Evaluating the Influence of Pseudo Tree Crown (PTC) Input Alternatives for Machine Learning and Deep Learning Models on Individual Tree Classification Performance

1. Introduction

Individual tree classification is a fundamental task in forest remote sensing and serves as a critical parameter for both urban tree management [1] and natural mature forest ecosystems [2,3]. Accurate identification at the individual tree level enables a wide range of applications, including species-specific inventory, biomass estimation, biodiversity monitoring, and precision forestry practices [4].

In recent years, the research focus has increasingly shifted toward the adoption of machine learning (ML) and deep learning (DL) techniques in forest species and individual tree classification [3,5,6,7,8]. A wide variety of models, ranging from classical classifiers to deep learning architectures, have been proposed to improve classification performance [9,10]. However, most of these approaches are heavily data-driven and often lack an explicit consideration of the underlying physical and biological characteristics of trees. As a result, model development tends to emphasize architectural complexity and incremental gains in accuracy on relatively small, curated datasets, rather than fostering a deeper understanding of tree structure, spectral behaviour, and ecological context [11]. Technically speaking, the same approaches can be applied to any target objects, e.g., crop mapping [12,13], animal detection [14], mineral exploration [15], and iris recognition [16]. This trend has led to a proliferation of increasingly “fancy” models, where improvements are sometimes marginal and may not generalize well beyond the experimental setup [17]. Meanwhile, traditional methods such as watershed segmentation [18] and region-growing [19] approaches, which are more closely tied to geometric and structural interpretations of tree crowns, have gradually been sidelined in contemporary research or used as the comparison benchmark [20] .

Despite the rapid evolution of ML and DL algorithms, the format and quality of data remain the true foundation of individual tree classification [21]. Historically, major advances in remote sensing have been driven not only by model innovation but also by the introduction of new data types and acquisition platforms, ranging from ground-based systems to airborne and satellite observations, and, more recently, to UAVs [22]. Similarly, transitions from RGB imagery to multispectral and hyperspectral data [23], as well as the integration of LiDAR [19], thermal [24], and synthetic aperture radar (SAR) [25], have each introduced distinct physical characteristics that enabled breakthroughs in Earth observation and opened new research directions.

In this context, some studies have moved beyond purely model-centric approaches, seeking to align ML/DL frameworks with a more physically grounded understanding of Earth observation processes. At the same time, recent work has reported notable progress in multi-branch, multip-label [26], multi-source data fusion [27,28,29] and multimodal integration [30,31]. Some researchers have also reported issues with data quality, such as noise affecting classification accuracy and uncertainty, particularly for multimodal data [32].

Over more than a decade of our research, we have focused on establishing a direct estimation of an individual tree from two-dimensional (2D) top-down imagery to the true three-dimensional (3D) structures, progressing from multi-angular analysis [33], to 2D longitudinal profiles [34,35], and to 3D mesh-based crown top surfaces [36,37]. We have also examined the presentation of available data when used directly as input to ML/DL models.

In 2024, we introduced a novel data representation method, Pseudo Tree Crown (PTC), that provides a pseudo-3D pixel-value perspective, which enhances the informational richness of imagery. The original implementation was successfully tested on urban and deciduous trees [38], and later extended to Canadian conifer species in 2025 [39]. However, the initial PTC formulation relied on the green band, limiting its applicability primarily to green-leaf species.

In this study, on top of continuing to test the PTC approach on new tree species and different datasets. We investigate the influence of alternative inputs on different original data variations and transformations. Specifically, we employ the Green Red Vegetation Index (GRVI) [40] and Principal Component Analysis (PCA) to generate new PTC transformations. The results are compared between PTC and non-PTC inputs, as well as among different PTC alternatives. Selected classifiers, including Random Forest, ResNet50, and YOLOv10, are adopted to evaluate performance.

The key contributions and implications of this study are: First, we explicitly demonstrate the significant influence of input data alternatives on the performance of both ML/DL models, highlighting the critical role of data design in classification tasks; Second, the results confirm the effectiveness of the proposed PTC transformation, which consistently improves classification accuracy without introducing additional computational cost or model complexity; Third, the study shows that PTC exhibits robust and stable performance across diverse data forms, indicating strong generalizability, resilience, and reliability in varying experimental conditions.

2. Material and methods

2.1. The study area and data description

The original dataset comprises RGB images collected at the Chenglong Campus of Sichuan Normal University. The study area is located in the eastern part of Chengdu, Sichuan, China (104°12’5.76"E, 30°33’54.64"N), as illustrated in Figure 1. It contains five main tree species: Tree of Heaven (Ailanthus altissima), which has even-pinnate compound leaves with alternately arranged leaflets; the leaflets are ovate-lanceolate with shallow serrations and a relatively thin texture. Osmanthus tree (Osmanthus fragrans) has leathery, opposite, elongated-oval leaves with fine serrations or smooth edges and a smooth, glossy surface. Big-leaved Fig (Ficus virens) has even-pinnate compound leaves with alternately arranged leaflets; the leaflets are ovate-lanceolate with shallow serrations and a relatively thin texture. Chinese Banyan (Ficus macrocarpa) has leaves that are large, thin-leathery, elongated oval or ovate-lanceolate, with entire margins and a tapering tip; the surface is smooth. Camphor Tree (Cinnamomum camphora) has alternate, thin-leathery leaves that are ovate or ovate-elliptical, with three prominent veins from the base and smooth margins. A sample top-down view of tree crowns is shown in Figure 2.

A total of 639 trees were on-site observed, confirmed and manually labelled, including 117 Tree of Heaven, 151 Osmanthus, 107 Big-leaved Fig, 147 Chinese Banyan, and 117 Camphor Tree.

The survey was conducted using a UAV, the DJI Mavic 3 Enterprise Survey Edition (Shenzhen, China), a professional-grade drone designed for surveying and mapping. It is equipped with two cameras: a 4/3-inch CMOS 20 MP wide-angle camera and a 1/2-inch CMOS 12 MP telephoto camera. The wide-angle 4/3-inch camera was used to capture a wider field of view. The survey was conducted at an average flight altitude of approximately 80 m, with the shutter speed set to 1/500 s. The overlap rate was 80% and side overlap was 70% to balance the image mosaic and data volume. The fly speed was about 15 m/s. All images were georeferenced to the WGS84 coordinate system and saved in JPEG format. All aerial imagery was acquired on November 28, 2025, and the data are available upon request.

2.2. Methods: PTC, GRVI and PCA

As mentioned in the introduction, this study had one primary objective: to investigate the influence of data alternatives in original data source variations and transformations of the input data on different ML/DL models. We conducted the work in two separate groups of experiments: (1) directly feeding all alternatives of the data sources into the selected models; and (2) transforming them into PTC and then feeding them into the models.

The individual tree crowns were manually segmented out from the original imagery. Similar to our findings in 2025 [39], we reduced significant background noise in the imagery by clipping the surrounding area to focus only on the tree crowns, as shown in Figure 3, the RGB original image. In addition, because tree sizes varied across the dataset, all samples were standardized to a uniform size to ensure consistency during model training. All inputs were processed the same way, which should not affect the final outcomes.

Based on the UAV RGB imagery, several derived data variations were designed, including the Green band only, the Green-Red Vegetation Index (GRVI), as shown in Equation 1, and Principal Component Analysis (PCA) components. The Green band was used to transform into PTC in our original work [38,39]. The GRVI is a vegetation index that serves as an alternative to the Normalized Difference Vegetation Index (NDVI), while PCA represents information concentration. We used Principal Component (PC) band 1, which is the image intensity, and contained over 97% of the information; we refer to it as PCA1 throughout the remainder of the manuscript.

Therefore, the first group of experiments, as illustrated in Figure 3. We used the original RGB data to extract the Green band, calculate the GRVI (Equation 1), and perform PCA transformation (PC1 was used). The Green band was then used to generate the original version PTC. The original PTC used only the Green band, which is shown in Figure 4. As a result, a total of five data types were used in this study: RGB, GRVI, PCA1, Green, and PTC, where RGB is the only true-colour data input.

$$GRVI={\displaystyle \frac{{\rho }_{Green}-{\rho }_{Red}}{{\rho }_{Green}+{\rho }_{Red}}}$$

(1)

The second part of the experiment, which is shown as Figure 5 a sample. We used the same tree RGB data, along with its variations and transformations, to generate PTCs.

Three ML/DL models were selected and employed for classification: Random Forest (RF) [41], ResNet50 [42], and YOLOv10 [43]. Our PTC is independent of the models themselves. RF represents a classic and stable machine learning approach, ResNet50 represents a reliable deep learning approach, and YOLOv10 is currently one of the most robust object detection algorithms. We acknowledge that other YOLO variants exist, such as YOLOv11 and YOLOv12. However, the version number does not necessarily reflect the performance increment, and we prioritize stability and robustness. And the models are not the focus of this study. These models were used to provide a relative comparison of the data alternatives rather than the approaches themselves. The dataset was divided into training and validation subsets using an 8:2 ratio, and all models were run multiple times with random assignment of trees into training and testing.

3. Results

In this study, model performance was evaluated using multiple metrics, including Accuracy, Precision, Recall, F1 Score, and Intersection over Union (IoU). For each training run, detailed outputs were recorded, including per-species accuracy matrices, training time, confusion matrices, and training loss and validation accuracy curves, to provide a comprehensive and intuitive assessment of model performance. The species-specific classification results are presented in Table 1.

The results indicate that PTC consistently achieves strong and stable performance across all cases, outperforming other data variations and transformations across all three models. This consistency further highlights the effectiveness of PTC in enhancing feature representation and improving classification reliability.

$$\mathrm{Accuracy}={\displaystyle \frac{TP+TN}{TP+FN+FP+TN}}$$

(2)

$$\mathrm{Precision}={\displaystyle \frac{TP}{FP+TP}}$$

(3)

$$\mathrm{Recall}={\displaystyle \frac{TP}{TP+FN}}$$

(4)

$$\mathrm{F}1\mathrm{Score}={\displaystyle \frac{2·\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

(5)

$$\mathrm{IoU}={\displaystyle \frac{TP}{TP+FP+FN}}$$

(6)

As noted in the methods, all results were obtained from multiple runs, and the mean and standard deviation are reported in the tables. The accuracy results are summarized in Table 2 for different variations and transformations directly used as input to the models. The PTC had clearly outperformed other data alternatives in all three models, 0.668 %, 0.865% and 0.793%, respectively.

The confusion matrix species classification of different data alternatives using ResNet50 is illustrated in Figure 6, providing a detailed view of class-wise performance. The matrix shows that most classes are well distinguished, with relatively low misclassification rates, further confirming the effectiveness of the PTC-based features in capturing discriminative structural information. The RF and YOLOv10 confusion matrices are provided in Appendix A ( Figure A1 and Figure A2).

The training loss and validation accuracy (val acc) over epochs using different data alternatives using ResNet50 are shown in Figure 7 as a sample. The remaining results for RF and YOLOv10 are provided in Appendix A ( Figure A3 and Figure A4). The curves indicate that ResNet50 and YOLOv10 have a very similar stable convergence behaviour, while RF had a much lower loss rate. However, no significant signs of overfitting or instability.

Table 4 presents the performance of all transformed PTCs. Although no single PTC variation stands out as a clear dominant performer, all configurations yield highly comparable results, with only minor differences among them. This further supports the consistency and stability of the PTC representation across different variations. The confusion matrix from ResNet50 is illustrated in Figure 8. The RF and YOLOv10 confusion matrices are provided in Appendix A ( Figure A5 and Figure A6). The learning curve of training loss and validation accuracy of different PTCs is shown in Figure 9. The other curves are listed in Appendix A ( Figure A7 and Figure A8).

The computation time comparison is presented in Table 5. Despite the additional PTC transformation step, the overall computational overhead remains minimal. In the RF, PTC even achieved the minimum time.

Table 3. Species classificationresults of different PTCs transformations using RF, ResNet50 and YOLOv10 models

Model	Data	Tree Species	Precision	Recall	F1-score	IoU
RF	RGB	Tree of Heaven	0.5700	0.5938	0.5819	0.4199
		Osmanthus Tree	0.6013	0.6636	0.6234	0.4646
		Chinese Banyan	0.7211	0.7000	0.7100	0.5111
		Big-leaved Fig	0.5824	0.6667	0.6243	0.4332
		Campphor Tree	0.5882	0.4909	0.5441	0.3636
	Green	Tree of Heaven	0.6296	0.6084	0.6169	0.4498
		Osmanthus Tree	0.6340	0.6501	0.6404	0.4749
		Chinese Banyan	0.6907	0.6834	0.6854	0.5283
		Big-leaved Fig	0.6239	0.6739	0.6459	0.4810
		Campphor Tree	0.5387	0.5341	0.5347	0.3739
	GRVI	Tree of Heaven	0.8283	0.8179	0.8176	0.7127
		Osmanthus Tree	0.7279	0.7595	0.7342	0.5896
		Chinese Banyan	0.7271	0.6928	0.7041	0.5528
		Big-leaved Fig	0.6292	0.6928	0.6509	0.4905
		Campphor Tree	0.6699	0.5552	0.5961	0.4448
	PCA1	Tree of Heaven	0.6014	0.6000	0.6003	0.4312
		Osmanthus Tree	0.6510	0.6375	0.6438	0.4803
		Chinese Banyan	0.6374	0.6708	0.6529	0.4915
		Big-leaved Fig	0.5844	0.5970	0.5902	0.4206
		Campphor Tree	0.5889	0.5483	0.5669	0.3975
ResNet50	RGB	Tree of Heaven	0.9091	0.8333	0.8696	0.7692
		Osmanthus Tree	0.9231	0.8000	0.8571	0.7500
		Chinese Banyan	0.9333	0.9333	0.9333	0.8750
		Big-leaved Fig	0.7917	0.9048	0.8444	0.7308
		Campphor Tree	0.7692	0.8696	0.8163	0.6897
	Green	Tree of Heaven	0.9130	0.8750	0.8936	0.8077
		Osmanthus Tree	0.8846	0.7667	0.8214	0.6970
		Chinese Banyan	0.9000	0.9000	0.9000	0.8182
		Big-leaved Fig	0.8000	0.9524	0.8696	0.7692
		Campphor Tree	0.7500	0.7826	0.7660	0.6207
	GRVI	Tree of Heaven	0.9565	0.9167	0.9362	0.8800
		Osmanthus Tree	0.6970	0.7667	0.7302	0.5750
		Chinese Banyan	0.9231	0.8000	0.8471	0.7500
		Big-leaved Fig	0.7619	0.7619	0.7619	0.6154
		Campphor Tree	0.4800	0.5217	0.5000	0.3333
	PCA1	Tree of Heaven	0.8947	0.7083	0.7907	0.6538
		Osmanthus Tree	0.9524	0.6667	0.7843	0.6452
		Chinese Banyan	0.8486	0.9333	0.8889	0.8000
		Big-leaved Fig	0.6774	1.0000	0.8077	0.6207
		Campphor Tree	0.7500	0.7826	0.7660	0.6774
YOLOv10	RGB	Tree of Heaven	0.8119	0.8056	0.7936	0.6631
		Osmanthus Tree	0.7971	0.6968	0.7068	0.5833
		Chinese Banyan	0.7941	0.6896	0.7381	0.5966
		Big-leaved Fig	0.7046	0.7190	0.7049	0.6107
		Campphor Tree	0.6657	0.5000	0.5542	0.3923
	Green	Tree of Heaven	0.8360	0.7986	0.7962	0.6675
		Osmanthus Tree	0.8656	0.6783	0.6816	0.5736
		Chinese Banyan	0.8444	0.7673	0.8031	0.6781
		Big-leaved Fig	0.7366	0.6440	0.6711	0.5536
		Campphor Tree	0.7258	0.6556	0.6610	0.5276
	GRVI	Tree of Heaven	0.8537	0.7461	0.7927	0.6681
		Osmanthus Tree	0.8644	0.7821	0.8171	0.6968
		Chinese Banyan	0.7576	0.7408	0.7475	0.6033
		Big-leaved Fig	0.7404	0.6765	0.6966	0.5641
		Campphor Tree	0.6820	0.7582	0.7114	0.5827
	PCA1	Tree of Heaven	0.8731	0.8334	0.8469	0.7346
		Osmanthus Tree	0.6243	0.6482	0.6340	0.5493
		Chinese Banyan	0.7972	0.7069	0.7491	0.6245
		Big-leaved Fig	0.7365	0.7965	0.7618	0.6318
		Campphor Tree	0.6304	0.6130	0.6161	0.4638

Table 3. Species classificationresults of different PTCs transformations using RF, ResNet50 and YOLOv10 models

Model	Data	Tree Species	Precision	Recall	F1-score	IoU
RF	RGB	Tree of Heaven	0.5700	0.5938	0.5819	0.4199
		Osmanthus Tree	0.6013	0.6636	0.6234	0.4646
		Chinese Banyan	0.7211	0.7000	0.7100	0.5111
		Big-leaved Fig	0.5824	0.6667	0.6243	0.4332
		Campphor Tree	0.5882	0.4909	0.5441	0.3636
	Green	Tree of Heaven	0.6296	0.6084	0.6169	0.4498
		Osmanthus Tree	0.6340	0.6501	0.6404	0.4749
		Chinese Banyan	0.6907	0.6834	0.6854	0.5283
		Big-leaved Fig	0.6239	0.6739	0.6459	0.4810
		Campphor Tree	0.5387	0.5341	0.5347	0.3739
	GRVI	Tree of Heaven	0.8283	0.8179	0.8176	0.7127
		Osmanthus Tree	0.7279	0.7595	0.7342	0.5896
		Chinese Banyan	0.7271	0.6928	0.7041	0.5528
		Big-leaved Fig	0.6292	0.6928	0.6509	0.4905
		Campphor Tree	0.6699	0.5552	0.5961	0.4448
	PCA1	Tree of Heaven	0.6014	0.6000	0.6003	0.4312
		Osmanthus Tree	0.6510	0.6375	0.6438	0.4803
		Chinese Banyan	0.6374	0.6708	0.6529	0.4915
		Big-leaved Fig	0.5844	0.5970	0.5902	0.4206
		Campphor Tree	0.5889	0.5483	0.5669	0.3975
ResNet50	RGB	Tree of Heaven	0.9091	0.8333	0.8696	0.7692
		Osmanthus Tree	0.9231	0.8000	0.8571	0.7500
		Chinese Banyan	0.9333	0.9333	0.9333	0.8750
		Big-leaved Fig	0.7917	0.9048	0.8444	0.7308
		Campphor Tree	0.7692	0.8696	0.8163	0.6897
	Green	Tree of Heaven	0.9130	0.8750	0.8936	0.8077
		Osmanthus Tree	0.8846	0.7667	0.8214	0.6970
		Chinese Banyan	0.9000	0.9000	0.9000	0.8182
		Big-leaved Fig	0.8000	0.9524	0.8696	0.7692
		Campphor Tree	0.7500	0.7826	0.7660	0.6207
	GRVI	Tree of Heaven	0.9565	0.9167	0.9362	0.8800
		Osmanthus Tree	0.6970	0.7667	0.7302	0.5750
		Chinese Banyan	0.9231	0.8000	0.8471	0.7500
		Big-leaved Fig	0.7619	0.7619	0.7619	0.6154
		Campphor Tree	0.4800	0.5217	0.5000	0.3333
	PCA1	Tree of Heaven	0.8947	0.7083	0.7907	0.6538
		Osmanthus Tree	0.9524	0.6667	0.7843	0.6452
		Chinese Banyan	0.8486	0.9333	0.8889	0.8000
		Big-leaved Fig	0.6774	1.0000	0.8077	0.6207
		Campphor Tree	0.7500	0.7826	0.7660	0.6774
YOLOv10	RGB	Tree of Heaven	0.8119	0.8056	0.7936	0.6631
		Osmanthus Tree	0.7971	0.6968	0.7068	0.5833
		Chinese Banyan	0.7941	0.6896	0.7381	0.5966
		Big-leaved Fig	0.7046	0.7190	0.7049	0.6107
		Campphor Tree	0.6657	0.5000	0.5542	0.3923
	Green	Tree of Heaven	0.8360	0.7986	0.7962	0.6675
		Osmanthus Tree	0.8656	0.6783	0.6816	0.5736
		Chinese Banyan	0.8444	0.7673	0.8031	0.6781
		Big-leaved Fig	0.7366	0.6440	0.6711	0.5536
		Campphor Tree	0.7258	0.6556	0.6610	0.5276
	GRVI	Tree of Heaven	0.8537	0.7461	0.7927	0.6681
		Osmanthus Tree	0.8644	0.7821	0.8171	0.6968
		Chinese Banyan	0.7576	0.7408	0.7475	0.6033
		Big-leaved Fig	0.7404	0.6765	0.6966	0.5641
		Campphor Tree	0.6820	0.7582	0.7114	0.5827
	PCA1	Tree of Heaven	0.8731	0.8334	0.8469	0.7346
		Osmanthus Tree	0.6243	0.6482	0.6340	0.5493
		Chinese Banyan	0.7972	0.7069	0.7491	0.6245
		Big-leaved Fig	0.7365	0.7965	0.7618	0.6318
		Campphor Tree	0.6304	0.6130	0.6161	0.4638

Table 4. Accuracy Comparison of Different Models Across Different PTC Transformations

Input Data	Random Forest	ResNet50	YOLOv10
RGB-PTC	$0.6367\pm 7.6\times {10}^{-4}$	$0.8809\pm 2.8\times {10}^{-3}$	$0.8132\pm 1.3\times {10}^{-4}$
Green-PTC	$0.6680\pm 1.5\times {10}^{-3}$	$0.8652\pm 3.8\times {10}^{-4}$	$0.7928\pm 5.8\times {10}^{-3}$
GRVI-PTC	$0.6328\pm 1.9\times {10}^{-4}$	$0.8281\pm 1.2\times {10}^{-2}$	$0.8083\pm 6.9\times {10}^{-4}$
PCA1-PTC	$0.7266\pm 1.0\times {10}^{-4}$	$0.8594\pm 3.2\times {10}^{-4}$	$0.7715\pm 8.1\times {10}^{-4}$

Table 4. Accuracy Comparison of Different Models Across Different PTC Transformations

Input Data	Random Forest	ResNet50	YOLOv10
RGB-PTC	$0.6367\pm 7.6\times {10}^{-4}$	$0.8809\pm 2.8\times {10}^{-3}$	$0.8132\pm 1.3\times {10}^{-4}$
Green-PTC	$0.6680\pm 1.5\times {10}^{-3}$	$0.8652\pm 3.8\times {10}^{-4}$	$0.7928\pm 5.8\times {10}^{-3}$
GRVI-PTC	$0.6328\pm 1.9\times {10}^{-4}$	$0.8281\pm 1.2\times {10}^{-2}$	$0.8083\pm 6.9\times {10}^{-4}$
PCA1-PTC	$0.7266\pm 1.0\times {10}^{-4}$	$0.8594\pm 3.2\times {10}^{-4}$	$0.7715\pm 8.1\times {10}^{-4}$

Table 5. Computation time with different data variations and transformation on different models

Model	Input	Data Variation	time (seconds)
RF	Directly	RGB	76.90
		Green	78.07
		GRVI	79.79
		PCA1	77.29
		PTC	75.28
	PTCs	RGB	81.15
		Green	75.28
		GRVI	83.58
		PCA1	77.14
ResNet50	Directly	RGB	45.79
		Green	23.55
		GRVI	29.49
		PCA1	43.05
		PTC	45.49
	PTCs	RGB	29.48
		Green	45.49
		GRVI	27.31
		PCA1	36.75
YOLOv10	Directly	RGB	409.53
		Green	412.97
		GRVI	416.11
		PCA1	406.36
		PTC	413.26
	PTCs	RGB	416.58
		Green	413.26
		GRVI	412.83
		PCA1	422.48

Table 5. Computation time with different data variations and transformation on different models

Model	Input	Data Variation	time (seconds)
RF	Directly	RGB	76.90
		Green	78.07
		GRVI	79.79
		PCA1	77.29
		PTC	75.28
	PTCs	RGB	81.15
		Green	75.28
		GRVI	83.58
		PCA1	77.14
ResNet50	Directly	RGB	45.79
		Green	23.55
		GRVI	29.49
		PCA1	43.05
		PTC	45.49
	PTCs	RGB	29.48
		Green	45.49
		GRVI	27.31
		PCA1	36.75
YOLOv10	Directly	RGB	409.53
		Green	412.97
		GRVI	416.11
		PCA1	406.36
		PTC	413.26
	PTCs	RGB	416.58
		Green	413.26
		GRVI	412.83
		PCA1	422.48

4. Discussion

The comparison of PTC with RGB variations and transformations is presented in Table 2. The results show that PTC consistently outperforms all original data variations and transformations as the direct input for all three models. This finding aligns with our previous studies [38,39], which demonstrate that PTC enhances information representation by incorporating vertical structural characteristics.

A key finding of Table 2 is that the RF exhibits substantially lower accuracy compared to ResNet50 and YOLOv10. For non-PTC data alternatives inputs, classification accuracy remains below 50%, with the exception of the Green band, which achieves a marginal improvement to 53%. This limitation can be rooted in the fundamental mechanism of RF, which relies on the feature vectors and is not well-suited to modelling image data, particularly in capturing spatial relationships. However, the PTC transformation demonstrates notably superior performance, achieving an accuracy of 66.8% when used with RF. These results provide strong evidence that the choice of input data representation plays a critical role in determining the performance of both ML/DL classification models, which is our primary research objective of this study.

The Green band ranked second across all three models and had the minimum difference between PTC 7% in ResNet50 and YOLOv10, while the other alternatives had 12 - 14%. As illustrated in Figure 3, the Green band effectively highlights sunlit tree crowns while suppressing shadowed canopy regions. This enhances the visibility of key crown structures. In comparison, the GRVI reduces shadow effects but also alters the representation of tree crown structures. Meanwhile, PCA exaggerates grayscale variations, leading to an oversimplified depiction of the tree crown. This finding is consistent with our previous studies [38,39]. The Green band is the most effective and simplest transformation for generating PTC in any green-leaved species, including deciduous and conifers. The differences between Green and other alternatives except PTC ranged from 7 - 9%, and it is greater than or equal to ResNet50 and YOLOv10, which is 5 - 7%. This echoes the influence of the input data, which is absolutely critical and can not be ignored.

Although some improvement was expected from PCA due to its ability to concentrate information, it does not produce meaningful performance gains. This suggests that merely enhancing spatial resolution or pixel-level correlation does not necessarily benefit machine learning or deep learning models. In contrast, features with more discernible shapes are easier to extract from both image processing and interpretation perspectives.

Since PTC does not introduce information loss and does not require complex computation, as shown in Table 5, the computation time is comparable to the other input alternative, even less in the RF model. Hence, the additional PTC transformation step is justified, especially considering the observed accuracy improvement of at least 7%. Furthermore, following our previous studies that examined different view angles, species, and tree ages, the effectiveness of PTC is once again approved.

These findings also suggested the flexibility of PTC, as it can be applied to a wide range of tree species and integrated with various ML/DL models to produce more robust results. The next stage of this research involves automating the application of PTC to entire images, which currently represents its primary limitation. More importantly, PTC serves as an intermediate bridge for retrieving true tree crown structures from top-down imagery.

The second group of experiments, representing the second objective of this study and shown in Table 4, suggests that PTC is highly robust and resilient to variations and transformations in the original data. Unlike the first group of experiments, where performance varied noticeably across different input alternatives, the PTC-based results exhibit a high degree of consistency. There is no single variation or transformation that consistently outperforms the others; however, the differences among them remain minimal, at approximately 3% or less.

This consistency indicates that once data are processed through PTC transformation, the influence of input variation is significantly reduced. In other words, PTC effectively normalizes or stabilizes the feature space, making downstream ML/DL models less sensitive to differences in input data, including resolution and data quality. It may be worth exploring the RGB converted PTC transformation further. The RGB-to-mono grayscale is essentially the average of the RGB values and has a marginally higher accuracy, 1.5-1.9% better than the Green band original.

Given that PTC has already demonstrated clear superiority over non-PTC approaches, these findings further suggest its robustness and resilience. This characteristic is particularly valuable in real-world applications, where data variability is inevitable, as it ensures stable and reliable model performance across different conditions.

5. Conclusion

In this study, we systematically evaluated the effectiveness of the PTC transformation in comparison with various original data variations and transformations. The results demonstrate that PTC consistently outperforms all original input alternatives across ML/DL models, achieving an accuracy improvement of at least 7%. This reinforces previous findings that incorporating vertical structural information significantly enhances feature representation.

This study also shows that structurally meaningful and visually discernible features play a more critical role in effective learning and interpretation compared to merely enhancing pixel-level correlations.

Furthermore, PTC exhibits strong flexibility and generalizability. It can be readily applied across different tree species, imaging conditions, and model architectures without requiring complex computation or introducing information loss. The second group of experiments demonstrated that PTC-based inputs produce highly consistent results across different variations and transformations, with only marginal performance differences. This indicates that PTC effectively stabilizes the feature space and reduces sensitivity to input data variability, including differences in resolution and data quality.

Overall, as the third study in the PTC research line, we reinforce the general conclusion that PTC is not only a high-performing representation but also a robust and reliable one for real-world applications. Its ability to bridge top-down imagery with meaningful tree crown structures represents an important step toward more accurate and interpretable remote sensing analysis.

In our future work, we will focus on automating the PTC generation process at the full-image scale, addressing its current limitations and enabling broader practical deployment, as well as establishing a direct correlation between PTC and true tree crown structures.