The timely and efficient acquisition of plant height information is a critical component in modern precision agriculture. For winter wheat, plant height serves as a vital phenotypic trait strongly correlated with growth vigor, biomass, lodging risk, and ultimately, yield. Traditional manual measurement methods are labor-intensive, destructive, and incapable of supporting large-scale, high-throughput phenotyping. In this context, UAV drones have emerged as a revolutionary platform, offering the ability to collect high-resolution, multi-modal remote sensing data with unparalleled flexibility. My research focuses on developing and validating a robust methodology for winter wheat plant height estimation by synergistically fusing spectral, structural, and textural features extracted from UAV drone systems. The core hypothesis is that combining information from multiple data sources—specifically vegetation indices from multispectral imagery, direct canopy height models from photogrammetry, and spatial texture features—can overcome the limitations of single-source models and significantly improve estimation accuracy across different growth stages.
The foundation of this study is a comprehensive field experiment. Data collection was synchronized with key phenological stages: the jointing stage and the heading stage. A UAV drone platform was equipped with two primary sensors: a high-resolution RGB camera for capturing orthophotos and a six-band multispectral camera. Flights were conducted at 100 meters above ground level under clear sky conditions to ensure consistent illumination. Concurrently, ground-truth plant height measurements were meticulously taken from 41 georeferenced sample plots within the field, providing the essential data for model training and validation.
Data Processing and Feature Extraction
The raw data from the UAV drone missions underwent a systematic preprocessing pipeline. Multispectral images were radiometrically calibrated and orthomosaicked to generate seamless reflectance maps for six bands: Blue (B1, 450 nm), Green (B2, 555 nm), Red (B3, 660 nm), Red Edge (B4, 720 nm), Red Edge 750 (B5, 750 nm), and Near-Infrared (B6, 840 nm). From these reflectance values, a suite of twelve vegetation indices (VIs), known for their sensitivity to vegetation biophysical parameters, were calculated. The formulas for key indices are listed below:
- Normalized Difference Vegetation Index (NDVI): $$NDVI = \frac{B_6 – B_3}{B_6 + B_3}$$
- Blue Normalized Difference Vegetation Index (BNDVI): $$BNDVI = \frac{B_6 – B_1}{B_6 + B_1}$$
- Green Normalized Difference Vegetation Index (GNDVI): $$GNDVI = \frac{B_6 – B_2}{B_6 + B_2}$$
- Modified Simple Ratio (MSR): $$MSR = \frac{(B_6 / B_3) – 1}{\sqrt{(B_6 / B_3) + 1}}$$
- Triangular Vegetation Index (TVI): $$TVI = 0.5 \times [120 \times (B_6 – B_2) – 200 \times (B_3 – B_2)]$$
Simultaneously, the RGB imagery from the UAV drone was processed using structure-from-motion photogrammetry to generate high-density point clouds and subsequently, Digital Surface Models (DSMs). A critical step was acquiring a bare soil DSM after sowing. The Winter Wheat Plant Height (WWPH) raster was then derived by subtracting this bare soil DSM from the crop canopy DSMs obtained at the jointing and heading stages:
$$WWPH_{d} = DSM_{d} – DSM_{0}$$
where \(WWPH_{d}\) is the estimated plant height layer for day \(d\), \(DSM_{d}\) is the canopy surface model, and \(DSM_{0}\) is the bare soil reference model.
To capture the spatial heterogeneity and pattern of the wheat canopy, texture features were extracted from each band of the multispectral imagery using the Gray-Level Co-occurrence Matrix (GLCM) method. A moving window of 7×7 pixels was applied to compute eight common texture metrics: Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment, and Correlation.
Feature Selection and Correlation Analysis
Prior to model building, a correlation analysis was performed between all extracted features (VIs, texture metrics, WWPH) and the ground-measured plant height. This step was crucial for identifying the most informative predictors and reducing dimensionality. The correlation patterns differed between growth stages.
| Growth Stage | Top Correlated Vegetation Indices (in order) | Typical Correlation Coefficient Range |
|---|---|---|
| Jointing | TVI, RDVI, RVI, MSR, BNDVI, GR | 0.42 – 0.69 |
| Heading | BNDVI, MSR, NDVI, RVI, RDVI, GNDVI | 0.43 – 0.72 |
For texture features, the Mean from the NIR (B6) and Red Edge 750 (B5) bands consistently showed the strongest positive correlations with plant height at both stages, often exceeding 0.60. This suggests that the average brightness in these longer-wavelength bands, which is influenced by canopy density and structure, is a robust textural indicator of plant height. The derived WWPH itself showed a very high correlation with measured height, validating the photogrammetric approach, but it also contained noise from residual soil micro-topography and non-plant objects.
| Growth Stage | Texture Feature (Band) | Correlation Coefficient |
|---|---|---|
| Jointing | Mean (B6: NIR) | 0.70 |
| Jointing | Mean (B5: Red Edge 750) | 0.60 |
| Heading | Mean (B6: NIR) | 0.69 |
| Heading | Mean (B5: Red Edge 750) | 0.63 |
Model Development and Comparison
Based on the correlation analysis, different feature combinations were constructed as input variables for predictive modeling. The primary combinations tested were:
- Combination A (VI+WWPH): Top 6 Vegetation Indices + WWPH layer.
- Combination B (VI+WWPH+Tex): Top 4 Vegetation Indices + WWPH layer + Top 2 Texture Features (Mean of B5 and B6).
To evaluate performance, three distinct machine learning regression algorithms were employed: Random Forest (RF), Support Vector Regression (SVR), and Partial Least Squares Regression (PLSR). The dataset from the 41 sample points was randomly split into a training set (80%) and an independent testing set (20%). Model performance was evaluated using the Coefficient of Determination (R2) and the Root Mean Square Error (RMSE).
The results unequivocally demonstrated the superiority of the Random Forest (RF) algorithm for this task, regardless of the input combination. RF’s inherent ability to handle non-linear relationships, its robustness to overfitting through ensemble learning, and its insensitivity to the scale of input features made it ideally suited for the complex, multi-source data provided by the UAV drone.
| Input Combination | Algorithm | R2 | RMSE (cm) |
|---|---|---|---|
| VI + WWPH | Random Forest (RF) | 0.850 | 1.767 |
| VI + WWPH | Support Vector Regression (SVR) | 0.802 | 2.101 |
| VI + WWPH | Partial Least Squares (PLSR) | 0.587 | 2.525 |
| VI + WWPH + Texture | Random Forest (RF) | 0.872 | 1.731 |
| VI + WWPH + Texture | Support Vector Regression (SVR) | 0.820 | 2.030 |
| VI + WWPH + Texture | Partial Least Squares (PLSR) | 0.633 | 2.217 |
| Input Combination | Algorithm | R2 | RMSE (cm) |
|---|---|---|---|
| VI + WWPH | Random Forest (RF) | 0.869 | 1.405 |
| VI + WWPH | Support Vector Regression (SVR) | 0.842 | 1.615 |
| VI + WWPH | Partial Least Squares (PLSR) | 0.653 | 2.237 |
| VI + WWPH + Texture | Random Forest (RF) | 0.887 | 1.335 |
| VI + WWPH + Texture | Support Vector Regression (SVR) | 0.784 | 2.140 |
| VI + WWPH + Texture | Partial Least Squares (PLSR) | 0.662 | 2.443 |
The incremental benefit of fusing all three data sources is clear. For the jointing stage, adding texture features to the VI+WWPH model improved the R2 of the best RF model from 0.850 to 0.872 and reduced RMSE. A similar, though slightly less pronounced, improvement was observed for the heading stage. This confirms that texture features provide complementary spatial information not fully captured by spectral indices or the DSM-derived height alone, likely related to canopy roughness, row spacing visibility, and overall canopy architecture. The photogrammetric WWPH layer provided a strong direct structural signal, while the VIs captured the photosynthetic activity and greenness. The UAV drone‘s ability to seamlessly collect these complementary data streams is what enables this powerful fusion.
Spatial Estimation and Map Generation
Following the model selection, the optimal pipeline—Random Forest algorithm trained on VI+WWPH+Texture features—was applied to the entire study area. The required input rasters (VIs, WWPH, texture layers) were generated for each pixel, and the trained RF model predicted plant height values across the entire field, producing high-resolution spatial distribution maps.
The generated maps revealed distinct spatial patterns. At the jointing stage, estimated plant height varied considerably, ranging from approximately 40.5 cm to 52.6 cm. Clear spatial heterogeneity was observed, with lower heights in the central-southern sectors and taller crops along the eastern and western edges. This variability can be attributed to differences in soil properties, micro-topography affecting water distribution, or slight variations in seeding and early management. By the heading stage, the canopy had become more uniform, with heights ranging from 60.3 cm to 71.9 cm. The spatial variability was reduced, though lower-height patches persisted in specific locations and field borders, potentially indicating areas of lingering stress or poorer soil conditions.
The process can be conceptually summarized by the following function, where \(RF_{model}\) is the trained Random Forest regressor:
$$PH_{estimated}(x,y) = RF_{model}\big( VI_1(x,y), …, VI_n(x,y), WWPH(x,y), Tex_1(x,y), Tex_2(x,y) \big)$$
Here, \((x,y)\) denotes the spatial coordinates of each pixel in the UAV drone-derived rasters.
Discussion and Implications
The success of the fused model underscores a critical principle in remote sensing phenotyping: multi-source data integration mitigates the limitations of individual techniques. Vegetation indices can saturate in dense canopies. Photogrammetric height can be noisy at the individual plant level and requires a bare soil reference. Texture features alone may be ambiguous. However, when combined within a powerful non-linear learner like Random Forest, these data streams inform and correct each other, leading to a more robust and accurate estimation. The portability and flexibility of UAV drones are fundamental to this approach, as they allow for the cost-effective and synchronous collection of RGB, multispectral, and even hyperspectral or LiDAR data on demand.
The choice of Random Forest proved decisive. Compared to SVR and PLSR, RF required minimal hyperparameter tuning to achieve high performance, handled the mixed-type input features effortlessly, and provided built-in estimates of feature importance. Analysis of the RF model’s internal metrics showed that the WWPH layer was consistently the most important feature, followed by key vegetation indices like MSR and BNDVI, with the texture features providing a valuable, though secondary, contribution. This feature importance ranking offers practical guidance for optimizing future UAV drone sensing campaigns.
While this study demonstrates a highly effective framework, further research is warranted. The model was developed on a single wheat variety under specific agronomic conditions. Validation across diverse cultivars, management practices, and environmental conditions is necessary to assess its generalizability. Furthermore, exploring temporal fusion—building models that can intelligently integrate data from multiple growth stages to predict final yield or detect stress—represents a promising future direction. The integration of data from UAV drones with other platforms, such as satellite imagery for broader context, could also enhance large-scale monitoring applications.
Conclusion
This research presents and validates a novel, high-accuracy methodology for estimating winter wheat plant height by fusing multi-source remote sensing data acquired from UAV drones. The core findings are threefold. First, a feature combination comprising selected Vegetation Indices, a photogrammetrically-derived Plant Height layer (WWPH), and spatial Texture features provides the most informative input for prediction models. Second, among the tested algorithms, the Random Forest regressor consistently delivers the highest estimation accuracy and robustness at both jointing and heading stages. Third, the application of this optimized model enables the generation of detailed, high-resolution plant height maps that reveal significant spatial heterogeneity within fields, information that is invaluable for precision crop management.
The study conclusively demonstrates that the synergistic use of spectral, structural, and textural data from UAV drones, processed through advanced machine learning, offers a powerful, non-destructive, and scalable solution for high-throughput plant phenotyping. This approach provides agronomists and farmers with a critical tool to monitor crop growth, identify in-field variability, and make data-driven decisions to optimize inputs and maximize productivity, paving the way for more sustainable and efficient agricultural systems.