SEX DISCRIMINATION OF INDIVIDUAL TREES USING UAV IMAGERY

: The sex ratio is the proportion of male to female trees, which has a substantial impact on reproductive success and conservation status. Appropriate sex-related differences in dioecious trees commonly result in leading to a robustly structured population. Fieldwork for sex discrimination is time-consuming and labor-required. Benefiting from the unmanned aerial vehicle (UAV) and SfM techniques, the present study aims to detect male and female Caspian poplar ( Populus caspica ) trees. In March 2021, a heterogeneous forest in Noor city located in Mazandaran province was photographed, then 3D point clouds were extracted from the images using structure from motion algorithm (SfM) to generate an orthomosaic and a point cloud. The field survey was carried out to record the species, sex, and position of the overstory trees which were identifiable on the orthomosaics. A random forest classification algorithm was applied using R software to classify the trees into male and female. By assessing the producer's accuracy, user's accuracy, and overall accuracy, the classification accuracy for identified trees was computed using 10-fold cross-validation. The results showed an accuracy of 83% for identifying Caspian poplar trees and 52% accuracy for Sex discrimination. Overall, our effort to evaluate sex discrimination of dioecious trees using UAV imagery represents a


INTRODUCTION
Flowering trees have a wide range of sexuality. Some dioecious trees (for example ginkgo, holly, yew, willow and Caspian poplar) have male and female parts on distinct trees, with one tree being exclusively female and the other male. In dioecious trees, females devote more resources to sexual reproduction and spend less on growth and maintenance than males (Matsushita et al., 2016). The sex ratio (SR) of these trees is the ratio of male to female trees, which has a substantial impact on reproductive success and conservation status. The sexes' geographical distribution, which is typically impacted by environmental variation, is particularly crucial for reproductive success. In hostile situations with limited resources, environmental variables are accountable for the sexes' spatial segregation. The reproductive biology may also contribute to dioecious trees' sensitivity to global change, as females require more resources for reproduction than males and thus develop more slowly. On the poorer sites, males dominate. On the other hand, climate change may affect the availability of resources, resulting in a concentration of females in climatically advantageous areas (Garbarino et al., 2015). This issue shows the importance of determining the male and female trees distribution. Populus caspica Bornm (Caspian poplar) is a dioecious tree that contributes significantly to Hyrcanian forest biodiversity especially in the lowland area. It is an endangered and endemic tree species that are declining due to land use changing as well as its habitat fragmentations (Fallah et al., 2011;Yousefzadeh et al., 2018). This species is endemic to the Hyrcanian forests which are identified as a refuge for many Arco-Tertiary relict plants (Sagheb Talebi et al., 2014) Sex determination of these tree is very difficult due to the high trees height and the short flowering season. That's why a gender * Corresponding author: hsohrabi@modares.ac.ir study of this species has never been performed. Because fieldwork efforts are costly and time-consuming, some activities on providing the required data for plants evaluation has been carried out by remote sensing (Rominger & Meyer, 2019). Compared to fieldwork, remote sensing puts out data for mapping and monitoring at reduced costs. While RS based mapping can cover larger and inaccessible areas, it can be done at a less cost and faster (Gülci, 2019;Madera et al., 2019) which intrigues the search for more suitable data sources. In recent years, unmanned aerial vehicles (UAVs) have risen in popularity, in particular in forestry management. Because of the fine spatial (achieved by the low flying height) and temporal resolution, RGB cameras placed on UAV platforms are becoming cost-effective tools for forest monitoring and conservation projects. Light UAVs fitted with consumer-grade cameras have recently attracted a lot of attention for forest monitoring due to their low cost and operating flexibility (Guerra-Hernández et al., 2017;Rees et al., 2018). Various studies have shown UAV potential for forest trees classification and identification (Gini et al., 2018;Kampen et al., 2019;Kuzmin et al., 2017;Onishi & Ise., 2018;Sadeghi & Sohrabi., 2018). One of the prominent traits that are widely used for classification is the tree species color. Success in detection, identification, and classification generally depends on the main color differences between the species (Kuzmin et al., 2017;Rominger & Meyer., 2019;Sadeghi & Sohrabi., 2018). Therefore, color can be used as a phenotypic trait to classify tree species. The color of flowers has been studied in estimating the volume and flowering rate of plants (Carl et al., 2017;Wan et al., 2018), but the study of this trait in dioecious trees to discriminate male and female trees has not been performed so far. In this study, the discrimination of male and female trees of Caspian poplar based on RGB-UAV images was.

Study Area
The test site is in Noor, Mazandaran Province in northern Iran, in an area of Hyrcanian forests (Figure 1). With a mean elevation of -27 meters, the area is situated in an eastern latitude of 36 ⸰ 34՛43.96՛ and a northern longitude of 52 ⸰ 02՛29.14՛՛՛. The forest is described as a deciduous and temperate forest with multilayered stands and high tree species diversity. The research region has slopes ranging from 0 to 4 percentage, and the climate is humid (Hadiani., 2015). The average annual precipitation is 997 millimetres, and the average annual temperature is 16.4 °C. Different tree species such as Populus caspica (Caspian Poplar), Parrotia persica (Persian ironwood), Ulmus minor (Common Elm), and Quercus castaneifolia (Chestnut-leaved oak) can be found in this area.

Field Data
Species, crown diameter, and the gender of Caspian poplar trees were recorded for all the trees that their crowns were visible on the orthomosaic. The trees' geographic positions were determined using an RTK GPS (Gintec G10) or if the GPS signal was too weak, an azimuth-distance technique utilizing a laser distance meter and a hand-bearing compass were used. Male and female Caspian poplar trees were also identified using binoculars and observing the shape and color of flowers ( Figure 2).

UAV Image Acquisition and Processing
The raw images were taken in March 2021 with a UAV Phantom 4 PRO equipped with a mechanical shutter, capable of recording a 20-megapixel RGB image in JEPG format (focal length of 8.8 millimetres). With a clear sky and no wind, the image acquisition was accomplished within a half-hour of the solar zenith (Cardil et al., 2019). The summary of flight parameters is shown in Table  1. First, the quality of images was checked manually. As a result, images with evident artifacts, blurring with a high level, or images are taken too near to the ground were not processed (Brovkina et al. 2018). Then, processing the UAV images was performed by Agisoft Metashape v1.6.3 using the structure from motion (SfM) technique. Based on 2D overlapping pictures, the SfM approach generates 3D point clouds. It uses key points in each image in accordance with the same points in other sets of images over the same region (Otero et al., 2018). The parameters setting used for processing are shown in Table 2.

Processing
Step Finally, the Random Forest algorithm was applied in the randomForest package in the R environment (3.6.0) for classification. Random Forest classification is an ensemble learning method that generates many decision trees from a randomly selected subset of training samples and variables. The RF classifier has become well-known in the field of remote sensing due to its performance and precision in classifying (Zhang et al., 2019). We used default parameters (i.e., number of trees (ntree) set to 500) to run the RF classifier because prior research have shown that default settings produce decent results (Daryaei et al., 2020;Immitzer et al., 2012Immitzer et al., , 2016. A k-fold cross-validation method was used to validate the classification models with a set of k = 10. When there are a minimal number of samples, the k-fold approach is a dependable and robust method that does not rely on a specific collection of samples ( Sothe et al., 2019;Nevalainen et al., 2017). The overall accuracy (OA), producer's accuracies (PA), and user's accuracies (UA) were calculated using the confusion matrix. In this method, all data are considered as test data and training data. In such a way that the process is repeated 10 times. In each repetition, 9 parts of the data are used as training data and one part as test data.

RESULTS AND DISCUSSION
As can be seen in Figure 4, Caspian poplar trees were identified from other trees i.e., Persian ironwood, Chestnut-leaved oak, and Common elm with an overall accuracy of 83%. The reason for this high accuracy can be considered in the different color, Caspian poplar due to having five-lobed with silver-green colored leave are quitley recognizable by the other trees, and geometry of Caspian poplar trees (Miraki et al., 2021). Because, the success of detection generally depends on a major color difference between target species (Kuzmin et al., 2017;Rominger & Meyer, 2019;Sadeghi & Sohrabi, 2018).

Figure 4. Accuracy of Caspian poplar trees identification compared to other trees
To evaluate the results of RF classification calculated the producer's and user's accuracies for each class. Caspian poplar class producer's accuracy is 97% which is lower than other trees class accuracy (99%). Also, the Caspian poplar class user's accuracy is 48% ( Figure 5). Many studies have been conducted to identify tree species using UAV images (Feng et al., 2015;Gini et al., 2018;Leduc et al., 2018;Maschler et al., 2018;Nevalainen et al., 2017). All of this research has shown that the spectral signeture of each species is different from the other, which helps to differentiate the species using remote sensing data. Therefore, the difference in the leaf optical properties, i.e., color of the flowers at the crown level of the dioecious trees also causes a different spectrum of those species and they can be used to the sex discrimination. In this study, the sex discrimination of dioecious trees i.e., Caspian poplar was tested as a pilot. Caspian poplar trees were identified into two classes i.e., male and female trees with a total of 7 female trees and 14 male trees. 3 out of 7 female trees and 8 out of 14 male trees were correctly identified. This result showed that there is a relationship between vegetation indices (RGB indices) and flowers spectral reflectance of a tree and it can be used to discriminate male and female trees of the same species. (Wan et al., 2018) also showed a relationship between vegetation indices and number of flowers. The identification of male and female trees was performed with a small number of trees (unexpected quarantines associated with COVID-19 pandemic constrained the possibility of providing a large data set), for this reason, the results may not have the necessary generalizability. Overall, taking of images time is very important in gender segregation, and the pandemic made it difficult to collect the data with the best timing of flower densities on trees. Also, at the time of flowering of Caspian poplar trees, there were no leaves on the trees, and in crown polygons for sex identification, spectral mixing is observed due to a combination of branches, flowers, herbaceous plants of forest floor and soil, which reduced the accuracy of identification. To our knowledge, this is the first time that the sex discrimination of male and female trees has been performed, so this method can be improved in future research. In the further studies, the following cases can also be studied to improve results.
1. Taking images at different flights altitude to determine the appropriate flight altitude for the sex discrimination of the trees 2. Sex discrimination evaluation by the multispectral data 3. Estimating biochemical properties using UAV imagery and examining its relationship with the sex of trees 4. Sex discrimination of evergreen dioecious trees, i.,e, taxus bacata

CONCLUSION
This paper demonstrated the usefulness of the RGB-UAV imagery for the detection of sex of Caspian poplar individuals in a mixed broadleaf forest. The results showed that UAV-based RGB imagery gives enough information to detect trees. Also, the random forest classification produced acceptable accuracies, indicating that this is a viable method for detecting individual trees. But it did not provide high accuracy for identifying the male and female individuals. User's acc.