Yuki Hirose and Naoto Matsumura: Effectiveness of tree count estimation in Sugi and Hinoki plantations using aerial orthophotos and deep learning. Jpn. J. For. Plann.　In this study, we focused on high-resolution aerial orthophoto data, which are being increasingly acquired by local governments, in order to explore efficient methods for obtaining forest resource information in medium- to large-scale regions, particularly with the aim of updating nationwide laser survey data. Using the aerial orthophotos acquired during laser surveys, combined with a deep learning model (EfficientNet) that has demonstrated excellent performance in the field of remote sensing, we estimated the number of standing trees in Sugi (Cryptomeria japonica) and Hinoki (Chamaecyparis obtusa) plantations (15×15m plots). The results showed an overall accuracy of approximately 80%. Notably, the Sugi plantations, due to their distinct crown shapes, maintained high estimation accuracy even in densely populated areas, showing superiority over the hinoki plantations. Additionally, it is expected that applying the proposed method to high-resolution satellite images and past aerial photographs will enable the acquisition of a wider range of spatiotemporal forest resource data.

Naoki Aihara and Yasushi Mitsuda: Automatic forest road extraction using a deep-learning model with a CS topographical map. Jpn. J. For. Plann. 　Utilizing a CS topographical map derived from a Digital Elevation Model (DEM) acquired through airborne laser scanning (ALS), we developed a methodology to automatically extract forest roads using a deep learning model. The deep-learning model, employing a CS topographical map as the input, reproduced 75.1% of the route length for actual forest roads. Of the forest roads extracted by the model, 91.7% were consistent with the existing forest roads. In contrast, the deep-learning model developed using elevation, slope, and plane curvature as inputs could reproduce only 0.1% of the route length for actual forest roads. In forested areas, the automatic extraction of forest roads using conventional photographs is challenging due to vegetation cover. Our forest road detection model can overcome this limitation because it utilizes an ALS-derived DEM as the input. Furthermore, the reproduction rate of forest roads in our model employing the CS topographical map as the input image was significantly higher than that of the model that used elevation, slope, and plane curvature as the input images. This result indicates the efficacy of the CS topographical map for automatic forest road extraction using deep-learning models.