3.1. Individual Tree Detection
The implementation of ITD from LiDAR data gave satisfactory results for the 236 trees analysed. The overall recall rate (r) was 0.73, ranging from 0.53 to 1, the overall precision rate (p) was 0.74, also ranging from 0.53 to 1, and the overall F-score was 0.74, with a range from 0.55 to 1 (
Table 2). These metrics indicate a robust performance for tree detection. Higher F-scores were obtained in stands with lower tree density and reduced canopy overlap, while denser canopies and greater crown overlap resulted in lower F-scores. This trend is common in deciduous mixed forests, where complex crown shapes and vertical structures can reduce the detection accuracy [
43,
50]. The lower accuracy of some plots may be linked to the use of an LMF to define the treetop, which can lead to overcounting in trees with complex canopy structures, such as in oak–hornbeam forests, where it is often not straightforward to identify a single crown centre treetop [
66,
67]. Furthermore, in dense forest stands with a closed canopy, crown overlap further complicates the search for and the correct identification of local maxima [
68,
69].
Our accuracy results were achieved by testing many processing parameter combinations, such as local maxima window size (3, 5, and 7 m diameter) and filtering (e.g., Gaussian). In fact, the choice of point cloud processing parameters, such as CHM resolution [
70], filtering method [
44], size and shape of the local maxima window [
46,
71], and segmentation algorithm used [
47,
72] strongly influence the ITD accuracy.
Our results are consistent with similar studies conducted in structurally complex mixed forests. Reference [
12] reported an average F-score of 0.66 ± 0.01 using a similar algorithm in a mixed broadleaved forest. The higher accuracy achieved in our study might have been determined by the use of an unprocessed CHM, which, according to [
44], is more reliable than the smoothed CHM because of the more accurate reproduction of the crown shapes. Ref. [
47] achieved, in mixed natural forest plots, an r of 0.72–0.85, a p of 0.58–0.51, and an F-score of about 0.64 using a local maximum algorithm. Notably, the stands analysed by [
47] are made up of a mixture of deciduous and coniferous species, a feature that simplifies the treetop identification process and therefore probably determines the higher precision of tree identification in some stands compared to our results. This was also the case of [
49]. They applied a similar workflow to mixed plots dominated by longleaf pine (
Pinus palustris) and turkey oak (
Quercus laevis Walter), achieving an F-score of 0.86. Their higher tree detection performance is likely due to the inclusion of both deciduous and evergreen coniferous species. In fact, coniferous species are generally easier to detect and distinguish due to their characteristic conical crown shape, which makes treetop identification easier compared to broadleaved trees [
43,
47]. Moreover, as highlighted by [
50], most segmentation algorithms assume a conical crown structure, thus ensuring better results when applied to coniferous species. On the contrary, tree identification for deciduous forests is still an open challenge, as there is no widely accepted and accurate method.
In addition to the species composition, variations in ITD accuracy can be linked to other ecological factors such as canopy closure [
46,
47,
73]. Reference [
46] tested different LM window sizes in forest types with different density conditions, finding that the optimal LM window size is highly species- and density-specific, significantly influencing detection accuracy (F-scores = 0.82–0.91). In our study, we used a single LM window size across all species and canopy density conditions due to the presence of mixed-species plots in our study area, which likely explains our comparatively lower accuracy. Ref. [
72] achieved an r of 0.85, a p of 0.70, and an F-score of 0.77 in a mixed plot dominated by sycamore (
Acer pseudoplatanus L.) and English oak using an MCWS algorithm. Overall, the better F-score value may be related to the number of reference trees used, which was significantly lower than ours.
3.2. Classification of Forest Species
The random forest classification achieved a high overall accuracy of 84% and a Kappa coefficient of 0.74 (
Table 3). Black locust was classified with 100% accuracy, while some misclassification occurred between white hornbeam and English oak (PA = 75% and 88%, respectively) due to their higher spectral similarity. The algorithm was trained with 133 tree crown objects and validated with 57 tree crown objects. The resulting drone-based species map (
Figure 4) illustrates the spatial distribution of English oak, white hornbeam, and black locust.
Comparing our results with other studies is not so straightforward due to differences in the number of species classified, training size, and the different features taken into consideration. However, examining comparable research on object-based tree species classification reveals several factors likely contributing to the high accuracy in our study.
Firstly, the inclusion of textural metrics in addition to spectral features in our classification process likely enhanced accuracy. For instance, our classification accuracy was slightly higher than similar studies that did not incorporate textural metrics, such as [
74], who reported 78% accuracy with four species. Similar findings were reported by [
75] (73% of accuracy), [
76] (64.85% of accuracy), [
77] (77% of accuracy), and [
7] (78% of accuracy) who consistently observed improved results when texture-based metrics were included in the classification process. Secondly, the fact that MAIA S2 is one of the multispectral sensors with the highest number of spectral channels likely contributed to our improved classification performance compared to studies that used lower-resolution sensors, such as [
74,
75].
However, our accuracy was lower compared to studies focusing on conifer species. Ref. [
74] reported a user accuracy of 87% in a coniferous forest using multispectral drone images. Ref. [
78] achieved 95% accuracy and a Kappa coefficient of 0.95 in an RF classification in a temperate forest, where two of the four species were coniferous, using the eight spectral bands of the WorldView-2 satellite (comparable to those of the MAIA sensor used in this study). In general, conifers are characterised by easier tree top identification and a greater spectral differentiation between species [
50]. These findings were also confirmed by [
79], who achieved an overall species classification accuracy of 90% for conifers and 80% for broadleaves when analysing multi-temporal datasets of Sentinel-2 images.
Concerning the species distribution in our study area,
Figure 4 shows how English oak is the dominant species in the xerophilic area of the Fagiana Forest (highlighted in red in
Figure 1), although it is also evident that English oak density is relatively low. This likely reflects the trees’ adaptation to the limited water availability. Indeed, reduced forest density increases the trees’ resistance to water scarcity by minimising competition for scarce resources [
80].
In the mesophilic areas of the Fagiana Forest (highlighted in yellow in
Figure 1), a noticeably denser English oak canopy cover can be observed (
Figure 4), indicating more favourable soil moisture conditions, which also support the sporadic presence of white hornbeam [
81]. The meso-hygrophilic areas of the Fagiana Forest (highlighted in green in
Figure 1) are dominated by white hornbeam trees, which, mixed with English oaks, form an oak–hornbeam association, typical of the Ticino Forest. This association is indicative of moister soil conditions, as evidenced by the taller tree and denser foliage compared to the xerophilic zones. In contrast, black locust is the least represented species, confined to distinct patches mainly along the edges of the mapped area, and with few individual trees scattered throughout the different forest microclimatic conditions (
Figure 4).
3.3. Plant Trait Retrieval
The cross-validation statistics of the models calculated between the measured field data and the estimated data retrieved from the PRISMA dataset resampled to the MAIA S2 spectral configuration (n = 50) are shown in
Table 4. Overall, both the LAI and CCC showed very high performance in cross-validation. More specifically, the LAI was estimated with high accuracy with all the investigated MLRA (R
2 = 0.84–0.90, nRMSE = 8.66–11.01%), whereas CCC showed a more variable performance depending on the MLRA used (R
2 = 0.68–0.83, nRMSE = 9.17–13.08%). Among the MLRA, the kernel-based algorithms (i.e., SVR, GPR) and PLSR showed the highest predictive capacity.
The goodness-of-fit metrics calculated between the field dataset collected near-simultaneously the drone acquisitions and the drone-based retrievals obtained by applying the developed MLRA models to the MAIA S2 spectra (n = 40) are shown in
Table 5. As expected, the results obtained using an independent validation dataset are slightly worse than those obtained in cross-validation on the PRISMA dataset resampled to the MAIA S2 spectral resolution. Still, both the LAI and CCC were accurately estimated. GPR, SVR, and PLSR showed the highest predictive capacity for both the LAI (R
2 = 0.81–0.83, nRMSE = 14.18–16.39%) and CCC (R
2 = 0.79–0.80, nRMSE = 22.45–27.7%). Using the fully independent dataset and the actual MAIA S2 data, all models showed a slight to moderate tendency to overestimate compared to the field data (rbias = 4.8–43.23%), especially for CCC. The scatter plots showing the measured and estimated values obtained from the MAIA S2 sensor for the LAI and CCC are shown in
Figure 5 and
Figure 6, respectively.
Overall, the results obtained in this study are promising towards the development of effective and consolidated retrieval schemes for the estimation of forest traits using drone-based sensors. The LAI and CCC were accurately estimated using an effective approach based on an MLRA trained on a reasonable amount of data, which could be easily applied to similar conditions or updated by adding more training samples to extend its applicability. Previous studies using drone data to estimate plant traits have mainly focused on crops, e.g., [
82,
83,
84,
85], while only a few studies have dealt with forests, e.g., [
25,
26]. The literature reports promising results in the retrieval of leaf or canopy chlorophyll content using a look-up table (LUT) or MLRA-based approaches. Ref. [
86] used a hybrid approach based on radiative transfer simulations coupled with an artificial neural network to estimate the LCC and CCC of apple orchards from DJI Phantom 4 multispectral data, achieving an R
2 of 0.73 and 0.79 and RMSE of 6.63 and 28.48 μg cm
−2, respectively. Ref. [
84] used the same sensor to estimate the LCC of sugarcane using the MLRA applied on vegetation indices, with R
2 = 0.68–0.98. Ref. [
82] used MicaSense Dual multispectral data to retrieve the LCC and CCC of maize using an LUT-based approach, obtaining RMSE = 3.74–4.92 μg cm
−2 and RMSE = 33.1 μg cm
−2, respectively. Such results are in line with ours (R
2 = 0.80, RMSE = 0.33 g m
−2, nRMSE = 24.02%), though we targeted a forest canopy which adds complexity to the retrieval because of the structure. In these ecosystems, previous studies have found contrasting results: [
26] estimated the LCC of Norway spruce from Parrot Sequoia data with moderate accuracy (R
2 = 0.45–0.49), while [
25] achieved very good results in retrieving the LCC of Himalayan pine from MicaSense RedEdge data using an LUT-based approach (R
2 = 0.94, RMSE = 6.20 μg cm
−2).
The retrieval of the LAI is often reported to be more challenging, especially in complex canopies with mixed sunlit and shaded pixels [
82,
86]. The presence of shadows, rows, and varying leaf angles in fact confounds the signal, posing difficulties in the accurate quantification of the LAI. Ref. [
82] obtained an RMSE of 0.61–0.7 m
2 m
−2 in the estimation of the LAI of maize, which has a more complex geometry compared to turbid medium crops, and [
86] achieved R
2 = 0.74 and RMSE = 0.28 m
2 m
−2 in the retrieval of the LAI of apple orchards. In beech forests, [
87] estimated the LAI from a drone-based RGB camera with R
2 = 0.59–0.7. In our study, the good results obtained (R
2 = 0.83, RMSE = 0.44 m
2 m
−2) indicate that the problem of different lighting conditions at the high spatial resolution of the drones is probably mitigated by using the average crown reflectance instead of a pixel-based retrieval, which is in line with the findings of [
26].
Figure 7 shows the LAI and CCC maps of the Fagiana Forest obtained by applying the best performing MLRA to the MAIA S2 segmented images acquired on 01/07/22 and 31/08/22. A significant reduction in the LAI and CCC (indicative of biomass loss and chlorosis) is evident between these two dates (
Figure 7e,f), likely as a result of the persistent drought of summer 2022 [
28,
29]. In the Fagiana Forest, CCC exhibited a stronger decline (
Figure 7f) compared to the LAI (
Figure 7e). This suggests that CCC may be a more sensitive indicator in detecting the effects of water shortage on vegetation functionality compared to the LAI. This result is consistent with previous studies showing the higher sensitivity of chlorophyll content compared to the LAI for assessing the condition of English oak trees in Ticino Park [
88,
89].
3.4. Functional Trait Analysis
The ANOVA results showed that LAI values were significantly influenced by time, forest microclimatic conditions, and species (
Table 6). The pairwise interaction between time and forest microclimatic conditions indicated a poorly significant interaction, suggesting that the LAI varies consistently through time across the three different microclimatic conditions. Differently, the pairwise interaction analysis indicated a highly significant interaction between time and species, suggesting that temporal changes in the LAI differ considerably across different species (
Table 6,
Figure 8a). Specifically, the Tukey post hoc test revealed a significant decrease in the LAI between 1 July 2022 and 31 August 2022 for both white hornbeam (−16%) and English oak trees (−12%). In contrast, no statistically significant change was observed for black locust during this period (
Figure 8a). Notably, black locust already had a low mean LAI value in June compared to the other two species.
The ANOVA analysis also revealed a statistically significant interaction between forest microclimatic conditions and species, indicating that the LAI differs significantly across various forest microclimatic conditions for the same species (
Table 6,
Figure 8b). For the black locust species, the LAI of the mesophilic forest is 22% lower than the meso-hygrophilic one and 55% lower when comparing the meso-hygrophilic and xerophilic environments (
Figure 8b). In the case of white hornbeam, only the differences in the LAI between meso-hygrophilic and xerophilic forest were significant (with a reduction of 20%) (
Figure 8b). For English oak, the differences in the LAI among the different forest microclimatic conditions were also significant, with a reduction of 7% between meso-hygrophilic and mesophilic, and a reduction of 18% when comparing the meso-hygrophilic and xerophilic forest.
The CCC analysis and, in this case, the ANOVA results, indicated that CCC values were significantly influenced by time, forest microclimatic condition, and species (
Table 7). A significant interaction between time and species was observed, suggesting that temporal changes in CCC differ substantially between species (
Table 7,
Figure 9a). The Tukey post hoc test revealed a significant decrease in CCC between 1 July 2022 and 31 August 2022 for black locust (−14%), white hornbeam (−21%), and English oak trees (−18%). Similar to the LAI results, black locust had a lower initial CCC compared to the other two species (
Figure 9a). However, in this case, the decrease in CCC for black locust between July and the end of August was statistically significant.
This suggests that CCC may be a more sensitive indicator of drought-induced vegetation stress compared to the LAI. In fact, as the product between LCC and the LAI, CCC can capture information on vegetation plant chlorosis and canopy biomass, thus providing a deeper insight into the physiological status of plants.
The ANOVA analysis also revealed a statistically significant interaction between forest microclimatic conditions and species, indicating that CCC differs for the same species significantly across various forest microclimatic conditions (
Table 7,
Figure 9b). The Tuckey post hoc test confirmed a statistically significant difference in CCC between the meso-hygrophilic forest and xeric environments for the three species analysed (
Figure 8b). In the case of black locust, the reduction in CCC between meso-hygrophilic and xeric forest was −56%, whereas, for white hornbeam and English oak, it was −18%.
Overall, both white hornbeam and English oak showed a decrease in the LAI and CCC when comparing the values obtained on 1 July 2022 and 31 August 2022, thus confirming the efficiency of the drone-based estimation of functional traits to detect a drought-induced variation. This result is in line with what was observed by PRISMA in the Ticino Forest between June and early September [
63]. Moreover, in the case of English oak, our results align with [
18], who described discolouration episodes in 10 different English oak stands in Ticino Park during the 2003 summer heatwave, highlighting the susceptibility of English oak to water scarcity and high temperatures and the effectiveness of remotely sensed CCC in detecting English oak stress.
In contrast, black locust showed a less pronounced decline in functional traits between June and the end of August 2022 compared to the other two species analysed. However, it already exhibited an unusually low LAI in June, with a mean value of 1.9 m
2 m
−2, for a broadleaf species at the peak of the vegetative season. Field observations during the sampling campaign confirmed that black locust trees had relatively small canopies and appeared to be in poor health. This pre-existing condition could explain the lack of a significant reduction in the LAI over time, as the species was already under stress. Thus, although black locust is generally recognised for its stress tolerance, its non-significant reduction in functional traits under drought conditions in the Fagiana Forest context could be due to its compromised health, rather than a distinctive tolerance to the lack of water. Ref. [
90] reported the decline of this species in the Ticino Forest, underlying the species’ vulnerability to climate change in this region. Reference [
91] further supports this by suggesting that black locust distribution models predict its decline in Southern Europe, potentially leading to a northward range shift favoured by future warmer climatic conditions.
The drone-based functional trait retrieval effectively captured the variation in the LAI and CCC among the three analysed species within the different forest microclimatic conditions. Both LAI and CCC values were higher in the meso-hygrophilic and mesophilic forest area, compared with the xerophilic one. This reflects the general smaller tree leaf size and reduced canopy in the drier forest area, as a form of adaptation to the lower water moisture availability [
92].
3.5. Strength and Limitations
The results of this study highlighted the importance of integrating drone-based multiple sensors, which together provided a comprehensive and accurate analysis of the Fagiana forest ecosystem. On the one hand, we obtained the precise spatial reconstruction of both forest structure and species composition. On the other hand, the proposed integrated approach allowed the accurate quantification of forest functional traits (the LAI and CCC). The drone-based high-resolution tree-level data obtained offer valuable and detailed insights into forest structure and ecological processes, accounting for variations related to species and forest microclimatic conditions. The processing workflow was optimised for automation and based entirely on open-source software, ensuring both efficiency and accessibility. Therefore, this approach can be applied in similar contexts.
The developed methodology, by providing an automated workflow for the accurate reconstruction of forest structure and plant trait retrieval, represents an important step forward in the understanding and parameterisation of process-based ecological models for estimating the gross primary productivity of forest ecosystems, which is fundamental for assessing and predicting fluctuations in carbon storage due to inter- and intra-seasonal variations in climate variables.
Despite the promising results, certain limitations were encountered in the proposed processing workflow. Firstly, the tree detection accuracy was not always high due to challenges in segmenting complex canopies and correctly identifying treetops. To address this, advances in segmentation algorithms for broadleaf species are needed to improve the automatic identification of treetops and tree crowns. Moreover, varying the processing parameters (e.g., CHM resolution, CHM smoothing, and LM window size) according to canopy conditions (dense or sparse) and species types could further improve detection accuracy [
46]. Secondly, the species classification accuracy could potentially be enhanced through the use of multitemporal multispectral data [
93,
94]. Ref. [
93] demonstrated that utilising data from three different acquisition dates significantly improved species classification accuracy compared to relying on data from a single date (no further benefits were observed beyond three dates). Finally, although the plant traits were accurately quantified, the retrieval approach used has some limitations that may hinder the application of the workflow in different contexts. Data-driven approaches based on machine regression algorithms, although remarkably powerful, typically come at the expense of transferability [
95,
96]. In addition, they can be biassed by the characteristics of the sensor used in the training phase, in this case, the PRISMA data. To broaden the applicability of the developed methodology, the use of hybrid approaches based on the combination of radiative transfer simulations and machine learning regression could be explored. However, this task is not trivial in forest ecosystems and at the high spatial resolution of drones due to the complexity of the canopy structure, which requires the use of geometric models that are difficult to parameterise and computationally expensive [
97,
98]. The proposed solution provides a relatively simple retrieval workflow, which represents a trade-off between generalisability and operability.