Monitoring Nitrogen Nutrition in Ginkgo Using Unmanned Aerial Vehicle RGB Imagery and the Gaussian Process Regression Model

Abstract

Nitrogen nutrition monitoring is crucial in agriculture and forestry. With the development of Unmanned Aerial Vehicle (UAV) imaging technology, its application in nitrogen nutrition monitoring has gained attention. Traditional regression methods often struggle to accurately capture the nonlinear relationships between image features and nitrogen nutrition parameters. This study introduces Gaussian regression models to better model the relationship between UAV image features and nitrogen nutrition in Ginkgo. UAV RGB imagery of three-year-old Ginkgo biloba L. seedlings was used to extract nitrogen-related image features. Gaussian regression models were employed to select and model these features, creating regression models for nitrogen accumulation and nitrogen content in Ginkgo. The accuracy of the models was validated. Results indicated that the optimal canopy type for monitoring nitrogen accumulation in Ginkgo was the shadowed canopy, with the color feature BMR being the most important feature. For monitoring nitrogen content, sunlight and shadow canopy types were suitable, with BMR and b* being the key features. Gaussian regression demonstrated superior accuracy and robustness compared to traditional regression models. This study emphasizes the potential of Gaussian regression models to improve nitrogen monitoring through UAV imagery, offering valuable applications in precision agriculture and forestry management, particularly in supporting nitrogen fertilization and nutrition management for Ginkgo.

1. Introduction

Ginkgo biloba L., Ginkgoaceae is a deciduous tree with considerable economic and ecological importance [1]. This species has multifaceted utility encompassing roles in nutrition, medicine, construction, landscaping, and aesthetics. In China, Ginkgo biloba extract can treat circulatory or brain diseases and respiratory diseases [2]. Notably, China is the foremost producer and exporter of Ginkgo worldwide and has over 70% of the world’s Ginko resources. Achieving high yields and superior quality in Ginkgo biloba cultivation constitutes a pivotal challenge in scientific research and production [1]. As the demand for Ginkgo biloba escalates across various sectors, industrial standards continually evolve, necessitating a corresponding enhancement in its quality. Nitrogen is a vital mineral nutrient needed in substantial amounts for the growth and development of plants. It participates in various physiological and metabolic processes within plants and is a key component of proteins, nucleic acids, and most enzymes. Nitrogen fertilizer is crucial for plant production, as it majorly contributes to optimal plant growth, promotes enhanced photosynthesis, and improves nutrient absorption [3]. Nevertheless, empirical evidence underscores a crucial caveat: the adage “more is better” does not invariably apply to nitrogen fertilization in Ginkgo cultivation. Indeed, excessive nitrogen application can precipitate adverse consequences, degrade photosynthetic enzymes, hasten leaf senescence, and compromise nitrogen use efficiency, engendering environmental contamination [4].

Traditional methods for diagnosing seedling nitrogen nutrition primarily fall into two categories: morphological and chemical index diagnosis. However, both methods require labor- and time-intensive processes involving destructive plant sampling [5]. This limitation hinders prompt nitrogen nutrition monitoring and the formulation of accurate fertilization recommendations. Currently, numerous instruments and devices can directly or indirectly assess nitrogen content in seedlings, such as chlorophyll meters [6], GreenSeeker^® spectrometers [7], and Crop Circle active canopy sensors [8]. Although these monitoring methods offer advantages (e.g., simplicity, practicality, speed, and non-destructiveness), their applicability is limited by restricted determination areas. Consequently, they fail to meet the demands of high-throughput, rapid, and large-area field information gathering [9].

Drones, commonly referred to as unmanned aerial vehicles (UAVs), hold significant value due to their numerous advantages over other remote sensing technologies [10]. Due to its numerous benefits, including high data collection efficiency, cost-effectiveness, non-destructiveness, ease of use, flexibility, and remarkably high spatial and temporal resolution [11], UAV remote-sensing technology has been widely applied across various fields. The utilization of UAV-mounted RGB cameras has become prevalent in monitoring the nitrogen nutrition of crops such as rice, soybean, wheat, and maize [12,13,14,15]. However, research on nitrogen nutrition diagnosis of Ginkgo and other seedlings in economic forests remains relatively limited. Thus, further investigation into the method of nitrogen nutrition monitoring of Ginkgo using UAV-based RGB imagery is warranted. In addition to color information, commonly used RGB images include texture information, which reflects the internal characteristics of ground surface elements, such as terrain, geomorphic, vegetation, and hydrological features [16]. Existing research indicates that leaf color and leaf texture can directly reflect the absorption, reflection, refraction, and utilization of different light by crops. These characteristics are also a comprehensive response to a series of metabolic results of nitrogen within seedlings. Zhang [17] utilized spectral and texture features from UAV images to create an enhanced model for monitoring winter wheat leaf nitrogen content. This approach offered a valuable reference for assessing wheat nitrogen status and enabling precise management. Li [18] extracted spectral and texture features of citrus trees from UAV multispectral images and developed a semi-supervised collaborative regression model combining Ridge regression, support vector regression, and random forest to estimate nitrogen content in citrus leaves.

Introducing Gaussian process regression into monitoring models offers several key advantages, particularly in handling nonlinear relationships between variables. Gaussian process regression (GPR) excels at modeling complex, nonlinear data patterns, which traditional regression models often find challenging [19]. Additionally, GPR is well-suited for small sample sizes, providing reliable predictions even with limited data. One of its most notable strengths is its ability to quantify prediction uncertainty, thereby offering a measure of confidence in model outputs. Compared to conventional regression models, GPR demonstrates greater flexibility and higher predictive accuracy, making it an ideal choice for complex, real-world applications such as nitrogen nutrition monitoring in forestry and agriculture [20].

Recent studies have highlighted that UAV flight altitude significantly influences image resolution and, consequently, the accuracy of nutrient estimation models. Lower altitudes tend to yield higher spatial resolution and more reliable feature extraction, which are essential for accurate assessment of crop nitrogen status [21].

Previous studies on nitrogen nutrient monitoring based on digital image features have often overlooked further classification of canopy layers. This study proposes applying Gaussian regression combined with UAV imagery to analyze nitrogen content in Ginkgo. The innovation of this approach lies in using Gaussian regression to integrate multidimensional image feature data from different canopy types and flight altitudes, enabling a more accurate and flexible analysis of nitrogen nutrition in Ginkgo and overcoming limitations of conventional methods.

2. Materials and Methods

2.1. Study Area

The study area of the Baima base of the Nanjing Forestry University is situated in the Lishui District, Nanjing, Jiangsu Province. The climate in this region is characterized by mild humidity, abundant rainfall, and ample sunlight, with an average annual temperature of 15.4 °C, approximately 2240 h of sunshine annually, and an average rainfall of 1087.4 mm over the years. The area boasts a high vegetation coverage rate and diverse species. The study area encompasses approximately 800 m², with the sampled Ginkgo seedlings being 3 years old and exhibiting heights ranging between 0.5 and 1.5 m. Each sample plot consists of three rows and six columns, totaling 18 plots. The spacing between plants within rows and between rows ranges from 0.7 m to 1.4 m. Each plot contains 120 plants and covers a surface area of 72 m². The width of the dividing line between plots is 0.5 m (see Figure 1).

2.2. Experimental Design

To investigate how varying nitrogen application rates affect Ginkgo seedling growth, 18 plots were designated with the following nitrogen treatments: N0 (0 kg·ha⁻¹), N1 (225 kg·ha⁻¹), N2 (450 kg·ha⁻¹), N3 (675 kg·ha⁻¹), N4 (900 kg·ha⁻¹), and N5 (1125 kg·ha⁻¹). For detailed analysis, six representative plots corresponding to these nitrogen levels (N0–N5) were selected to investigate growth variations under different fertilization regimes. The fertilizers utilized in the study comprised urea (46% N), calcium superphosphate (12% P₂O₅), and potassium chloride (60% K₂O). The base fertilizers were applied at rates equivalent to 833 kg·ha⁻¹ of P₂O and 233 kg·ha⁻¹ of K₂O. Nitrogen fertilizer was applied in three doses annually: 40% of the total in furrows in March, another 40% in shallow bands in May, and the remaining 20% in shallow bands in July. Phosphorus and potash-based fertilizers were applied once as basal fertilizer and incorporated into furrows in March.

2.3. Plant Sampling and Nitrogen Analysis

Plant sampling was conducted on 25 April, 25 May, 25 June, and 10 September 2021, with the 10 September sampling designated as the final leaf yield assessment. On each sampling date, three representative Ginkgo biloba seedlings were randomly selected from each experimental plot for destructive sampling to assess growth, physiological parameters, and nutrient content.

For nitrogen analysis, leaf samples were collected from the selected seedlings and immediately transported to the laboratory. The leaves were oven-dried at 80 °C until reaching a constant weight. Dry biomass was then measured using an electronic balance. The dried leaf samples were ground into a fine powder, and total nitrogen content (LNC, %) was determined using the Kjeldahl digestion method following standard protocols performed by a commercial testing company [22].

In this study, three nitrogen-related indicators were analyzed: nitrogen accumulation (LNA, g), nitrogen content per unit dry weight (LNC, %), and nitrogen content per unit leaf area (LNCarea, g·m⁻²).

Nitrogen accumulation (LNA, g) was calculated by multiplying the nitrogen content (expressed as a decimal fraction, g·g⁻¹) by the dry biomass of the corresponding leaf samples:

$$\mathrm{L}\mathrm{N}\mathrm{A}=\mathrm{L}\mathrm{N}\mathrm{C}\left(\mathrm{\%}\right)\times \mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{D}\mathrm{r}\mathrm{y} \mathrm{L}\mathrm{e}\mathrm{a}\mathrm{f} \mathrm{W}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t} \mathrm{o}\mathrm{f} \mathrm{a}\mathrm{l}\mathrm{l} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{d} \mathrm{l}\mathrm{e}\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{s} \left(\mathrm{g}\right)$$

Additionally, Nitrogen content (g·m⁻² leaf area) was calculated to assess the nitrogen status at the leaf surface level using the following formula:

$${\mathrm{L}\mathrm{N}\mathrm{C}}_{\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}}={\displaystyle \frac{\mathrm{L}\mathrm{N}\mathrm{C}\left(\mathrm{\%}\right)\times \mathrm{D}\mathrm{r}\mathrm{y} \mathrm{l}\mathrm{e}\mathrm{a}\mathrm{f} \mathrm{w}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t} \mathrm{o}\mathrm{f} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{d} 9 \mathrm{l}\mathrm{e}\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{s} \left(\mathrm{g}\right)}{\mathrm{L}\mathrm{e}\mathrm{a}\mathrm{f} \mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a} \mathrm{o}\mathrm{f} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{d} 9 \mathrm{l}\mathrm{e}\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{s} \left({\mathrm{m}}^2\right)}}$$

Note: LNC (%) is converted to a decimal fraction (g/g) for the calculation.

2.4. UAV Image Data Acquisition and Processing

This study employed a DJI Phantom 4 UAV (Figure 2) equipped with a 20-megapixel RGB camera (model FC6310R; resolution: 5472 × 3648 pixels). The flight campaign was conducted on 19 September 2021 under clear weather conditions between 10:00 and 14:00. UAV flights were carried out at altitudes of 30 m, 60 m, and 90 m, with respective flight durations of 6, 5, and 4 min. The resulting images had ground sampling distances (GSDs) of approximately 0.7 cm/pixel, 1.5 cm/pixel, and 2.3 cm/pixel. Both flight direction and side overlap were set to 80% to ensure sufficient overlap for image stitching and reconstruction.

This study utilized a DJI Phantom 4 UAV (Figure 2) equipped with a 20-megapixel RGB camera (model FC6310R; resolution: 5472 × 3648 pixels). The UAV mission was conducted on 10 September 2021 under clear weather conditions between 10:00 and 14:00, coinciding with the end of the vegetation season. This timing was chosen to align with the peak nitrogen accumulation stage in Ginkgo leaves, which, according to previous studies [4] and field observations, occurs in early autumn. Nitrogen levels in Ginkgo increase steadily throughout the growing season and reach their maximum in this period, making it ideal for assessing final nitrogen status under different fertilization treatments.

The selection of these flight altitudes aimed to balance spatial resolution, flight efficiency, and computational workload. All flights were completed within a consistent 30 min window to minimize variability in solar illumination and maintain consistent image quality [23]. This approach aligns with previous findings [24], which demonstrated that lower UAV altitudes improve spatial detail and enhance multi-view image reconstruction by increasing the number of reliable feature matches. In addition, lower-altitude imagery has been reported to provide superior prediction accuracy in estimating crop yield based on RGB-derived color and texture features [21].

DJI TERRA software (SZ DJI Technology Co., Shenzhen, China) was used for image processing, including a 3D reconstruction of the acquired UAV RGB images and the extraction of RGB image data. The extracted image features included color and texture features. Using ENVI 5.6 software, pixel values from the red, green, and blue channels of the RGB images in the Ginkgo biloba test area were extracted and denoted as R, G, and B, respectively. Subsequently, various color indices based on R, G, and B—such as the HSI color model, Lab* color space, and other composite color indices—were calculated in Excel, totaling 26 indices. Additionally, eight texture features were extracted using the Filter tool in ENVI 5.6. First, principal component analysis (PCA) was performed on the RGB image in ENVI 5.6, producing four principal components. As shown in

Table 1. Image features.

Image Features	Calculation Formula or Description	Origin
R	R	[25,26]
G	G	[25,26]
B	B	[25,26]
G/R	G/R	[27,28]
G/B	G/B	[27,28]
R/B	R/B	[27,28]
GMR	G − R	[28]
GMB	G − B	[28]
BMR	B − R	[28]
NRI	R/(R + G + B)	[29]
NGI	G/(R + G + B)	[29]
NBI	B/(R + G + B)	[29]
VIG, R	(G − R)/(G + R)	[27]
VIG, B	(G − B)/(G + B)	[30]
VIB, R	(R − B)/(B + R)	[31]
H	$2\pi -{\cos }^{-1}\left\{{\displaystyle \frac{0.5\times \left[\left(r-g\right)+\left(r-b\right)\right]}{{\left[{\left(r-g\right)}^2+\left(r-b\right)\left(g-b\right)\right]}^{\frac{1}{2}}}}\right\}$	[32]
S	1–3 × min (R, G, B)	[32]
I	(R + G + B)/3	[32]
L*	116 × (0.299 R + 0.587 G + 0.114 B)1/3 − 16	[33]
a*	500 × [1.006 × (0.607 R + 0.174 G + 0.201 B)1/3 − (0.299 R + 0.587 G + 0.114 B)1/3]	[33]
b*	200 × [(0.299 R + 0.587 G + 0.114 B)1/3 − 0.846 × (0.066 G + 1.117 B)1/3]	[33]
ExG	2 × NGI − NRI − NBI	[34,35]
ExR	1.4 × NRI − NGI	[34,35]
ExGR	ExG − ExR	[35]
VARI	(NGI − NRI)/(NGI + NRI − NBI)	[36]
GLI	(2 × NGI − NBI − NRI)/(2 × NGI + NBI + NRI)	[35]
MEA	Mean reflects the average of gray levels	[37,38]
VAR	Variance reflects the magnitude of gray change	[37,38]
HOM	Homogenetity reflects the local homogeneity	[37,38]
CON	Contrast reflects the clarity of the texture, as opposed to homogeneity	[37,38]
DIS	Dissimilarity is similar to contrast, used to detect pixel similarity	[37,38]
ENT	Entropy reflects the diversity of pixel values	[37,38]
SEM	Second moment reflects the uniformity of image gray distribution	[37,38]
COR	Correlation reflects the extension length of a certain gray value along a certain direction	[37,38]

, texture features were subsequently extracted from the first principal component using the Filter tool with a 3 × 3 window.

2.5. Model Construction and Testing

Gaussian Process Regression (GPR) is a nonparametric Bayesian regression method particularly well-suited for modeling complex nonlinear relationships and small datasets. Its core component is the kernel function—commonly the Radial Basis Function (RBF)—which measures data similarity and controls model flexibility. GPR provides probabilistic predictions along with adaptive hyperparameter optimization, thereby enhancing model reliability compared to traditional regression approaches.

In this study, we employed the Gaussian Process Regression-Backward Additive Tree (GPR-BAT) algorithm for feature selection and model construction. This iterative backward elimination technique evaluates the significance of input variables and progressively refines the model by removing the least influential features at each step, leading to improved predictive performance. Modeling and optimization were conducted using MATLAB R2017b. Model performance was assessed via leave-one-out cross-validation. The accuracy of nitrogen monitoring models was quantified using widely accepted metrics, including the coefficient of determination (R²), root mean square error (RMSE), and normalized root mean square error (nRMSE), in line with established validation standards [39].

3. Results

3.1. Analysis of Nitrogen Accumulation and Nitrogen Content of Ginkgo bilobaUnder Different Nitrogen Application Levels

Figure 3 provides an overview of nitrogen accumulation and nitrogen content dynamics in Ginkgo biloba leaves under different nitrogen application levels. Leaf nitrogen content is expressed both per unit dry weight (%) and per unit leaf area (g·m⁻²).

The average nitrogen accumulation per plant increased from 0.43 g in May to 0.80 g in July, remaining similar at 0.81 g in September, indicating a progressive accumulation over the growing season. Fertilization rate substantially influenced per-plant nitrogen accumulation, which rose with increasing nitrogen levels and peaked under N2 and N3 treatments.

Regarding nitrogen content (% dry weight), values rose from 1.96% in May to 2.35% in July before declining slightly to 2.06% in September, with the maximum observed in July. Nitrogen content (g·m⁻² leaf area) exhibited a consistent upward trend, increasing from 1.6 g·m⁻² in May to 2.7 g·m⁻² in July, and further to 2.8 g·m⁻² in September. Detailed numerical values for nitrogen accumulation and content across different treatments and time points are presented in Supplementary Table S1.

In summary, both nitrogen accumulation and content in G. biloba leaves were positively correlated with fertilization rate, with the highest levels observed under the N3 and N4 treatments.

3.2. Correlation Analysis Between Image Features and Nitrogen Accumulation and Content in Ginkgo biloba

To account for the baseline variation in image brightness and color composition across different UAV flight altitudes and canopy types, average RGB values are provided in Supplementary Table S2, illustrating the effects of lighting conditions and altitude on image features.

The detailed correlation analyses between Ginkgo nitrogen accumulation, nitrogen content, and various image features across different flight altitudes and canopy types are provided in Supplementary Tables S3–S5 in the Supplementary Materials. To enhance readability, only the key findings are discussed here.

Supplementary Table S3 summarizes the correlation between nitrogen accumulation in Ginkgo leaves and various image features across different flight altitudes and canopy types. Color features exhibited stronger and more consistent correlations with nitrogen traits than texture features. Among the 26 color indices, G/R, VARI, GMB, and BMR consistently showed the strongest associations with nitrogen accumulation across all flight altitudes. Notably, G/R and VARI reached the highest correlation at 60 m (R² = 0.55), highlighting the sensitivity of green reflectance-based indices to nitrogen status in the canopy.

Supplementary Tables S4 and S5 provide insights into the associations between unit dry weight nitrogen content and unit leaf area nitrogen content with image features. Regarding unit dry weight nitrogen content, among the 26 color indices, VIB, R, GMB, BMR, and b* showed the strongest correlations. These features maintained stable and high correlations across all canopy types and flight altitudes, indicating their robustness and reliability in different imaging conditions. Specifically, at a flight altitude of 30 m, GMB and b* demonstrated the strongest correlation (R² = 0.64). For Nitrogen content (g·m⁻² leaf area), NRI, R/B, BMR, and b* exhibited high correlations with image features across all flight altitudes and canopy types. The b* color features, in particular, showed a notable correlation under shadowed canopy conditions at 30 m (R² = 0.58).

Among the eight texture features analyzed, the mean features consistently exhibited better correlations than the others across all nitrogen indicators. It showed moderate associations with nitrogen content (% dry weight) and per unit leaf area under various canopy types and flight altitudes. In contrast, other texture features displayed relatively low and irregular correlations.

3.3. Construction and Testing of a Monitoring Model of Nitrogen Accumulation and Nitrogen Content in Ginkgo biloba Based on All Image Features

As shown in Figure 4 and Supplementary Tables S6–S8, the GPR-based models exhibited varying estimation accuracies across canopy types and flight altitudes for nitrogen accumulation and nitrogen content metrics.

Specifically, the nitrogen accumulation models (NA) showed their highest accuracy at 60 m altitude, with the shadowed canopy consistently outperforming the total and light canopies. At this altitude, the shadowed canopy model achieved the highest accuracy (R² = 0.52, RMSE = 0.33 g, nRMSE = 20.71%), indicating that nitrogen accumulation estimation is most reliable under shaded conditions at mid-altitude flights, as shown in Supplementary Table S6.

For nitrogen content (% of dry weight), the highest accuracy was achieved over the shadowed canopy at 30 m (R² = 0.43, RMSE = 0.23%, nRMSE = 16.61%), as shown in Supplementary Table S7. In contrast, models for the total and light canopies showed lower performance, with R² values ranging from 0.22 to 0.38 and 0.03 to 0.31, respectively.

Supplementary Table S8 presents the results for nitrogen content (g·m⁻² leaf area). The light canopy model at 30 m outperformed others, achieving R² = 0.50, RMSE = 0.05 g·m⁻², and nRMSE = 17.39%. The shadowed and total canopy models showed intermediate performance (R² = 0.12–0.47 and 0.22–0.38, respectively).

Overall, these findings highlight that model performance is influenced by both canopy type and flight altitude. The shadowed canopy consistently provides optimal conditions for estimating nitrogen accumulation and nitrogen content (% dry weight), particularly at mid to low flight altitudes. In contrast, the sunlight canopy is better suited for accurate estimation of leaf-area-based nitrogen content, especially at lower altitudes. This suggests that tailored modeling approaches based on canopy characteristics and flight parameters can improve nitrogen status assessments in forest ecosystems.

3.4. Construction and Testing of the Monitoring Model of Nitrogen Accumulation and Nitrogen Content in Ginkgo biloba Based on Preferred Image Features

Table 2. Monitoring model and validation of Ginkgo nitrogen accumulation based on selected features.

Region of Interest	Flight Altitude (m)	Selected Features	R2	RMSE (g)	nRMSE (%)
Total canopy	30	G/R, b*	0.49	0.34	21.27
	60	G/R	0.53	0.33	20.41
	90	G/R	0.38	0.37	23.34
Shadow canopy	30	GMB	0.51	0.33	20.70
	60	BMR, L*	0.64	0.28	17.72
	90	Blue, G/B, G − R	0.39	0.38	23.48
Light canopy	30	b*, SEM	0.47	0.35	21.68
	60	Green, Blue, H	0.40	0.37	23.1
	90	(B − R)/(B + R)	0.36	0.38	23.77

presents the GPR-BAT (Gaussian Process Regression-Backward Additive Tree) model validation results for Ginkgo nitrogen accumulation using the selected (preferred) image features. Compared to models based on all features, the selected features models achieved higher accuracy, better fit, and lower variability.

For the selected features models, the highest R² for the total and shadow canopies occurred at 60 m, whereas the light canopy peaked at 30 m. Specifically, the shadow canopy model attained R² = 0.64 at 60 m—surpassing the total canopy (R² = 0.53) and light canopy (R² = 0.47)—with RMSE = 0.28 g and nRMSE = 17.72% under the same conditions. Overall, the shadow canopy at 60 m provided the best performance in terms of estimation accuracy, model fit, and consistency.

Figure 5 shows the predicted versus observed nitrogen accumulation for the selected features models. Figure 6 displays the two most influential features, BMR and L*, with sigma values of 0.55 and 0.95, respectively, indicating that BMR contributes more to the model than L*.

Table 3. Monitoring model and validation of Ginkgo nitrogen content (% dry weight) based on selected features.

Region of Interest	Flight Altitude (m)	Selected Features	R2	RMSE (%)	nRMSE (%)
Total canopy	30	GMB	0.57	0.20	14.30
	60	BMR, H, CON, SEM	0.58	0.20	14.28
	90	BMR	0.48	0.22	15.82
Shadow canopy	30	BMR	0.54	0.20	14.81
	60	NRI, VIB, R, b*, VAR	0.47	0.22	15.93
	90	BMR	0.36	0.24	17.45
Light canopy	30	BMR, b*, VAR, COR	0.61	0.19	13.60
	60	GMB, ExGR, COR	0.48	0.22	15.73
	90	BMR, VAR, COR	0.28	0.26	18.70

and

Table 4. Monitoring model and validation of Ginkgo Nitrogen content (g·m⁻² leaf area) based on selected features of multiple regression.

Region of Interest	Flight Altitude (m)	Selected Features	R2	RMSE (g·m−2)	nRMSE (%)
Total canopy	30	H	0.52	0.05	17.07
	60	VIB, R	0.50	0.05	17.34
	90	NRI, COR	0.49	0.05	17.56
Shadow canopy	30	b*	0.53	0.05	16.86
	60	R/B	0.52	0.05	17.08
	90	R/B, BMR, VAR CON	0.37	0.06	19.50
Light canopy	30	b*	0.52	0.05	16.98
	60	BMR	0.43	0.05	18.32
	90	BMR	0.34	0.06	19.89

present the modeling and validation results for monitoring nitrogen content (% dry weight) and per unit leaf area of Ginkgo biloba using preferred image features. Consistent with previous findings based on all image features, the GPR model’s estimation accuracy declined with increasing flight altitude.

For nitrogen content (% dry weight), the highest accuracy, best model fit, and lowest variation were achieved under the light canopy at a flight altitude of 30 m (R² = 0.61, RMSE = 0.19%, nRMSE = 13.6%), as illustrated in Figure 7. Figure 8 highlights the Selected Features—BMR, b*, VAR, and COR—with corresponding Sigma values of 2.50, 2.34, 5.73, and 2.09, respectively. Lower Sigma values indicate greater contribution to the model; texture feature COR contributed the most, followed by BMR and b*, while VAR had the least impact.

For nitrogen content (g·m⁻² leaf area), the highest estimation accuracy, best model fit, and minimal variance were observed under the shadowed canopy at a 30 m flight altitude (R² = 0.53, RMSE = 0.05 g·m⁻², nRMSE = 16.86%), as shown in Figure 9. Figure 10 identifies the selected features as b*, with a Sigma value of 2.16.

These findings underscore the significance of flight altitude and canopy type in optimizing nitrogen estimation accuracy.

4. Discussion

4.1. Influence of Flight Altitude on Ginkgo biloba Nitrogen Monitoring Models

The spatial resolution and acquisition efficiency of UAV images vary with flight altitude. Increasing the flight altitude can reduce data acquisition and processing time, thereby improving monitoring efficiency. However, this comes at the cost of decreased spatial resolution, which may affect the accuracy of nitrogen status monitoring. Therefore, it is essential to explore the optimal flight altitude for different nitrogen indicators in Ginkgo.

In this study, UAV flights were conducted at three altitudes (30 m, 60 m, and 90 m) to evaluate their impact on the performance of nitrogen monitoring models using RGB imagery. Models were constructed using the best-performing features and the Gaussian process regression (GPR) method. The results indicated that 60 m was the optimal altitude for nitrogen accumulation monitoring, while 30 m was optimal for nitrogen content monitoring.

These findings are consistent with previous research showing that 30 m is suitable for monitoring nitrogen nutrition in winter wheat, providing better model stability [40]. The differing optimal altitudes for nitrogen accumulation and content observed in this study may be attributed to variations in canopy structure information. Although imagery at 60 m has lower resolution than that at 30 m, it can better capture structural features such as canopy biomass and leaf area index [41], which are essential for nitrogen accumulation estimation.

However, at 90 m, the resolution was too low to accurately distinguish the seedling canopy, leading to a significant drop in accuracy. While lower-altitude flights provide higher-resolution images, they also result in longer acquisition and processing times [42]. Therefore, selecting a suitable flight altitude for practical nitrogen monitoring should balance accuracy and efficiency.

4.2. The Role of Shadow Canopies in Monitoring Ginkgo biloba Nitrogen Nutrition

Due to differences in light exposure, shaded and sunlit canopies exhibit variations in physiology, structure, and spectral response [41]. Despite this, limited research has been conducted to classify and utilize different canopy illumination conditions in nitrogen monitoring. Some studies have attempted to eliminate shadow effects; for instance, comparative analyses showed that shadows weaken the spectral reflectance of fruit tree canopies [43].

In this study, we classified canopy regions into three types—whole canopy, shaded canopy, and sunlit canopy—and assessed their effectiveness in monitoring Ginkgo nitrogen status. The results indicated that image features derived from the entire canopy exhibited the strongest correlation with nitrogen indicators. However, when considering nitrogen accumulation and nitrogen content (g·m⁻² leaf area), the shaded canopy yielded better model performance. In contrast, nitrogen content (% dry weight) was best predicted using features from the sunlit canopy.

These findings underscore the critical role of shaded canopy regions in capturing reliable nitrogen-related spectral information and suggest that future monitoring frameworks should incorporate canopy illumination differentiation.

4.3. Selected Features for Nitrogen Nutrition Monitoring in Ginkgo biloba

The results of this study indicate that most color-based image features exhibited strong correlations with Ginkgo nitrogen indicators. Among texture features, only the mean value showed a significant correlation, whereas other texture metrics had relatively limited predictive power.

Monitoring models based on selected image features outperformed those using all available features in estimating nitrogen accumulation and content (both per unit dry weight and per unit leaf area). Specifically, the selected color feature BMR was most effective for both nitrogen accumulation and nitrogen content (% dry weight), while the color feature b* was optimal for predicting nitrogen content (g·m⁻² leaf area).

These results demonstrate the value of feature selection in enhancing model performance and highlight the dominant role of specific color channels in nitrogen status estimation using UAV RGB imagery.

4.4. Innovative Points of the Thesis and Directions for Future Research Direction

This study, however, has certain limitations that future research aims to address and expand upon. The model used in this study was solely based on Gaussian Process Regression; future research could compare other modeling techniques, such as deep Gaussian Process Regression and convolutional neural networks, to identify the most effective method for monitoring nitrogen status in Ginkgo. This study explored the monitoring method of Ginkgo nitrogen nutrient composition from the perspectives of flight height and canopy type. To enhance the model’s generalizability, further research should focus on modeling Ginkgo nitrogen nutrient monitoring across various spatial scales, regions, and forest ages.

In addition, this study was based on a single UAV mission conducted near the end of the vegetation season, which may not fully capture the temporal dynamics of nitrogen accumulation. Although this timing coincides with the peak nitrogen status of Ginkgo leaves—making it optimal for evaluating final fertilization effects—we acknowledge the potential value of multi-temporal UAV observations. Incorporating UAV campaigns at different phenological stages (e.g., early growth, mid-season) would enable more comprehensive insights into nitrogen uptake dynamics and facilitate adaptive fertilization strategies. Future work will explore the integration of multi-temporal data to enhance model robustness and extend its applicability across growth stages.

5. Conclusions

This study evaluated the optimal combinations of flight altitude, canopy type, and image features for monitoring nitrogen nutrition in Ginkgo biloba, demonstrating the high precision of Gaussian Process Regression models in this context. The optimal flight altitude was 60 m for nitrogen accumulation monitoring and 30 m for nitrogen content monitoring. Practical applications should consider the trade-off between efficiency and accuracy when selecting flight altitude.

Regarding canopy types, the shadowed canopy was most suitable for monitoring nitrogen accumulation and nitrogen content (g·m⁻² leaf area), while the light canopy was optimal for nitrogen content (% dry weight). These results highlight the critical role of the shadowed canopy in capturing nitrogen nutrition status.

The color feature BMR was the best predictor for nitrogen accumulation and nitrogen content (% dry weight), whereas the color feature b* was most effective for nitrogen content (g·m⁻² leaf area). This study confirms the robustness of GPR in identifying nitrogen nutrition variations across different flight altitudes and canopy types, illustrating its adaptability to complex environments and heterogeneous data.

Future research should extend the application of GPR methods at larger spatial scales, incorporate diverse forest types, and explore additional imaging modalities such as multispectral and hyperspectral sensors to validate scalability and improve model generalizability. Overall, these findings provide valuable theoretical guidance for precision fertilization management in Ginkgo biloba, contributing to advancements in modern, intelligent, and precision forestry.