EGU General Assembly 2023
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

River coastline detection using maximum likelihood classification

Matylda Witek, Grzegorz Walusiak, and Tomasz Niedzielski
Matylda Witek et al.
  • Department of Geoinformatics and Cartography, Faculty of Earth Sciences and Environmental Management, University of Wroclaw, pl. Uniwersytecki 1, 50-137 Wrocław, Poland

In frame of the project no. 2020/38/E/ST10/00295 carried out within the Sonata BIS programme, financed by the National Science Centre of Poland, we consider different approaches to delineate the river coastline based only on close-range visible light aerial images (RGB) acquired by unmanned aerial vehicles (UAVs).  Among the scrutinized methods, we use automatic mapping of extent of water bodies  by means of image classification. It was found that the best results of reconstructing inundation extent (even 95%) were obtained using the supervised methods, in particular the maximum likelihood algorithm. The accuracy assessment of this classification, using the confusion matrix visualization, allowed us to notice that the areas incorrectly classified as "water underestimation" (surface where there is real inundation which was not indicated by the classifier) are located mainly on the borders of the water bodies.

In most of the analyzed cases, the incorrectly classified "water underestimation" areas form a narrow zone around the inundated areas, which can be interpreted as the water-land interface zone. Therefore, it is possible to delineate, with high probability, the approximate  water-land boundary line. In low-altitude aerial photographs or orthophotomaps with visible fragments of river channels, the designation of such "water underestimation" zone allowed us to delineate the approximate course of the river channel coastline. This approach was tested on several UAV images acquired over the middle Odra River channel in western Poland. We analyzed several images representing various terrain situations: (1) the river channel completely visible, without vegetation, where the visual determination of the reference coastline by the human expert was not difficult, (2) the river channel partially shaded, where significant classifier errors can be expected, (3) the river channel partially covered, for example by vegetation, where the course of the real coastline is uncertain. The obtained results confirm that the proposed approach allows to reconstruct some courses of the coastline for channels with a width of at least 100 m using the "water underestimation" areas.

How to cite: Witek, M., Walusiak, G., and Niedzielski, T.: River coastline detection using maximum likelihood classification, EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-13135,, 2023.