Study site
Takaragaike Park (Sakyo-ku, Kyoto, Japan) is an urban planned park managed by the City of Kyoto with a publicly recognized area of 128.9 ha. The study area was 109.4 ha of forest in the park area (hereafter referred to as “Takaragaike Forest”) (Fig. 1). In the study area, Pinus densiflora forests are distributed around the ridges, and Quercus serrata forests are distributed around the valleys. Deciduous broad-leaved forests account for 88% of the forests (Niwa et al. 2020). Once used as satoyama, these forests are now largely unmanaged, and problems such as progressive vegetation succession and increased feeding damage by deer have emerged, making forest management problematic. In March 2022, the Takaragaike Forest Conservation and Restoration Council formulated a “Forest Management Vision” for Takaragaike Forest, and forest management efforts have begun, with the restoration of biodiversity as one of the goals.
Target species
In this study, Rhododendron reticulatum was selected as the target species. R. reticulatum is a species of wild azalea that grows naturally in many secondary forests in the Kansai region of Japan (Morimoto and Yoshida 1998). R. reticulatum is a familiar plant in the Kansai region, and its conservation is significant because it is a symbolic species of secondary forests (Nakajima et al. 2016). R. reticulatum grows in light forests dominated by deciduous broadleaf trees and blooms in spring before the deciduous broadleaf trees in the taller tree layer have expanded their leaves. Therefore, the flowers can be photographed from above if the blooming period is targeted.
In Japan, the impact of feeding damage on vegetation is becoming more noticeable due to the increase in deer population density. R. reticulatum is subject to feeding damage by deer (Hashimoto and Fujiki 2014), and in the study area it is declining due to feeding damage by deer. In addition, the density of evergreen broad-leaved trees is increasing in the study area due to vegetation succession caused by decreased human management; this has led to a decrease in the number of R. reticulatum, which prefer light forests. Therefore, R. reticulatum is a suitable indicator for capturing changes in vegetation and the effects of forest management in the study area.
Identifying the distribution of flowers of R. reticulatum
A method was established by Niwa (2022) to identify the distribution of R. reticulatum flowers using a UAV platform. Following that method, the distribution of R. reticulatum flowers was monitored between 2019 and 2023. The blooming of R. reticulatum in the study area was observed in early April, and UAV photographs were taken on days when the flowers were judged to be close to full bloom (Table 1). A DJI Phantom 4 Pro or DJI Mavic 2 Pro was used to take images at an altitude of 130 m, with a front wrap rate of 80% and a side wrap rate of 60%. Although the aircrafts used were different, all images were 5472 × 3078 pixels in size. The acquired images were processed using SfM-MVS Photogrammetry (hereafter referred to as “photogrammetry”) to create an ortho-mosaic image. The xyz coordinates of the ground control points surveyed by real-time kinematics were imported into the photogrammetry software, and their positions were corrected. The photogrammetry software used in this study was Agisoft Photoscan Pro. Ver. 2.0.0. The ground resolution of the generated ortho-mosaic images is shown in Table 1.
Picterra (https://picterra.ch/), an online, artificial intelligence-based image recognition service, was used to extract the flowers of R. reticulatum. Picterra uses machine learning to create a detector of the image, which can then be used to extract the desired geographic feature from the ortho-mosaic image. A major feature of this technology is that geographic features can be detected with high accuracy from a small number of training data and can be exported as geographic information such as polygons and points. In the training area where R. reticulatum flowers were distributed, R. reticulatum flowers were traced by visual interpretation, and patch polygons were created. Training area additions were repeated several times while the detection results were monitored. The final detector was created from randomly selected six training areas of 100 m2 (five with flowering individuals and one without flowering individuals) (Fig. 2), with 43 patch polygons of R. reticulatum flowers. The detector used was the same as that used by Niwa (2022) and was created with ortho-mosaic images from 2019. The same detector was used in other years to extract patch polygons of R. reticulatum flowers throughout the study area.
Separate from the training area, four 100 m × 100 m accuracy verification areas were set up (Fig. 2). Patch polygons of R. reticulatum flowers exported from Picterra and ortho-mosaic images from the same shooting date were overlaid in a geographic information system (GIS) at a scale of 1:100 to verify the accuracy of the extraction results. The patch polygons exported from Picterra were classified into R. reticulatum flower and others (i.e., false positives), and if there were undetected R. reticulatum flowers (i.e., false negatives), patch polygons were manually created. The area of the patch polygons was tabulated to verify the detection accuracy. This accuracy verification area was unique to this study and was the same for all five years. The GIS software used was Esri ArcGIS.
Data analysis
A 10 m × 10 m mesh polygon was created in the analysis area as a spatial unit for comparing the abundance of flowers of R. reticulatum in multitemporal periods. The patch polygon area of R. reticulatum flowers for each year was calculated for each mesh. Although the UAV was used to take photographs on days when the flowers were judged to be near full bloom, the total patch polygon area of R. reticulatum flowers was considered to vary from year to year due to flower richness and phenology. Therefore, the area calculated for each mesh was standardized by year. To evaluate the abundance of R. reticulatum flowers over time, the standardized values (Z values) for the five years were summed. To evaluate the change over time, the Z value in 2023 was subtracted from the Z value in 2019.
In the Takaragaike Forest, there is an area where a fence was installed in March 2017 to prevent deer from entering in order to protect R. reticulatum from deer feeding damage (Fig. 2). The area protected by the fence measures 3,453 m2, of which 1,269 m2 on the east side had all evergreen broadleaf trees cut down in June 2018. To verify whether the conservation effect of forest management on R. reticulatum could be evaluated, we compared the mesh totals by dividing the fenced area by whether evergreen broadleaf trees had been cut or not.
In 2020, Niwa et al. applied a UAV platform to create a vegetation map of the Takaragaike Forest; this map had high accuracy and spatial resolution of community boundaries (Fig. 3). Vegetation types reflecting the increase in evergreen broadleaf trees due to vegetation succession were presented in the map. To evaluate the relationship between the increase in evergreen broadleaf trees due to vegetation succession and the abundance of R. reticulatum flowers, mesh totals were compared for each vegetation type.