enhance the spectral contrast between open water and land surfaces. Second, a simple OTSU classification is used to delineate the large rivers. Next, a top-hat transformation and multi-scale matched filters are used to enhance the small linear rivers. Then, the OTSU classification is conducted again to delineate the small linear rivers, in concert with a multi-points fast marching method to rejoin the resulting river segments. Finally, the complete river networks are delineated by combining the small and large rivers. A comparison of this procedure with manual digitization when applied to two Landsat-5 TM images demonstrates the former procedure’s value in delineating rivers. It achieves accurate results and outperforms the other three alternative approaches (large river classification, maximum likelihood classifier, and support vector machine classifier) in accuracy, true positive rate, and Kappa coefficient, while also yielding a comparable false positive rate.