As a researcher deeply involved in precision agriculture and remote sensing, I am acutely aware of the critical role that the Leaf Area Index (LAI) plays in characterizing the canopy structure and growth status of winter wheat. LAI is a fundamental biophysical parameter that governs photosynthesis, transpiration, and carbon cycling. Accurate and timely monitoring of LAI is therefore essential for assessing crop health, predicting yield, and implementing site-specific field management strategies such as irrigation scheduling and precise fertilizer application. However, traditional field-based LAI measurement methods are labor-intensive, destructive, and incapable of capturing spatial heterogeneity across large agricultural fields. To address these limitations, I have turned to unmanned aerial vehicle (UAV) remote sensing, which offers a flexible, cost-effective, and non-destructive means of acquiring high-resolution multispectral imagery. In China, the adoption of UAV drones for agricultural monitoring has accelerated rapidly, providing a powerful tool for data-driven crop management. Yet, a significant challenge persists: the accuracy of LAI inversion models can be unstable and exhibit poor generalization across different UAV flight altitudes. This scale effect, driven by changes in spatial resolution and the mixing of spectral signals from soil, shadows, and vegetation, introduces variability in both reflectance and textural features. My central objective in this study is to systematically investigate the impact of flight altitude on the performance of LAI inversion models and to develop a robust framework that can deliver high-precision and generalizable results. By integrating multi-altitude multispectral data with advanced ensemble learning techniques, I aim to overcome the limitations of single-feature and single-model approaches, thereby advancing the capabilities of China UAV drones in precision agriculture.
To achieve this goal, I designed and conducted a comprehensive field experiment during the 2024-2025 winter wheat growing season in the Yangling Demonstration Zone of Shaanxi Province, China. The experimental site is located at coordinates 108°4’20″E, 34°17’42.17″N, at an altitude of 525 meters. This region is characterized by a temperate semi-arid to semi-humid continental monsoon climate, with an average annual precipitation of approximately 640 mm, concentrated between July and September. The soil type in the 0-60 cm layer is medium loam, with an average field capacity of 26% and a wilting coefficient of 8.6% (by mass). I established two separate field plots to ensure the diversity and robustness of my dataset. Plot I consisted of 12 plots, each 4 m × 4 m, with four different irrigation treatments (T1: 40-50% field capacity, T2: 50-65% field capacity, T3: 65-80% field capacity, T4: 80-95% field capacity), each replicated three times. Plot II consisted of 50 plots, each 4.5 m × 4.5 m, with five irrigation treatments including a rain-fed control (T0), each replicated ten times. All plots were equipped with water meters and drip irrigation systems to ensure precise and uniform water application. The winter wheat variety ‘Xinong 9112’ was sown on October 12, 2024, and harvested on June 1, 2025, with a row spacing of 25 cm and a seeding rate of 40 grams per row. To minimize border effects, I maintained a 1.5 m buffer zone between plots.
A Matrice 300 RTK UAV, a high-performance platform commonly used in China for agricultural monitoring, was employed for data acquisition. This UAV drone was equipped with an AQ600 Pro multispectral camera, which captures imagery in five discrete spectral bands: Blue (450 nm), Green (555 nm), Red (660 nm), Red Edge (720 nm), and Near-Infrared (NIR, 840 nm), in addition to an RGB sensor. To systematically investigate the scale effect, I programmed the UAV to hover at five different flight altitudes: 20 m, 40 m, 60 m, 80 m, and 100 m. At each altitude, the camera was pointed directly nadir to capture images of the winter wheat canopy. All flights were conducted under clear, sunny weather conditions around 13:00 local time to minimize variations in solar illumination. The acquired images were processed using Pix4Dmapper software for stitching and ground control point (GCP) georeferencing to ensure spatial accuracy. Subsequently, ENVI 5.6 software was used to extract the average reflectance for each band within each plot. To minimize the influence of soil background and shadows on the canopy reflectance, I applied a supervised maximum likelihood classification to segment the imagery, isolating pure winter wheat pixels for subsequent analysis.
Concurrently, I collected ground truth data to validate the remote sensing measurements. For LAI measurement, three representative wheat plants were randomly selected from each plot. The leaves were separated from the stems, and the leaf area was determined using a threshold-based image segmentation method in MATLAB R2022b. The LAI was then calculated as the product of the average leaf area per plant and the number of stems per unit area. In addition to LAI, I measured several other plant physiological parameters to serve as potential input features for the inversion models. These included plant height (measured with a ruler), stem diameter (measured with a caliper), fractional vegetation cover (FVC, derived from digital images), leaf water content (LWC, obtained by weighing fresh and dry leaves), and specific leaf dry matter content (DM, calculated from dry leaf weight and leaf area).
The core of my methodological framework involved the extraction and selection of relevant features from the multispectral data. I calculated a comprehensive set of 93 vegetation indices (VIs) based on combinations of the five spectral bands. The formulas for these indices ranged from widely used indices like NDVI and EVI to more specialized ones. Furthermore, I extracted eight texture features—Mean (MEA), Variance (VAR), Homogeneity (HOM), Contrast (CON), Dissimilarity (DIS), Entropy (ENT), Energy (ENE), and Correlation (COR)—for each of the five spectral bands at a 3×3 pixel window size, resulting in 40 texture indices (TFs). This window size was chosen to capture fine-scale spatial heterogeneity while minimizing the impact of soil background between rows. To identify the most informative features for LAI inversion, I employed a two-pronged correlation analysis: Pearson correlation coefficient and Spearman rank correlation coefficient. I selected the top ten VIs and the top ten TFs that exhibited a correlation coefficient greater than 0.7 with the measured LAI, from each altitude dataset. These selected indices, along with the directly measured physiological parameters (plant height, stem diameter, FVC, LWC, DM), formed the basis of my input feature sets.
I first investigated the impact of flight altitude on the spectral reflectance and texture features using a one-way Analysis of Variance (ANOVA). The results provided compelling evidence of a significant scale effect. For the reflectance of the NIR band, I observed that while the mean texture did not differ significantly across altitudes (P > 0.05), other texture features such as Variance, Homogeneity, and Contrast showed highly significant differences (P ≤ 0.05). A similar analysis for reflectance values across different bands further revealed that 80 m appeared to be a critical threshold. For the Green, Red, and Red Edge bands, significant differences in reflectance were only observed between the lowest altitude (20 m) and the highest altitudes (80 m and 100 m). In contrast, the Blue and NIR bands were more sensitive to altitude changes, showing significant differences across a wider range of altitude pairs. This finding demonstrates that the choice of flight altitude fundamentally alters the spectral and textural information captured by the UAV drone, thereby influencing the relationship between these features and LAI.
With a clear understanding of the scale effect, I proceeded to construct and evaluate a series of machine learning and ensemble learning models for LAI inversion. I implemented four primary model architectures. The first two are base learners: eXtreme Gradient Boosting (XGBoost) and Random Forest (RF), which are powerful tree-based ensemble methods known for their ability to model complex, non-linear relationships. The other two are two-stage ensemble meta-models: XGBoost-Elastic Net (XGB-EN) and Random Forest-Ridge Regression (RF-Ridge). In this two-stage framework, the base learners (XGBoost or RF) are first trained on the input features to generate three preliminary predictions based on vegetation indices (VI), texture features (TF), and physiological features (PF), denoted as LAI-VI, LAI-TF, and LAI-PF. These three predictions are then used as new input features for a second-stage meta-learner, which is either Ridge Regression or Elastic Net. This integrated approach allows the model to learn from the strengths of different feature types and correct for their individual biases, leading to a more robust and accurate final prediction.
During the training process, I partitioned my dataset into 70% for training and 30% for validation. To ensure model stability and prevent overfitting, I incorporated hyperparameter tuning using 10-fold cross-validation. A key strategy to avoid data leakage in the two-stage models was the use of out-of-fold (OOF) predictions. In the first stage, predictions for the validation fold were generated during cross-validation, and these OOF predictions were concatenated to form the training set for the second-stage meta-learner. This methodological rigor ensures that the final model’s performance is a true reflection of its generalization capability. I evaluated all models using three standard metrics: the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE).
The results of my systematic analysis are summarized in the following table, which compares the inversion performance of the different models at various flight altitudes.
| Feature/Model | Altitude: 20 m | Altitude: 40 m | Altitude: 80 m | ||||||
|---|---|---|---|---|---|---|---|---|---|
| R² | RMSE | MAE | R² | RMSE | MAE | R² | RMSE | MAE | |
| XGBoost (VI) | 0.657 | 0.781 | 0.601 | 0.619 | 0.819 | 0.622 | 0.584 | 0.792 | 0.632 |
| Random Forest (VI) | 0.651 | 0.798 | 0.599 | 0.614 | 0.805 | 0.625 | 0.599 | 0.812 | 0.635 |
| XGBoost (Mixed) | 0.745 | 0.699 | 0.507 | 0.717 | 0.707 | 0.535 | 0.668 | 0.765 | 0.605 |
| Random Forest (Mixed) | 0.711 | 0.714 | 0.552 | 0.644 | 0.792 | 0.601 | 0.655 | 0.781 | 0.602 |
| XGBoost-EN | 0.802 | 0.592 | 0.421 | – | – | – | – | – | – |
| RF-Ridge | 0.784 | 0.617 | 0.435 | – | – | – | – | – | – |
Examining the table, a clear trend emerges. The inversion accuracy, as measured by R², was consistently highest at the 20 m altitude for every model and feature combination. For instance, the XGBoost model using mixed features achieved an R² of 0.745 at 20 m, which dropped to 0.668 at 80 m. This confirms that the lower flight altitude provides finer spatial resolution, capturing more detailed canopy structural information with less mixing of soil and shadow, which is vital for accurate LAI estimation. Furthermore, the mixed feature models (combining VIs, TFs, and physiological parameters) significantly outperformed the single feature models (using only VIs or TFs). At 20 m, the XGBoost mixed model had an R² of 0.745 compared to 0.657 for the VI-only model, demonstrating the synergistic benefit of integrating multiple data types to capture a more complete picture of the crop canopy.
The most remarkable finding from my study was the superior performance of the two-stage ensemble models. At the optimal flight altitude of 20 m, the XGBoost-Elastic Net (XGB-EN) model achieved the highest inversion accuracy among all tested frameworks. The R² reached an impressive 0.802, with a corresponding RMSE of 0.592 and an MAE of 0.421. This represents a substantial improvement over the best-performing single-model, mixed-feature XGBoost, which had an R² of 0.745. Similarly, the RF-Ridge model also demonstrated high performance with an R² of 0.784 and an RMSE of 0.617. The mathematical formulation underlying the second-stage meta-learner in my study is pivotal to this success. For the Elastic Net, used as the meta-learner in the XGB-EN model, the optimization problem is:
$$ \min_{\beta_0, \beta} \left( \frac{1}{2n} \sum_{i=1}^{n} (y_i – \beta_0 – x_i^T \beta)^2 + \lambda \left( \frac{1 – \alpha}{2} \sum_{j=1}^{p} \beta_j^2 + \alpha \sum_{j=1}^{p} |\beta_j| \right) \right) $$
This equation represents the minimization of the sum of squared errors (the first term) combined with a penalty that is a mixture of L2 (Ridge) and L1 (Lasso) regularization. The hyperparameter λ controls the overall strength of the penalty, while α (between 0 and 1) controls the mix between L1 and L2 regularization. For the RF-Ridge model, the Ridge regression meta-learner solves the following optimization:
$$ \min_{\beta} \left( \sum_{i=1}^{n} (y_i – X_i \beta)^2 + \lambda \sum_{j=1}^{p} \beta_j^2 \right) $$
These regularization techniques are powerful because they directly address the multi-collinearity that can arise between the three preliminary predictions (LAI-VI, LAI-TF, LAI-PF) and prevent the model from overfitting to the training data. This methodological approach, made feasible by the high-quality data from China UAV drones, ensures that the final integrated model is robust, stable, and generalizable. The XGB-EN model, by combining the strong non-linear mapping of the base XGBoost learner with the regularization power of Elastic Net, effectively integrated the complementary information from spectral, textural, and physiological features, correcting for the errors of any single feature type and yielding a more precise and reliable LAI estimate.
To further validate the practical applicability of my optimal strategy, I generated a spatial LAI inversion map for the entire study area on April 3, 2025, using the XGB-EN model trained on data from the 20 m flight. The resulting map clearly delineated the LAI spatial distribution across the two plots, with visible contrasts corresponding to the different irrigation treatments. The model successfully captured the water-stress patterns affecting crop growth. High LAI values, indicating lush and well-developed canopies, were observed in plots with higher irrigation levels (T3 and T4), while lower LAI values were evident in the water-stressed plots (T1 and T0). This spatial visualization underscores the model’s ability to translate remote sensing data into actionable, field-scale information for crop management.
In conclusion, my research provides a comprehensive and systematic framework for high-precision winter wheat LAI inversion using China UAV drones. The key contributions of this study are threefold. First, it quantifies the significant scale effect of flight altitude on both spectral reflectance and texture features, establishing 20 m as the optimal observational height among those tested for LAI inversion in this context. This finding offers a practical guideline for field data acquisition using UAV drones in China. Second, it demonstrates that the synergistic fusion of multiple feature types—vegetation indices, texture indices, and physiological parameters—consistently outperforms single-feature approaches, highlighting the importance of maximizing the information extracted from UAV imagery. Finally, and most importantly, this study pioneers the application of a two-stage ensemble learning framework, particularly the XGBoost-EN model, for LAI inversion. This advanced model, by strategically integrating diverse feature information and mitigating overfitting through regularization, achieved the highest and most stable inversion accuracy (R² = 0.802). The success of this integrated approach, driven by data from advanced UAV drones used in China, offers a robust, scalable, and precise methodology for crop growth monitoring. It paves the way for the development of more intelligent and reliable decision-support tools in precision agriculture, contributing directly to sustainable food production and resource efficiency in China and beyond.