UAV Hyperspectral Detection of Citrus Huanglongbing

UAV-based hyperspectral remote sensing technology combines the advantages of hyperspectral cameras and UAV platforms, enabling both spectral and imaging data collection across large plantation areas. This technology provides efficient and non-destructive disease monitoring, accelerating the transition towards precision agriculture and smart farming.

Based on extensive ground-based hyperspectral plant disease detection research, several crops have already adopted UAV hyperspectral imaging for disease diagnosis. For example, Photochemical Reflectance Index (PRI) has been used to detect wheat stripe rust, demonstrating its potential in early disease identification. Researchers also used multi-temporal hyperspectral images to extract sensitive spectral bands for winter wheat, constructing vegetation disease indices and establishing multivariate linear regression models. By integrating Normalized Difference Spectral Index (NDSI), Difference Spectral Index (DSI), Ratio Spectral Index (RSI), and wheat take-all disease indices, they developed a grading model for wheat take-all disease, where Radial Basis Function Support Vector Machine (RBF-SVM) achieved classification accuracy of 90.35% with a kappa coefficient of 0.86.

First, ground-based and aerial hyperspectral data were simultaneously collected. The ground data obtained by a field spectroradiometer was used as a reference to evaluate the quality of the UAV-based hyperspectral data. Correlation analysis demonstrated a strong relationship between the two datasets, with a coefficient of determination (R²) exceeding 0.96.

Second, different methods for extracting canopy spectral samples were compared. Extracting canopy spectra at the single-pixel level ensured sample diversity and improved the robustness of the detection model. An isolation forest algorithm was applied to identify and remove anomalous samples, reducing noise interference and enhancing the detection model's performance.

Third, both the Savitzky-Golay filter and the Daubechies wavelet transform were effective for smoothing and denoising hyperspectral data. However, the Daubechies wavelet method incorporated noise type identification and feature reconstruction during the denoising process. Using a particle swarm optimization algorithm to optimize threshold selection further enhanced the balance between noise reduction and spectral feature preservation.

Fourth, the spectral characteristics of citrus Huanglongbing (HLB)-infected trees differed significantly from healthy trees, showing a higher green peak, lower near-infrared reflectance, and a red edge shift towards shorter wavelengths. Key spectral bands identified using an improved genetic algorithm with optimized selection operators were 468 nm, 504 nm, 512 nm, 516 nm, 528 nm, 536 nm, 632 nm, 680 nm, 688 nm, and 852 nm. Detection models based on vegetation indices calculated from these key bands demonstrated higher classification accuracy.

Finally, it was found that as the input data became more feature-rich, the performance of the HLB detection model significantly improved. A multi-feature input HLB detection model achieved a classification accuracy of 99.33% on the training set (loss = 0.0783) and 99.72% on the validation set (loss = 0.0585). When applied to the test area, the model effectively identified HLB-infected citrus trees and mapped the disease distribution within the tree canopy, enabling further assessment of disease severity.

Ground-based and aerial hyperspectral data were collected simultaneously, using the field spectroradiometer data as a reference to evaluate the quality of the UAV-based hyperspectral data. Correlation analysis showed a significant correlation between the hyperspectral camera data acquired by the UAV and the ground data, with R² exceeding 0.96.

Canopy spectral sample extraction methods were analyzed and compared. Single-pixel-based sample extraction preserved sample diversity, improving the model’s robustness. An isolation forest was used to detect and remove outliers, effectively reducing noise interference and enhancing the detection model’s performance.

Both Savitzky-Golay filtering and Daubechies wavelet transform effectively smoothed and denoised the hyperspectral data, with the Daubechies wavelet method offering advantages by considering noise characteristics and reconstructing spectral features. A particle swarm optimization algorithm was used to optimize the denoising threshold, improving noise suppression while maintaining spectral features.

The spectral characteristics of HLB-infected citrus canopies differed from healthy ones, including a higher green peak, lower NIR reflectance, and a red edge shift. Key spectral bands (468 nm, 504 nm, 512 nm, 516 nm, 528 nm, 536 nm, 632 nm, 680 nm, 688 nm, and 852 nm) were selected using a genetic algorithm with optimized selection operators. Detection models built using vegetation indices derived from these bands achieved superior classification accuracy. It was also observed that incorporating richer spectral features into the model significantly enhanced the detection performance.

The hyperspectral data extracted from the canopy carried abundant spectral information. By applying spectral transformation methods, transformed data were generated that better highlighted spectral characteristics. Based on both original and transformed spectra, key spectral features were extracted to facilitate the analysis of HLB-infected canopy spectral characteristics.

In the visible range, the maximum first derivative value within each spectral region defines the "edge" for that region. Common features in hyperspectral vegetation research include the blue edge, yellow edge, and red edge. Physiologically important parameters like the green peak and red valley are closely linked to vegetation health. The green peak reflects chlorophyll (Chl-a) content, while the red valley relates to the strong absorption of red light by vegetation, both of which can be quantified by green peak reflectance and red valley absorption.

In the near-infrared (NIR) region, the red edge is a distinctive feature of healthy green vegetation. Parameters such as the red edge position (λre), red edge amplitude, and red edge area are widely used. Physically, λre is the wavelength where the first derivative of reflectance reaches its maximum between 670 nm and 760 nm. The red edge amplitude is the highest first derivative value in this range, and the red edge area is the area between the first derivative curve and the x-axis within this band.

The field spectroradiometer used in this study covered a spectral range of 325 nm to 1075 nm with 1 nm spectral resolution, providing 750 spectral bands. In contrast, the UAV hyperspectral camera covered 450 nm to 950 nm with a spectral resolution of 4 nm, providing 125 bands.

To ensure comparability, the higher-resolution ground spectra were resampled. First, data outside the 450 nm to 950 nm range were removed. Then, within this range, the spectra were resampled by averaging adjacent bands to match the hyperspectral camera's spectral resolution. The resampling process ensured the ground-based and UAV datasets were spectrally compatible for subsequent analysis.

This study applied a Support Vector Machine (SVM) classifier with a quadratic kernel function to develop the classification model. The model used the first three components derived from Minimum Noise Fraction (MNF) transform of the full-band UAV hyperspectral imagery as input features. Regions of Interest (ROIs) for canopy and non-canopy areas were defined using ENVI software, and these ROIs were used to train the classification model. The trained SVM model was then applied to the UAV hyperspectral panoramic image, producing a classified output map.

For the canopy classification model, the SVM kernel parameters and r were set to their default values, while the penalty coefficient (C) was tuned through a gradient search from 10 to 50 with a step size of 10. The optimal penalty coefficient was determined to be C = 40. The final canopy classification result, shown in Figure 5, indicates that the black areas represent the background, while the white areas correspond to the extracted citrus canopy.

This approach demonstrates the effective use of UAV hyperspectral imaging and MNF transformation combined with SVM classification to accurately identify and extract citrus canopy areas, laying a solid foundation for further citrus Huanglongbing detection using hyperspectral camera data.