Estimation of cotton leaf area index under Verticillium wilt stress using UAV-based multispectral remote sensing
2026 年 19 巻 1 号 p. 51-59
詳細
2026 年 19 巻 1 号 p. 51-59
This study utilized unmanned aerial vehicle (UAV)-based multispectral remote sensing to estimate the leaf area index (LAI) of cotton under Verticillium wilt stress. Key spectral bands (B12, B9, B8) and vegetation indices—transformed vegetation index (TVI), difference vegetation index (DVI), and enhanced vegetation index (EVI)—were identified as strongly correlated with LAI. A support vector regression (SVR) model utilizing these features achieved the best estimation performance (validation: R2 = 0.877, RMSE = 0.284). Furthermore, a radial basis function kernel support vector machine (SVM-RBF) classifier attained the highest accuracy in mapping canopy parameters (overall accuracy = 94.05 %, Kappa = 0.916). The proposed framework offers a viable technical solution for large-scale, real-time monitoring of cotton Verticillium wilt.
Leaf area index (LAI) is a key plant physiological parameter that reflects photosynthetic capacity and canopy architecture, and has a significant impact on crop growth assessment and yield forecasting (Wang et al., 2005). There have been many studies using near earth remote sensing and satellite remote sensing to monitor crop leaf area index (Chen et al., 2020; Dong et al., 2019; Zhao, 2014). With the rapid development of drone remote sensing, using drones for real-time and non-destructive monitoring of LAI has become a new trend. Wang et al. (2021) used drone multi-spectral remote sensing data to obtain index regression models and support vector machine models, which can quickly invert soybean LAI. Jiang et al. (2022) used quinoa as the object, monitored crop phenotype information using drones, and estimated LAI through random forest regression method. They found that the determination coefficient R2 of LAI was 0.977‒0.980, indicating a significant estimation effect. Feng et al. (2022) used a six-rotor drone with a multi spectrometer and thermal imaging system to extract vegetation indices, texture features, and canopy temperature information from remote sensing images of wheat powdery mildew. Then a monitoring model for wheat powdery mildew disease index was established. With the breakthrough of drone remote sensing and machine learning algorithms, LAI has developed from a theoretical indicator to a core tool for quantitative monitoring of diseases, especially n crop disease resistance evaluation and early diagnosis of stress (Alchemi et al., 2022; Li et al., 2023; Liu et al., 2023).
Verticillium wilt (Verticillium dahliae Kleb.) is one of the most devastating diseases in cotton cultivation. Investigating the changes in LAI under the stress of cotton wilt disease is beneficial for rapid, accurate, and non-destructive monitoring of cotton wilt disease (Chen et al., 2007; Dadd-Daigle et al., 2021). Studies have shown that the physiological phenotype of cotton leaves can be effectively detected by spectral data, which could provide a solid theoretical basis for the rapid detection of Verticillium wilt stress (Wu et al., 2025). Tian et al. (2025) utilized multi-spectral data acquired from the UAV, and systematically analyzed the adaptability of traditional machine learning (RF, ELM), the deep learning (CNN) in LAI prediction to establish the optimal modeling paradigm. Although many scholars have studied the estimation of cotton LAI, there is still limited research on the inversion of LAI under disease stress, especially the estimation of LAI under cotton wilt stress (Zhao et al., 2022). Targeting typical cotton fields north of the Tianshan Mountain, we investigate LAI variations under different levels of disease. UAV multispectral data is used to invert sensitive bands and vegetation indices of cotton LAI. Then a multiple linear regression model, partial least squares regression model, principal component regression model, and support vector machine estimation model were validated. The results can provide a significant reference for real-time monitoring of cotton growth and effective field management under Verticillium wilt conditions.
Field tests were conducted in the Verticillium wilt nursery (Xinjiang Academy of Agricultural and Reclamation Sciences, Shihezi; 44°18'N, 86°02'E) from 2021 to 2022. It is continental climate with 168‒171 frost-free days. Cotton cv. Xinluzao 8 was planted in mid-April using precision sowing (66 + 10 cm wide-narrow row spacing) and drip irrigation. The fertilization is 375 kg·ha‒1 N, 150 kg·ha‒1 P2O5, and 75 kg·ha‒1 K2O.
2.2. Experimental designAfter random sampling, 30 samples were selected from 46 experimental sites nationwide, totaling 1318 tested cultivars. A grid-based point distribution method was used to arrange survey points in the experimental field, with 60 survey points in 2021 (model training set) and 30 in 2022 (model validation set). Handheld GPS was used for latitude and longitude positioning. Targeting Verticillium wilt, the disease naturally occurred and spread from mid-July 2021to 2022 without manual intervention. Based on disease progression (showed in Table 1), cotton plants were classified into five severity grades: healthy (b0), slight (b1), moderate (b2), serious (b3), and critical (b4). Disease severity was recorded at each growth stage, with measurements of three physiological parameters: aboveground fresh biomass, LAI, and plant water content.
| Disease severity | Disease Index | Disease Division Standard |
|---|---|---|
| b0 (Health) | 0 | Plant health, disease-free leaves |
| b1 (Slight) | 0 < DI ≤ 25 % | Less than 1/4 of the leaves show symptoms, with light yellow or yellow irregular lesions between the main veins of the leaves |
| b2 (Moderate) | 25 % < DI ≤ 50 % | 1/4‒1/2 of the leaves show symptoms, most of the disease spots are yellow or yellowish brown, and the edge of the leaves is slightly curled and withered |
| b3 (Serious) | 50 % < DI ≤ 75 % | 1/2‒3/4 of the leaves show disease symptoms, and a few leaves fall off |
| b4 (Critical) | 75 % < DI ≤ 100 % | More than 3/4 leaf disease, mostly brown withered spots, cotton plant leaves fall off for light pole or even death |
UAV Six-rotor Hawk-6 X, Shanghai Huanying Aviation Technology Co., Ltd., China) image acquisition was conducted between 10:00 and 14:00 during 2021‒2022 under clear, cloudless, and low-wind conditions, with a data collection interval of 15‒20 days. Data in the key growth stages of cotton was collected, 2021: June 30 (full bud stage), July 17 (full bloom stage), August 7, August 17 (peak boll stage), September 3 (boll opening stage); 2022: June 30 (full bud stage), July 18 (full bloom stage), August 5, August 18 (peak boll stage), September 5 (boll opening stage). Data from 2021 were used for modeling, 2022 data for model validation. The flight Parameters are 100 m altitude, 75 % forward overlap, 70 % side overlap.
A multispectral camera (Micro MCA12 Snap, Tetracam Ltd., USA) was mounted on the UAV, providing 12 spectral bands. As shown in Table 2, bands T1‒T5 cover the visible spectrum; T6‒T7 cover the red-edge, and T8‒T12 cover the near-infrared range. Multispectral data was preprocessed (format conversion, image stitching, radiometric calibration) using Pixel Wrench software (Pixel Wrench 2 x64, Beijing Servirst Technology Co., Ltd., China). Flight route planning was conducted using Ground Control Station 1.0 software (Shanghai Huanying Aviation Technology Co., Ltd.), with the UAV operating autonomously along a predefined path. Raw images in ‘RAW’ format were converted to ‘TIF’ using PixelWrench 2 x64. A 12-band ortho-mosaic was then generated through image stitching in Pix4Dmapper (Pix4D S.A., Switzerland). The subsequent preprocessing, including radiometric calibration and image cropping, was performed using ENVI Classic 5.3 (Exelis Inc., USA).
| Master | Wavelength-width (nm) | Band characteristics and application |
|---|---|---|
| T1 | B1 (470 nm-10 nm) | Differentiation of vegetation from soil and rock surfaces |
| T2 | B2 (515 nm-10 nm) | Green peak in the visible spectrum |
| T3 | B3 (550 nm-10 nm) | Sensitivity to water turbidity variations |
| T4 | B4 (610 nm-10 nm) | Initial red band in the vegetation spectral reflectance curve |
| T5 | B5 (656 nm-10 nm) | Renormalized difference vegetation index (RDVI) |
| T6 | B6 (710 nm-10 nm) | Initial upslope position of the vegetation spectral reflectance curve (typically the Red-Edge region) |
| T7 | B7 (760 nm-10 nm) | Initial near-infrared (NIR) position of the vegetation spectral reflectance curve |
| T8 | B8 (800 nm-10 nm) | Normalized difference vegetation index (NDVI), Renormalized difference vegetation index (RDVI) |
| T9 | B9 (830 nm-10 nm) | Discrimination of vegetation species |
| T10 | B10 (860 nm-10 nm) | Significant correlation with total plant chlorophyll content |
| T11 | B11 (900 nm-20 nm) | Calculation of the crop water stress index (CWSI) |
| T12 | B12 (950 nm-20 nm) | Calculation of the water band index (WBI) |
Disease assessment was conducted according to the relative disease index method in ‘Plant disease research methods’ (Song et al., 2023). Before disease outbreak, a total of 60 survey points (2021) and 30 points (2022) were geotagged using GPS and uniformly distributed in the experimental area in 2021‒2022, respectively. Field surveys were carried out at 15-day intervals, with sampling conducted at each critical growth stage until harvest. Disease severity recorded at these monitoring points was subsequently used to predict the progression of Verticillium wilt in the entire field.
2.3.3. Agronomic physiological parameter measurementLAI was determined using the punching method. Cotton leaves were punched to calculate the LAI of different disease severity. The leaf area is calculated as follows.
| (1) |
Where ML represents the dry weight of cotton leaves after punching, MP represents the dry weight of cotton pore leaves, SP represents the single hole area of cotton leaves, n is the number of holes drilled, and S represents the footprint of 4 cotton plants.
The measured values of cotton LAI under different growth stages of Verticillium wilt stress (Table 3) showed that under Verticillium wilt stress, the LAI of cotton decreases with increasing disease severity. At different growth stages, the LAI of healthy cotton fields (b0) increased and then decreased, reaching its maximum value during the period of the peak boiling. During the same reproductive period, as the disease severity increases, the LAI decreased. The critical disease (b4) occurs in the peak bell period and boll opening period.
| Growth period | Disease grade | Average value | Average value | Maximum value | Minimum value | Standard deviation |
|---|---|---|---|---|---|---|
| Full budding stage | b0 | 60 | 1.67 | 2.34 | 1.29 | 0.30 |
| Blooming stage | b0 | 8 | 4.39 | 4.59 | 4.18 | 0.16 |
| b1 | 41 | 4.20 | 4.55 | 3.93 | 0.18 | |
| b2 | 11 | 3.46 | 3.76 | 3.23 | 0.18 | |
| Peak boiling stage | b0 | 3 | 4.95 | 5.00 | 4.90 | 0.05 |
| b1 | 4 | 4.14 | 4.29 | 4.03 | 0.12 | |
| b2 | 39 | 3.20 | 3.63 | 2.98 | 0.18 | |
| b3 | 14 | 2.23 | 2.51 | 2.03 | 0.14 | |
| Boll opening stage | b0 | 3 | 4.86 | 4.94 | 4.81 | 0.07 |
| b1 | 3 | 4.04 | 4.09 | 4.00 | 0.11 | |
| b2 | 21 | 3.11 | 3.40 | 2.97 | 0.10 | |
| b3 | 28 | 2.16 | 2.47 | 1.99 | 0.14 | |
| b4 | 5 | 1.36 | 1.49 | 1.25 | 0.11 |
Based on the correlation analysis between the reflectance of different spectral bands and LAI of cotton under Verticillium wilt stress (Table 4), B3, B6‒B12 are highly positively correlated with LAI at the 0.01 level; band B2 is highly negatively correlated with LAI at the 0.01 level, band B4 at the 0.05 level, and bands B1 and B5 are correlated with LAI. In the visible light band (B1‒B5), the correlation between band and LAI is relatively small, while bands B1 and B5 are not significantly correlated with LAI. Band B2 has a highly negative correlation, and band B3 has a highly positive correlation. |r| fluctuates within 0.014‒0.712. In the red edge band (B6‒B7), bands B6 and B7 show a highly positive correlation, with |r| fluctuating in 0.667‒0.854. In the near-infrared band (B8‒B12), there is a highly significant positive correlation between bands B8‒B12, with |r| fluctuating in 0.809‒0.868. The correlation coefficients rank as: B12 > B9 > B8 > B7 > B11 > B10 > B3 > B6 > B2 > B4 > B5 > B1. B12 has the highest correlation coefficient, |r| = 0.868.
| Wave band | Correlation coefficient | |r| | Sequence | |
|---|---|---|---|---|
| B1 | 0.014 | NS | 0.014 | 12 |
| B2 | −0.466 | ** | 0.466 | 9 |
| B3 | 0.712 | ** | 0.712 | 7 |
| B4 | 0.186 | * | 0.186 | 10 |
| B5 | −0.036 | NS | 0.036 | 11 |
| B6 | 0.667 | ** | 0.667 | 8 |
| B7 | 0.854 | ** | 0.854 | 4 |
| B8 | 0.858 | ** | 0.858 | 3 |
| B9 | 0.866 | ** | 0.866 | 2 |
| B10 | 0.809 | ** | 0.809 | 6 |
| B11 | 0.835 | ** | 0.835 | 5 |
| B12 | 0.868 | ** | 0.868 | 1 |
** is significant correlation at 0.01 level; * is significant correlation at 0.05 level; NS is not significant.
Band B12 is the optimal band for UAV monitoring of cotton LAI under Verticillium wilt stress. Bands B8, B9 and B12 can be used as estimation models for constructing cotton leaf area index under drone Verticillium wilt stress.
3.3. Optimal inversion vegetation index selection for LAI estimation under Verticillium wilt stressBased on the correlation analysis between 11 vegetation indices and LAI (Table 5), it can be seen that NDVI, RVI, DVI, RDVI, NDGI, TVI, SAVI, OSAVI, and MSAVI are significantly positively correlated with LAI at the 0.01 level. The correlation coefficients rank as: TVI > DVI > EVI > RDVI > SAVI > MSAVI > OSAVI > NDGI > ARI > NDVI > RVI. TVI, DVI, EVI have the highest correlation, and |r| values are 0.886, 0.884, and 0.867, respectively. The TVI is determined as the optimal vegetation index for cotton LAI monitoring under Verticillium wilt stress using UAV. In addition, these three vegetation indices, TVI, DVI, and EVI with the highest correlations can be incorporated into estimation models for predicting cotton LAI under this disease stress.
| Vegetation index | Correlation coefficient | |r| | Sequence | |
|---|---|---|---|---|
| Normalized difference vegetation index (NDVI) | 0.513 | ** | 0.513 | 10 |
| Ratio vegetation index (RVI) | 0.492 | ** | 0.492 | 11 |
| Difference vegetation index (DVI) | 0.884 | ** | 0.884 | 2 |
| Re-normalized difference vegetation index (RDVI) | 0.850 | ** | 0.850 | 4 |
| Normalized difference greenness index (NDGI) | 0.655 | ** | 0.655 | 8 |
| Triangular vegetation index (TVI) | 0.886 | ** | 0.886 | 1 |
| Soil adjusted vegetation index (SAVI) | 0.848 | ** | 0.848 | 5 |
| Optimized soil adjusted vegetation index (OSAVI) | 0.744 | ** | 0.744 | 7 |
| Modified type of soil adjusting the vegetation index (MSAVI) | 0.838 | ** | 0.838 | 6 |
| Anthocyanin emission index (ARI) | −0.597 | ** | 0.597 | 9 |
| Enhanced vegetation index (EVI) | 0.867 | ** | 0.867 | 3 |
** is significant correlation at 0.01 level; * is significant correlation at 0.05 level; NS is not significant.
We constructed four regression models: a multiple linear regression model (Table 6), a least squares regression model (Table 7), a principal component regression model (Table 8), and a support vector machine regression model (Table 9) for the LAI using UAV multispectral bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI) as independent variables.
| Variable | Equation | R2 | RMSE |
|---|---|---|---|
| B8, B9, B12, DVI, TVI, EVI | LAI = 2.953 − 0.55B8 − 0.526B9 + 6.498B12 − 10.325DVI + 4.789TVI − 0.275EVI | 0.814 | 0.345 |
| B12, TVI | LAI = 0.051 + 5.34B12 + 1.905TVI | 0.810 | 0.344 |
| Variable | Equation | R2 | RMSE |
|---|---|---|---|
| B8, B9, B12, DVI, RDVI, SAVI |
LAI = −2.882 + 2.298B8 + 2.530B9 + 2.048B12 + 2.288DVI + 1.949RDVI + 2.030SAVI |
0.803 | 0.347 |
| B12, TVI | LAI = 1.799 + 0.712B12 + 2.905TVI | 0.757 | 0.339 |
| Variable | Equation | R2 | RMSE |
|---|---|---|---|
| B8, B9, B12, DVI, TVI, EVI | LAI = 0.803 + 0.385B8 + 0.407B9 + 0.331B12 + 0.391DVI + 1.546TVI + 1.326EVI | 0.785 | 0.362 |
| B12, TVI | LAI = 1.812 + 0.676B12 + 2.913TVI | 0.757 | 0.339 |
| Variable | Radial basis function Kernel (Sun, 2004) | R2 | RMSE |
|---|---|---|---|
| B8, B9, B12, DVI, TVI, EVI | K(x, xi) = exp(−γ‖x − xi‖2) | 0.825 | 0.328 |
| B12, TVI | 0.794 | 0.311 |
Validation of the multiple linear regression models (Fig. 1) showed that the model using band B12 and the transformed vegetation index (TVI) as predictors achieved the highest R2 (0.853) and the lowest RMSE (0.311). These results demonstrate that a multiple linear regression model based on B12 and TVI can effectively estimate the leaf area index under Verticillium wilt stress.
The least squares regression models were validated (Fig. 2). The combinations were ranked in descending order of R2 as: band (B12) with vegetation index (TVI) > bands (B8, B9, B12) with vegetation indices (DVI, TVI, EVI) > bands (B2, B7, B12). The combination of band (B12) and vegetation index (TVI) achieved the highest R2 (0.851) and the lowest RMSE (0.313). Conversely, the combination of bands (B2, B7, B12) achieved the lowest R2 (0.79) and the highest RMSE (0.366). The results demonstrate that the principal component regression model constructed using bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI) can effectively estimate the LAI of cotton under Verticillium wilt stress.
The principal component regression models were validated (Fig. 3). Band B12 and vegetation index (TVI) have the highest R2 and the lowest RMSE at 0.851 and 0.313, respectively. The results indicate that a principal component regression model using bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI) can effectively estimate the LAI under Verticillium wilt stress.
The support vector machine regression models were validated (Fig. 4). R2 ranks as: band (B12) and vegetation index (TVI) < band (B8, B9, B12) and vegetation index (DVI, TVI, EVI). The vegetation indices (DVI, TVI, EVI) of the bands (B8, B9, B12) with the highest R2 and the lowest RMSE are 0.877 and 0.284, respectively. The results indicate that a support vector machine model using bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI) can effectively estimate the LAI under Verticillium wilt stress.
All four regression models achieved high performance, with determination coefficients (R2) consistently falling within a narrow range of 0.75 to 0.83, indicating only minor accuracy differences in the various spectral parameter combinations. In the three combinations of spectral indices, the support vector machine regression model constructed with bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI) was the best (modeling set R2 = 0.825, RMSE = 0.328). The validation set R2 = 0.877, RMSE = 0.284), band (B12), and vegetation index (TVI) are used to construct the optimal multiple linear regression model (modeling set R2 = 0.810, RMSE = 0.344; Validation set R2 = 0.853, RMSE = 0.311). Therefore, the support vector machine regression model based on bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI) is an optimal model for monitoring cotton leaf area index under Verticillium wilt stress.
3.5. Monitoring the spatial distribution of cotton LAI under Verticillium wilt stressA support vector machine classifier with linear, polynomial, radial basis function, and sigmoid kernels was applied to features from the combination of spectral bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI). The classification results and their corresponding accuracies were compared to identify the optimal method for monitoring cotton LAI under Verticillium wilt stress using UAV multispectral imagery. This optimal method was subsequently used to generate an LAI spatial distribution map.
3.5.1. Separability analysis of training and validation samplesThe separability between training and validation samples at different disease severity levels was assessed using Jeffries‒Matusita (JM) distance and transformed divergence (TD) (Tables 10 and 11). The high calculated values confirmed that the samples were well-separated, providing a reliable foundation for accurate classification.
| Training samples | Other training samples | Jeffries‒Matusita distance | Transformed Divergence |
|---|---|---|---|
| b0 (4.5‒5.0) | b1 (4.0‒4.5) | 1.982 | 2.000 |
| b2 (3.0‒3.9) | 1.998 | 2.000 | |
| b3 (2.0‒2.9) | 1.999 | 2.000 | |
| b1 (4.0‒4.5) | b0 (4.5‒5.0) | 1.982 | 2.000 |
| b2 (3.0‒3.9) | 1.991 | 2.000 | |
| b3 (2.0‒2.9) | 1.999 | 2.000 | |
| b2 (3.0‒3.9) | b0 (4.5‒5.0) | 1.998 | 2.000 |
| b1 (4.0‒4.5) | 1.991 | 2.000 | |
| b3 (3.0‒3.9) | 1.973 | 2.000 | |
| b3 (2.0‒2.9) | b0 (4.5‒5.0) | 1.999 | 2.000 |
| b1 (4.0‒4.5) | 1.999 | 2.000 | |
| b2 (3.0‒3.9) | 1.973 | 2.000 |
| Verification sample | Other validation samples | Jeffries‒Matusita distance | Transformed Divergence |
|---|---|---|---|
| b0 (4.5‒5.0) | b1 (4.0‒4.5) | 1.985 | 2.000 |
| b2 (3.0‒3.9) | 1.999 | 2.000 | |
| b3 (2.0‒2.9) | 1.999 | 2.000 | |
| b1 (4.0‒4.5) | b0 (4.5‒5.0) | 1.985 | 2.000 |
| b2 (3.0‒3.9) | 1.989 | 2.000 | |
| b3 (2.0‒2.9) | 1.999 | 2.000 | |
| b2 (3.0‒3.9) | b0 (4.5‒5.0) | 1.999 | 2.000 |
| b1 (4.0‒4.5) | 1.989 | 2.000 | |
| b3 (3.0‒3.9) | 1.995 | 2.000 | |
| b3 (2.0‒2.9) | b0 (4.5‒5.0) | 1.999 | 2.000 |
| b1 (4.0‒4.5) | 1.999 | 2.000 | |
| b2 (3.0‒3.9) | 1.995 | 2.000 |
After separability analysis, SVM classification is performed and LAI maps are drawn based on different Verticillium wilt severity. Figure 5 shows the LAI classification maps obtained from different SVM kernels based on combined band and vegetation index data.
Using independent validation samples for verification, the overall accuracy and Kappa coefficient were obtained (Table 12). The classification accuracy of different SVM kernels has the smallest variation, all reaching a Kappa coefficient of 0.9 or above. The performance of the four kernel functions ranks as: SVM-RBF > SVM sigmoid >SVM polynomial > SVM linear. The SVM-RBF classifier has the highest accuracy, with an overall classification accuracy of 94.05 % and a Kappa coefficient of 0.916, making it the best method for estimating cotton leaf area index under Verticillium wilt stress.
| Class method | Overall classification accuracy (%) | Kappa coefficient | Sequence |
|---|---|---|---|
| SVM-Linear classification | 93.04 | 0.902 | 4 |
| SVM-Polynomial classification | 93.26 | 0.902 | 3 |
| SVM-Sigmoid classification | 93.82 | 0.913 | 2 |
| SVM-Radial Basis Function classification | 94.05 | 0.916 | 1 |
Based on the combination of bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI), support vector machine classification method (linear kernel function, polynomial kernel function, Sigmoid kernel function, and radial basis kernel function) are used for image classification (Fig. 5). Spatial analysis shows that a substantial proportion of the field is affected by moderate to severe Verticillium wilt infection, corresponding to disease severity levels b2 and b3. In these areas, the LAI values of cotton plants were significantly lower (< 4.0). The spatial distribution map clearly shows the changes in LAI under different disease severity levels. The classification results from support vector machine (SVM) models exhibited marked variation depending on the kernel function used. The linear and polynomial kernels misclassified certain areas of b1-level as b0, and the sigmoid kernel confused some regions of b3-level into the b2 category. In contrast, the radial basis function (RBF) kernel generated a classification map most consistent with the actual field conditions. Together, these results confirm the feasibility of employing UAV-based multispectral imagery for mapping the spatial distribution of cotton Verticillium wilt.
This study uses field survey data of cotton under the stress of Verticillium wilt to conduct descriptive statistics and explore the changing trends during the disease period. The results are drawn as follows:
1) Based on UAV multispectral remote sensing, the optimal band for cotton LAI under Verticillium wilt stress was determined to be B12, and the optimal vegetation index was determined to be TVI.
2) A support vector machine regression model based on bands (B8, B9, B12) and vegetation indices (DVI, TVI, EVI) was constructed as the optimal model for monitoring cotton LAI under Verticillium wilt stress (model set R2 = 0.825, RMSE = 0.328); Validation set R2 = 0.877, RMSE = 0.284).
3) The support vector machine radial basis kernel function classification method is the best classification method for the main canopy parameters of cotton under Verticillium wilt stress (classification accuracy = 94.05 %, Kappa = 0.916).
This study only used the multispectral parameters of UAV to invert the cotton leaf area index in the study area, without incorporating with texture features in the analysis. In future work, a multi spectral texture feature collaborative inversion of cotton LAI under Verticillium wilt stress will be carried out, and deep neural network (DNN) and other algorithms will be introduced to compare with existing models.
The research of this thesis was supported by The Science and Technology Innovation Talents Program of Xinjiang Corps (2022CB003-05); National Natural Science Foundation of China (41961054); Intelligent cotton field innovation team of Xinjiang Academy of Agricultural Reclamation Sciences (NCG202304).
The authors declare no conflicts of interest.
The authors Zijie CHEN and Xiu WANG contributed equally.
(URLs on references were accessed on 29 January 2026.)
