Real-time, accurate, and non-destructive monitoring of plant nutrient status is fundamental for precision agriculture and forestry management. For perennial oil crops like Camellia oleifera, optimizing the application of nitrogen (N), phosphorus (P), and potassium (K) is crucial for enhancing fruit yield and oil quality. Traditional methods for assessing leaf nutrient content are labor-intensive, destructive, and cannot provide spatially continuous information. Unmanned Aerial Vehicle (UAV)-based remote sensing, particularly hyperspectral imaging, offers a revolutionary approach by capturing detailed spectral signatures across hundreds of narrow, contiguous bands. These rich spectral data contain subtle information related to leaf biochemical composition, enabling the potential for modeling and mapping macronutrient content across entire orchards. This study focuses on developing and evaluating high-precision estimation models for leaf N, P, and K content in Camellia oleifera canopies using UAV-borne hyperspectral data, comparing the performance of a classical linear method, Partial Least Squares Regression (PLSR), and a robust non-linear ensemble method, Random Forest (RF).
1. Materials and Methods
1.1 Data Acquisition and Preprocessing
The study was conducted in a primary Camellia oleifera production area. Three experimental plots were established, each subjected to a different fertilization regime to create a gradient in soil and plant nutrient availability. Each plot contained multiple rows of mature tea-oil trees of the same cultivar. This setup ensured variability in the target leaf nutrient contents for robust model development.
1.1.1 Leaf Sampling and Nutrient Analysis
In September 2022, approximately 100 mature leaves were randomly collected from the middle section of each tree row within the three plots. The samples were transported to the laboratory, oven-dried, and ground. The concentrations of total nitrogen (TN), phosphorus (P), and potassium (K) were determined using standard protocols: the Kjeldahl method (GB 5009.5-2016) for N, the molybdenum blue spectrophotometric method (GB 5009.87-2016) for P, and flame atomic absorption spectrometry (GB 5009.91-2017) for K. The descriptive statistics of the measured leaf NPK content are summarized in Table 1.
| Plot | Nitrogen (N) | Phosphorus (P) | Potassium (K) | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Mean | Std. Dev. | Range | Mean | Std. Dev. | Range | Mean | Std. Dev. | Range | |
| 1 | 12.9 | 1.1 | 4.2 | 1.27 | 0.067 | 0.22 | 1.317 | 0.186 | 0.664 |
| 2 | 13.5 | 1.0 | 3.0 | 1.30 | 0.064 | 0.18 | 1.624 | 0.247 | 0.77 |
| 3 | 14.3 | 0.6 | 1.7 | 1.44 | 0.100 | 0.27 | 2.054 | 0.352 | 1.08 |
1.1.2 Unmanned Aerial Vehicle Hyperspectral Image Acquisition
Hyperspectral imagery was captured using a GaiaSKY-mini2 sensor mounted on a DJI M600 unmanned drone. The sensor collects data across 176 spectral bands within the 394.5–996.5 nm wavelength range, with a spectral resolution of approximately 3.2 nm. Flights were conducted between 10:00 and 14:00 under clear sky conditions at an altitude of 120 meters, ensuring sufficient overlap between images. Radiometric calibration was performed before and during flights using reference panels. The acquired images were processed through a dedicated pipeline including radiometric calibration, atmospheric correction, image alignment, and mosaicking using the sensor’s proprietary software (SpecStitcher). Images from two different months were normalized and fused to create a seamless, high-resolution (0.04 m) hyperspectral orthomosaic of the study area.
1.1.3 Image Classification and Spectral Feature Extraction
A critical step in linking canopy spectral data to leaf chemistry is isolating pixels belonging to the target vegetation. The high spectral resolution of the unmanned drone data allows for the discrimination of canopy components. An ISO unsupervised classification algorithm was applied to the hyperspectral mosaic to identify and subsequently mask out non-vegetation pixels, primarily soil background and deep canopy shadows, which could introduce noise into the spectral-phenotype relationship. Post-classification refinement was performed in a GIS environment. For each tree row corresponding to the leaf sampling units, the spectral reflectance values for all vegetation-classified pixels were extracted. After removing outliers, the average reflectance across all pixels for each row was calculated for every spectral band, generating a dataset where each row is represented by a single spectral signature (176-band reflectance) paired with its corresponding measured average leaf N, P, and K content. This row-level aggregation effectively scales the point-based leaf measurements to the spatial resolution of the unmanned drone observations.
1.2 Correlation Analysis and Model Development
Pearson correlation analysis was initially conducted between each hyperspectral band’s reflectance and the measured leaf N, P, and K content to identify bands with significant linear relationships. This step provides a baseline understanding of spectral sensitivity.
Two distinct machine learning regression techniques were employed to build the estimation models:
Partial Least Squares Regression (PLSR): This method is particularly suited for datasets with a large number of collinear predictors (spectral bands). PLSR reduces the dimensionality of the spectral data by constructing a set of orthogonal latent components (factors) that maximize the covariance with the response variable (nutrient content). It effectively handles multicollinearity and noise.
Random Forest Regression (RF): An ensemble learning method that operates by constructing a multitude of decision trees during training. The final prediction is the average prediction of the individual trees. RF is robust to overfitting, can model complex non-linear relationships, and provides an internal measure of variable importance.
Both models were implemented using standard packages in the R programming environment (`pls` for PLSR and `randomForest` for RF). The input features for both models were the reflectance values from all 176 hyperspectral bands.
1.3 Model Validation and Accuracy Assessment
To prevent overfitting and obtain a realistic estimate of model performance on unseen data, the Leave-One-Out Cross-Validation (LOOCV) method was employed. For a dataset with n samples (tree rows), LOOCV involves training the model n times, each time using n-1 samples for training and the remaining single sample for validation. This process is repeated until each sample has been used once as the validation data. The predictions from all folds are aggregated to compute overall performance metrics.
The accuracy of the estimation models was evaluated using the following statistical metrics:
Root Mean Square Error (RMSE): Measures the average magnitude of the estimation errors.
$$E_{RMSE} = \sqrt{\frac{\sum_{i=1}^{n} (\hat{Y_i} – Y_i)^2}{n}}$$
Adjusted Root Mean Square Error (RMSEadj): Adjusts the RMSE for the number of predictors (or latent components in PLSR), providing a more conservative error estimate, especially for models with many parameters.
$$E_{RMSEadj} = \sqrt{\frac{\sum_{i=1}^{n} (\hat{Y_i} – Y_i)^2}{n – k – 1}}$$
Coefficient of Determination (R²): Represents the proportion of variance in the measured nutrient content that is explained by the model.
$$R^2 = 1 – \frac{\sum_{i=1}^{n} (\hat{Y_i} – Y_i)^2}{\sum_{i=1}^{n} (Y_i – \bar{Y})^2}$$
Where \(Y_i\) is the measured nutrient content for sample \(i\), \(\hat{Y_i}\) is the predicted nutrient content, \(\bar{Y}\) is the mean of the measured values, \(n\) is the total number of samples, and \(k\) is the number of latent components in the PLSR model.
2. Results and Analysis
2.1 Correlation between Hyperspectral Data and Leaf NPK Content
The Pearson correlation analysis revealed distinct spectral regions associated with each nutrient. Leaf N content showed significant positive correlations with reflectance in the blue-green to red-edge regions (approximately 413-651 nm). For leaf P, significant positive correlations were observed in the visible region (394-693 nm), while significant negative correlations appeared in the near-infrared (NIR) region beyond 716 nm. Leaf K exhibited a similar pattern to P, with positive correlations in the visible and red-edge and negative correlations in parts of the NIR. The correlation coefficients, while statistically significant, were moderate, highlighting the complexity of directly relating canopy spectra to leaf chemistry and underscoring the need for multivariate modeling techniques like PLSR and RF.
2.2 Performance of PLSR and RF Estimation Models
The LOOCV results for the PLSR models, optimized for the number of latent components, are presented in Table 2. The optimal model complexity differed for each nutrient: 6 components for N, 8 for P, and 8 for K, chosen based on the best trade-off between low RMSE and high R².
| Nutrient | Optimal Components | RMSE (mg/g) | RMSEadj (mg/g) | R² |
|---|---|---|---|---|
| Nitrogen (N) | 6 | 0.70 | 0.70 | 0.74 |
| Phosphorus (P) | 8 | 0.094 | 0.093 | 0.85 |
| Potassium (K) | 8 | 0.295 | 0.290 | 0.86 |
The RF models were trained and validated using the same LOOCV procedure. Their performance is summarized below:
– Leaf N (RF): RMSE = 0.70 mg/g, R² = 0.84
– Leaf P (RF): RMSE = 0.09 mg/g, R² = 0.79
– Leaf K (RF): RMSE = 0.349 mg/g, R² = 0.81
Model Comparison: At a comparable RMSE level, the RF model demonstrated a superior explanatory power (higher R²) for estimating leaf N and P content compared to the PLSR model. Conversely, for leaf K content, the PLSR model achieved both a lower RMSE and a higher R² than the RF model, indicating its better performance for this specific nutrient.
2.3 Important Spectral Bands for Estimation
A key advantage of machine learning models is their ability to identify which input features (spectral bands) are most relevant for prediction. For the PLSR model, this is indicated by the magnitude of the regression coefficients for the latent components. For the RF model, variable importance is measured by the increase in the model’s prediction error when a given variable is permuted (%IncMSE). The top important bands identified by each optimal model are listed in Table 3.
| Nutrient | Optimal Model | Top Important Bands (nm) |
|---|---|---|
| Nitrogen (N) | PLSR (6 comp.) | 593.8, 597.1, 590.4, 600.5, 603.8 |
| RF | 637.7, 497.6, 661.5, 468.3, 458.6 | |
| Phosphorus (P) | PLSR (8 comp.) | 723.6, 720.1, 727.1, 716.7, 730.6 |
| RF | 734.0, 397.7, 583.7, 587.0, 590.4 | |
| Potassium (K) | PLSR (8 comp.) | 394.5, 397.7, 423.2, 400.9, 426.4 |
| RF | 494.4, 504.2, 500.9, 497.6, 897.5 |
3. Discussion
3.1 Characteristic Spectral Bands for NPK Estimation
The spectral bands identified as important by both PLSR and RF models align with known plant biophysics and chemistry. For N estimation, the significant bands fell within the visible green-red edge regions (e.g., 590-660 nm). Nitrogen is a fundamental component of chlorophyll and proteins; therefore, its spectral signature is closely tied to pigment absorption features in the red (chlorophyll absorption) and the reflectance slope in the red-edge, which is highly sensitive to chlorophyll content and canopy structure. The inclusion of blue-edge bands (~468-498 nm) by the RF model may be linked to carotenoid absorption, which can also correlate with plant nitrogen status and photosynthetic efficiency.
For P and K estimation, the models highlighted bands in the visible, red-edge, and near-infrared. Phosphorus does not have direct absorption features in the 400-1000 nm range but influences numerous physiological processes, including energy transfer (ATP) and enzyme activation, which indirectly affect pigment content, leaf structure, and water relations. The strong weighting of red-edge and NIR bands (e.g., ~716-734 nm) for P suggests an indirect link to canopy structural parameters or biomass. Potassium, primarily involved in osmotic regulation and enzyme activation, also lacks direct spectral features. Its estimation likely relies on its correlation with other traits that do influence spectra, such as leaf water content (hence the importance of NIR water absorption bands like ~897 nm in the RF model) and overall plant vigor. The prominence of blue/violet-edge bands (~394-426 nm) for K in the PLSR model, a region also sensitive to leaf surface characteristics and non-photosynthetic compounds, indicates a potentially complex, indirect relationship captured by the linear model.
3.2 Comparative Analysis of PLSR and RF Models
The differential performance of PLSR and RF models underscores that the optimal algorithm depends on the nature of the relationship between the spectral data and the target nutrient. The non-linear, ensemble-based RF model excelled at estimating N and P content. This suggests that the underlying relationships between canopy hyperspectral reflectance and these nutrients may involve complex, non-linear interactions that RF’s decision trees can effectively capture. The ability of the unmanned drone to capture high-resolution spatial data may contribute to this complexity, as spectral mixing and shadow effects within the canopy create non-linearities.
In contrast, the linear PLSR model performed best for K estimation. This implies that the spectral information relevant to leaf K status within this dataset might be more linearly related or that PLSR’s dimensionality reduction was more effective at isolating a stable linear signal for K from the hyperspectral data. The difference in the most important bands selected by each model for the same nutrient (Table 3) further illustrates how each algorithm “views” the spectral feature space differently—PLSR through linear combinations maximizing covariance, and RF through hierarchical, non-linear splits.
3.3 Uncertainties and Advantages of UAV Hyperspectral Estimation
While the models achieved high accuracy, several sources of uncertainty are inherent when using unmanned drone hyperspectral imagery for nutrient estimation in tree canopies. First, the measured canopy reflectance is an integrated signal from leaves at different ages, orientations, and vertical positions, mixed with varying degrees of shadow and occasional background (soil, grass). Although image classification mitigated this, residual mixing can affect spectral purity. Second, atmospheric conditions, illumination geometry, and sensor calibration during unmanned drone flights introduce variability that must be carefully corrected. Third, the models are empirically calibrated; their transferability to different cultivars, growth stages, or environmental conditions requires further validation.
Despite these challenges, the approach offers transformative advantages. It is completely non-destructive, enabling frequent monitoring throughout the growing season. The high spatial resolution of the unmanned drone platform allows for the detection of within-field variability, facilitating site-specific nutrient management rather than uniform whole-field applications. This study demonstrates that by combining advanced unmanned drone remote sensing with robust machine learning analytics, it is feasible to rapidly and accurately assess the nutritional status of Camellia oleifera orchards, providing a powerful tool for precision agroforestry aimed at sustainable yield enhancement.
4. Conclusion
This study successfully developed and validated models for estimating leaf nitrogen, phosphorus, and potassium content in Camellia oleifera using hyperspectral imagery acquired by an unmanned drone. The research demonstrates that machine learning techniques, specifically Random Forest (RF) and Partial Least Squares Regression (PLSR), are highly effective in leveraging the rich spectral information for nutrient prediction, outperforming simple linear correlation analysis. The RF model proved superior for estimating leaf N and P content (R² = 0.84 and 0.79, respectively), while the PLSR model was more accurate for leaf K content (R² = 0.86). The analysis identified key spectral regions important for each nutrient, primarily within the visible, red-edge, and near-infrared wavelengths, linked to pigments, canopy structure, and indirect physiological relationships. The implemented methodology, incorporating rigorous pre-processing to isolate canopy signals and a leave-one-out cross-validation strategy, ensures robust and reliable model performance. This work establishes a practical, non-destructive framework for high-frequency nutritional monitoring in tea-oil tree plantations. By enabling the creation of detailed nutrient distribution maps, this unmanned drone-based approach provides a critical decision-support tool for implementing precision fertilization strategies, ultimately contributing to optimized resource use, improved crop health, and increased yield potential in Camellia oleifera cultivation.