2025 Volume 81 Issue 16 Article ID: 24-16060
This paper tried to improve the versatility of flood flow analysis by using the river land covers classified by machine learning with normalized indices of a new satellite image at different times. Here, we chose random forests (RF) for machine learning. The target river was a forested gravel-bed channel in Kinu River, Japan. For river land cover classification, RF was learned using normalized indices from the satel-lite image taken in October 2016, and the trained RF was applied to new satellite images taken at differ-ent times. The river land cover classification results indicated an effective classification with an accuracy over 0.76 in F-measure, even for untrained images. Then, a flood flow analysis was performed using the roughness coefficients concerted from the RF's land covers. The flood water levels were compared with those of a correct data analysis, resulting in a difference of less than 0.1 m. The average F-measure cal-culated for each river cross-section indicated over 0.8, which noted that the proposed condition in the previous research for ensuring accuracy in flood flow analysis was also verified for the gravel river of Kinu River. These results strongly supported that the normalized indices of satellite images could be helpful for high-frequency quantitative monitoring of the impacts of river land cover changes in flood risk management.
