Author(s): Andre Araujo Fortes; Masakazu Hashimoto; Keiko Udo; Ken Ichikawa; Shosuke Sato

Linked Author(s): Andre Araujo Fortes, Masakazu Hashimoto, Ken Ichikawa

Keywords: Hydraulic roughness; 3D point cloud; Riparian Vegetation; Point cloud classification

Abstract: River management in urban areas is of fundamental importance for the safety control of these environments. Mainly located in the floodplain areas, the riparian vegetation plays an important role on the conveyance capacity of rivers, as it is directly related to hydraulic roughness. One way to reduce flood risk, is to decrease the floodplain roughness by controlling the vegetation population. The wide variation of characteristics, such as height, foliage density and branch configuration, are important factors in defining roughness, making the identification of the many vegetation species in river environments an important task for studying the river’s hydrological processes. Obtaining these characteristics by traditional surveying methods is time consuming and financially expensive. Recent techniques combining the use of Unmanned Aerial Vehicles (UAV) and optical cameras have been implemented as a more time- and cost-effective way to survey riverine areas. UAV photogrammetry derived 3D point clouds are powerful datasets for digital recreation of 3 dimensional terrains, with practical uses for river management, e.g., for feature classification of river environments. Moreover, techniques such as Machine Learning (ML) have become popular for image classification. Different algorithms have been developed and applied for that objective, achieving overall good accuracy. The objective of this study is to combine the use of classification ML techniques to classify the vegetation species in a 2 km stretch of the Nanakita river, in Miyagi prefecture of Japan, from the 3D point cloud data available for the area, and to use the species information in a 2D hydraulic model adapted to interpret the vegetation information and dynamically estimate the hydraulic roughness using the real case of typhoon Hagibis of 2019. The point classification was performed in two parts. First, to differentiate the vegetated points from ground points and other objects, color-based region growing semantic segmentation was used. With the vegetation points completely filtered, a sample number of points were labeled into different species and set as training area for the ML algorithm Random Forest (RF), which, once trained, was used to classify the species in the entire UAV observed area. Once the classification was completed, the information regarding the vegetation was divided into the terrain grid cells and prepared for the 2D hydraulic model. The accuracy of the model was observed by comparing the results with a different simulation not considering the vegetation effect. The species differentiation highlighted the different plant characteristics that could influence the hydraulic roughness. The scenario simulated in the adapted 2D hydraulic model considering the vegetation characteristics obtained more accurate results when compared with the simulation with observed roughness values, without the estimated roughness from the vegetation.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221003

Year: 2022