Tobacco disease and pest monitoring method based on double-time phase remote sensing features

Abstract

The application discloses a tobacco disease and pest monitoring method based on double-time-phase remote sensing features, comprising the following steps: S1, acquiring double-time-phase satellite remote sensing image data of a key growth period of tobacco; S2, extracting vegetation indexes, wavelet features and texture features from the double-time-phase satellite remote sensing image data respectively to form a single-time-phase feature set; S3, calculating double-time-phase features based on the single-time-phase feature set, wherein the double-time-phase features comprise feature difference features, relative change features and ratio features, and a high-dimensional feature set fusing the single-time-phase feature set and the double-time-phase features is constructed; and S4, performing feature screening on the high-dimensional feature set based on a recursive feature elimination algorithm to obtain a sensitive feature subset; the application has the beneficial effects that: by introducing the double-time-phase feature set, the application quantifies the change trend of the features in the key growth period, directly converts the dynamic development information of the diseases into identifiable model inputs, and enhances the distinguishing ability of the model to early diseases and different disease types and the early warning potential.

Description

Technical Field

[0001] This invention relates to the field of tobacco pest and disease monitoring technology, specifically a method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features. Background Technology

[0002] Tobacco pests and diseases are key factors restricting the improvement of tobacco yield and quality, making large-scale, high-efficiency, and precise monitoring an urgent need for modern agricultural management. Traditional methods relying on manual field surveys have inherent drawbacks such as low efficiency, strong subjectivity, and limited coverage, making it difficult to meet the timeliness and objectivity requirements of modern precision agriculture. Against this backdrop, satellite remote sensing technology, with its advantages of macroscopic, objective, and periodic observation, has become an indispensable technical means for monitoring the health status of crops over large areas.

[0003] Currently, remote sensing-based crop monitoring methods have been widely used. Their core lies in using image data to extract various features to characterize crop growth status. These commonly used methods mainly rely on image data from a single time phase, and the extracted features can be summarized into three categories: First, vegetation indices, such as NDVI and EVI, which enhance vegetation information by combining reflectance in different bands, but lack specificity for early disease responses. For example, Yuan, L. et al. used vegetation indices such as NDVI and TVI from SPOT-6 satellite imagery to construct a disease mapping method for winter wheat powdery mildew. Second, texture features, such as correlation and contrast based on the gray-level co-occurrence matrix (GLCM), are used to describe the heterogeneity of canopy spatial structure caused by diseases, but have low sensitivity in the early stages of diseases. For example, Shafi U et al. achieved early monitoring of wheat powdery mildew based on texture features. Third, wavelet features, which can capture subtle local deformations of spectral curves by performing wavelet transforms in specific bands and are more sensitive to early stresses, but have high feature dimensionality and are easily affected by noise. For example, X Cheng et al. verified the feasibility of combining continuous wavelet transform features with machine learning for rubber tree powdery mildew detection.

[0004] However, current research on remote sensing monitoring of tobacco diseases and pests mainly revolves around single-temporal images, with relatively limited exploration at the satellite scale. Most studies focus on the leaf and canopy scales. For example, at the leaf scale, Lin J et al. proposed a multi-feature fusion network based on collaborative attention (CAMFFNet) for tobacco disease identification under field conditions, achieving rapid analysis of canopy images. At the canopy scale, this paper proposes a lightweight disease detection algorithm based on YOLOv10. At the regional scale, relevant research is even more lacking. Chiringa et al. used Sentinel-2 and Landsat satellite imagery to calculate NDVI and conducted a preliminary assessment of tobacco health status.

[0005] Current research on tobacco pest and disease monitoring methods based on single-temporal images or single-type features has significant limitations. First, single-temporal features only reflect the physiological and spectral state of crops at a specific point in time, failing to depict the dynamic process of changes in spectral and textural characteristics under pest and disease stress over time. Furthermore, the occurrence and development of diseases exhibit a clear temporal sequence, with spectral and textural features showing dynamic evolutionary patterns at different stages; single-temporal analysis struggles to effectively capture this crucial change information. Second, single-type features (such as using only spectral or textural features) express limited information dimensions, making it difficult to comprehensively reflect the integrated characteristics of diseases in spectral response and spatial structure, thus limiting identification accuracy. In addition, research on tobacco pest and disease monitoring at the satellite scale is still relatively limited; existing methods are mostly concentrated at the leaf or canopy scale, lacking adaptability and scalability for large-scale remote sensing monitoring. Summary of the Invention

[0006] The purpose of this invention is to provide a method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features, comprising the following steps:

[0008] S1. Acquire dual-temporal satellite remote sensing image data of tobacco during its critical growth period;

[0009] S2. Extract vegetation index, wavelet features and texture features from dual-temporal satellite remote sensing image data to form a single-temporal feature set;

[0010] S3. Based on the single-phase feature set, calculate the dual-phase features, which include feature difference features, relative change features, and ratio features, and construct a high-dimensional feature set that integrates the single-phase feature set and the dual-phase features;

[0011] S4. Based on the recursive feature elimination algorithm, the high-dimensional feature set is filtered to obtain a sensitive feature subset;

[0012] S5. Train a random forest classification model using a subset of sensitive features, and optimize the hyperparameters of the random forest classification model through grid search to obtain an optimized tobacco pest and disease monitoring model.

[0013] S6. Use the optimized tobacco pest and disease monitoring model to identify and classify pests and diseases in the tobacco area to be monitored.

[0014] Preferably, in step S2, the vegetation index includes one or more of the following: plant carotenoid reflectance index, normalized vegetation index, ratio vegetation index, optimized soil-adjusted vegetation index, improved soil-adjusted vegetation index, improved triangular vegetation index 2, simple carotenoid reflectance index, photosynthetic capacity sensitive ratio index A, photosynthetic capacity sensitive ratio index B, enhanced vegetation index, and modified simple ratio.

[0015] Preferably, in step S2, wavelet features are extracted by continuous wavelet transform, which uses Gaussian wavelet or Morlet wavelet as the mother wavelet and calculates wavelet coefficients at scale 1 or 2 at the 940nm band.

[0016] Preferably, in step S2, the texture features are extracted using a gray-level co-occurrence matrix. The texture features are correlation features. The extraction window size is 3×3, the gray-level quantization level is 64, the step size is 1, and one or more texture features are extracted from the wavelengths of 820nm, 836nm, 866nm, 880nm, and 896nm.

[0017] Preferably, in step S3, the feature difference feature is the difference between the same feature in two time phases; the relative change feature is the ratio of the feature change between two time phases to the feature value in the initial time phase; and the ratio feature is the ratio between the features in the two time phases.

[0018] Preferably, in step S4, the recursive feature elimination algorithm uses a support vector machine as the base learner, and through iterative training and feature importance evaluation, gradually eliminates redundant features and selects the feature subset that is most sensitive to the identification of tobacco pests and diseases.

[0019] Preferably, in step S5, the hyperparameters include at least one of the following: number of decision trees, maximum depth, minimum number of leaf node samples, minimum number of split samples, and maximum number of features.

[0020] Preferably, in step S5, the grid search uses cross-validation to optimize the hyperparameters and selects the optimal hyperparameter combination based on preset performance evaluation indicators.

[0021] Preferably, in step S6, the pests and diseases include mosaic virus and bollworm, and the classification results include healthy plants, plants infected with mosaic virus, and plants infected with bollworm.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1. The core advantage of this invention lies in its systematic solution to the bottlenecks of traditional single-phase monitoring methods, which suffer from limited information dimensions and model performance limitations, through multi-temporal dynamic feature construction and automated feature optimization-model tuning processes. Existing methods often rely on vegetation indices, textures, or wavelet features from a single phase, making it difficult to capture the temporal evolution of crop spectra and structures under disease stress. Furthermore, high-dimensional feature inputs can easily lead to model redundancy and overfitting. This invention innovatively introduces a dual-temporal feature set, quantifying the changing trends of features during critical growth periods. This directly transforms the dynamic development information of diseases into identifiable model inputs, thereby fundamentally enhancing the model's ability to distinguish early diseases and different disease types, and its early warning potential.

[0024] 2. This invention constructs a complete optimization chain from features to model. Through recursive feature elimination (RFE), the most sensitive subset of core features to diseases is automatically selected, effectively eliminating information redundancy and reducing data dimensionality and computational overhead. Based on this, a grid search-optimized random forest model is combined to achieve automatic optimization of hyperparameter combinations, ensuring the model always operates at its best performance. This technical solution not only significantly improves the accuracy and robustness of disease identification but also reduces human intervention and reliance on professional knowledge through automated processes. This results in a lower-cost, more efficient, and easier-to-promote solution for large-scale field monitoring, providing reliable technical support for precise control of tobacco diseases at the regional scale. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention;

[0026] Figure 2 This is an overall flowchart of the tobacco pest and disease monitoring method based on dual-temporal remote sensing features of the present invention;

[0027] Figure 3 This is a schematic diagram illustrating the random forest classification principle of this invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Please see Figure 1-3 As shown, the present invention provides a technical solution: a method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features, comprising the following steps:

[0030] S1. Acquire dual-temporal satellite remote sensing image data of tobacco during its critical growth period.

[0031] It should be noted that in this embodiment, high-resolution multispectral satellite remote sensing images (e.g., WorldView-3, Sentinel-2, etc.) covering the target tobacco planting area and occurring during two phases of the tobacco's critical growth stages (e.g., the vigorous growth stage and the budding stage when diseases are more likely to occur) are required. Data preprocessing is the foundation for subsequent analysis and includes:

[0032] Radiometric calibration: Converting the raw digital quantization values ​​of satellite sensors into surface reflectance to eliminate the influence of the sensor itself;

[0033] Atmospheric correction: Atmospheric correction models (such as FLAASH, 6S model) are used to eliminate the effects of atmospheric scattering and absorption to obtain true surface reflectance data;

[0034] Orthorectification: Using digital elevation models and ground control points, the images are geometrically corrected to eliminate geometric distortions caused by terrain and sensor attitude, and to ensure spatial registration accuracy between images of different time phases;

[0035] Study area cropping and sample point collection: Based on the GPS coordinates obtained from the field survey, healthy tobacco samples, samples infected with mosaic virus, and samples infected with bollworm were marked on the preprocessed images. These points will serve as "ground values" for subsequent model training and validation.

[0036] S2. Extract vegetation indices, wavelet features, and texture features from dual-temporal satellite remote sensing image data to form a single-temporal feature set. Vegetation indices include one or more of the following: plant carotenoid reflectance index, normalized vegetation index, ratio vegetation index, optimized soil-adjusted vegetation index, improved soil-adjusted vegetation index, improved triangular vegetation index 2, simple carotenoid reflectance index, photosynthetic capacity sensitive ratio index A, photosynthetic capacity sensitive ratio index B, enhanced vegetation index, and modified simple ratio. Wavelet features are extracted through continuous wavelet transform, using Gaussian wavelet or Morlet wavelet as the mother wavelet, and wavelet coefficients are calculated at a scale of 1 or 2 at the 940nm band. Texture features are extracted through gray-level co-occurrence matrix, and texture features are correlation features. The extraction window size is 3×3, the gray-level quantization is 64, the step size is 1, and one or more texture features are extracted from the bands of 820nm, 836nm, 866nm, 880nm, and 896nm.

[0037] It should be noted that in this embodiment, for the two temporal preprocessed images obtained in step S1, three types of features sensitive to plant stress are extracted independently: spectral index, wavelet features, and texture features.

[0038] In spectral index extraction, a series of spectral indices sensitive to vegetation physiological parameters (such as chlorophyll content, biomass, and water stress) are calculated based on the spectral reflectance data of the corresponding pixels for each sample point. This invention preferably uses, but is not limited to, the following 11 indices, and their calculation formulas are shown in Table 1:

[0039] Table 1. Preferred spectral indices and their calculation methods according to the present invention.

[0040] Spectral index Calculation method Plant carotenoid reflectance index (PSRI) Normalized Difference Vegetation Index (NDVI) Ratio Vegetation Index (RDVI) Optimized Soil-Adjusted Vegetation Index (OSAVI) Improved Soil Adjusted Vegetation Index (MSAVI) Improved Triangular Vegetation Index 2 (MTVI2) Simple carotenoid reflectance index (SRPI) Photosynthetic capacity sensitivity ratio A (PSSRA) Photosynthetic capacity sensitivity ratio index B (PSSRB) Enhanced Vegetation Index (EVI) Modified Simple Ratio (MSR)

[0041] in, Indicates in Reflectivity value in the nanometer band;

[0042] In wavelet feature extraction, continuous wavelet transform is used to analyze the reflectance spectrum curve of each sample point. Continuous wavelet transform can simultaneously localize the signal in both the frequency and time (spectral) domains, effectively capturing subtle spectral deformations caused by disease. Based on continuous wavelet transform, correlation analysis is performed between the original spectral curve and the Gaussian function at different locations and scales to generate a series of continuous wavelet energy coefficients. These energy coefficients can extract weak information from the spectra of different diseases. The CWT calculation formula is as follows: ,

[0043] in, The original signal, ( () is the mother wavelet, This is the scaling parameter (which controls the scaling of the wavelet). For translation parameters (controlling the position of the wavelet on the time axis);

[0044] This invention preferentially uses the Gaussian third derivative wavelet (gaus3) and Morlet wavelet (morl) as mother wavelets. The Gauss3 wavelet is suitable for extracting the smoothing trend of the spectrum and suppressing noise, while the Morl wavelet is sensitive to local frequency changes and can effectively characterize local disturbances caused by diseases. Considering the characteristics of tobacco diseases, the 940 nm band is selected as the analysis center. This band is located in the near-infrared platform and is highly sensitive to the internal cell structure and water content of leaves, making it a sensitive indicator band for disease stress. Wavelet coefficients near the 940 nm band are calculated at scale parameter 1 (for the Gauss3 wavelet) and scale parameter 2 (for the Morl wavelet). These coefficients are the extracted wavelet features.

[0045] Texture features of images were extracted using the gray-level co-occurrence matrix (GLCM) to quantify the heterogeneity of canopy spatial structure caused by diseases. The GLCM calculation window was set to 3×3, the gray-level quantification level to 64, and the step size to 1. The calculation direction was set to 0°, and its average value was calculated to maintain rotation invariance. Various statistics of the GLCM were calculated, and this invention found that the correlation feature was most sensitive to texture changes caused by tobacco diseases. Correlation features were calculated in five near-infrared bands: 820 nm, 836 nm, 866 nm, 880 nm, and 896 nm. Ultimately, five texture features were obtained for each sample point at each time phase. Thus, for two time-phase images, each sample point could obtain a single-phase feature vector containing spectral indices, wavelet features, and texture features.

[0046] S3. Based on the single-phase feature set, calculate the dual-phase features, which include feature difference features, relative change features, and ratio features, and construct a high-dimensional feature set that integrates the single-phase feature set and the dual-phase features; the feature difference feature is the difference between the same feature in two phases; the relative change feature is the ratio of the feature change between two phases to the initial phase feature value; the ratio feature is the ratio between the two phase features.

[0047] It should be noted that, in order to depict the dynamic changes during the occurrence and development of the disease in this embodiment, the present invention constructs three types of dual-temporal features based on single-temporal features:

[0048] ① Difference feature: Calculate the difference between the same feature in time phase 2 and time phase 1, reflecting the absolute change, i.e., difference - feature time phase 2 - feature time phase 1;

[0049] ② Relative rate of change characteristic: Calculate the percentage change rate of the difference relative to the characteristic value of phase 1, reflecting the relative change magnitude and reducing the influence of background differences. The formula is: Relative rate of change (%) = (Characteristic phase 2 - Characteristic phase 1) / Characteristic phase 1 ;

[0050] ③ Ratio characteristic: Calculate the ratio of the characteristic value of phase 2 to the characteristic value of phase 1. The formula is: Ratio = Characteristic phase 2 / Characteristic phase 1;

[0051] The dual-phase features obtained from the above calculation are combined with the single-phase features of the two phases to form a high-dimensional fusion feature set containing temporal information, which is used as the input for subsequent model training.

[0052] S4. Based on the recursive feature elimination algorithm, the high-dimensional feature set is screened to obtain a sensitive feature subset. The recursive feature elimination algorithm uses support vector machine as the base learner, and through iterative training and feature importance evaluation, it gradually eliminates redundant features and selects the feature subset that is most sensitive to the identification of tobacco pests and diseases.

[0053] It should be noted that, in this embodiment, due to the high dimensionality and redundancy of the fused feature set, directly using it for modeling can easily lead to the curse of dimensionality and overfitting. Therefore, this invention uses a recursive feature elimination algorithm for feature optimization.

[0054] ① Base learner selection: Linear support vector machine is used as the base learner for RFE, and its weight coefficients are used to evaluate feature importance;

[0055] ② Screening process: Input all fused features to train the SVM model, sort them according to feature weights, and remove the features with the smallest weights. Retrain the SVM with the remaining features, and repeat this process until the number of features is reduced to the preset target or the model performance is optimal;

[0056] ③ Screening results: As shown in Table 2, RFE screened out 8 core features that are most discriminative in distinguishing between healthy and diseased tobacco, including single-phase features and dual-phase features.

[0057] Table 2 Features selected by RFE

[0058] Feature Name Feature weights Feature type Is it a dual-temporal characteristic? ndvi_1 0.399509 Vegetation Index no sipi_1 0.345768 Vegetation Index no pssrb_1 0.399509 Vegetation Index no morl_2_940_rel_change(%) 0.325956 Wavelet index yes correlation_band_24_1 0.650450 Texture features no correlation_band_24_2 0.423072 Texture features no correlation_band_27_ratio 0.436804 Texture features yes correlation_band_29_1 0.348214 Texture features no

[0059] S5. Train a random forest classification model using a subset of sensitive features, and optimize the hyperparameters of the random forest classification model through grid search to obtain an optimized tobacco pest and disease monitoring model. The hyperparameters include at least one of the following: number of decision trees, maximum depth, minimum number of leaf node samples, minimum number of split samples, and maximum number of features. The grid search uses cross-validation to optimize the hyperparameters and selects the optimal combination of hyperparameters based on preset performance evaluation indicators.

[0060] It should be noted that in this embodiment, a filtered subset of sensitive features is used to construct and optimize the random forest classification model. All labeled sample data are randomly divided into training set and independent test set in an 8:2 ratio. A grid search combined with cross-validation strategy is used to optimize the key hyperparameters of the random forest. The optimized parameters include: number of decision trees (n_estimators), maximum tree depth (max_depth), minimum number of samples required for internal node split (min_samples_split), minimum number of samples for leaf nodes (min_samples_leaf), and maximum number of features considered for each split (max_features).

[0061] Using classification accuracy as the evaluation metric for grid search, an exhaustive search is performed in the parameter space to find the parameter combination that optimizes cross-validation performance. The optimal model parameters are shown in Table 3.

[0062] Table 3 Optimal RF model parameters obtained from grid search

[0063] Parameter name Optimal parameters 'class_weight' 'None' 'max_depth' 3 'max_features 'log2' 'min_samples_leaf' 1 'min_samples_split' 5 'n_estimators' 200

[0064] The random forest model was retrained on the full training set using the optimal parameters, and its performance was evaluated on an independent test set. Evaluation metrics included overall precision, precision for each category, and recall. As shown in Table 4, the model demonstrated good recognition ability for each category.

[0065] Table 4. Precision and Recall of the Optimal RF Model on the Test Set

[0066] Sample categories Precision Recall Number of test samples healthy 0.90 0.95 20 Mosaic Disease 0.95 0.90 20 bollworm infestation 0.88 0.85 20 Overall accuracy 0.91

[0067] The trained optimal model was applied to dual-temporal remote sensing images of the entire study area. For each pixel in the image, its corresponding sensitive feature subset was extracted, input into the model for prediction, and finally a classification map showing the spatial distribution of "healthy", "mosaic disease", and "cotton boll pests" was generated, realizing large-scale tobacco pest and disease monitoring.

[0068] S6. Use the optimized tobacco pest and disease monitoring model to identify and classify pests and diseases in the tobacco area to be monitored; the pests and diseases include mosaic virus and bollworm, and the classification results include healthy plants, plants infected with mosaic virus and plants infected with bollworm.

[0069] It should be noted that this invention, based on satellite data, accurately distinguishes and identifies tobacco mosaic disease and cotton bollworm. The process involves multi-feature extraction, feature dimensionality reduction, and the construction of a pest and disease monitoring model. Specifically, three types of features are extracted from tobacco: spectral index, wavelet feature extraction, and texture feature extraction. The overall process is as follows: Figure 2 As shown.

[0070] This invention constructs and filters a sensitive feature set specifically for tobacco pest and disease identification. Specifically, it systematically extracts three types of heterogeneous features—vegetation indices, texture features, and wavelet features—from multi-temporal satellite imagery, and further calculates their temporal differences, relative rates of change, and ratios to construct a dual-temporal dynamic feature set, effectively quantifying temporal changes under pest and disease stress. Based on this, a recursive feature elimination algorithm is used to automatically filter out the most sensitive subset of key features from the constructed high-dimensional feature set. This feature engineering process, by fusing multi-source information and temporal changes, and through efficient feature dimensionality reduction, provides the model with input features that are more discriminative and less redundant, fundamentally improving the accuracy, efficiency, and stability of the monitoring model.

[0071] Furthermore, this invention encompasses a complete classification model construction process. Its innovation lies in using the aforementioned filtered sensitive feature subset as input, employing random forest as the basic classifier, and integrating a grid search strategy to automatically and globally optimize the model's key hyperparameters to obtain the optimal parameter combination. Ultimately, it constructs a classification model suitable for large-scale regions, capable of accurately distinguishing between healthy and infected tobacco plants affected by bollworms and mosaic virus. The key protection of this technical solution lies in the end-to-end modeling method combining "optimized feature input—automatic hyperparameter tuning," which effectively coordinates the representational capabilities of multi-source temporal features with the classification performance of the random forest model, significantly improving the accuracy, robustness, and application value of tobacco pest and disease identification at the regional scale.

[0072] In summary, this invention, through the overall design of dual-temporal feature construction, feature recursive optimization, and highly robust classification modeling, effectively compensates for the shortcomings of existing technologies in the utilization of temporal information and feature fusion, and realizes efficient, stable, and scalable remote sensing monitoring of tobacco pests and diseases.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features, characterized in that, Includes the following steps: S1. Acquire dual-temporal satellite remote sensing image data of tobacco during its critical growth period; S2. Extract vegetation index, wavelet features and texture features from dual-temporal satellite remote sensing image data to form a single-temporal feature set; S3. Based on the single-phase feature set, calculate the dual-phase features. It includes feature difference features, relative change features, and ratio features, and constructs a high-dimensional feature set that integrates single-phase feature sets and dual-phase features; S4. Based on the recursive feature elimination algorithm, the high-dimensional feature set is filtered to obtain a sensitive feature subset; S5. Train a random forest classification model using a subset of sensitive features, and optimize the hyperparameters of the random forest classification model through grid search to obtain an optimized tobacco pest and disease monitoring model. S6. Use the optimized tobacco pest and disease monitoring model to identify and classify pests and diseases in the tobacco area to be monitored.

2. The method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features according to claim 1, characterized in that: In step S2, the vegetation index includes one or more of the following: plant carotenoid reflectance index, normalized vegetation index, ratio vegetation index, optimized soil-adjusted vegetation index, improved soil-adjusted vegetation index, improved triangular vegetation index 2, simple carotenoid reflectance index, photosynthetic capacity sensitive ratio index A, photosynthetic capacity sensitive ratio index B, enhanced vegetation index, and modified simple ratio.

3. The method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features according to claim 1, characterized in that: In step S2, wavelet features are extracted through continuous wavelet transform, which uses Gaussian wavelet or Morlet wavelet as the mother wavelet and calculates wavelet coefficients at scale 1 or 2 at the 940nm band.

4. The method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features according to claim 1, characterized in that: In step S2, texture features are extracted using a gray-level co-occurrence matrix. The texture features are correlation features. The extraction window size is 3×3, the gray-level quantization level is 64, the step size is 1, and one or more texture features are extracted from the wavelengths of 820nm, 836nm, 866nm, 880nm, and 896nm.

5. The method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features according to claim 4, characterized in that: In step S3, the feature difference feature is the difference between the same feature in two time phases; the relative change feature is the ratio of the change in the feature between two time phases to the feature value in the initial time phase; the ratio feature is the ratio between the features in two time phases.

6. The method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features according to claim 1, characterized in that: In step S4, the recursive feature elimination algorithm uses support vector machine as the base learner. Through iterative training and feature importance evaluation, it gradually eliminates redundant features and selects the feature subset that is most sensitive to the identification of tobacco pests and diseases.

7. The method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features according to claim 1, characterized in that: In step S5, the hyperparameters include at least one of the following: number of decision trees, maximum depth, minimum number of leaf node samples, minimum number of split samples, and maximum number of features.

8. The method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features according to claim 1, characterized in that: In step S5, the grid search uses cross-validation to optimize the hyperparameters and selects the optimal combination of hyperparameters based on preset performance evaluation indicators.

9. The method for monitoring tobacco pests and diseases based on dual-temporal remote sensing features according to claim 1, characterized in that: In step S6, the pests and diseases include mosaic virus and bollworm, and the classification results include healthy plants, plants infected with mosaic virus, and plants infected with bollworm.

Images