Classification of forest vegetation types considering forest vertical stratification using airborne laser scanning data in Japan

• We derived metrics for forest vertical structure and topography from public ALS data, adding to existing canopy surface features. • We classified forest vegetation types in 5 m pixel units using phytosociological survey methods and evaluated the classification accuracy. • Vertical structure and topography were key to classifying five vegetation types in natural forests, achieving 0.882 accuracy. Recently, there has been a growing interest in broad-leaved forest classification, with proposals to introduce multiple-layered forests and broad-leaved forests in unsuitable forestry areas. Previous studies on vegetation-type classification using airborne laser scanning have focused on artificial coniferous forests and employed metrics related to the canopy surface, such as reflectance intensity and canopy slope. In this study, in addition to those used in previous studies, we considered metrics of vertical stratification, such as the vertical complexity index and the canopy’s euphotic volume, as well as topographic indices, to classify artificial coniferous forests and natural forest types. We constructed ground-truth samples for forest vegetation based on field surveys and UAV orthophotos. We then used the random forest algorithm to perform vegetation classification, verified its accuracy, and evaluated the importance of each feature. The results showed that two classification models achieved high accuracy: Model A (three classes: cedar, cypress, and natural forest) had an overall accuracy (OA) of 0.909, and Model B (five vegetation types in natural forests) had an OA of 0.882. For Model A, the euphotic volume of the canopy (Euphotic), a vertical stratification metric, was the most important feature. For Model B, the topographic slope and Euphotic were the most important features. This suggests that the vertical stratification of natural forests is more complex than that of artificial forests, and that adding vertical stratification metrics can effectively capture this complexity.