Author(s): Yiwei Guo; Michael Nones; Yitian Qi

Linked Author(s): Michael Nones

Keywords: Dynamic vegetation roughness; Unmanned aerial vehicle; Random forest; Multispectral indices; Po River and Vistula River

Abstract: The presence of vegetation introduces hydrodynamic resistance during water-vegetation interactions in fluvial systems, with a significant impact on the peak flood discharge and peak flood arrival time. The use of fixed roughness in flood modeling leads to uncertainty, since the changes in bed roughness due to dynamic vegetation are ignored. Aiming to quantify long-term vegetation roughness for supporting accurate flood process representation , this research combines remote sensing and deep learning to train a model that can estimate vegetation height and roughness by inputting multispectral indices. Specifically, vegetation height measurements and multispectral imagery were collected along the Po River (Italy) and Vistula River (Poland) using a DJI Mavic 3 Multispectral drone. Five indices, including the Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), Leaf Chlorophyll Index (LCI), Normalized Difference Red-Edge Index (NDRE), and Optimized Soil Adjusted Vegetation Index (OSAVI), were used as input features for a random forest model. The trained model was validated using a three-fold cross-validation method. The results showed that the Random Forest model can explain the vegetation well, with the residual-enhanced random forest consistently improving the vegetation height estimation across all regions, showing an increase in R² from 0.48 to 0.53 and a reduction of RMSE from 4.50 m to 4.30 m, thereby establishing a robust and practical baseline for cross-regional multispectral index to vegetation height modeling.

DOI: https://doi.org/10.64697/iahr.proc.ise2026.abs.240

Year: 2026