Assessing the success of forest crops using UAVs

. The article presents data on the growth and development of 22-year-old forest crops created by sowing and planting seedlings with a closed root system on the territory of the Republic of Karelia (Russia). Field surveys showed that young forests with a predominance of Pinus sylvestris were formed in all experimental plots. The share of forest plantations on the plots in terms of timber stock amounted to 38 - 44% of the total stock. The UAV data processing method made it possible to build an orthophotomap of the area and calculate the quantitative distribution of tree species: 60% (plots without tillage) -80% (plots with tillage) - Pinus sylvestris , 10% - 22% - betula pendula . These indicators are consistent with the field survey of the area (differences less than 10%). As a result of running the algorithm for automatically searching for trees using point clouds using the lidR package, it was possible to detect about 90% of trees in all areas and determine their heights. At the same time, most of the trees (85%) found by the algorithm were identified correctly. The number of false positives and the number of missing trees were quite low, and the weighted average quality score was 0.89, which indicates a high efficiency of tree search. The heights measured from the UAV data were in good agreement with the heights measured by the ground method.


Introduction
The experience of creating forest plantations in the Republic of Karelia (Russia) has several decades.Forest plantations were created with and without tillage, sowing or planting.The work [1] analyzes the effectiveness of artificial reforestation with conifers.In [2], the higher efficiency of forest plantations created by seedlings or saplings is reflected.As a rule, the biometric characteristics of forest plantations created with planting material with a closed root system are significantly higher than crops created with an open root system or sowing.
The classic method for assessing the success of reforestation is the method of natural ground survey [3], which is quite laborious and is not always available for use on large areas or at a considerable distance.Alternative methods, such as interpretation of space images, are not sufficiently developed or limited by low spatial resolution and do not allow a full assessment of the quality of forest plantations.However, some studies show high accuracy (84%) in assessing the success of reforestation using multispectral images from the Landsat satellite using the vegetation indices NDVI and SWVI [4].
The use of unmanned aerial vehicles (UAVs) for the purposes of reforestation analysis can significantly reduce labor costs in comparison with field surveys and, in addition, obtain important information and images of high spatial resolution in real time [5].At the same time, the use of UAVs has its drawbacks in comparison with satellite remote sensing data: dependence on weather conditions (strong wind, rain, low temperature) and small coverage of the shooting area due to limited battery life (from several minutes to 10 hours).
It was shown in that when studying a forest, the combination of combined research methods is most effective: the use of UAVs and field survey of the area.As a rule, during field surveys, circular areas of 10 m 2 are laid with a detailed description of the vegetation.Further comparison of the results obtained from the survey sites and materials of high spatial resolution makes it possible to reliably characterize the vegetation in the study area.The purpose of the study is to give a comparative assessment of the growth and development of forest plantations created by sowing and planting with and without tillage, by comparing aerial photography data from UAVs with the results of a field survey.

Materials and methods
The object of the study is the plots of forest plantations of Scots pine (Pinus sylvestris L.), created in quarter No. 49 of the Pryazhinsky central forestry of the Republic of Karelia (Russia).Geographic coordinates of the plots: 61°45'15.7"N33°45'40.8"E.The soils in the plots are sandy, weakly podzolic, coarse-humus on the sands.Forest plantations are located on the site of the former lingonberry pine forest, on which in 1991 felling was carried out on an area of 6 hectares for the main use.Nine years after the felling (in 1999), the reconstruction of the species composition was carried out by the complete removal of all hardwoods, leaving natural pine, as well as soil cultivation.
The next year after the reconstruction, forest plantations were created -4 experimental plots planted in different ways: 1) row sowing with tillage; 2) ordinary plantings with tillage; 3) ordinary plantings without tillage; 4) planting by a biogroup without tillage.
Planting material was used with a closed root system at the age of one year from a local forest seed nursery.All landings were made using the Pottiputki landing tube.The number of pine seedlings per ha: row plantings -3600 pcs/ha, biogroup planting -4000 pcs/ha.
Soil preparation before planting in plots No. 1 and 2 was carried out using a PDN-1 cover stripper.The total area of crops was 2.2 ha, of which 0.8 ha were created without tillage, and 1.4 ha with tillage.The characteristics of the sites are given in Table 1.At the moment, the sites are overgrown with deciduous forests (mainly drooping birch and mountain ash).When performing the work, a combined method was used: the classical method for assessing the success of natural reforestation [3] and aerial photography from a UAV with photogrammetric data processing [6].
The counting of trees in young stands was carried out in 2-cm thickness steps.The average diameter was determined as a weighted average.The average height is obtained from a height curve constructed from the results of measuring the height and diameter of 9 trees of different sizes, selected from the prevailing thickness grades.Accounting for undergrowth, undergrowth, and living ground cover (LTC) was carried out on circular accounting plots of 10 m2.For the undergrowth and undergrowth, the number, composition, and structure in terms of height and condition categories were indicated.For the VNP, the species, the frequency of occurrence, and the projective cover for each species were recorded in the grassshrub and moss-lichen layers.
For aerial photography, a DJI Mavic Mini 2 quadrotor UAV was used, which has the following technical characteristics: a standard RGB camera FC7303 (the number of effective pixels is 12 million; image resolution is 4000 × 3000 pixels) and GPS and GLONASS satellite positioning systems (positioning accuracy in the vertical plane ± 0.1 m -visual positioning system, ± 0.5 m -positioning by satellites; in the horizontal plane ± 0.3 m -visual positioning system, ± 1.5 m -positioning by satellites).
The UAV flight took place at a height of 80 m from the level of the take-off area throughout the survey object.The initial resolution is 2.5-3 cm per pixel, depending on the local relief heights.The drone was manually controlled by the pilot using the DJI Fly mobile app without using a flight mission.The flight took place in parallel lines with at least 70% overlap of images.
The data obtained were processed by photogrammetric algorithms using Agisoft Metashape Professional Version 1.5.4 software.The data processing algorithm included the implementation of the solution proposed in the user manual: image alignment, building a dense point cloud, building a textured model, a height map and an orthomosaic.
The search for trees and the calculation of dendroparameters were performed using photogrammetric point clouds using the statistical programming environment R using the functions of the specialized package lidR v. 3.2.3.At the first stage, the procedure for classifying the set of points into 2 types was carried out: points of the earth's surface and other points located above.Ground surface points were identified using the cloth simulation filtering algorithm implemented in the lasground() function.Next, point clouds were normalized with respect to points on the earth's surface using the tin algorithm in the lasnormalize() function.At the final stage, a digital height model was built using the pitfree algorithm implemented in the grid_canopy() function.After that, using the tree_detection() function, we performed an automatic search for tree vertices in the point cloud.Then, using the obtained coordinates of the tree vertices, the crowns of the corresponding trees were segmented using the segment_trees() function.For verification, the results of the automatic search for trees were compared with the orthophotomap of the area and field counts.

Results
Field surveys showed that 22 years after the creation of forest plantations, young forests with a predominance of Scots pine were formed in all experimental plots (Table 2).It has been established that the share of forest plantations in the areas of wood stock is 38 -44% of the total stock, therefore, tree species of natural origin predominate in the areas of forest plantations.
The composition of the forest stand in the area of forest plantations created by planting material with a closed root system is 44Slk53Sev2B2Ols + I, the total stock is 189m 3 /ha.The E3S Web of Conferences 458, 08028 (2023) EMMFT-2023 https://doi.org/10.1051/e3sconf/202345808028composition of the forest stand on the site with pine sowing is 38Slk56Sev4E2B+Os, the total stock is 128 m 3 /ha.
It follows from the survey data that the creation of forest crops by planting (experimental plots No. 2-4) is preferable to sowing (plot No. 1), since the indicators are higher in terms of the composition of young stands and the stock of forest crops.The stock of pine forest crops created from briquetted seedlings is significantly higher compared to crops created by sowing, respectively -83 and 48 m 3 /ha, planting without tillage -76 m 3 /ha.At the same time, the mortality of pine in the area with sowing is greater than in the areas of forest crops created by planting -320-488 ind./ha, respectively, which is less than 3% in terms of stock.
Under the canopy of forest crops in all experimental plots, undergrowth of forest-forming species appears with a total number of 620-790 pcs/ha.On the plots without tillage (No. 3 and 4), spruce predominates, and on the plots on cultivated soil (No. 1 and 2), deciduous species predominate.Presumably, this may be due to the characteristics of soil conditions.The undergrowth is involved in the formation of forest phytocenoses in all areas.The total number of undergrowth is from 220 to 410 pcs/ha.In the composition of the undergrowth, mountain ash predominates in all cases.
As a result of aerial photography, 651 images were obtained.Photogrammetric processing made it possible to obtain 336 thousand tie points with an average discrepancy of 2.1 pixels (reprojection error).As a result of photogrammetric processing, an RGB orthomosaic with a spatial resolution of 2 cm/pixel (Fig. 1) and a dense cloud of points in the amount of 55 million units and an average density of 3.9 thousand units/m2 were obtained.
To analyze the distribution of the species and quantitative composition of vegetation over the territory, trees were marked on an orthophoto map for experimental plots, as a result of which the quantitative distribution of tree species was determined: from 60% (plots without tillage) to 80% (plots with tillage) -pine, 10 %-22% -birch.These indicators are consistent with the field survey of the area (differences less than 10%).For the analysis of three-dimensional point clouds, points were used that belong only to individual experimental plots.The characteristics of the point clouds used for each of the sections are presented in Table .2. As a result of running the algorithm for automatically searching for trees using point clouds using the lidR package, it was possible to detect (Fig. 2) a significant part of the trees in all areas and determine their heights (Table 2).As a result of manual detection according to the orthophotomap of the area, 250 trees were counted in plot 1, 353 trees in plot 2, 277 trees in plot 3, and 130 trees in plot 4, and automatic detection revealed 215, 313, 240 and 118 trees, respectively, which , in general, is about 90% of the total number of trees.At the same time, most of the trees (85%) found by the algorithm were identified correctly.The number of false positives and the number of missed trees were quite low, and the weighted average quality score was 0.89, which indicates a high efficiency of tree search.When comparing the heights of individual trees on photogrammetric point clouds with field measurements, it was possible to reliably determine only the height for the main forestforming species in the areas (pine, birch).At the same time, the heights measured according to the UAV data were in good agreement with the heights measured by the ground method.The relationship between the heights of individual trees obtained by different methods is very significant (R 2 = 0.96), and no significant differences were found between the heights obtained by different methods.The maximum height of Scotch pine trees in the studied areas did not exceed 13 m, and the average values varied from 10.1 m to 12.6 m.

Discussion
Plantations pine created by sowing and planting vary in density, average height and diameter, and stock.The best results -on the site of forest crops, created by ordinary plantings of briquetted seedlings on the treated soil.These results can be used in drawing up projects for the development of forests and choosing a method of reforestation on the territory of the Republic of Karelia (Russia).
The use of UAVs in assessing the success of reforestation made it possible to obtain an orthophoto map with a high spatial resolution and a visible projection of the crowns of individual trees, to understand their spatial distribution, but there are difficulties in determining its abundance and species composition.Small undergrowth is not identified on the orthophotomap.
It was also not possible to determine the height of undergrowth for small and medium groups (from 0.5 to 1.5 m) using 3D point clouds due to the difficulty of separating 3D point clouds belonging to different classes ("earth surface" and "low vegetation").To solve this problem, one can probably use multispectral imaging with near infrared (NIR) and/or far infrared (RedEdge) spectrum channels.
At the same time, the results of the automatic search for individual trees (in pieces) using 3D point clouds show that the selected tree detection algorithm showed good results in practice, which indicates the possibility of its application for counting and estimating the height of trees in forest crops.These results confirm the already published data obtained by scientists in areas with natural regeneration [7].

Conclusion
The applicability of photogrammetric aerial photography to assess the success of the growth and development of forest plantations is shown on the example of studying experimental forest plantations created in different ways on the territory of the Republic of Karelia.The actual data obtained during the survey of the plots made it possible to analyze the state of forest plantations.On all experimental plots, young stands were formed with a predominance of Scotch pine.The share of forest plantations on the plots in terms of timber stock amounted to 38 -44% of the total stock.The UAV data processing method made it possible to build an orthophotomap of the area and calculate the quantitative distribution of tree species: 60% (plots without tillage), 80% (plots with tillage) -pine, 10%-22% -birch.These indicators are consistent with the field survey of the area (differences less than 10%).It has been established that the method of photogrammetric aerial photography can be successfully used to assess the biometric characteristics of individual trees.To increase the correctness of tree detection by photogrammetric point clouds, the methods used need to be improved.

Fig. 1 .
Fig. 1.Orthophotomap of the area.The boundaries of the experimental plots are marked with contours.

Fig. 2 .
Fig. 2. Detected tree tops in section No. 2 (marked with dots) on a fragment of the orthophotomap of the area (left) and height map (right).

Table 2 .
Results of tree detection by point cloud