2024 Volume 80 Issue 16 Article ID: 23-16056
A method for classifying land cover of riverine regions into categories such as grassland, woodland, bare land, and water bodies was investigated using a random forest (RF) algorithm. The target sections of the Kamanashi and Kuzuryu rivers were divided into two domains and training and test data sets were generated. The classification accuracies of the RFs were tested in four cases by combining the data sets. Upon applying the method, the highest macro-F1 unweighted mean of F1 scores calculated per class was 73%, and the F1 scores were approximately 90% for water bodies and bare land and approximately 70% for grasslands and woodlands. Furthermore, the accuracy was affected less by the training data sets used to develop the RFs, which may partly be explained by the fact that the digital surface model, vegetation height distribution, and reflected laser intensity are the features that make the highest contributions in the RFs.
