Author(s): Seonggi An; Chanjoo Lee; Yongmin Kim; Hun Choi
Linked Author(s): SEONGGI AN, Chanjoo Lee, Hun Choi
Keywords: Random forest Surface cover classification Remote sensing Woody vegetation
Abstract: In recent years, vegetation establishment has been observed in several rivers within mid-latitude monsoon regions. Vegetation colonizes exposed riverbeds, such as sandbars, during low-flow periods, eventually transforming previously white rivers into green rivers. This vegetation establishment affects river landscapes, geomorphology, and formative processes, which increases flood risks and impacts ecological functions (Gurnell, 2014). Consequently, understanding the current state of surface cover within rivers is essential for balancing river management with ecological conservation, particularly in rivers constrained by levees. However, traditional survey methods, such as vegetation mapping by specialists and vegetation indices, are limited in their ability to capture the extensive and dynamic nature of river environments. Against this backdrop, this study employed satellite-based surface cover classification using the Random Forest algorithm to analyze 24 rivers in South Korea. Using synthetic imagery of the Naeseong Stream in 2016, the spatial distribution of woody vegetation was quantitatively assessed with an accuracy of 85.1% (AN et al., 2024). Based on this, datasets for each river were developed for the period 2016 to 2023. Surface cover was classified into four categories: open water, bare bars, herbal vegetation, and woody vegetation. The changes in area and distribution across categories were analyzed. Additionally, the spatial distribution of woody vegetation within river channels was evaluated by calculating woody density and analyzing river sections at 1 km intervals. Woody vegetation is known to increase flood risk (Rood et al., 2019). Vegetation near the channel center was assigned higher scores because of its greater flood risk potential (Bae et al., 2024). The results showed successful classification, regardless of the individual characteristics of each river, allowing for clear detection of surface cover within the channels. Of the 6,115 extracted sections, four were identified as having high woody density and high scores, while 409 sections were assessed as having the potential to develop into high-density, high-score states in the future. This study utilized satellite imagery and Random Forest techniques to classify the surface cover and assess the distribution of woody vegetation in 24 rivers across South Korea. Applying machine learning techniques to river environment analysis can significantly reduce both time and costs. Furthermore, a quantitative evaluation of the current state of woody vegetation within channels is expected to contribute to decision-making regarding its removal.
Year: 2025