論文ID: 23.002
In the present study, we aimed to develop an efficient prediction model for the site index of hinoki (Chamaecyparis obtusa) plantations in Higashi Yoshino Village, Nara Prefecture, Japan. For this purpose, we trained a convolutional neural network (CNN) model, then investigated the accuracy of site index prediction and the reproducibility of topographic factors from digital elevation model (DEM) image data acquired using an aerial laser scanner. We also examined terrain red–green–blue (RGB) images, derived from the DEM, as an alternative form of input data. The terrain RGB images outperformed DEM images in terms of prediction accuracy and reproducibility for all evaluation indicators. Reproducibility analysis of elevation, slope, and orientation revealed particularly high accuracy (coefficient of determination > 0.95). Although further improvements to the model are needed, our results emphasize the practicality of using a CNN model combined with terrain RGB images for site index predictions. This approach has the potential to achieve prediction accuracy close to or even surpassing those of existing methods, while requiring fewer types of input data.
