Tobacco chemical quality index nondestructive monitoring method based on multi-source remote sensing feature fusion

Abstract

The invention discloses a tobacco chemical quality index nondestructive monitoring method based on multi-source remote sensing feature fusion. The tobacco chemical quality index nondestructive monitoring method comprises the following steps: S1, selecting two different tobacco varieties; s2, acquiring each piece of remote sensing data of each planting plot by using an unmanned aerial vehicle; s3, extracting multi-type remote sensing features, wherein the remote sensing features comprise a vegetation index, a color index, a texture feature and a three-dimensional structure feature; s4, calculating each quality index, wherein the quality index comprises a plurality of evaluation parameters; screening out each remote sensing feature sensitive to the tobacco chemical quality index based on correlation analysis and unary linear regression, fusing the screened remote sensing features through a machine learning algorithm, and constructing an optimal prediction model; and S5, drawing a spatial distribution diagram by using the optimal prediction model and a GIS (Geographic Information System). According to the method, the prediction precision of the tobacco chemical quality is remarkably improved through multi-source remote sensing fusion and a machine learning algorithm, and the generated spatial distribution map provides a decision basis for precise fertilization and intelligent management.

Description

Technical Field

[0001] This invention relates to the technical field of monitoring chemical quality indicators of tobacco, and specifically to a non-destructive monitoring method for chemical quality indicators of tobacco based on multi-source remote sensing feature fusion. Background Technology

[0002] Tobacco is an economic crop with a wide planting area. The chemical composition of tobacco leaves is a key intrinsic factor determining its quality. Common components mainly include nicotine, nitrogen, and potassium; excessively high or low levels of these components are detrimental to the overall quality of tobacco. The amount of nitrogen applied has a significant impact on the content and proportion of chemical quality components in flue-cured tobacco leaves. Current research indicates that within a certain range, as the nitrogen supply level increases, the content of total nitrogen, nicotine, and potassium in tobacco leaves rises accordingly. Appropriate nitrogen application helps improve tobacco yield and grade, but excessive nitrogen fertilizer application reduces tobacco quality. Therefore, achieving precise management and quality control of tobacco urgently requires the development of real-time, non-destructive intelligent monitoring technology for tobacco leaf chemical quality indicators.

[0003] Monitoring of nitrogen and other quality indicators in tobacco leaves during growth still relies primarily on traditional manual ground surveys, which are inefficient and labor-intensive. Remote sensing technology, based on electromagnetic spectrum detection of material properties, offers advantages over conventional ground surveys, including wider coverage, faster information acquisition, shorter cycles, stronger real-time performance, and no limitations imposed by ground conditions. It has been widely applied in monitoring the growth and predicting the quality of crops such as rice, wheat, and corn. Common spectral data acquisition platforms include satellites, aircraft, drones, and ground-based remote sensing platforms. Among these, drones are easy to operate, can fly autonomously along preset routes, and offer higher continuity and stability in data acquisition. Drone platforms can carry multispectral or RGB cameras and other sensors, and the acquired multi-source spectral features are widely used for monitoring crop nitrogen nutrition and quality indicators.

[0004] Existing research includes early identification of waterlogging in tobacco fields, detection of tobacco plant numbers, and yield estimation based on UAV spectral imagery. It also includes using UAV multispectral and hyperspectral data to extract reflectance and construct vegetation indices to predict nitrogen and nicotine content in tobacco leaves. Texture features are a comprehensive reflection of vegetation canopy structure and leaf color changes in UAV remote sensing imagery. During tobacco growth and development, differences in nitrogen absorption and consumption cause changes in leaf color and canopy horizontal structure, leading to differences in image texture features. Texture features can enhance the sensitivity of remote sensing data to crop biophysical characteristics, and their fusion with spectral reflectance data helps improve the accuracy of flue-cured tobacco classification. Furthermore, crop surface models can be constructed based on UAV remote sensing imagery to further extract structural parameters such as point cloud density and height for monitoring tobacco growth and quality components. Multi-source spectral features reflect important information such as tobacco plant canopy structure, leaf shape, leaf color, and texture.

[0005] However, current research on monitoring chemical quality indicators of tobacco leaves has not yet deeply analyzed and explored the four types of spectral features mentioned above (vegetation index, color index, texture features, and structural features), and also lacks effective methods for multi-source spectral feature fusion. Therefore, by comprehensively utilizing multi-source spectral information obtained from UAV remote sensing platforms, exploring combinations of spectral features that respond consistently to and have strong monitoring correlations with chemical quality indicators of tobacco leaves, it is hoped that the accuracy of tobacco leaf quality component monitoring models can be further improved.

[0006] Feature fusion algorithms are crucial for mining and integrating spectral features to construct monitoring models for chemical quality indicators of tobacco leaves. Currently, most studies employ simple regression and machine learning methods to analyze spectral features and build quality component monitoring models. While simple regression is easy to operate, its model accuracy and generalizability are usually limited. In contrast, machine learning algorithms can fuse multiple spectral feature parameters sensitive to quality components, deeply mining quality-related spectral characteristics. Therefore, spectral monitoring models built based on machine learning have higher accuracy and stronger generalization ability. The first step in building a high-precision machine learning model is feature selection. For example, spectral and spatial depth features of hyperspectral images are extracted using one-dimensional and two-dimensional convolutional neural networks. After fusing multi-source features using partial least squares, support vector machines, and random forest algorithms, the results show that the monitoring model that fuses spectral and spatial depth features outperforms single-information models. Similar studies have also found that combining UAV spectral indices and texture features as machine learning input variables can improve the accuracy of crop quality indicator monitoring models. In monitoring the content of tobacco quality components, different studies have used various machine learning modeling algorithms, resulting in varying model performance. For example, when monitoring tobacco starch content, the support vector machine regression algorithm outperformed partial least squares regression, multiple linear regression, and backpropagation artificial neural networks. However, in predicting tobacco tar content, the random forest algorithm demonstrated higher prediction accuracy and stability than support vector machines. Some studies have also shown that the random forest model performs better in predicting the chlorophyll content of flue-cured tobacco.

[0007] Currently, the chemical quality assessment of tobacco faces at least the following problems: 1) Monitoring mainly relies on traditional manual ground surveys, which are inefficient, labor-intensive, and cumbersome; 2) When combined with drones and various model algorithms, the monitoring accuracy of tobacco is insufficient, and it is difficult to comprehensively reflect the characteristic quality of tobacco. Summary of the Invention

[0008] To address the aforementioned problems, the purpose of this invention is to propose a non-destructive monitoring method for tobacco chemical quality indicators based on multi-source remote sensing feature fusion. By fusing multi-source information, the method significantly improves estimation accuracy, breaking through the limitations of single-spectral analysis. It innovatively integrates four types of remote sensing features: spectrum, color, texture, and three-dimensional structure, fully leveraging the complementary advantages of each feature. This multi-source data fusion strategy overcomes the limitations of single information, thereby significantly improving the accuracy and robustness of the tobacco chemical quality prediction model. Through systematic comparison of three machine learning algorithms—RF, LASSO, and PLSR—feature fusion is performed, and the correlation between each evaluation parameter and the spectrum is analyzed in detail, providing optimal algorithmic support for achieving efficient and accurate quality estimation. Furthermore, a spatial distribution map is generated, intuitively predicting field variability under different fertilization conditions. This provides direct decision-making basis for precise fertilization, differentiated field management, and quality traceability in tobacco production, powerfully promoting intelligent and refined management of tobacco planting.

[0009] This was achieved through the following technical solutions: A non-destructive monitoring method for chemical quality indicators of tobacco using multi-source remote sensing feature fusion, comprising the following steps: S1. Select two different tobacco varieties and cultivate them separately in each planting area; S2. Acquire each sample in each planting plot, measure the quality parameters of each sample, and record the measurement results; use a drone to acquire remote sensing data for each planting plot. S3. Extract the corresponding multi-type remote sensing features from each remote sensing data. The multi-type remote sensing features include vegetation index, color index, texture features and three-dimensional structure features. S4. Based on multi-type remote sensing features, an optimal prediction model is constructed using machine learning algorithms. The construction process includes: Each quality index was calculated using the random forest algorithm, LASSO regression algorithm, and partial least squares regression algorithm. The quality index includes multiple evaluation parameters, such as the coefficient of determination R for nicotine content, nitrogen content, and potassium content. 2 The root mean square error (RMSE) was calculated. Based on correlation analysis and univariate linear regression, each remote sensing feature sensitive to tobacco chemical quality indicators was selected. The selected remote sensing features were then fused using machine learning algorithms to construct the optimal prediction model. S5. Use the optimal prediction model to predict the tobacco content of each planting plot, and use GIS to draw spatial distribution maps of nicotine, nitrogen and potassium content under different nitrogen application treatments.

[0010] Preferably, in step S1, when cultivating any tobacco variety, the row spacing is set to 50 cm, and five different nitrogen application levels (N0, N1, N2, N3, and N4) are set. N0 serves as the control, and each of N1 to N4 has three replicate planting plots, with each plot measuring 5 meters × 5 meters. By setting multiple nitrogen application levels and replicating the experiment, the influence of nitrogen fertilizer on tobacco quality can be systematically revealed. The standardized row spacing and plot planning ensure the uniformity of experimental conditions and the comparability of data, laying the foundation for building a reliable model.

[0011] Preferably, in step S2, when acquiring each sample, three tobacco plants are selected from each planting plot, and quality parameters are measured according to each leaf position; the leaf positions include the lower, middle, and upper parts. Leaf position sampling and measurement can accurately reflect the quality differences at different canopy locations, enabling the model to monitor the stereochemical quality of tobacco more precisely. This method ensures that the ground-based measured data and UAV remote sensing data correspond accurately in space, which is a key prerequisite for establishing a high-precision estimation model.

[0012] Preferably, the process for determining quality parameters in step S2 is as follows: first, perform blanching at 105℃ for 30 minutes, then dry at 80℃ to constant weight, and finally pulverize and use... The digestion process was then performed, and the nitrogen content was determined using the Kjeldahl method, the phosphorus content using the vanadium molybdenum yellow colorimetric method, the potassium content using the flame photometry method, and the nicotine content using the extraction decolorization method.

[0013] Preferably, in step S2, the UAV is equipped with a multi-channel multispectral camera to acquire remote sensing data for each planting plot. Each remote sensing data includes RGB imagery, multispectral imagery, and 3D point cloud data of the tobacco canopy. Simultaneous acquisition of RGB, multispectral, and 3D point cloud data enables comprehensive capture of information on the spectral, texture, and three-dimensional structure of the tobacco canopy. The acquisition of multi-source data provides a rich data foundation for subsequent extraction of diverse features and fusion modeling to comprehensively improve monitoring accuracy.

[0014] Preferably, when extracting each remote sensing feature in step S3, each remote sensing data is first imported into image processing software for stitching and radiometric correction to generate each corresponding orthophoto; vegetation index extraction, color index extraction, texture feature extraction and three-dimensional structure feature extraction are performed on each orthophoto.

[0015] Preferably, when extracting vegetation indices, reflectance of each band is extracted from each sampling point in each orthophoto corresponding to each multispectral image to construct the spectral vegetation index; when extracting color indices, the mean gray values ​​of the R, G, and B channels are extracted from each orthophoto corresponding to each tobacco canopy RGBV influence, and then normalized to obtain the corresponding r, g, and b; when extracting texture features, a gray-level co-occurrence matrix is ​​used with an analysis window of 3×3 pixels to calculate the texture features of each band in each multispectral image. Each band includes blue, green, red, red edge, and near-infrared. Texture features include mean (MEA), variance (VAR), homogeneity (HOM), contrast (CON), dissimilarity (DIS), entropy (ENT), second moment (SEM), and correlation parameter (COR); when extracting three-dimensional structural features, a dense point cloud data is generated using a motion recovery structure algorithm to construct a tobacco surface model and extract multiple structural parameters from the sampling point cloud corresponding to each sampling point. By employing multiple indices and texture features and extracting them from multiple bands, it can comprehensively capture canopy information related to tobacco quality from different dimensions. Using a 3x3 small window to calculate texture can better capture subtle changes in canopy images, which helps to detect early physiological state differences.

[0016] Preferably, the multiple structural parameters include maximum height, average height, height quantiles (H99 to H80, 1% intervals), canopy coverage, projected leaf area, plant area index, three-dimensional profile index, and volume; wherein, the height quantiles start from H99 and are spaced 1% to H80 at intervals of 1%. Extracting 27 structural parameters covering multiple dimensions such as height, density, and volume allows for a quantitative description of the complex three-dimensional spatial structure of the tobacco canopy. These detailed structural features are closely related to crop growth and physiological activities, providing crucial three-dimensional information support for quality prediction models.

[0017] Preferably, in step S4, when calculating each quality index, the random forest algorithm, LASSO regression algorithm, and partial least squares regression algorithm are used to calculate the evaluation parameters between vegetation index, color index, texture features, and three-dimensional structure features and nicotine content, nitrogen content, and potassium content, respectively; among the evaluation parameters, , P i and O i These represent the predicted and observed values ​​of the sample, respectively. R represents the arithmetic mean of the sample observations, where n represents the sample size. 2 The coefficient of determination is RMSE, which is the root mean square error RMSE.

[0018] Preferably, when fusing each selected remote sensing feature using a machine learning algorithm, each merged feature is calculated based on the lower, middle, and upper parts of the leaf position, and the results of each merged feature are verified by ten-fold cross-validation to select the optimal prediction model.

[0019] The beneficial effects of this invention compared to the prior art are: The technical solution of this invention significantly improves estimation accuracy through multi-source information fusion, breaking through the limitations of single spectral analysis. It innovatively integrates four types of remote sensing features: spectrum, color, texture, and three-dimensional structure, fully leveraging the complementary advantages of each feature. This multi-source data fusion strategy overcomes the limitations of single information, thereby significantly improving the accuracy and robustness of the tobacco chemical quality prediction model. By systematically comparing three machine learning algorithms—RF, LASSO, and PLSR—feature fusion and detailed analysis of the correlation between each evaluation parameter and the spectrum are performed, providing optimal algorithmic support for achieving efficient and accurate quality estimation. Furthermore, it generates spatial distribution maps, intuitively predicting field variability under different fertilization conditions. This provides direct decision-making basis for precision fertilization, differentiated field management, and quality traceability in tobacco production, powerfully promoting the intelligent and refined management of tobacco planting. Attached Figure Description

[0020] Figure 1 A flowchart of a non-destructive monitoring method for tobacco chemical quality indicators by fusing multi-source remote sensing features; Figure 2 A schematic diagram of a drone and each tobacco planting plot; Figure 3 This is a schematic diagram illustrating the correlation between spectral characteristics and quality indicators. Figure 4 Spatial distribution maps of nicotine, nitrogen, and potassium content under different nitrogen application treatments, generated using a GIS. Detailed Implementation

[0021] The following will be based on embodiments of the present invention. Figures 1 to 4 The technical solutions in the embodiments of the present invention will be described in detail below.

[0022] like Figure 1 The diagram shows a flowchart of a non-destructive monitoring method for tobacco chemical quality indicators based on multi-source remote sensing feature fusion. Using multi-source remote sensing image data from unmanned aerial vehicles (UAVs), spectral, color, texture, and three-dimensional structural features are fused to construct estimation models for key chemical components in tobacco, such as nicotine, nitrogen, and potassium. Multi-source feature fusion significantly improves model accuracy; the predictions for nicotine, nitrogen, and potassium are more accurate after the fusion of the four types of features. Furthermore, based on the fused estimation model, a spatial distribution map related to quality can be predicted and drawn, clearly reflecting the spatial variation characteristics of tobacco quality under different nitrogen application levels.

[0023] The method specifically includes the following steps: S1. Select two different tobacco varieties and cultivate them separately in each planting plot. For example, select two varieties of tobacco, "Yunyan 87" (V1) and "ZY157" (V2), and set the row spacing and plant spacing at 50 cm. At the same time, set N0 (0 kg·mu) -1 N1 (14kg·mu) -1 N2 (17kg·mu) -1 N3 (21kg·mu) -1 N4 (24kg·mu) -1 Five different nitrogen application levels were used, with N0 serving as the control. Each of N1 to N4 had three replicate planting plots, with each plot measuring 5 meters × 5 meters. By setting up multiple nitrogen application levels and replicate experiments, the influence of nitrogen fertilizer on tobacco quality could be systematically revealed. Standardized row and plant spacing and plot planning ensured the uniformity of experimental conditions and the comparability of data, laying the foundation for building a reliable model.

[0024] S2. Obtain each sample in each planting plot (e.g., select 3 tobacco plants in each planting plot), measure the quality parameters of each sample, and record the measurement results; use a drone to obtain remote sensing data for each planting plot.

[0025] When acquiring each sample, three tobacco plants were selected from each planting plot, and quality parameters were measured according to each leaf position; leaf positions included the lower (1-5 leaves), middle (6-13 leaves), and upper (14-20 leaves). Sampling and measurement by leaf position accurately reflects the quality differences at different canopy locations, enabling the model to monitor the stereochemical quality of tobacco more precisely. This method ensures a precise spatial correspondence between ground-based measured data and UAV remote sensing data, which is a key prerequisite for establishing a high-precision estimation model.

[0026] Then, quality parameters were measured: first, the plants were blanched at 105℃ for 30 minutes, then dried at 80℃ to constant weight, and finally pulverized. The digestion process was then performed, and the nitrogen content was determined using the Kjeldahl method, the phosphorus content using the vanadium molybdenum yellow colorimetric method, the potassium content using the flame photometry method, and the nicotine content using the extraction decolorization method.

[0027] like Figure 2The diagram illustrates a drone and a tobacco planting plot. The drones used can be a DJI Phantom 4 Multispectral or a Mavic 3T, used to acquire remote sensing data for each planting plot. All remote sensing data includes RGB imagery, multispectral imagery, and 3D point cloud data of the tobacco canopy. Simultaneous acquisition of RGB, multispectral, and 3D point cloud data enables comprehensive capture of the tobacco canopy's spectral, texture, and three-dimensional structure. The acquisition of multi-source data provides a rich data foundation for subsequent extraction of diverse features and fusion modeling, thereby comprehensively improving monitoring accuracy.

[0028] It should be noted that the DJI Phantom 4 multispectral drone is equipped with a 5-channel multispectral camera, capable of collecting reflectance data in the blue (450±16nm), green (560±16nm), red (650±16nm), red-edge (730±16nm), and near-infrared (840±26nm) bands. It has a 62.7° field of view, a 2-megapixel sensor, and an image resolution of 1600×1300 pixels. The Mavic 3T weighs 920g and features a visible light camera with dual wide-angle and telephoto lenses. The wide-angle sensor is a 1 / 2-inch CMOS with 48 megapixels, and the telephoto sensor is a 1 / 2-inch CMOS with 12 megapixels. The maximum image size is 8000×6000 pixels. Data collection was conducted between 11:00 AM and 2:00 PM in clear weather with low wind speeds. A standard reflectance chart was photographed before each flight for radiometric correction. The flight altitude is uniformly set at 30 m, the overlap between the heading and the lateral direction is 80%, and the flight path remains consistent.

[0029] S3. Extract corresponding multi-type remote sensing features from each remote sensing data point. These features include vegetation indices, color indices, texture features, and 3D structural features. To extract each feature, first import each remote sensing data point into image processing software (e.g., Pix4Dmapper AG software) for mosaicking and radiometric correction to generate corresponding orthophotos. Then, extract vegetation indices, color indices, texture features, and 3D structural features from each orthophoto.

[0030] When extracting vegetation indices, reflectance of each band is extracted from each sampling point in each orthophoto corresponding to each multispectral image to construct the spectral vegetation index. When extracting color indices, the mean grayscale values ​​of the R, G, and B channels are extracted from each orthophoto corresponding to each tobacco canopy RGBV influence, and then normalized to obtain the corresponding r, g, and b values. When extracting texture features, a gray-level co-occurrence matrix (GLCM) with a 3×3 pixel analysis window (to capture subtle image changes) is used to calculate the texture features of each band in each multispectral image. Each band includes blue, green, red, red edge, and near-infrared. Texture features include mean (MEA), variance (VAR), homogeneity (HOM), contrast (CON), dissimilarity (DIS), entropy (ENT), second moment (SEM), and correlation parameter (COR). When extracting 3D structural features, a dense point cloud data is generated using the structure-in-motion (SIM) algorithm to construct a tobacco surface model and extract multiple structural parameters from the sampling point cloud corresponding to each sampling point. Employing multiple indices and texture features extracted from multiple bands, this method comprehensively captures canopy information related to tobacco quality from different dimensions. Using a 3x3 small window to calculate texture allows for better capture of subtle changes in canopy images, aiding in the detection of early physiological state differences. Vegetation indices and color indices can be obtained using Table 1 below. Table 1:

[0031] Wherein, NIR represents near-infrared reflectance, and λ is set as the reflectance of blue, green, red, and red edges, respectively, to calculate different types of vegetation indices; R, G, and B represent the red, green, and blue DN values ​​of the RGB image, respectively. The specific meanings of the index column and corresponding formulas in the table above are well-known and will not be elaborated here.

[0032] Multiple structural parameters were extracted, including maximum height, average height, height quantiles (H99 to H80, 1% intervals), canopy coverage, projected leaf area, plant area index, 3D profile index, and volume. Height quantiles were defined starting from H99 and increasing to H80 at 1% intervals. Twenty-seven structural parameters covering multiple dimensions such as height, density, and volume were extracted, quantifying the complex three-dimensional spatial structure of the tobacco canopy. These detailed structural features are closely related to crop growth and physiological activities, providing crucial three-dimensional information support for quality prediction models.

[0033] S4. Based on multi-type remote sensing features, an optimal prediction model is constructed using machine learning algorithms. The construction process includes: calculating each corresponding quality index using random forest, LASSO regression, and partial least squares regression algorithms respectively. The quality index includes multiple evaluation parameters, including the coefficient of determination R for nicotine content, nitrogen content, and potassium content. 2The root mean square error (RMSE) was calculated. Based on correlation analysis and univariate linear regression, the optimal algorithm for each leaf position in each sample was determined. Remote sensing features sensitive to tobacco chemical quality indicators were screened out, and each screened remote sensing feature was fused through machine learning algorithms to construct the optimal prediction model.

[0034] It's important to note that Random Forest (RF) is an ensemble learning algorithm that improves accuracy and robustness by constructing multiple decision trees and integrating their predictions. It's suitable for high-dimensional data and nonlinear regression. Its main parameters include the number of trees (Ntree) and the number of candidate features (Mtry) at each tree split. The modeling steps include: data preparation, bootstrap sampling, tree construction (based on random feature subset splitting), prediction ensemble, and model evaluation. The Random Forest model uses an ensemble of multiple regression trees for modeling. The core parameter tuning parameters include: the number of trees (Ntree) (which usually has the greatest impact on model stability) and the number of candidate features randomly selected (Mtry) at each node split. To obtain optimal model performance, Ntree is set to perform a grid search with a step size of 100 within the range of 300–1500, and the optimal Ntree is ultimately determined to be 1000. Candidate values ​​for Mtry are selected within the range of p and p / 3 (where p is the number of features), and cross-validation is used to evaluate the best parameter combination. Other parameters, such as the minimum number of samples per node (min_samples_split=2–10) and the minimum number of samples per leaf node (min_samples_leaf=1–5), are also adjusted within a controllable range to prevent overfitting.

[0035] The LASSO regression algorithm employs L1 regularization for variable selection and parameter estimation, making it suitable for high-dimensional modeling. Its key parameter is the regularization coefficient λ, which controls model complexity and sparsity. The modeling process includes data standardization, λ-value cross-validation selection, model training, and evaluation. The LASSO regression model achieves feature sparsity through L1 regularization, and the core parameter for parameter tuning in this study is the regularization coefficient λ. To ensure optimal regularization strength, this study uses 10-fold cross-validation (CV) to search for λ on a logarithmic scale, with a candidate range set to 10⁻⁻⁶. 4 –10² (100 equally spaced logarithmic points in total). During the parameter tuning process, λ_min (λ corresponding to the minimum CV error) and λ_1se (the simplest model within 1 standard deviation of error) were evaluated simultaneously, and finally, a λ value that balances high accuracy and high sparsity was selected. In addition, the data were Z-score standardized before modeling to avoid the penalty term being biased by the difference in the dimensions of the variables.

[0036] Partial Least Squares Regression (PLSR) achieves dimensionality reduction by extracting latent variables that maximize the covariance between independent and dependent variables. It is suitable for multicollinear data, and its key parameter is the number of latent variables. The modeling steps include: data standardization, latent variable extraction, regression construction, and cross-validation evaluation. PLSR achieves dimensionality reduction by extracting latent variables (LVs) that have the largest covariance between independent and dependent variables. The number of latent variables is used as the main parameter for tuning, and the optimal number of LVs is searched using 10-fold cross-validation, with a candidate range of 1–20. Before modeling, the PLSR model performs Z-score standardization on all variables and selects the number of latent variables that minimizes the RMSECV through cross-validation to avoid overfitting caused by excessive latent variable extraction.

[0037] The performance of each algorithm is evaluated using the coefficient of determination (R², Equation 1) and the root mean square error (RMSE, Equation 2). The higher the R² and the lower the RMSE, the higher the prediction accuracy and the stronger the robustness.

[0038] When calculating each quality index, the random forest algorithm, LASSO regression algorithm, and partial least squares regression algorithm were used to calculate the evaluation parameters between vegetation index, color index, texture features, and 3D structure features and nicotine content, nitrogen content, and potassium content, respectively; among the evaluation parameters, , P i and O i These represent the predicted and observed values ​​of the sample, respectively. This represents the arithmetic mean of the sample observations, where n represents the sample size.

[0039] like Figure 3 The diagram illustrates the correlation between spectral features and quality indicators, showing the quantitative correlation between quality indicators and vegetation indices (3a), color indices (3b), texture features (3c), and three-dimensional features (3d). Here, YJS, YJZ, YJX, NS, NZ, NX, KS, KZ, and KX represent the nicotine content of upper leaves, middle leaves, lower leaves, nitrogen content of upper leaves, nitrogen content of middle leaves, nitrogen content of lower leaves, potassium content of upper leaves, potassium content of middle leaves, and potassium content of lower leaves, respectively. By analyzing the quantitative correlation between spectral features (i.e., multiple types of remote sensing features) and quality indicators, the spectral features most sensitive to each quality parameter (i.e., with the highest correlation) can be effectively screened. Combined with... Figure 3As shown, among the vegetation indices, except for lower leaf nicotine which showed the highest correlation with BWDRVI, all other leaf quality indices showed the highest correlation with REOSAVI, with correlation coefficients ranging from 0.54 to 0.86. Among the color indices, VEG, RGBVI, INT, WI, RGBVI, ExG, RGBVI, RGBVI, and CIVE showed the highest correlations with each quality index. Regarding texture features, upper leaf nicotine showed the highest correlation with red light Data Range texture, middle leaf nicotine showed the highest correlation with green light Data Range texture, and lower leaf nicotine showed the highest correlation with blue light Mean texture; for nitrogen and potassium content, the Data Range and Variance texture features of green light images were the most sensitive. In the three-dimensional structural features, H80 showed the highest correlation with the three quality indices of upper and middle leaves, while the ps feature showed the highest correlation with the three quality indices of lower leaves, where ps represents point resolution.

[0040] Then, based on Figure 3 The results of the correlation analysis were validated using 10-fold cross-validation with univariate linear regression (see Table 2 below), which analyzed the differences in quality indicators of the three leaf positions of tobacco under different remote sensing features. As shown in Table 2, for nicotine in the middle and upper leaves, the color index VEG (R² = 0.71) and RGBVI (R² = 0.67) showed the best monitoring effect; for the lower leaves, the blue light Mean texture feature (R² = 0.41) showed the best monitoring effect. Regarding nitrogen content in the three leaf positions, the vegetation index REOSAVI (R² = 0.27–0.73) showed the best monitoring effect. For potassium content in the middle and upper leaves, the green light Data Range texture feature (R² = 0.55–0.62) showed the best monitoring effect, while for potassium content in the lower leaves, TG4 (R² = 0.37) showed the best monitoring effect.

[0041] Table 2:

[0042] Based on the results of 10-fold cross-validation, the LASSO algorithm showed the best performance across different feature fusions, with the highest R², while the PLSR algorithm performed relatively poorly. Among the four feature combinations, texture feature combination showed the best monitoring effect on nicotine (R² = 0.67–0.87), followed by vegetation index (R² = 0.62–0.74), while the three-dimensional structural feature combination performed slightly worse (R² = 0.32–0.44).

[0043] Therefore, the four types of features can be fused into a model, which further improves the monitoring effect, with R² reaching 0.68–0.89 and RMSE ranging from 0.96–2.68. The results of the 10-fold cross-validation after the fusion of the four types of features are shown in Table 3 below: Table 3:

[0044] The monitoring effects varied across different leaf positions, with the upper and middle leaves showing significantly better monitoring results than the lower leaves. For monitoring nitrogen content in the upper leaves, the vegetation index, color index, and 3D feature models constructed using the RF algorithm performed better. Based on the table above, the texture feature model and fusion feature model constructed using the LASSO algorithm showed the best performance. For nitrogen content in the middle leaves, the model constructed by the RF algorithm fusing various features had the highest R²; while for nitrogen content monitoring in the lower leaves, the model constructed by the LASSO algorithm fusing various features had the highest R². In the upper and middle leaves, the combination of vegetation index (R² = 0.85–0.87) and texture features (R² = 0.84–0.89) showed good monitoring results for nitrogen content; after fusing the four types of features into a model, the monitoring effect was further improved, with R² reaching 0.87–0.89 and RMSE 3.42–3.66. For the lower leaves, the vegetation index showed relatively better monitoring results (R² = 0.49). Overall, the monitoring results for the upper and middle leaves were significantly better than those for the lower leaves.

[0045] In monitoring potassium content in the upper leaves, the models constructed using the RF algorithm performed best across various models. For potassium content in the middle leaves, the vegetation index and color index models constructed using the RF algorithm performed better, while the texture feature, 3D feature, and fusion feature models constructed using the LASSO algorithm showed the best performance. In the lower leaves, the models constructed using the RF and LASSO algorithms performed well. Consistent with the monitoring results for nicotine and nitrogen content, the four-feature fusion model showed the best performance, with R² of 0.62–0.89 and RMSE of 3.06–6.20; the monitoring results for the upper and middle leaves were comparable, while the monitoring results for the lower leaves remained poor.

[0046] S5. Use the optimal prediction model to predict the tobacco content of each planting plot, and use GIS to draw spatial distribution maps of nicotine, nitrogen and potassium content under different nitrogen application treatments.

[0047] like Figure 4 The image shows a spatial distribution map of nicotine, nitrogen, and potassium contents under different nitrogen application treatments, generated using GIS. Based on the optimal prediction model, the field tobacco quality can be accurately predicted. The GIS digital plot mapping function was used to generate the spatial distribution maps of nicotine, nitrogen, and potassium contents under different nitrogen application treatments. The validation results show significant spatial variability in field quality under different nitrogen application treatments. The no-nitrogen treatment (N0) had the lowest nicotine, nitrogen, and potassium contents, the N1 treatment showed a slight increase, while the N3–N4 treatments had higher contents of each quality indicator. Therefore, the spatial distribution map of quality indicators can visually demonstrate the spatial differentiation characteristics of tobacco quality under different growth conditions.

[0048] In summary, this invention significantly improves estimation accuracy through multi-source information fusion, breaking through the limitations of single-spectral analysis. It innovatively integrates four types of remote sensing features: spectrum, color, texture, and three-dimensional structure, fully leveraging the complementary advantages of each feature. This multi-source data fusion strategy overcomes the limitations of single information, thus significantly improving the accuracy and robustness of the tobacco chemical quality prediction model. By systematically comparing three machine learning algorithms—RF, LASSO, and PLSR—and performing feature fusion and detailed analysis of the correlation between each evaluation parameter and the spectrum, it provides optimal algorithmic support for achieving efficient and accurate quality estimation. Furthermore, it generates spatial distribution maps, intuitively predicting field variability under different fertilization conditions. This provides direct decision-making basis for precision fertilization, differentiated field management, and quality traceability in tobacco production, powerfully promoting intelligent and refined management of tobacco planting, and demonstrating significant progress.

[0049] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A non-destructive monitoring method for tobacco chemical quality indicators based on multi-source remote sensing feature fusion, characterized in that, Includes the following steps: S1. Select two different tobacco varieties and cultivate them separately in each planting area; S2. Acquire each sample in each planting plot, measure the quality parameters of each sample, and record the measurement results; use a drone to acquire remote sensing data for each planting plot. S3. Extract the corresponding multi-type remote sensing features from each remote sensing data. The remote sensing features include vegetation index, color index, texture features and three-dimensional structure features. S4. Based on multi-type remote sensing features, an optimal prediction model is constructed using machine learning algorithms. The construction process includes: Each quality index was calculated using the random forest algorithm, LASSO regression algorithm, and partial least squares regression algorithm. The quality index includes multiple evaluation parameters, such as the coefficient of determination R for nicotine content, nitrogen content, and potassium content. 2 The root mean square error (RMSE) was calculated. Based on correlation analysis and univariate linear regression, each remote sensing feature sensitive to tobacco chemical quality indicators was selected. The selected remote sensing features were then fused using machine learning algorithms to construct the optimal prediction model. S5. Use the optimal prediction model to predict the tobacco content of each planting plot, and use GIS to draw spatial distribution maps of nicotine, nitrogen and potassium content under different nitrogen application treatments.

2. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 1, characterized in that, In step S1, when cultivating any tobacco variety, the row spacing is set to 50 cm, and five different nitrogen application levels are set: N0, N1, N2, N3, and N4. N0 is used as a control, and planting plots with three replicates are set for each of N1 to N4. The size of each planting plot is 5 meters × 5 meters.

3. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 1, characterized in that, In step S2, when obtaining each sample, three tobacco plants are selected for each planting plot and quality parameters are measured according to each leaf position; the leaf position includes the lower, middle and upper parts.

4. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 1, characterized in that, The procedure for determining quality parameters in step S2 is as follows: first, perform blanching at 105℃ for 30 minutes, then dry at 80℃ to constant weight, and finally pulverize and use... The digestion process was then performed, and the nitrogen content was determined using the Kjeldahl method, the phosphorus content using the vanadium molybdenum yellow colorimetric method, the potassium content using the flame photometry method, and the nicotine content using the extraction decolorization method.

5. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 1, characterized in that, In step S2, the drone is equipped with a multi-channel multispectral camera to acquire remote sensing data for each planting plot. Each remote sensing data includes tobacco canopy RGB image, multispectral image, and three-dimensional point cloud data.

6. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 5, characterized in that, When extracting each remote sensing feature in step S3, each remote sensing data is first imported into image processing software for stitching and radiometric correction to generate each corresponding orthophoto; vegetation index extraction, color index extraction, texture feature extraction and three-dimensional structure feature extraction are performed on each orthophoto.

7. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 6, characterized in that, When extracting vegetation indices, reflectance of each band is extracted from each sampling point in each orthophoto corresponding to each multispectral image to construct the spectral vegetation index. When extracting color indices, the mean gray values ​​of the R, G, and B channels are extracted from each orthophoto corresponding to each tobacco canopy RGBV influence, and then normalized to obtain the corresponding r, g, and b values. When extracting texture features, a gray-level co-occurrence matrix is ​​used with an analysis window of 3×3 pixels to calculate the texture features of each band in each multispectral image. Each band includes blue, green, red, red edge, and near-infrared. Texture features include mean (MEA), variance (VAR), homogeneity (HOM), contrast (CON), dissimilarity (DIS), entropy (ENT), second moment (SEM), and correlation parameter (COR). When extracting three-dimensional structural features, a dense point cloud data is generated using the structure-in-motion algorithm to construct a tobacco surface model and extract multiple structural parameters from the sampling point cloud corresponding to each sampling point.

8. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 7, characterized in that, Multiple structural parameters include maximum height, average height, height quantiles (H99 to H80, at 1% intervals), canopy coverage, projected leaf area, plant area index, three-dimensional profile index, and volume; among which, the height quantiles start from H99 and are at 1% intervals to H80.

9. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 8, characterized in that, In step S4, when calculating each quality index, the random forest algorithm, LASSO regression algorithm, and partial least squares regression algorithm are used to calculate the evaluation parameters between vegetation index, color index, texture features, and three-dimensional structure features and nicotine content, nitrogen content, and potassium content, respectively; among the evaluation parameters, , P i and O i These represent the predicted and observed values ​​of the sample, respectively. R represents the arithmetic mean of the sample observations, where n represents the sample size. 2 The coefficient of determination is RMSE, which is the root mean square error RMSE.

10. The method for non-destructive monitoring of tobacco chemical quality indicators by multi-source remote sensing feature fusion according to claim 9, characterized in that, When fusing each selected remote sensing feature using machine learning algorithms, each merged feature is calculated based on the lower, middle, and upper parts of the leaf position. The results of each merged feature are verified using ten-fold cross-validation to select the optimal prediction model.

Images