1. Introduction
The Intergovernmental Panel on Climate Change (IPCC) has identified climate change as a major threat to agriculture, food security and the livelihood of millions of people across the globe [
1]. Forests are particularly at risk, as their long life span does not permit rapid adaptation to environmental change [
2]. Climate change is predicted to alter: relative abundance of tree species within forests; tree phenology and frequency and intensity of disturbance mechanisms, including wind and forest pests [
3]. Planted forests have been identified as being especially susceptible to climate change as they are generally dominated by only one or a few tree species with limited genetic diversity [
4].
In its origins around the world, plantation forestry was an entity focused on producing wood products such as fuelwood or timber [
5]. In the 1980s, a new paradigm emerged which emphasised the potential to manage forests for range of ecosystem services alongside more traditional forest commodities [
6]. In the UK this approach is named under the collective umbrella of Continuous Cover Forestry (CCF). CCF is a management approach that maintains forest cover over time, by the selective felling of single or small patches of trees, which aims to produce a more diverse forest structure [
7]. CCF silvicultural principles, such as promoting structural diversity is well suited to assist making plantations less vulnerable to the effects of climate change [
8]. Transformation of even-aged plantations to CCF is a long and difficult task which needs frequent management interventions [
9]. A single management intervention can have considerable, long term impacts on forest structure [
10] and has been suggested to be comparable to natural pathogen infestation and sporadic fire [
11]. Therefore, new methods are needed for monitoring transformation which are cost and time effective when compared against traditional field data collection methods to assist in informing management interventions.
Unmanned aerial vehicle (UAV) offer opportunities for rapid collection of remotely sensed data from which forest metrics can be achieved [
12,
13]. Progression of computer power and dense image matching techniques has enabled the development of dense photogrammetric points clouds from UAVs [
14]. Photogrammetric generated points clouds are passive and produce detailed Canopy Height Models (CHM) but are limited in their ability to penetrate the forest canopy [
15,
16]. However, photogrammetry is a low cost alternative to Light Detection and Ranging (LiDAR), and enhances the possibility of using high temporal and spatial resolution to monitor of forest [
17]. This development has opened up the opportunity for Individual Tree Detection (ITD), which enables detailed information to be retrieved on a tree by tree basis [
18], which is vital in assisting with monitoring forests through transformation to CCF [
19]. ITD methods revolve around three sequential steps which involve tree detection, feature extraction and estimation of tree attributes [
20]. A variety of approaches have been developed to delineate individual trees from both LiDAR and photogrammetry [
21], but generally they can be divided into two categories, i.e., raster based and point based methods. Point based methods are usually considered more computationally demanding but allows the whole three-dimensional data to be considered. To-date raster based approaches have been preferred in ITD studies [
22]. A variety of approaches to delineating individual trees from the CHM have been developed: local maxima detection, valley following, template matching and marked point processes [
23,
24,
25]. Raster based approaches are relatively more simple, computationally efficient over large areas and the CHM can be generated using cost effective photogrammetry.
The performance of ITD methods has been found to vary with forest structure [
26], species [
27] and sensor type [
28]. Achieving optimal ITD accuracies is heavily dependant upon two factors: (i) the extraction methods [
29], and (ii) the parameter combinations [
30,
31]. Investigation into parameter choice in ITD is in its infancy, with the majority of studies conducting sensitivity analyses by changing one parameter at a time whilst keeping the other constant (instead of globally) [
31,
32], thereby not integrating multi-way parameter interaction [
33,
34]. Monitoring transformation using ITD methods is made additionally challenging by rapid structural alterations which occur during selective tree felling to promote natural regeneration. Selective felling has been found to alter the allometric relationships of the remaining trees and therefore ITD parameterisation is necessary to incorporate these changes and is vital in determining transferability of ITD methods across stages of transformation [
35,
36]. Therefore, ITD methods used to monitor transformation to CCF must be able to efficiently and accurately self-parametrise to different forest structures to maintain cost effectiveness.
This study address the aforementioned practical challenges by: (i) presenting a novel Bayesian optimisation approach of parameterising an ITD framework that uses external tree data, removing the need for site specific allometric models, (ii) demonstrate the effectiveness of this framework against combinations of fixed sized local maxima search windows which is a well established ITD approach and, (iii) investigate the transferability of this approach to retrieve tree biophysical variables from forest stands in a range of transformation stages.
4. Discussion
This research developed a novel self-parametrising, transferable photogrammetric based UAV method capable of monitoring transformation to CCF at the individual tree level, which could also be transferred to any high resolution CHM, such as delivered from LiDAR. Our results demonstrate a robust and efficient framework for monitoring even-aged plantations to CCF, capable of providing accurate biophysical estimates from a range of forest structures. In the following sections, we discuss our findings in relation to other studies to identify the effectiveness of this approach and outline future improvements.
4.1. Accuracy of Individual Tree Detection
Our findings indicate that the BOVW can self-parametise to achieve respectable tree detection accuracies (mean test F-score = 0.88) and resulted in a significant improvement of F-score in comparison to the OFW approach (
= 0.05). Improvements of 5% in tree detection, as found in this study in comparison to the OFW method, can have significant enhancements on forest inventory estimates as tree detection has been found to be the primary source of error in forest inventory estimates when using ITD [
66]. Despite the inclusion of a range of forest structure and species compositions, we achieved segmentation results comparable with studies applied in similar forest types with minimal input changes between sites [
23]. Vega et al. [
67], using the PTrees ITD method reported to segment individual trees in a conifer plantation in south western France to an F-score of 0.95. Lahivaara et al. [
68] was able to achieve a reported accuracy of 70.2% in a managed boreal forest in Eastern Finland and Maltamo et al. [
69] detected only 39.5% of all trees in southern Finland using local maxima techniques.
The independent test site results were consistent with Sites 1–5 in which the BOVW achieved a higher test tree detection accuracy in comparison to the OFW. While providing evidence for transferability of this approach, the results for the independent test site could be further improved by advancements in the raw data collection and the inputs to the BOVW. The low point densities (4/m
2) used to create the CHM for the independent test site meant that smaller diameter trees were less well represented in the CHM and therefore increased the difficulty in their detection. The h-cr quantiles used in the VW were derived from Douglas fir data and not the species present in the test site. Therefore, with access to individual tree measurements of the species present at the site would allow the integration of the correct allometric relationships into the optimisation process. Whilst the BOVW results are viewed favourably, caution should be taken when comparing ITD results across studies given the differences in stand complexity and species composition [
22].
4.2. Biophysical Estimates
Our results confirmed the general statement that tree heights are mostly underestimated from UAV photogrammetry, but provided a strong relationship with the observed tree height (R
2 = 0.93). On the other hand, tree height errors (RMSE = 2.1 m) reported in this study exceeded the accuracy reported by other UAV photogrammetry research and could be attributed to several errors [
70,
71]. Firstly, DTMs derived from photogrammetry are based only on the canopy surface which is viewable from the sky. Therefore, in areas of high canopy cover there are significant areas of omission, which have been coined as dead ground [
72] and results in DTM extrapolation. The undulating topography and varying densities of canopy cover found in the study sites caused fluctuating difficulty in DTM production. Secondly, vertex measurements are also error prone with operator skill and canopy closure affecting the accuracy of the tree height measurements [
73], with discrepancies being reported up to 1.3 m in coniferous plots [
20]. Thirdly, the quality of the UAV photogrammetric product. Environmental conditions have a significant impact on the quality of the photogrammetric data set, with wind speed being shown to impact height by up to 0.5 m [
74]. Accumulatively, these factors could explain the discrepancy in tree height estimations. The addition of a LiDAR DTM it is expected to improve tree height estimations using this approach, as it has been found to provide accurate DTMs in forests with high canopy cover [
75].
Monitoring tree diameter distributions is a common method of discerning the stage of transformation and providing direction for future management interventions [
76]. Although a relatively simple model, quantile regression is rarely used in deriving diameter estimates in ITD studies. The results from this approach are encouraging (mean RMSE = 5.6 cm), especially when considering:(a) predictions were based upon h-d models parameterised using external data sets, (b) the method required limited training data to select the correct quantile (
Figure 9) and (c) used tree height as the sole predictor which had errors (RMSE = 2.1 m) exceeding that of other studies [
77,
78]. Allometric models are highly vulnerable to errors in their input data, especially when using one predictor and therefore improvements in diameter predictions are expected with enhancements in tree height accuracies. Similar mean diameter accuracies were achieved by developing site specific Weibull h-d models (RMSE = 6.8 cm) and using either approach would result in predictions within the broad range found in the wider literature which fluctuates from 15% to 23% rRMSE [
69]. Vastaranta et al. [
79], using a multi-source ITD inventory in a range of forest stands in Finland, reported dbh accuracies ranging from 10.9% to 19.9% RMSE.Valbuena-Rabadán et al. [
80] reported in a Scots pine (
Pinus sylvestris L). plantation dbh estimates of 18.97% RMSE. The comparison of ITD biophysical estimates is complex and caution must be taken as it is highly data dependant [
81]. Comparable diameter predictions results found in this study provides evidence that the quantile approach could be used as an alternative when site specific h-d relationships are not available and when collecting only a few tree heights possible due to the labour intensive nature of collecting tree heights.
4.3. BOVW Parametrisation
We consider the following aspects relevant advantages of the proposed BOVW approach: (i) the proposed approach is structured and follows the same steps regardless of the parameters being tuned or the person tuning it. In contrast, the steps taken, the depth and choice of parameter exploration is predetermined by the user when using approaches such as grid searches or manual parametrisation, requiring previous knowledge, (ii) parametrisation occurs globally. Incorporating multi-way parameter interactions, addresses weaknesses identified in previous studies which sought optimal parameter sets by altering one parameter whilst keeping others constant and thereby not integrating multi-way parameter interactions, (iii) transferability. This approach maintained successful segmentation accuracies with minimal input changes across sites of different forest structures, including the independent test site. The only input change between sites was selecting the dominant tree species and (iv) training sample size. The BOVW approach requires a limited number of training samples to achieve the same accuracy as using all the training data, implying that it could be cost efficient in the field. However, further research is required to investigate the influence of forest structure on the required training sample size. These advantages ensure the BOVW is easy to use, adaptable and efficient, making it highly attractive to forest managers and research projects which require ITD components. Thus far, BO has been is little exploited in relation to ITD parametrisation.
4.4. Future Work
Issues of automated parametrisation are addressed in this study, but the main unsolved obstacle to ITD is still present with the difficulty of delineating suppressed and understory trees [
63]. Improvements can be seen with the BOVW approach presented here when compared with the OFW approach, however, there is still a tendency to miss smaller suppressed trees. The tendency of missing trees will increase in multi-layered, dense forests [
82]. Some methods have been developed to improve segmentation accuracies of suppressed trees, however these are still experimental and generally computationally expensive over large areas [
83].
This framework could be further improved through including multiple segmentation methods in the optimisation stage, allowing BO to fit the segmentation method to the forest structure. Previous research has shown that tree segmentation approaches differ in their accuracies in regards to forest structure [
29] and therefore, allowing multiple segmentation approaches to be incorporated should improve tree detection accuracies. Integrating multiple segmentation methods into this approach would allow the whole processing chain to be optimised to the forest.
5. Conclusions
In this study, we compared a novel Bayesian optimisation individual tree approach to an established and effective individual tree detection (ITD) approach over five sites in transformation to Continuous Cover Forestry (CCF) and to a test site located in the Teakettle Experimental Forest, California, USA. The novel method (BOVW) uses a variable search window and a Bayesian optimisation approach of parameterisation which utilises external tree data to derive the correct allometric relationship. The established method (OFW) uses a combination of fixed sized search windows to smooth and identify individual trees from the canopy height model.
Our investigation demonstrated that the novel BOVW method was transferable between sites with limited manual parameterisation and was able to achieve advantageous results in comparison to the OFW ITD method. The BOVW was also able to achieve higher detection rates for smaller diameter trees than in comparison to the OFW. Currently, the BOVW approach uses a variable window search approach, however future research could be focused on integrating multiple segmentation methods and thereby optimising the entire processing chain to the forest and further improve tree detection accuracies.
This study presents a transferable, self-paramatising ITD method which is capable of adjusting to changing forest structures which occur during transformation to CCF. Requiring limited training data, little prior knowledge of optimal parameters and the use of external data sets to derive allometric relationships results in a highly effective, easily applied and cost effective ITD method which can be used to monitor transformation to CCF at the individual tree level.