1. Introduction
Unmanned aerial vehicles/systems (UAVs/UASs) have been becoming more common solutions for various forestry purposes, e.g., for 2D/3D mapping, environmental monitoring, health assessment, tree species classification, firefighting or SAR missions [
1]. The general availability of low-cost platforms has intensified both scientific research and the applicability of UAS solutions in forestry. As an alternative to manned aircraft and satellite-based remote sensing (RS) solutions, autonomous low-altitude systems can provide high spatial and temporal data resolution while offering relatively low operational costs and flexibility of data acquisition [
2,
3,
4,
5]. The main limitation of small UASs is their mapping range, mainly due to flight duration [
6,
7]. Based on the literature review conducted by Manfreda et al. [
8], the cost-effective range of UAV systems for environmental monitoring was only below 20 hectares. Nevertheless, this aspect has improved over the last six years due to both technological advances and industry-specific optimization research [
7,
9,
10,
11,
12].
The last decade has seen an increased interest in the research and application of precise forest inventories at stand or even single-tree level [
13,
14,
15]. In Poland, according to forest management plans [
16], stand-level inventories are conventionally planned once per 10 years—for midterm management—and 1 or 2 years before harvesting in order to evaluate the potential yield. Depending on the potential value of a particular stand, more or less accurate/time-consuming inventory methods can be applied. For instance, angle-count (relascopes) or visual assessment methods are generally used for homogeneous stands. If there are similar stands in the vicinity, the estimates can be transferred from one stand to another. For particularly valuable units, a total survey is planned, whereas for brief estimates, yield tables can be used. Due to the variety of available inventory options characterized by varying degrees of accuracy and subjectivity driven by stand complexity, some discrepancies between the inventory results as compared to post-harvest crop evaluation may occur. This problem could prove crucial for medium- and long-term forest management as well as for governmental and international reporting.
Remote sensing (RS)-assisted forest inventories enable the possibility of objectively estimating spatially continuous forest attributes for large and hardly accessible sites, with promising performance [
17,
18,
19,
20,
21,
22]. Among some popular inventory methods that utilize RS data, individual tree detection (ITD) has shown a potential to be useful for precision forestry [
14,
23,
24,
25,
26]. The core aspect of this method is the detection and delineation of single trees (segments) based on the following RS layers: point clouds, crown height models and/or orthomosaics. In the next step, RS metrics are assigned to each segment, which in turn serve as independent variables in some sort of model developed to estimate the dendrometric attributes of a single tree, for example, height, stem volume, DBH, etc. The low-altitude character of UAS operations enables the derivation of fine-resolution rasters and very dense point clouds, which can be useful for the accurate detection and delineation of the first canopy layer crowns and in some cases can outperform LiDAR-based tree detection [
27,
28] and airborne laser scanning (ALS) biomass estimates [
29]. In general, many studies show a high degree of agreement between UAV images and laser-based forest metrics [
30,
31,
32,
33].
As mentioned above, mapping large forest areas using UASs can still prove to be a challenging task. The main limitation of low-altitude UAV systems is the considerable amount of data required to cover vast areas. This entails requirements for flight duration, constant weather conditions during data acquisition and long-range signal transmission (especially in forest environments). In addition, big data usually has considerable redundancy and a lot of noise. Depending on the goal of the inventory, such a detailed representation can be an overkill. Therefore, optimization steps seem to be necessary when planning larger UAV inventory campaigns.
Ground sampling distance (GSD) as well as flight pattern density (photo overlap) are the factors that have a decisive influence on the resolution of the RS layers, their quality, as well as on the mapping range [
34]. Depending on the scale and objectives of the respective inventory, an operator must define these parameters accordingly. The GSD is closely linked to the camera parameters and flight altitude. As shown in some studies, lower flight altitudes do not necessarily mean better results [
10,
29,
34,
35]. Among other things, this could be due to the increased influence of the tree crowns moving with the wind on the projection of the camera image. Moreover, a low approach results in more homogeneous textures of the captured photos. This may confuse the digital aerial photogrammetry (DAP) algorithms during the aero-triangulation phase of aligning the photo blocks [
10]. On the other hand, high-altitude photogrammetric missions provide coarser resolutions and sparse RS material where the ITD approach may not be reliable for estimating some forest attributes—for instance the DBH [
36].
As outlined in the previous paragraph, there is a keen interest in optimizing image-based aerial forest inventory methods. Nevertheless, to the best of the authors’ knowledge, there is little work on optimizing single tree detection methods for growing stock volume (GSV) estimation, based on photogrammetric UAV data. Most research on optimization so far was related to tree height estimation [
10,
35,
36], which can be directly measured, tree detection and segmentation [
10,
27,
36], which are very often auxiliary or intermediate results, DBH [
36] or aboveground biomass [
29]. It is worth noting that GSV, as a traditional indicator of timber resources, carbon stock, management efficiency and sustainability in the forestry sector, is one of the most important forest attributes to be determined [
37,
38]. Moreover, GSV is closely related to aboveground biomass, which in turn becomes an important indicator of carbon stock assimilation. We found one study by Tuominen et al. [
39] which partially covers similar issues. Nevertheless, in that study, an area-based approach (ABA) was tested with two GSD values, i.e., 20 and 50 cm, which, however, were resampled in post-processing rather than originally captured.
Considering the above-mentioned aspects, this study aimed to test the influence of ground sampling distance and photo overlap on the efficiency of UAV-ITD stand-level growing stock inventories, with the perspective of presenting optimization possibilities that could further improve practical implementations.
4. Discussion
The presented UAV-ITD method for forest stock inventory provided robust (
Table 6) and stable estimates (
Table 4) for three of the four investigated stands. The results were neither directly dependent on the ground sampling distance nor on the overlap within the analyzed range. Pessacg et al. [
63] also showed a small difference in the quality of UAV-reconstructed DTMs based on aerial photographs taken at 50 and 150 m altitudes. In this case, the mapping range and time constraints could be further optimized at a coarser resolution of RS layers; as higher flight altitudes and wider flight patterns allow coverage of larger areas in a single flight. Similar results can be found in Grybas and Congalton [
35], where the differences in tree height estimates between individual flights with different altitudes and overlap configurations were less than 1%. This allows larger areas to be covered with fewer images. Smaller datasets, in turn, reduce the computational requirements, which are known to be quite high for photogrammetry. The promising results even for 6 cm GSD suggest that a broader resolution test is needed. The authors did not plan any flights at higher altitudes, i.e., above 120 m AGL, as these are prohibited in Poland under standard UAV operations. Nevertheless, some results for GSV estimated on the basis of coarser resolution DAP data can be found in the literature. For instance, Cao et al. [
32] reported R
2 and rRMSE values of 0.73 and 18.24% using 10 cm GSD photo sets for GSV prediction in poplar and dawn redwood stands. Based on 40 cm RS pushbroom DAP layers, the models and approaches proposed by Strunk et al. [
64] were able to explain between 65 and 76 of GSV variation among coniferous stands. A further reduction in the resolution of RS layers can be found in Tuominen et al. [
39], where negligible differences in GSV estimates in Pinus sylvestris/Norway spruce stands were found between UAV-DAP layers of 20 and 50 cm GSD (rRMSE 26.12 and 25.52, respectively). However, the compared layers were resampled from an original GSD of 10 cm, rather that independently acquired, and an area-based method was used in that study. Finally, Goodbody et al. [
31] have presented a comprehensive review showing a high degree of agreement between ALS and DAP-based volume estimates for various GSDs. As can be seen, the literature is abundant with research concerning coarser resolutions; therefore, our study focused more on potential performance gains when using finer pixels.
Regarding another important aspect of photogrammetric missions, namely, the photo overlap, three variants (80 × 80%, 90 × 80% and 90 × 90% front/side) were tested in this study, showing a very low influence on the GSV estimates at the flown altitudes. In six out of nine flights, the GSV estimates based on the 80 × 80% overlap were lower than the 90 × 90% counterparts only of about 13.7 m
3/ha. However, due to the relatively small sample, such a tendency could be a random effect. In this study, no sparse flight patterns were flown, as these are generally not recommended for low-altitude forest mapping [
65,
66,
67,
68] and can lead to significant gaps in the spatial continuity of the data [
63,
69,
70]. Photo overlaps of less than 70%–80% are more common in manned-level photogrammetry, with a threshold of 60% lateral overlap generally considered optimal [
71,
72]. Higher flight altitudes usually allow for less dense photo overlap [
67], as more common features can be detected on the images since the scenes are less homogeneous. In addition, the alignment algorithms do not suffer from object movements caused, for example, by wind [
69].We found one study by Lima et al. [
34] who analyzed lower overlaps for UAV operations and found that lateral overlap is more important than longitudinal with a 70 × 90% setting and a GSD of 3.1 cm being optimal for Scots pine stands. However, most research indicates rather denser flight scenarios, or simply does not consider more sparse flight patterns [
35]. For instance, Young et al. [
10] were more conservative and suggested that according to flight time constraints and the need to cover extensive areas, 90 × 80% front/side overlap may be optimal. Pessacg et al. [
63] show that a 95 × 85% overlap of nadir photos is generally best for forest conditions if accurate image-based DTM reconstruction is the main goal.
The stage of tree detection and crown segmentation is a core aspect of the post-processing pipeline in the ITD inventory method. However, this factor, which was not analyzed in this study (as it is a very complex aspect that deserves its own investigation), proved to be the factor most strongly associated with errors in the estimation of GSV at the herd level.Tree detection rates from different flight scenarios were in the range of 77% to 89%. Such a score is within the range of usually reported tree detection rates for the ITD method as follows: ~70% in Goldbergs et al. [
73], >80% in Krause et al. [
74] and 73%–94% in Mohan et al. [
75]. The most important drivers of ITD accuracy are usually stand density, tree clustering and tree vertical structure [
76,
77,
78]. Surprisingly, the fewer trees that were detected, the more accurate the estimates were (
Figure 10). Some plausible explanations for such a result could be as follows: (1) the omission of small trees during sampling for the model development, causing increased volume estimates for individual trees—the difference in the means of volumes for a single tree according to the model tree and reference datasets was 0.05 (
Table 2 and
Table 3), (2) higher fragmentation of pine crowns due to their structural characteristics, and (3) the presence of small second-layer trees which were not detected in the UAV-RS layers. Nevertheless, the auxiliary variable of the model, i.e., crown area, should leverage the influence of the tree detection rate. For example, a larger segment should have a similar estimated value as the sum of few smaller ones. Corresponding results were obtained by Yu et al. [
79], where under-segmentation yielded better results than over-segmentation, with estimation errors ranging from 0 to 55% and 12 to 161%, respectively. They pointed to similar reasons for this result. Nevertheless, as mentioned before, tree detection and segmentation were not the main subjects of this study. Moreover, in order to gain maximum control over the analyzed factors, i.e., ground sampling distance and photo overlap, we ensured that segmentations were as accurate as possible for each scenario. Nonetheless, this issue certainly requires further research.
Allometric attributes like tree height and crown area are significantly correlated with tree volume [
59,
60,
76]. Relatively low auto-correlation between these two variables makes them potential volume predictors when combined in some form of regression model. In this study, such a model was able to explain 82% of the variation in individual tree volumes, with an rRMSE = 17.1% for Scots pine stands. A similar model was able to explain 68% of the variation between individual tree volumes in Norway spruce-dominated stands [
76]; however, the tree heights were derived from ALS data. The same model, calibrated on UAV photogrammetric point clouds representing eucalypt plantations, provided estimates with an rRMSE = 20.3% [
80]. For some Quercus brantii stands, Fakhri and Latifi [
81] were able to provide aboveground biomass estimates with an rRMSE of less than 10 using a consumer-class UAV.
The results obtained at stand level show that the DAP-ITD method can be competitive with ALS inventories in terms of GSV estimation. In this study, the errors in GSV prediction were no more than 15.3% in relation to the conventional method as a reference, and 12.7% when compared to the TLS score (
Table 3). Similar error rates were obtained by Kotivuori et al. [
82] for Finnish conifer stands. By modeling drone-derived photogrammetric point cloud data, they were able to obtain a stand-level rRMSE of 13.1% for locally calibrated models and 20% when a nationwide model was applied. Parkitna et al. [
83] and Lisańczuk et al. [
21] reported rRMSE ranges of 16%–23% and 12%–20%, respectively, for Scots pine-dominated stands using the area-based ALS method. On the other hand, in our study, some estimation errors were as low as 0.6%. Similarly good results were obtained by Stolle et al. [
84] for Pinus taeda plantations (error rates of −3.5 to 1.8%). Although the study showed a high degree of repeatability of the UAV-based GSV estimates (coefficient of variation from 2 to 6%) and the maximum error obtained was still highly competitive compared to the alternative RS stock inventory methods, a remarkable variance (66% and 70% CV for the TLS and the traditional method, respectively) was observed in this narrow error interval (from −3.7 to 15.3%). Again, a more comprehensive study across more validation stands is needed to measure the strength of this random effect. Tuominen et al. [
39], for example, found only a 10% difference based on more than 80 conifer stands; however, they used only one flight repetition.
The above comparison can only give a brief overview of the possibilities. On this occasion, it is important to point out some limitations of our research. Regarding RS metrics, the height of the trees/canopy is one of the strongest and most common predictors of their volume [
76,
85,
86]. Therefore, it appears that high-quality DTMs are crucial for normalizing tree heights.
In our study design, we relied on self-generated image-based DTMs. However, the investigated studied sites were relatively flat, with some exposed patches of bare ground, which facilitated accurate DTM interpolation and alleviated its influence on errors in tree height measurements. However, in rugged and/or steep terrain, this issue can become more challenging [
87,
88,
89]. On the other hand, archived high-resolution LiDAR-based DTMs are available in many countries [
90], the use of which can reduce the influence of some site characteristics, including terrain and trees density, that can affect ITD accuracy [
76,
77]. Furthermore, Giannetti et al. [
91] showed that it is possible to achieve ALS-level GSV estimates by using photogrammetric DTM-independent metrics, even in mountainous regions. Nonetheless, further optimization encompassing more diverse site and terrain features seems to be an appropriate research focus.
In order to keep the investigated factors, i.e., GSD and photo overlap under maximum feasible control, only one camera system was used to capture the images of all flights. Nevertheless, Domingo et al. [
87] showed little influence of sensor resolution on biomass prediction in tropical forests, which are much more complex in structure than even-aged single-layer pine stands in Poland. They compared two camera types: the Canon IXUS127 HS with 1/2.3″ sensor size and 16 megapixels resolution and the Canon S110 with 1/1.7″ sensor and 12 megapixels. They obtained R2 and rRMSE differences of 0.01 and 0.7% for 10 cm GSD and 0.11/5.8% for 15 cm GSD. As we can see, the difference between the 12 and 16 megapixel sensor resolution was negligible for 10 cm GSD. However, their study also showed some variance in the repeatability of results for the same camera with different flight parameters, i.e., the rRMSE ranged from 31.5 to 37.7% from 70 to 80% side overlap.
No trends were found with regard to the influence of weather conditions; however, this element was relatively constant during the image collection (
Table 6). If this premise is true, it testifies to a further universality of the inventory method tested. However, in order to substantiate this statement, the following aspects should still be investigated on a broader scale: influence of tree detection and segmentation accuracy, terrain diversity, stand structure and species complexity, as well as oblique images and gridded/multi-scale flight patterns.