Enhanced Estimation of Crown-Level Leaf Dry Biomass of Ginkgo Saplings Based on Multi-Height UAV Imagery and Digital Aerial Photogrammetry Point Cloud Data

Abstract

Ginkgo is a multi-purpose economic tree species that plays a significant role in human production and daily life. The dry biomass of leaves serves as an accurate key indicator of the growth status of Ginkgo saplings and represents a direct source of economic yield. Given the characteristics of flexibility and high operational efficiency, affordable unmanned aerial vehicles (UAVs) have been utilized for estimating aboveground biomass in plantations, but not specifically for estimating leaf biomass at the individual sapling level. Furthermore, previous studies have primarily focused on image metrics while neglecting the potential of digital aerial photogrammetry (DAP) point cloud metrics. This study aims to investigate the estimation of crown-level leaf biomass in 3-year-old Ginkgo saplings subjected to different nitrogen treatments, using a synergistic approach that combines both image metrics and DAP metrics derived from UAV RGB images captured at varying flight heights (30 m, 60 m, and 90 m). In this study, image metrics (including the color and texture feature parameters) and DAP point cloud metrics (encompassing crown-level structural parameters, height-related and density-related metrics) were extracted and evaluated for modeling leaf biomass. The results indicated that models that utilized both image metrics and point cloud metrics generally outperformed those relying solely on image metrics. Notably, the combination of image metrics obtained from the 60 m flight height with DAP metrics derived from the 30 m height significantly enhanced the overall modeling performance, especially when optimal metrics were selected through a backward elimination approach. Among the regression methods employed, Gaussian process regression (GPR) models exhibited superior performance (CV-R² = 0.79, rRMSE = 25.22% for the best model), compared to Partial Least Squares Regression (PLSR) models. The common critical image metrics for both GPR and PLSR models were found to be related to chlorophyll (including G, B, and their normalized indices such as NGI and NBI), while key common structural parameters from the DAP metrics included height-related and crown-related features (specifically, tree height and crown width). This approach of integrating optimal image metrics with DAP metrics derived from multi-height UAV imagery shows great promise for estimating crown-level leaf biomass in Ginkgo saplings and potentially other tree crops.

1. Introduction

Ginkgo is a versatile economic tree species (e.g., valuable for its medicinal and timber properties) native to China and is regarded as one of the oldest surviving plant species. Ginkgo trees are elegant, tall, adaptable, and robust [1]. The extracts from Ginkgo leaves are rich in terpenoids, flavonoids, and proanthocyanidins, offering both medicinal and edible benefits [2]. Biomass, an ecological term, refers to the total quantity of living organic matter (expressed as dry weight) per unit area or for an individual plant, indicating the capacity of Ginkgo and other forest trees to harness energy within the ecosystem. Biomass is a critical indicator of Ginkgo’s growth, playing a key role in determining economic yield, especially for leaf-harvesting. Consequently, accurately measuring the leaf biomass is essential for the effective cultivation and management of Ginkgo saplings.

Traditional biomass estimation methods include both destructive and non-destructive approaches [3]. Destructive methods require the felling and weighing of trees, while non-destructive methods involve developing allometric growth equations using predictor variables such as tree height or diameter at breast height [4]. However, destructive methods are limited in their ability to provide efficient biomass monitoring over large areas. These methods are time-consuming and difficult to implement consistently over time and space. Current non-destructive approaches to estimating vegetation biomass primarily focus on mature forest trees, crops, and grasslands, utilizing airborne or spaceborne remote sensing sensors. Yang et al. [5] utilized satellite remote sensing images in combination with non-destructive techniques like object-oriented image analysis and regression analysis to obtain forest structural parameters, including crown width, tree height, and density of poplar, at a regional scale. Using these parameters, the above-ground biomass of forests was calculated through growth equations. Satellite remote sensing significantly reduces the labor and time required for biomass surveys and measurements, while ensuring the preservation of spatial integrity and temporal consistency of data with a high degree of accuracy [6]. However, satellite remote sensing has limitations, including susceptibility to cloud cover, high costs, and low spatial and temporal resolution. In contrast, airborne LiDAR technology offers greater penetration capability and can be used to extract biomass data from above-ground vegetation [7]. Despite its advantages, airborne LiDAR data remains expensive, and it lacks spectral information on vegetation.

In recent years, the widespread adoption of consumer-grade UAVs has greatly facilitated the acquisition of high-resolution aerial survey images and digital aerial photogrammetry (DAP) point cloud data at a lower operational cost, with equipping RGB digital cameras [8]. UAVs, with their inherent mobility and flexibility, are less influenced by environmental factors like air control and climate, allowing for rapid image acquisition [9]. Additionally, UAVs offer a high level of safety; in the event of equipment failure, there are no risks to human life. Unlike satellite and conventional aerial photogrammetry, which may be obstructed by cloud cover, drones can maneuver at low altitudes to bypass these limitations, yielding higher-resolution images [10].

Previous research utilizing UAV RGB cameras to estimate vegetation biomass often relies on a specific flight height to collect image metrics (encompassing color parameters and texture features) and DAP parameters including height- and density-related metrics, as well as structural parameters (

Table 1. Summary of the biomass estimation utilizing UAV RGB cameras.

UAV	Sensors	Flight Heights	Remote Sensing Metrics	Estimation Method	Monitoring Levels	Plant Species
Octocopter	Cyber-shot DSC- QX10	55 m	Color parameters	MLR	Plot-level AGB	Wheat [12]
DJI Phantom 3 and 4	FC350 and FC6310	30 m	Color parameters; Height-related metrics	MLM	Plot-level AGB	Wheat [19]
DJI Phantom 4 PRO	FC6310	30 m	Color parameters; Height-related metrics	MLR	Plot-level AGB	Maize [20]
DJI MAVIC 3M	Hasselblad L2D-20c	30 m, 60 m and 90 m	Color parameters; Texture features; Height-related metrics	MLM	Plot-level AGB	Wheat [11]
E-DO (A fixed-wing UAV)	Canon EOS 5D Mark II	800 m	Color parameters; Height-related metrics; Density-related metrics	MLR	Plot-level AGB	Beech and paper mulberry trees [21]
E-DO (A fixed-wing UAV)	Canon EOS 5D Mark II	500 m	Height-related metrics; Density-related metrics	MLR	Plot-level AGB	Dawn redwood and poplar trees [16]; Dawn redwood and poplar trees [17]
DJI Phantom 4 PRO and 4 Advanced	FC6310	40 m	Structural parameters	AE	Individual crown-level AGB	Dawn redwood trees [13]; Holm oak saplings [14]
DJI Phantom 4	FC6310	20 m	Structural parameters	AE	Individual crown-level wood biomass	American sweetgum [15]
DJI Phantom 4 PRO	FC6310	10 m	Color parameters; Texture features; Structural parameters	MLR	Individual crown-level AGB	Desert shrub [18]

Note: MLR, MLM, and AE represent multiple linear regression, machine learning methods, and allometric equations, respectively. AGB represents the above-ground biomass.

). Notably, limited research has utilized imagery captured at varying flight heights solely for model accuracy comparisons [11], without integrating data from different altitudes for joint biomass modeling. Studies have typically focused on estimating biomass for crops, young trees, and forests, utilizing either a single type [12,13,14,15] or a combination of two to three types of these parameters, such as height-related and density-related parameters [16,17], color, texture, and structure [18], and color or texture and height- or density-related metrics [11,19,20,21]. Thus, UAV remote sensing emerges as a more versatile and accessible alternative for estimating the crown-level leaf biomass of Ginkgo saplings compared to satellite and aerial remote sensing. However, there are limited studies focusing on the extraction of leaf biomass from Ginkgo saplings [22], especially for using multi-height UAV RGB imagery and DAP point cloud data to enhance modeling biomass. Specifically, there has been limited assessment of how various image metrics and DAP metrics contribute to the development of robust leaf biomass models. Furthermore, existing studies have primarily concentrated on plot-level leaf biomass estimation, neglecting the critical need for individual crown-level leaf biomass estimation, which is crucial for growth monitoring and precise management based on individual sapling conditions.

Biomass prediction models developed using machine learning algorithms offer a rapid and lossless estimation of sapling leaf biomass. Notably, these models include feature-extraction-based techniques, such as Partial Least Square Regression (PLSR) and random forest (RF). These machine learning models can automatically identify the contribution weights of features related to biomass, by taking into account the covariance between predictor variables and responses for PLSR or considering the weighted averages from all regression trees for RF. For instance, Jiang [23] utilized PLSR and RF algorithms to model predictions based on multidimensional data extracted from UAV RGB digital images. However, these methods can face challenges related to overfitting issues and often demand substantial training data. In contrast, Gaussian process regression (GPR), a nonparametric Bayesian modeling method, effectively addresses issues related to small sample sizes and provides uncertainty estimates [24]. Therefore, extracting Ginkgo sapling biomass using UAV RGB digital images in conjunction with GPR modeling warrants further exploration.

Given this need, this study investigates the combined use of image metrics and DAP metrics through two widely adopted multivariate regression methods, PLSR and GPR, to determine whether multi-height UAV imagery can enhance the estimation of Ginkgo leaf biomass. To the best of our knowledge, few studies have assessed the contribution of multi-height UAV imagery and associated DAP point cloud datasets on improving individual crown-level leaf biomass modeling in tree saplings, especially for Ginkgo saplings. The primary objectives of this study are (1) to explore the potential of multi-height UAV imagery in strengthening the synergetic integration of image and DAP metrics for modeling crown-level leaf biomass; (2) to evaluate tree top detection and sapling crown segmentation performance using multi-height DAP point cloud datasets; and (3) to identify the optimal image and DAP metrics that contribute most effectively, along with the best multivariate regression method.

2. Materials and Methods

Figure 1 illustrates the workflow for estimating crown-level leaf biomass of Ginkgo saplings, comprising three key steps: extracting metrics, selecting metrics, and modeling analysis. In the first step (extracting metrics), the 2D RGB (Red, Green, Blue) image mosaic and feature point matching, as well as the 3D reconstruction based on the Structure from Motion (SfM) algorithm, are performed to obtain image and DAP datasets [25]. These datasets are then preprocessed through point cloud normalization and sapling segmentation, allowing for the extraction of image metrics (including color feature parameters and texture features derived from all leaf (AL), the 50% brightest leaf (BL), and the 50% darkest leaf pixels (DL)), as well as DAP metrics (encompassing structural parameters, height-related and density-related metrics). The second step (selecting metrics) involves conducting a correlation analysis between metrics and leaf biomass. Metrics that are weakly correlated with biomass (r < 0.20) and those exhibiting high collinearity (absolute correlation coefficient > 0.95) are eliminated, resulting in a set of collinearity-removed metrics. These metrics are further refined using the backward elimination approach in Partial Least Squares Regression (PLSR) and Gaussian process regression (GPR) models. In the third step (modeling analysis), PLSR and GPR regression models are constructed based on the optimal image and DAP metrics. The best approach for estimating the crown-level leaf biomass of Ginkgo saplings is then identified.

2.1. Overview of the Study Area

This study was conducted in Baima Town, Lishui District, Nanjing City, Jiangsu Province. The base covers a total area of 220 ha and is situated in a transition zone between the northern subtropical and middle subtropical climates, characterized by mild and humid conditions, abundant rainfall, ample sunlight (average annual sunshine duration of 2240 h), and concurrent rain (average annual rainfall of around 1106.5 mm) and heat (average annual temperature of around 15.4 °C), making it suitable for planting Ginkgo saplings.

2.2. Experimental Design

The experimental area for Ginkgo saplings spans approximately 0.17 hectares, forming a total of 18 plots (three rows with six columns), each plot measuring 6 m by 12 m and covering approximately 72 m² (Figure 2). In 2020, 2-year-old Ginkgo saplings of a main leaf-harvesting cultivar called ‘Taixing Dafozhi’, with a uniform height of around 0.5 m, were initially planted in this area. The saplings were arranged with a planting density of 0.6 × 1.0 m, and a 0.5 m-wide pathway was established between each plot for separation. In 2021, to investigate the impact of varying nitrogen fertilization levels on the growth of 3-year-old Ginkgo saplings, different nitrogen treatments were administered across these 18 plots. The treatments included N0 (0 kg/hm²), N1 (225 kg/hm²), N2 (450 kg/hm²), N3 (675 kg/hm²), N4 (900 kg/hm²), and N5 (1125 kg/hm²), with each treatment replicated three times randomly following the completely randomized design. In particular, these nitrogen treatments included low (N0–N1), medium (N2–N3), and high (N4–N5) nitrogen levels applied at specific critical growth stages to assess their impact on leaf biomass accumulation and estimation models. Nitrogen was applied three times a year, with 40% of the annual amount placed in furrows in March (serve as a basal fertilizer), another 40% applied in shallow strips in May (serve as top-dressing to promote the growth of branches and leaves during the rapid growth period of Ginkgo), and the remaining 20% distributed in July (serve as top-dressing to enhance leaf thickness and increase the content of medicinal components) as shallow strips. The fertilizers utilized in the study were urea (46% N), calcium superphosphate (12% P₂O₅), and potassium chloride (60% K₂O). Additionally, phosphorus (6 kg/plot of calcium superphosphate) and potassium fertilizers (1.68 kg/plot of potassium chloride) were applied as a one-time basal fertilizer in March and incorporated into the furrows.

2.3. Field Data Collection and Processing

On 19 September 2021, three sampled representative saplings (which can represent the average growth status for each plot) were selected from each experimental plot. The height and crown width of these sampled representative saplings were measured, and leaf samples were collected. The leaves were then dried in an oven at 80 °C until reaching a constant weight. Finally, the dry biomass (g) of the leaves from each sampled sapling was measured using an electronic balance.

2.4. UAV RGB Image Acquisition and Processing

The RGB image data for this study were collected between 10:00 and 10:30 am on 19 September 2021, under sunny weather conditions. RGB images of Ginkgo sapling canopies were captured using a DJI Phantom 4 RTK UAV (SZ DJI Technology Co., Shenzhen, China), equipped with an RGB digital camera (model FC6310R) featuring a resolution of 20 megapixels (5472 × 3648 pixels) (Figure 3). In accordance with the existing literature (such as Zhai et al. [11]) and the resolution requirements for images captured at varying altitudes, gird size should be no more than one-fifth of the crown width of saplings, which is the recommended setting in the individual tree segmentation [26]. Accordingly, the flight of a UAV was conducted at altitudes of 30 m (flight duration of 6 min), 60 m (flight duration of 5 min), and 90 m (flight duration of 4 min), yielding spatial resolutions of 0.7 cm, 1.5 cm, and 2.3 cm for the acquired RGB images, with the flight direction and side overlap rates being both set to 80%. This strategy also ensured that all flights could be completed within a half-hour timeframe, capitalizing on relatively consistent sunlight conditions. In this study, DJI TERRA software (SZ DJI Technology Co., Shenzhen, China) was employed for 2D image mosaicking, providing detailed visual representations of the sampled saplings, while this software cannot be used for visual interpretation. To complement this, ArcGIS software (ESRI, Redlands, CA, USA) was used for the visual interpretation and delineation of individual canopy contours for sampled saplings (for extracting image and DAP metrics, as well as leaf biomass modeling), as well as all saplings within three typical plots (for comparing the tree top detection performance) under N0, N2, and N4 nitrogen level treatments. An area of interest was generated based on the extent of the canopy contours. To further mitigate the adverse effects of soil background on leaf biomass estimations within the area of interest in 2D images, soil pixels (such as those at the edges of the area of interest or captured through the sapling canopy gaps) were removed using threshold segmentation (ExG > 65) through ENVI+IDL software (EXELIS, Boulder, CO, USA) programming. This process facilitated the identification of the 50% darkest pure leaf pixels (the 50% of the sapling crown image pixels with relatively low RGB brightness) and the 50% brightest pure leaf pixels (the 50% of the sapling crown image pixels with relatively high RGB brightness). RGB values and subsequent color feature parameter and texture feature values were extracted for further analysis.

2.5. Remote Sensing Metric Extraction

2.5.1. Color Feature Parameters

To ensure consistency across all color feature parameters calculated from RGB images, we utilized DJI Terra software (SZ DJI Technology Co., Shenzhen, China) to stitch the photos into a high-resolution orthomosaic. This process corrected geometric distortions and maintained uniform spatial resolution, which is crucial for accurate metric calculations. While the use of reflectance panels for field calibration is standard practice to ensure accuracy when using vegetation indices, we adopted an alternative approach by ensuring consistent lighting conditions and applying image processing techniques, enabling us to achieve precise measurements without the use of calibration panels. Additionally, we took photos from various heights (i.e., 30 m, 60 m, and 90 m) but ensured that all images were captured within a half-hour period to keep sunlight conditions consistent. By controlling both the image stitching and the lighting conditions, this approach minimized lighting variations and helped in producing reliable and consistent color feature parameters presented in

Table 2. Table of color feature parameter names and related calculation formula.

Metrics Type	Characteristic Parameters	Calculation Formula	Reference
Single-band metrics	R	Red	[27]
	G	Green	[27]
	B	Blue	[27]
Ratio metrics	G/R	G/R	[28]
	G/B	G/B	[28]
	R/B	R/B	[28]
Normalized metrics	NRI	R/(R + G + B)	[27,29]
	NGI	G/(R + G + B)	[27,29]
	NBI	B/(R + G + B)	[27,29]
	VIG,R	(G − R)/(G + R)	[29,30]
	VIG,B	(G − B)/(G + B)	[29,31]
	VIB,R	(B − R)/(B + R)	[29,30]
	VARI	(NGI − NRI)/(NGI + NRI-NBI)	[32]
	GLI	(2 × NGI − NBI − NRI)/(2 × NGI + NBI + NRI)	[32]
Difference metrics	GMR	G − R	[28,29]
	GMB	G − B	[28,29]
	BMR	B − R	[28,29]
	ExG	2 × NGI − NRI − NBI	[32]
	ExR	1.4 × NRI − NGI	[32]
	ExGR	ExG − ExR	[32]
HSI color model metrics	H	arccos [0.5 × [(R − G) + (R − B)]/[(R − G)2 +(R − B)(G − B)]0.5]	[33]
	S	1 − (3 × [min(R, G, B)])/(R + G + B)	[33]
	I	(R + G + B)/3	[33]
Lab color model metrics	L*	116 × (0.299R + 0.587G + 0.114B)1/3 -16	[31,34]
	a*	500 × [1.006 × (0.607R + 0.174G + 0.201B)1/3 − (0.299R + 0.587G + 0.114B)1/3]	[31,34]
	b*	200 × [(0.299R + 0.587G + 0.114B)1/3 − 0.846 × (0.066G + 1.117B)]1/3	[31,34]

.

In this study, the Pearson correlation analysis was conducted between leaf biomass and the color feature parameters for all leaf pixels, the 50% darkest leaf pixels, and 50% bright leaf pixels derived from multi-height UAV RGB images. In particular, 26 commonly used color feature parameters (including single band, ratio, normalized, difference, HSI, and Lab color model metrics) are examined in this study, and the specific calculation formula is shown in Table 2. The ratio, difference, and normalized metrics are calculated by taking two or three channels from RGB images, where the first channel is sensitive to the target parameter, and the other channels serve as a reference to mitigate background effects, such as soil background or illumination conditions [27,28,29,30,31,32]. The HSI color model simplifies color analysis and processing in image processing and computer vision using three parameters: Hue (H), Saturation (S), and Intensity (I) to intuitively represent color characteristics [33]. The Lab color model enables device-independent color representation with three components: L* for lightness, a* from deep green to bright pink, and b* from bright blue to yellow, enabling vibrant color mixtures [31,34].

2.5.2. Texture Feature Parameters

The calculation of texture features using occurrence measures based on probability statistics was conducted in ENVI 5.3 (EXELIS, Boulder, CO, USA). Eight second-order matrix-based texture filters, derived from Co-occurrence Measures, can be applied. These texture features encompass a range of metrics such as mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation, all of which are calculated using second-order probability statistics. These statistics are obtained from a gray-toned spatial correlation matrix, representing the relative frequency of gray-level combinations in the image. This matrix reflects the frequency of occurrence of an image element value in two neighboring treatment windows that are separated by a specific distance and direction, indicating how often the relationship between an image element and its neighborhood occurs. Due to the small crown size of Ginkgo saplings, a 3 × 3 window was utilized for this analysis.

2.5.3. Extraction of Crown-Level Structural Parameters

In this study, DJI TERRA software (SZ DJI Technology Co., Shenzhen, China) was employed for reconstructing 3D aerial point clouds utilizing the Structure from Motion (SfM) method [25], resulting in DAP point cloud data. The raw DAP point cloud data were pre-processed using LiDAR360 software (Beijing Green Valley Technology. Co., Ltd., Beijing, China), which included steps such as pre-cropping, denoising, ground classification, and normalization based on ground surface points. This process ultimately facilitated the creation of a digital elevation model (DEM) and a digital surface model (DSM). In particular, DEM was built by the kriging interpolation of the classified ground points, which were extracted using an improved progressive encrypted triangular mesh filtering algorithm [35]. The DAP point cloud was normalized based on the DEM, and this normalized DAP point cloud was then interpolated to produce a canopy height model (CHM). Subsequently, a tree segmentation algorithm based on distance discriminant clustering [26] in LiDAR360 was employed to evaluate the performance of tree top detection and individual sapling segmentation using multi-height DAP data. In particular, Li et al. [26] developed a tree segmentation algorithm based on a small-footprint discrete return airborne LiDAR system to segment individual trees from the point cloud data. The segmentation begins with a seed point, which represents the highest detected point, expanding into a cluster based on spacing rules. In particular, this algorithm utilizes a top-to-bottom region-growing approach, allowing for the individual and sequential segmentation of trees, starting from the tallest to the shortest. In summary, the main steps for processing 3D DAP point cloud data include point cloud data denoising, ground point classification, individual tree segmentations using normalized DAP point cloud based on ground point normalization, and then extraction of tree height, crown width, crown area and crown volume of Ginkgo saplings. The parameter settings are displayed in

Table 3. The parameter settings in the tree segmentation using DAP point cloud data.

Parameters	Value Settings
Grid size (m)	0.0826
Buffer size (pixels)	50
Height above ground point (m)	0.4
Smallest tree height (m)	0.78

. Specifically, the grid size was set to 0.0826 m, which was the one-fifth value of the average crown widths for sampled saplings following the recommend setting in the manual of LiDAR360. The smallest tree height was set according to the measured smallest sampled sapling. Finally, the optimal height DAP data were selected to extract crown-level structural parameters (

Table 4. The crown-level structural parameters used in this study.

Parameters	Meaning
Tree height	The height for individual segmented saplings derived from DAP point clouds
Crown width	The crown width for individual segmented saplings derived from DAP point clouds
Crown area	The crown area for individual segmented saplings derived from DAP point clouds
Crown volume	The crown volume for individual segmented saplings derived from DAP point clouds

), as well as height-related and density-related metrics (

Table 5. Height-related and density-related metrics of DAP point cloud data for individual sapling crowns.

Metrics Name	Meaning
Height-related metrics
H25	The 25th, 50th, 75th, and 95th percentiles of the canopy height distributions for the first returns within the statistical cell of individual sapling crowns
H50
H75
H95
Hcv	The coefficient of variation (CV) represents the variability of Z-values across all points within a specific statistical cell of individual sapling crowns
Hmean	The average height is calculated by taking the mean of all the heights of all first returns within a specific statistical cell of individual sapling crowns
Density-related metrics
D3	The proportion of points exceeding the 30th, 40th, 50th, 70th, and 90th quantiles to the total number of points within a specific statistical cell of individual sapling crowns
D4
D5
D7
D9

).

2.5.4. Height-Related and Density-Related Metrics Extraction

The height-related and density-related metrics were extracted from the DAP point cloud data for individual saplings, utilizing LiDAR360 (Beijing Green Valley Technology. Co., Ltd., Beijing, China) as presented in Table 5 [16].

2.6. Pearson Correlation Analysis

The Pearson correlation coefficient is a metric used to assess the linear relationship between two variables, X and Y [36]. It varies from −1 to 1, indicating both the strength and direction of this relationship. This coefficient is prevalent across various scientific and technical disciplines [37]. It is mathematically expressed in Equation (1).

$$\mathrm{r}={\displaystyle \frac{\sum _i\left(X_i-\overline{X}\right)\left(Y_i-\overline{Y}\right)}{\sqrt{\sum _i\left(X_i-\overline{X}\right)\sum _i{\left(Y_i-\overline{Y}\right)}^2}}}$$

(1)

In this study, Pearson’s correlation coefficient serves to evaluate the relationship quantitatively between leaf biomass and image metrics (color and texture feature parameters), as well as DAP metrics (structural, height-related, and density-related parameters). Metrics with low correlation (r < 0.20) are excluded, following the methodology outlined by Cao et al. [17]. Subsequently, metrics demonstrating high collinearity (absolute Pearson’s correlation > 0.95) are also discarded. The remaining metrics undergo refinement through the backward elimination technique applied in PLSR and GPR models. Despite other correlation analysis methods like Spearman’s rank correlation preferring monotonic relationships, Pearson correlation can better track the linear relationships and could help minimize the risk of saturation issues in regression models [17].

2.7. Regression Analysis and Backward Elimination of Optimal Metrics

In this study, we constructed regression models employing PLSR and GPR for estimating leaf biomass of Ginkgo saplings based on optimal image metrics (including color feature parameters and texture feature parameters derived from all leaf, the 50% brightest leaf, and the 50% darkest leaf pixels of RGB images), as well as DAP metrics (encompassing structural, height-related, and density-related metrics). In particular, the selection of optimal image metrics and DAP metrics was initially conducted using the backward elimination strategy in both PLSR and GPR models. Subsequently, the implementation of PLSR and GPR models was carried out using the selected optimal metrics. All analyses were conducted in the MATLAB R2017b environment (MathWorks, Inc., Natick, MA, USA). A detailed overview of these methodologies is presented below.

2.7.1. GPR Models

Gaussian process regression (GPR) is a nonparametric method grounded in Bayesian theory and statistical learning principles. It is particularly effective for complex regression tasks involving high-dimensional data, small sample sizes, and nonlinear relationships. A critical component of GPR is the kernel function, also known as the covariance function, which determines the characteristics of the Gaussian process. One common example of a kernel function is the radial basis function (RBF), defined as follows:

$$K\left(x_i\right.,\left.x_j\right)={\sigma }_f^2\mathrm{e}\mathrm{x}\mathrm{p}(-{\displaystyle \frac{{‖x_i-x_j‖}_2^2}{2l^2}})$$

(2)

where ${\sigma }_f^2$ and $l^2$ are called hyperparameters, which are derived from the collected data, allowing for the definition of the kernel function based on these calculated hyperparameters. GPR offers several advantages over traditional machine learning methods such as neural networks and support vector machines, including ease of implementation, adaptive hyperparameter acquisition, and probabilistic output, which is beneficial for integration with predictive control, adaptive control, and Bayesian filtering. Additionally, training GPR models is generally simpler than with other machine learning techniques [37,38]. For more information, refer to Verrelst et al. [39].

The GPR-based band analysis tool (GPR-BAT) [40,41] was developed based on the backward elimination strategy. GPR-BAT features an iterative approach in which the influence of input variables on predictions is assessed either in isolation or in the presence of other predictors. During each iteration, the least significant variables are eliminated, and the model is retrained using the remaining variables. Theoretically, this process enhances the selection of parameters and improves modeling accuracy.

2.7.2. PLSR Models

Partial Least Squares Regression (PLSR) is a widely recognized parametric method for estimating structural parameters, such as biomass [9,42,43]. As a standard parametric regression method, PLSR effectively tackles multicollinearity issues often found in traditional regression approaches, like stepwise multiple linear regression [44]. This method transforms original datasets, such as spectral data, into a smaller set of latent variables to model target parameters, while taking into account the covariance between predictor variables and response [37,38,45]. Specifically, the optimal number of latent factors is quantified by minimizing the sum of squares of the prediction residuals. The PLSR models were implemented with the ‘plsregress’ function within the MATLAB R2017b environment (MathWorks, Inc., Natick, MA, USA).

2.8. Estimation Accuracy Evaluation Method

In this study, the performance of PLSR and GPR models was assessed with the leave-one-out cross-validation based on the commonly used cross-validation coefficient of determination (CV-R²), root mean square error (RMSE), and relative root mean square error (rRMSE). The corresponding equations are listed below:

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{N}\mathrm{u}\mathrm{m}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}^{\mathrm{’}}\right)}^2}{\mathrm{N}\mathrm{u}\mathrm{m}}}}$$

(3)

$$\mathrm{r}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}={\displaystyle \frac{\sqrt{\frac{\sum _{\mathrm{i}=1}^{\mathrm{N}\mathrm{u}\mathrm{m}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}^{\mathrm{’}}\right)}^2}{\mathrm{N}\mathrm{u}\mathrm{m}}}}{\overline{y}}}$$

(4)

$$\mathrm{C}\mathrm{V}-R^2=1-{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{N}\mathrm{u}\mathrm{m}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}^{\mathrm{’}}\right)}^2}{\sum _{\mathrm{i}=1}^{\mathrm{N}\mathrm{u}\mathrm{m}}{\left({\mathrm{y}}_{\mathrm{i}}-\overline{y}\right)}^2}}$$

(5)

where ${\mathrm{y}}_{\mathrm{i}}$ and ${\mathrm{y}}_{\mathrm{i}}^{\mathrm{’}}$ are the sample measured and predicted leaf biomass values, respectively. $\overline{y}$ represents the average value of the measured leaf biomass. Num is the number of samples. The smaller the RMSE and rRMSE values, the better the agreement between the predicted and measured values, the more accurate and reliable the simulation results of the model, and the higher the accuracy.

3. Results

3.1. The Leaf Biomass and Height of Ginkgo Saplings under Different Nitrogen Level Treatments

The results illustrated that leaf biomass, height, and crown width of Ginkgo saplings increased initially from N0 levels to the peak at N2~N3 levels and then subsequently declined with higher nitrogen fertilizer applications (Figure 4). Specifically, leaf biomass reached a maximum of 77.59 g/plant, height peaked at 185.00 cm, and crown width attained a maximum of 79 cm at the N2 or N3 levels.

3.2. Tree Top Detection of Ginkgo Saplings

Table 6. Summary of tree top detection accuracies based on multi-height DAP point cloud datasets for three typical plots under different nitrogen level treatments.

Height	N0 (Number = 105)				N2 (Number = 105)				N4 (Number = 106)
	CD	OE	CE	DT	CD	OE	CE	DT	CD	OE	CE	DT
30 m	105 (100.00%)	0 (0.00%)	2 (1.90%)	0 (0.00%)	104 (99.05%)	1 (0.95%)	3 (2.86%)	1 (0.95%)	106 (100.00%)	0 (0.00%)	1 (0.94%)	1 (0.94%)
60 m	98 (93.33%)	7 (6.67%)	4 (3.81%)	0 (0.00%)	99 (94.29%)	6 (5.71%)	3 (2.86%)	1 (0.95%)	98 (92.45%)	8 (7.55%)	2 (1.89%)	2 (1.89%)
90 m	81 (77.14%)	24( 22.86%)	5 (4.76%)	2 (1.90%)	92 (87.62%)	13 (12.38%)	3 (2.86%)	1 (0.95%)	86 (81.13%)	20 (18.87%)	5 (4.72%)	2 (1.89%)

Note: CD, OE, CE, and DT represent the proportion of the correctly detected trees, the omission errors, the commission errors, and the proportion of the duplicated detected trees. Number represents the total number of saplings within individual plots.

presents the tree top detection performance using multi-height DAP point cloud datasets (30 m, 60 m, and 90m) across three typical plots under low (N0), medium (N2), and high (N4) nitrogen treatments. The results indicated that 30 m DAP datasets obtained the best performance (CD = 99.05%~100.00%; CE = 0.94%~2.86%), followed by the 60 m DAP datasets (CD = 92.45%~94.29%; CE = 1.89%~3.81%), and then the 90 m DAP datasets (CD = 77.14%~87.62%; CE = 2.86%~4.76%). Regarding nitrogen fertilizer levels (Figure 5), the N2 plot (CD = 87.62%~99.25%; CE = 2.86%) generally outperformed both the N0 plot (CD = 77.14%~100.00%; CE = 1.90%~4.76%) and N4 plot (CD = 81.13%~100.00%; CE = 0.94~4.72%). In terms of the tree top detection for the sampled saplings, using 30 m DAP datasets successfully identified all sampled samplings (CD = 100.00%), whereas the 60 m DAP datasets (CD = 98.15%) and the 90 m DAP datasets (CD = 83.33%) missed 1~9 sampled saplings.

3.3. Correlation Analysis between Leaf Biomass and Remote Sensing Metrics

3.3.1. Correlation Analysis between Leaf Biomass and Image Metrics

Table 7. Summary of correlation coefficients between image metrics and leaf biomass.

Image Metrics	AL_30m	AL_60m	AL_90m	BL_30m	BL_60m	BL_90m	DL_30m	DL_60m	DL_90m
Color parameters
R	−0.58 **	−0.62 **	−0.59 **	−0.47 **	−0.46 **	−0.40 *	−0.60 **	−0.61 **	−0.56 **
G	−0.54 **	−0.54 **	−0.53 **	−0.51 **	−0.40 *	−0.35 *	−0.61 **	−0.46 **	−0.46 **
B	−0.23	−0.41 *	−0.47 **	−0.20	−0.18	−0.15	−0.38 *	−0.34 *	−0.37 *
NRI	−0.62 **	−0.69 **	−0.60 **	−0.39 *	−0.53 **	−0.56 **	−0.56 **	−0.69 **	−0.65 **
NGI	0.41 *	0.64 **	0.55 **	−0.15	0.01	0.01	0.27 *	0.45 **	0.43*
NBI	0.57 **	0.53 **	0.24	0.54 **	0.56 **	0.49 **	0.53 **	0.53 **	0.38 *
G/R	0.61 **	0.73 **	0.62 **	0.23	0.37 *	0.37 *	0.48 **	0.65 **	0.59 **
G/B	−0.49 **	−0.34 *	0.06	−0.54 **	−0.45 **	−0.36 *	−0.40 *	−0.27	−0.07
R/B	−0.60 **	−0.62 **	−0.46 **	−0.51 **	−0.58 **	−0.57 **	−0.59 **	−0.65 **	−0.58 **
VIG,R	0.61 **	0.73 **	0.62 **	0.23	0.37 *	0.37 *	0.49 **	0.65 **	0.60 **
VIG,B	−0.49 **	−0.34 *	0.08	−0.54 **	−0.45 **	−0.36 *	−0.40 *	−0.26	−0.07
VIB,R	0.60 **	0.62 **	0.45 **	0.50 **	0.58 **	0.57 **	0.57 **	0.64 **	0.57 **
GMR	0.59 **	0.71 **	0.63 **	0.22	0.37 *	0.36 *	0.43 *	0.60 **	0.58 **
GMB	−0.59 **	−0.54 **	−0.43 *	−0.62 **	−0.50 **	−0.47 **	−0.68 **	−0.48 **	−0.47 **
BMR	0.61 **	0.64 **	0.60 **	0.52 **	0.57 **	0.56 **	0.63 **	0.68 **	0.66 **
H	0.62 **	0.72 **	0.61 **	0.23	0.36 *	0.35 *	0.51 **	0.65 **	0.60 **
S	−0.60 **	−0.62 **	−0.44 **	−0.50 **	−0.58 **	−0.57 **	−0.57 **	−0.64 **	−0.56 **
I	−0.54 **	−0.57 **	−0.56 **	−0.45 **	−0.39 *	−0.33 *	−0.58 **	−0.52 **	−0.49 **
L*	−0.55 **	−0.57 **	−0.56 **	−0.49 **	−0.41 *	−0.35 *	−0.60 **	−0.52 **	−0.50 **
a*	−0.58 **	−0.73 **	−0.62 **	−0.08	−0.24	−0.22	−0.39 *	−0.56 **	−0.53 **
b*	−0.60 **	−0.59 **	−0.47 **	−0.59 **	−0.56 **	−0.53 **	−0.66 **	−0.59 **	−0.57 **
ExG	0.41 *	0.64 **	0.55 **	−0.15	0.01	0.01	0.27 *	0.45 **	0.43 *
ExR	−0.61 **	−0.72 **	−0.63 **	−0.27 *	−0.41 *	−0.42 *	−0.50 **	−0.66 **	−0.61 **
ExGR	0.57 **	0.72 **	0.60 **	0.07	0.22	0.20	0.40 *	0.58 **	0.53 **
VARI	0.61 **	0.73 **	0.62 **	0.23	0.37 *	0.37 *	0.48 **	0.65 **	0.59 **
GLI	0.41 *	0.64 **	0.55 **	−0.15	0.01	0.01	0.27 *	0.45 **	0.43 *
Texture parameters
Mean	0.55 **	−0.58 **	0.57 **	0.47 **	−0.41 *	0.35 *	0.59 **	−0.53 **	0.51 **
Variance	−0.01	−0.17	−0.02	0.05	0.24	0.27 *	−0.03	−0.09	0.15
Homogeneity	0.04	0.21	0.01	−0.01	−0.13	−0.19	0.12	0.25	0.01
Contrast	0.02	−0.16	−0.08	0.13	0.25	0.12	−0.03	−0.18	−0.02
Dissimilarity	0.00	−0.18	−0.07	0.10	0.20	0.14	−0.06	−0.23	−0.03
Entropy	0.04	−0.09	0.14	0.00	0.11	0.18	−0.18	−0.22	0.10
Second Moment	−0.06	0.09	−0.15	0.00	−0.09	−0.20	0.20	0.19	−0.15
Correlation	−0.29 *	−0.08	0.13	−0.34 *	−0.10	0.29 *	−0.31 *	−0.22	0.14

Note: ** represents p < 0.001, * represents p < 0.05. AL_30m, AL_60m, and AL_90m represent all leaf pixels derived from UAV imagery at the flight height of 30 m, 60 m, and 90 m, respectively. BL_30m, BL_60m, and BL_90m represent the 50% brightest leaf pixels derived from UAV imagery at the flight height of 30 m, 60 m, and 90 m, respectively. DL_30m, DL_60m, and DL_90m represent the 50% darkest leaf pixels derived from UAV imagery at the flight height of 30 m, 60 m, and 90 m, respectively. In each column, the strongest correlation coefficients are highlighted in bold.

illustrates the correlation coefficients between image metrics and leaf biomass. In terms of color parameters, the color parameters extracted from 60 m UAV imagery generally outperformed those from 30 m and 90 m UAV imagery. The color parameters extracted from all leaf pixels performed the best, followed by those from the 50% darkest leaf pixels and the 50% brightest leaf pixels. Notably, G/R, a*, and VARI derived from all leaf pixels in the 60 m UAV imagery yielded the strongest performance (|r| = 0.73). Regarding the stability of color parameters, BMR, R/B, b*, VI_B,R, and S showed significant correlations with leaf biomass (p < 0.001), regardless of different flight heights and different types of leaf pixels. In terms of texture parameters, the mean texture features exhibited a significant correlation with biomass (|r|= 0.35~0.59) across all cases examined in this study. However, most other texture parameters did not show significant correlations, except for the correlation feature derived from 30 m UAV imagery, which had notably lower significant correlations with leaf biomass. Given the superior performance of all leaf pixels compared to the 50% darkest leaf pixels and the 50% brightest leaf pixels, the color and texture parameters extracted from all leaf pixels were selected for further analysis and modeling.

3.3.2. Correlation Analysis between DAP Metrics and Leaf Biomass

The analysis of DAP point clouds generated from UAV imagery indicated that the 30 m height effectively detected all sampled saplings. In contrast, the 60 m height mostly missed saplings below 0.8 m, and the 90 m height mostly failed to identify those below 1.1 m. Consequently, this study performed the correlation analysis between leaf biomass and DAP metrics, including crown-level structural parameters and height- or density-related metrics obtained from the 30 m height DAP datasets (

Table 8. Summary of correlation coefficients between leaf biomass and DAP metrics derived from DAP point cloud data of 30 m UAV imagery.

DAP Metrics	Correlation Coefficients (r)
Structural parameters
Tree height	0.74 **
Crown width	0.72 **
Crown area	0.69 **
Crown volume	0.68 **
Height-related metrics
H25	0.40 *
H50	0.45 **
H75	0.50 **
H95	0.61 **
Hcv	0.42 *
Hmean	0.53 **
Density-related metrics
D3	0.26
D4	0.31 *
D5	0.18
D7	0.08
D9	−0.37 *

Note: ** represents p-value < 0.001, * represents p-value <0.05. In each column, the strongest correlation coefficients are highlighted in bold.

). Table 8 shows that all the structural parameters (r = 0.68~0.74) and height-related metrics (r = 0.40~0.61) are significantly related to leaf biomass, whereas only a few density-related metrics (D4 and D9) have significant correlations (|r| =0.31~0.37, p < 0.05). Notably, the tree height parameter exhibited the strongest correlation (r = 0.74, p < 0.001) with leaf biomass.

3.4. Estimation of Leaf Biomass Based on PLSR and GPR Models with Optimal Metrics

Table 9. The accuracies of PLSR and GPR models in the estimation of Ginkgo leaf biomass.

Metrics	PLSR Models				GPR Models
	CV-R2	RMSE (g/Plant)	rRMSE (%)	SelectedMetrics	CV-R2	RMSE (g/Plant)	rRMSE (%)	Selected Metrics
Image_AL_30m	0.33	17.01	45.12	B, NBI	0.33	17.00	45.09	BMR
Image_AL_60m	0.51	14.49	38.44	BMR, VIB,R, G, B, NGI, GMR, Homogeneity	0.57	13.57	35.99	B, GMR, Homogeneity
Image_AL_90m	0.38	16.34	43.36	b*, R/B, B, G/R	0.40	16.05	42.58	BMR, b*, NBI, G/R
DAP_30m	0.62	12.85	34.10	Tree height, Crown width, H25, H95, Hcv, Hmean, D9	0.60	13.05	34.63	Tree height, Crown width, D4
Image_AL_30m+DAP_30m	0.62	12.77	33.87	NBI, Tree height, Crown width, Crown volume, D9	0.64	12.44	32.99	G, B, Crown width, H50
Image_AL_60m+DAP_30m	0.71	11.18	29.67	NGI, Homogeneity, Tree height, Crown width, D9	0.79	9.63	25.56	G, NGI, Homogeneity, Crown width, H25, Hmean
Image_AL_90m+DAP_30m	0.65	12.32	32.68	G, B, Tree height, Crown width, Crown volume, D3, D9	0.67	11.95	31.69	G, Tree height, Crown width, H50

Note: Image_AL_30m, Image_AL_60m, and Image_AL_90m represent image metrics of all leaf pixels derived from UAV imagery at the flight height of 30 m, 60 m, and 90 m, respectively. DAP_30m represents the DAP metrics derived from DAP point cloud datasets with the UAV height of 30 m. In individual columns, the best approaches are highlighted in bold.

summarizes the leaf biomass estimation accuracies derived from PLSR and GPR models utilizing optimal metrics. The result illustrated that combining DAP metrics with image metrics (R² = 0.62~0.71 for PLSR; R² = 0.64~0.79 for GPR) generally improved the estimation performance compared to using image metrics (R² = 0.33~0.51 for PLSR; R² = 0.33~0.57 for GPR) or DAP metrics (R² = 0.62 for PLSR; R² = 0.60 for GPR) alone. Specifically, the integration of 60 m height image metrics with 30 m height DAP metrics yielded superior outcomes compared to other approaches. In terms of the performance of different regression models, GPR models mostly outperformed PLSR models, with the most effective model combining 60 m height image metrics and 30 m height DAP metrics, achieving an R² of 0.79 and an rRMSE of 25.56%. Notably, the scatterplots showed that PLSR models significantly underestimated leaf biomass for values exceeding 45 g/plant, particularly under medium nitrogen levels (N2~N3), while GPR models demonstrated considerably better performance (Figure 6).

Figure 7 depicts the frequency of optimal metrics selected in the PLSR and GPR models. Regarding the image metrics, G and B had substantially higher frequency (5~6 times) than other metrics, while NBI, BMR, NGI, Homogeneity, and b* obtained moderate frequencies (3 times). In terms of the DAP metrics, D₉, Tree height, and Crown width exhibited higher selection frequencies (4~8 times) than other DAP metrics, especially for the Crown width showing the highest frequency (8 times).

3.5. Mapping the Distribution of Crown-Level Leaf Biomass for Ginkgo Saplings

Given that the GPR model utilizing optimal image metrics from the height of 60 m and DAP metrics from the height of 30 m exhibited superior performance, this model was employed to map the crown-level leaf biomass of Ginkgo saplings (Figure 8). The average value of crown-level leaf biomass for Ginkgo saplings in the study area was found to be 45.72 g/plant, with a standard deviation of 11.34 g/plant. Figure 8 illustrates distinct spatial patterns in leaf biomass, with higher values (indicated in red) associated with medium nitrogen fertilizer levels (N2–N3), while lower values (shown in blue) are mostly distributed in plots under low (N0–N1) or excessive (N4–N5) nitrogen levels. Notably, some saplings in plots with insufficient or excessive nitrogen levels died, resulting in gaps in the field due to nutrient deficiencies or surplus fertilization.

4. Discussion

In this study, leaf dry biomass (i.e., leaf yield) of 3-year-old Ginkgo saplings was well mapped using GPR models built based on the combined datasets of 60 m height optimal image metrics and 30 m height optimal DAP metrics. To our knowledge, this represents the initial endeavor to refine estimations of crown-level leaf dry biomass in saplings, leveraging multi-height UAV imagery and DAP point cloud data.

4.1. Comparison of Measured Structural Parameters of Ginkgo Saplings under Different Nitrogen Treatments

In this study, the effects of varying nitrogen application levels on the leaf biomass, height, and crown width of Ginkgo saplings were determined by statistical analysis. The results indicated that as nitrogen application increased, leaf dry biomass, height, and crown width also increased. However, beyond a certain threshold (N application > 675 kg/hm²), both dry biomass and sapling height began to decline. This pattern aligned with findings from Wu et al. [46]. This phenomenon occurs because Ginkgo saplings require a specific critical level (e.g., medium levels of N2~N3 in this study) of nitrogen fertilization to optimize dry biomass accumulation. When exceeding this critical level, excessive nitrogen can result in toxic effects on the saplings, ultimately leading to a reduction in dry biomass.

4.2. Comparison of Image Metrics and DAP Metrics for Estimating Leaf Biomass

In this study, we analyzed the correlation between leaf biomass and image metrics (including color parameters and texture features) or DAP metrics (including structure parameters, height- and density-related metrics). Overall, DAP metrics demonstrated a stronger relationship with leaf biomass than image metrics. Notably, structure parameters and height-related metrics showed significant correlations with leaf biomass. In terms of image metrics, the results indicated that image metrics extracted from 60 m height UAV images performed better than 30 m and 90 m height images. Furthermore, metrics derived from all leaf pixels outperformed those from the 50% brightest or 50% darkest pixels. This may be because leaf biomass is a canopy indicator of population, and the relatively coarse resolution image (1.5 cm spatial resolution) of the 60 m UAV image (allowing for more multi-scattering photon chances within single pixels) can better capture the canopy population characteristics of the saplings than the finer resolution of the 30 m height images (0.7 cm spatial resolution). However, when using coarser resolution images with the height of 90 m (2.3 cm spatial resolution), these data cannot well segment tree crown pure pixels mixed with the confounding effects of the background materials, which resulted in worse performance.

In terms of the types of leaf pixels, the stable performance of all-leaf pixels can be attributed to their greater pixel count, offering more comprehensive biomass-related information compared to selectively using the 50% brightest or the 50% darkest leaf pixels. Additionally, image metrics constructed by the normalization or ratio transformations generally improved the performance compared to original RGB values. This supports the findings of Wang et al. [47,48] and Shi et al. [34], as normalized or ratio-type constructed image metrics help mitigate the impact of light intensity variations during image capture to a certain extent, enhancing the stability of biomass relationships, particularly for metrics like R/B, b*, VI_B,R, and S, and obtaining stable performance across different flight heights and different types of leaf pixels.

4.3. Comparison between PLSR and GPR Models and Optimal Metric Selection

Overall, GPR models (R² = 0.33~0.79, rRMSE = 25.56%~45.09%) outperformed PLSR models (R² = 0.33~0.71, rRMSE = 29.67%~45.12%) in estimating the leaf biomass of Ginkgo saplings. Specifically, the best GPR model achieved superior performance (R² = 0.79) compared to the best PLSR model (R² = 0.71) by utilizing optimal image metrics from 60 m height imagery and 30 m height DAP metrics. This improved performance may be attributed to the flexibility of GPR models in handling continuous, discrete, and categorical variables, along with their capability to manage nonlinear relationships more effectively [24,37,41]. In addition, the success of GPR and PLSR models based on the selected optimal image and DAP metrics was mainly attributed to the backward iterative feature selection method, which enhances modeling outcomes by addressing limitations associated with using Pearson correlation analysis alone for feature selection.

In terms of the selected optimal image metrics, G and B are the most contributing image metrics with the highest selection frequencies (5~6 times) than other image metrics. Additionally, their normalized counterparts, such as NBI, NGI, and b* obtained moderate frequency (three times) and relatively stronger relationships with leaf biomass. This could be explained by the fact that the normalized metrics can better track the variation in leaf biomass, especially for NGI, both being selected in the best case for GPR and PLSR models.

Regarding the optimal DAP metrics derived from 30 m height DAP point cloud datasets, Tree height and Crown width were the most frequently selected metrics (4~8 times). This finding is consistent with the existing literature that highlights the significant physiological relationship between these metrics and vegetation biomass [13,14,18], particularly for Ginkgo saplings. The results also highlighted the importance of accurate tree top detection and segmentation in improving leaf biomass estimation models. The 30 m height DAP point cloud datasets achieved superior tree top detection accuracies (CD = 99.05%~100.00%) with an average point density of approximately 700 points/m², which was significantly higher than the densities of 60 m (around 150 points/m²) and 90 m (around 60 points/m²) height DAP datasets. This is particularly important as high-density point clouds facilitate more precise assessments of vegetation growth status, reinforcing findings by Escolà et al. [49] and Torres-Sánchez et al. [50], who emphasize the advantages of higher-resolution data with lower flight altitudes in estimating vegetation canopy parameters. Moreover, the integration of multi-angle observations, such as nadir and oblique imagery, can further enhance the accuracy of 3D modeling. Studies have proved that combining various flight angles allows for better capturing of canopy complexity and increases the number of detectable underlying points [51,52]. The oblique perspectives can provide critical information about the canopy’s side and bottom features, which are often overlooked in traditional nadir-only approaches. This is particularly relevant for dense canopies where light penetration varies, impacting biomass estimation [53]. However, while lower flight altitudes yield finer resolution data, they also increase data storage and processing complexity. Escolà et al. [49] and Buunk et al. [52] note that although combing nadir and oblique imagery at a lower flight height improves the detection of finer objects, it also increases the burden of data processing. Therefore, when implementing these UAV systems on a larger scale, especially for estimating geometric canopy parameters, it is essential to strike a balance between the desired level of detail and the efficiency of the system configuration.

When combining the optimal image and DAP metrics, the best GPR and PLSR models can be obtained. This improvement might be caused by the fact that image metrics can be better related to the leaf biomass in low-nitrogen treatment plots, while DAP metrics strengthen the relationships with leaf biomass from the plots in high-nitrogen environments (Figure 9). These findings align with Mao et al. [18] and Zhai et al. [11], who demonstrated the effectiveness of integrating image and DAP point cloud datasets to enhance the estimation of vegetation biomass. While image-based metrics are more sensitive to low biomass values and can capture one-dimensional changes, they often encounter saturation issues when dealing with higher biomass levels [18]. In contrast, DAP metrics, particularly those related to structural parameters, effectively monitor both the horizontal and vertical characteristics of trees, thereby offering a more comprehensive description of biomass variation [49]. Despite these promising results, it is crucial to address the limitations associated with UAV-derived data. For instance, variations in point cloud density and spatial distribution can significantly affect model performance, particularly in heterogeneous environments. Additionally, excluding non-plant objects from the point cloud remains a challenge, often leading to underestimations of canopy structural parameters [51]. Furthermore, UAV-based DAP datasets derived from imagery with insufficient spatial resolution often struggle to accurately reconstruct thin branches, leading to underestimations of tree height [50]. To overcome these challenges, future studies should focus on refining data collection methodologies by incorporating advanced filtering techniques and utilizing machine learning algorithms to enhance data fidelity and accuracy.

4.4. Potentials and Limitations

This study investigated the estimation of leaf biomass in Ginkgo saplings using UAV RGB imagery. Our findings indicated that GPR models can enhance the estimation of Ginkgo sapling biomass based on optimal image metrics from 60 m height imagery and optimal DAP metrics from 30 m height DAP point cloud data. By employing the backward GPR-BAT feature selection method to select optimal metrics, certain weakly correlated parameters were eliminated, enhancing the modeling performance of the GPR model. Consequently, GPR-BAT presents a reliable method for the rapid and accurate extraction of biomass in Ginkgo saplings. The current optimization focus of this paper is on preferential parameter modeling, with plans to explore deep learning algorithms in future studies.

In future research, it is essential to integrate multi-angle RGB imaging data to enhance the observation capabilities of the middle and lower leaf layers. This integration will also facilitate the generation of more precise DAP point cloud data, thereby improving the accuracy of leaf biomass estimation models. Furthermore, it is important to evaluate the performance of image data collected during various key growth phases and under different lighting conditions, such as overcast days. This evaluation will help identify the optimal image and DAP metrics suitable for different growth phases and lighting scenarios. Consequently, this approach will expand the observational window and support timely fertilization management for Ginkgo saplings, particularly those experiencing nitrogen deficiency stress.

We acknowledge that there are still limitations in this study. First, the limited sample size used may hinder the GPR models’ ability to accurately estimate the function’s shape, leading to potential estimation errors. Second, when dealing with high-dimensional data, GPR faces overfitting risks, making feature selection critical. To enhance the model’s robustness and generalization capabilities, it is essential to carefully select relevant features during the regression process. Our analysis focused solely on backward elimination methods; other more comprehensive feature selection techniques might further improve model accuracy and reliability.

These limitations highlight several avenues for future research, such as increasing the number of samples, employing advanced feature selection methods, and examining alternative machine learning algorithms to enhance the precision and reliability of Ginkgo sapling leaf biomass estimations. Future studies could also investigate the application of deep learning algorithms for this purpose, such as deep Gaussian processes, which effectively address feature extraction issues [54]. Furthermore, techniques like Lasso, Elastic Net, or random forests could be utilized for feature selection prior to applying deep Gaussian process regression. This integrated approach allows for incorporating multi-method ensemble selection information into the modeling process [55], thereby increasing information content while ensuring modeling accuracy and achieving improved prediction results.

5. Conclusions

This study investigated the application of synthetic datasets of UAV imaging data alongside corresponding DAP point cloud datasets for estimating leaf biomass in Ginkgo saplings. PLSR and GPR models were evaluated using multi-height image metrics in combination with 30 m height DAP point cloud datasets. The main conclusions are presented below:

(1)

The integration of 60 m height image metrics with 30 m height DAP metrics, derived from the best-performing DAP point cloud datasets for tree top detection, significantly improves leaf biomass estimation.

(2)

GPR models generally outperformed PLSR models, regardless of using image metrics only or combined datasets, achieving a maximum R² of 0.79 in the best-case scenario.

(3)

The commonly selected optimal image and DAP metrics for both PLSR and GPR models included G, B, NGI, Tree height, and Crown width.

This study highlights the substantial potential of integrating multi-height UAV imagery with DAP point cloud datasets to strengthen the accurate predictions of leaf biomass in Ginkgo saplings. These findings offer valuable technical parameters for rapid and non-destructive estimation of Ginkgo sapling biomass, which is beneficial for effective monitoring and precise management of Ginkgo sapling growth.