A MULTIPLE-STEP MACHINE LEARNING APPLICATION FOR PREDICTING RIVER VEGETATION DISTRIBUTIONS AND THEIR PREDOMINANT FACTORS IN RECRUITMENT AND DISAPPEARANCE

Abstract

 This study tried to develop a new model that predicts the distribution of river vegetation by applying a machine-learning model in multiple steps and analyzing the factors behind vegetation recruitment and disappearance. This study used Random Forest (RF) as the machine-learning model. The multiple-step RF model was optimized for each river segment. Based on the input dataset, the optimized multiple-step RF model predicts the river vegetation distribution for the following year. As a result of applying the multiple-step RF model to several river channels of the Kinu River, the F1 scores improved by approximately 10% compared to the simple RF model, and they had over 0.8 in all river segments. The model could predict the distribution of vegetation on sandbars along the inner banks of meandering channels and the distribution of grass plants that frequently recruit and disappear. Furthermore, SHAP analysis of the model results successfully detected the main factors for vegetation recruitment and disappearance in each river segment. The results suggest that the multiple-step RF model in this study could help predict river vegetation distribution.