Wheat yellow mosaic disease unmanned aerial vehicle remote sensing monitoring method and system based on RGB image

Abstract

The invention discloses a wheat yellow mosaic disease unmanned aerial vehicle remote sensing monitoring method and system based on an RGB image, and relates to the field of intelligent agriculture, and the method comprises the steps: collecting RGB image data through an unmanned aerial vehicle for a breeding group naturally infected with wheat yellow mosaic disease, extracting a color index associated with the wheat yellow mosaic disease, and constructing a data set; a correlation analysis method is adopted, feature extraction is carried out through the data set, modeling is carried out through a machine learning algorithm, model parameter optimization is carried out through an optimization algorithm, and then a monitoring model is constructed and used for recognizing the wheat yellow mosaic disease through the unmanned aerial vehicle remote sensing technology. Technical support is provided for guaranteeing safe production of wheat and promoting development of intelligent breeding.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart agriculture, in particular to a wheat yellow mosaic disease unmanned aerial vehicle remote sensing monitoring method and system based on RGB images. BACKGROUND

[0002] Wheat, as an important food crop in the world, plays a strategic role in the global food supply system. The stability of its yield and quality is of great significance to food security. However, during the growth of wheat, it faces complex stress factors such as biotic and abiotic stress. Among them, wheat yellow mosaic disease, as a common and serious disease, seriously threatens the yield and quality of wheat. Wheat yellow mosaic virus (WYMV) is the pathogen that causes wheat yellow mosaic disease. This disease belongs to soil-borne virus disease, and its transmission medium is mainly Polymyxagraminis parasitic on gramineous crops. The infected wheat leaves turn yellow and twist, the plant is dwarfed, the tillering is reduced, the earing rate is reduced, and the yield is generally reduced by 10%-30%, and the yield of severely diseased fields can be reduced by more than 50%. For a long time, artificial field investigation is the main monitoring method for wheat yellow mosaic disease. Professional personnel record and evaluate by observing the symptoms of wheat disease in the field, but this method has significant limitations. It is high in manpower, material resources and time cost, low in work efficiency, and difficult to realize real-time dynamic monitoring in a large area; the survey results are greatly affected by personnel experience and subjective judgment, and the data consistency is poor; and it can only identify when the disease symptoms are obvious, and cannot provide early warning, which cannot meet the needs of modern precision agriculture, and new monitoring technologies that are efficient, accurate and real-time are urgently needed.

[0003] In recent years, unmanned aerial vehicle remote sensing technology has developed rapidly in the field of agriculture, bringing new opportunities for monitoring wheat yellow mosaic disease. Unmanned aerial vehicle low-altitude remote sensing, with the advantages of efficiency, flexibility, real-time and high resolution, can carry visible light, multispectral, hyperspectral and other sensors, quickly acquire high-resolution images and spectral data of farmland, and realize precise monitoring of diseases, resistance level identification and disease assessment by combining image processing technology and machine learning algorithms. Compared with traditional manual investigation, unmanned aerial vehicle remote sensing has the advantages of objectivity, wide coverage and fast data acquisition, and has achieved remarkable results in early identification and quantitative analysis of crop diseases, and has become an important research direction in the field of agriculture. From the perspective of the type of monitoring means, different types of unmanned aerial vehicle remote sensing images have their own characteristics and advantages, and play a key role in crop disease monitoring by combining advanced image processing technology and machine learning algorithms. Machine learning algorithms are increasingly widely used in crop disease monitoring, especially in handling high-dimensional spectral data and complex disease patterns. Although unmanned aerial vehicle remote sensing technology has made many progress in wheat disease monitoring, there are still deficiencies in monitoring multiple varieties of wheat. SUMMARY

[0004] In order to solve the above problems, the purpose of the present application is to provide a kind of wheat yellow leaf spot unmanned aerial vehicle remote sensing monitoring technology based on RGB image, focus on the breeding population wheat yellow leaf spot monitoring based on unmanned aerial vehicle RGB image, to explore efficient monitoring method, improve the accuracy and universality of monitoring model, to guarantee the safe production of wheat, promote the development of intelligent breeding and provide technical support.

[0005] In order to achieve the above technical purpose, the present application provides a kind of wheat yellow leaf spot unmanned aerial vehicle remote sensing monitoring method based on RGB image, comprising the following steps: The breeding population infected with natural wheat yellow leaf spot is taken as the object, RGB image data is collected by unmanned aerial vehicle, color index associated with wheat yellow leaf spot is extracted, and data set is constructed; Correlation analysis method is used for feature extraction by data set, machine learning algorithm is used for modeling, and model parameter optimization is carried out by using optimization algorithm, and then monitoring model is constructed, which is used for identifying wheat yellow leaf spot by unmanned aerial vehicle remote sensing technology.

[0006] Preferably, when the color index is extracted, the selected color index includes: super blue index ExB, super green index ExG, super red index ExR, green leaf index GLA, ikawa index IKAW, improved green red vegetation index MGRVI, normalized green red difference index NGRDI, red green blue vegetation index RGBVI, visible light atmospheric impedance index VARI, wobbe index WI, super green minus super red index ExGR, normalized difference index NDI, normalized green blue difference index NGBDI, vegetation extraction color index CIVE, deep green color index DEVI and super green red blue difference index EGRBDI.

[0007] Preferably, when the feature extraction is carried out, the correlation analysis method includes genetic algorithm or spearman.

[0008] Preferably, when the feature extraction is carried out, the selected spearman is used for feature extraction.

[0009] Preferably, when the modeling is carried out, the machine learning algorithm includes support vector machine, random forest and distributed gradient enhancement.

[0010] Preferably, when the modeling is carried out, the selected support vector machine is used for model construction, and model parameters are determined by using particle swarm optimization.

[0011] The present application discloses a kind of wheat yellow leaf spot unmanned aerial vehicle remote sensing monitoring system based on RGB image, for realizing the above-mentioned wheat yellow leaf spot unmanned aerial vehicle remote sensing monitoring method based on RGB image, including: The data acquisition module is used for collecting RGB image data by using a UAV for a breeding population naturally infected with wheat yellow mosaic disease, extracting color indexes associated with the wheat yellow mosaic disease, and constructing a data set. The recognition module is used for constructing a monitoring model by using a correlation analysis method to perform feature extraction through the data set, using a machine learning algorithm to perform modeling, and using an optimization algorithm to optimize model parameters.

[0012] Preferably, the data acquisition module selects color indexes including an excess blue index ExB, an excess green index ExG, an excess red index ExR, a green leaf area index GLA, an Ikarashi index IKAW, a modified green red vegetation index MGRVI, a normalized green red difference index NGRDI, a red green blue vegetation index RGBVI, a visible atmosphere resistance index VARI, a Wobbe index WI, an excess green minus excess red index ExGR, a normalized difference index NDI, a normalized green blue difference index NGBDI, a color index for vegetation extraction CIVE, a deep green color index DEVI, and an excess green red blue difference index EGRBDI.

[0013] Preferably, the recognition module selects Spearman for feature extraction, selects a support vector machine for model construction, and determines model parameters by using a particle swarm optimization.

[0014] The present application discloses the following technical effects: The present application takes a breeding population naturally infected with wheat yellow mosaic disease as an object, regularly collects UAV RGB image data, extracts color indexes after preprocessing, explores the influence of different feature selection methods and algorithms on the construction of disease grade classification models, and provides support for remote sensing monitoring, accurate prevention and control, and breeding screening of wheat yellow mosaic disease. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0016] Figure 1 is a schematic diagram of the method flow of the present application. DETAILED DESCRIPTION

[0017] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0018] As shown in the Figure 1 The present application provides a wheat yellow mosaic disease unmanned aerial vehicle remote sensing monitoring technology based on RGB images, and specifically includes the following contents: Step 1: data acquisition Step 1.1, unmanned aerial vehicle data acquisition The unmanned aerial vehicle remote sensing platform used in the present application is operated on a sunny day with low wind speed, the flight time is 11:00-14:00, the flight height is 12m, the airspeed is 1m / s, the lateral overlap rate is 70%, and the heading overlap rate is 80%. The unmanned aerial vehicle image acquisition time of wheat yellow mosaic disease is during the wheat green-up period to the booting period, and is collected every 7 days, i.e. March 2, March 8, March 15, March 22 and March 29. Before each flight task is performed, the shooting parameters and route planning are uniformly set to ensure that the shooting parameters and route planning of the flight task at each collection time remain consistent.

[0019] Before the unmanned aerial vehicle starts shooting, a multispectral radiation calibration board with reflectivity of 25%, 50%, 75% and 95% is placed in the image acquisition range, and the radiation calibration board is shot by the unmanned aerial vehicle to obtain the image of the reflection board, which is used to correct the reflectivity of the multispectral image band.

[0020] Step 1.2, for the artificial investigation wheat yellow mosaic disease identification method, specifically includes the following contents: from the initial stage of wheat yellow mosaic disease symptoms, artificial field wheat yellow mosaic disease grade investigation is carried out every 7 days, and at the same time, unmanned aerial vehicle is used to collect wheat yellow mosaic disease RGB and multispectral image data, five-point investigation method is used in each plot, 5 times are recorded, and the average value of 5 times is finally evaluated. The wheat yellow mosaic disease grade identification standard is shown in Table 1 as the wheat yellow mosaic disease grade classification standard.

[0021] Table 1

[0022] Step 1.3, for unmanned aerial vehicle image soil background, specifically includes the following contents: In the research of field crops, especially in the extraction and division of wheat disease severity area, accurate removal of soil background is of great importance. As part of the growing environment of field crops, the complex characteristics of soil background bring many interference factors to related research. The natural differences of soil color, texture and humidity will cause fluctuations in spectral reflectance characteristics, which will affect the accuracy of data like noise, and then interfere with the threshold segmentation process. In the segmentation, if the soil background cannot be effectively removed, these background disturbances will cause the segmentation result to deviate, making it difficult to accurately identify the wheat area, so that the subsequent extraction and division of wheat disease severity area lose the reliable data basis.

[0023] The present application adopts image threshold segmentation method, and selects OSAVI (Optimized Soil-Adjusted Vegetation Index) as the vegetation index for multispectral image threshold segmentation. The calculation method is as follows:

[0024] By calculating the OSAVI index, and using the maximum inter-class variance method, also known as Otsu's method, the threshold value is automatically determined. Otsu's method is based on the gray level histogram of the image, and the best threshold value is determined by maximizing the inter-class variance. The threshold value (such as 0.2) calculated by Otsu's method is used for binaryzation of the image, and the part greater than or equal to the threshold value is regarded as the vegetation area (the binary image takes value 1), and the part less than the threshold value is regarded as the soil background (the binary image takes value 0). The generated binary mask is applied to the original multispectral image to extract the vegetation area, and the result is saved as a multispectral image and a vector file for subsequent analysis. The soil removal of the RGB image uses the above obtained vector file to crop the original RGB image, and the result is saved as an RGB image.

[0025] Step 1.4, extract color index after background removal: Color indices (CI) are calculated by combining images from two or more spectral bands using different computational methods. They provide richer vegetation information, possessing not only clear physical meaning but also the ability to effectively detect changes in physiological and biochemical processes caused by diseases. Therefore, they are widely used in crop growth and disease monitoring. The basic principle behind vegetation index construction stems from the spectral characteristics of plant leaves: they exhibit significant absorption in the visible red light band (approximately 620–750 nm) and strong reflectance in the near-infrared band (approximately 750–1400 nm). Through combined calculations of different spectral bands, vegetation indices offer significant advantages over single-band data: on the one hand, they effectively reduce the interference of environmental factors on canopy reflectivity; on the other hand, they are more sensitive to vegetation changes.

[0026] Based on the above considerations and previous research results, this invention selects 16 representative color indices (CIs) for model construction. Using Python, the average digital number (DN) of the red (R), green (G), and blue (B) bands in the RGB images of each wheat plot's ROI (Region of Interest) is extracted, and the corresponding color indices are calculated according to the spectral index formula (as shown in Table 2). It is worth noting that, to eliminate the influence of changes in lighting conditions during shooting, the color indices (CIs) are calculated using the renormalized r, g, and b bands, as shown in the following calculation method:

[0027]

[0028]

[0029] Table 2

[0030] Step 2: Build and evaluate the model. Step 2.1, the methods for feature selection include: In practical machine learning modeling, it is often difficult to directly determine which features are effective for a specific learning algorithm, and high-dimensional data is frequently encountered in real-world applications. Selecting only a subset of key features to build the model can significantly shorten the computation time of the learning algorithm and improve the model's interpretability. To optimize model performance, it is usually necessary to reduce the number of features; this process is called feature selection or dimensionality reduction.

[0031] To screen for color indices that indicate susceptibility to wheat yellow mosaic disease, a genetic algorithm (GA) and Spearman correlation analysis were used for feature selection.

[0032] Genetic Algorithm (GA) is a heuristic search algorithm that simulates the natural evolutionary process of organisms. It was first proposed and systematically studied by John Holland. Based on the natural selection principle of "survival of the fittest," the algorithm encodes solutions to problems as chromosomes (individuals) and searches for the optimal solution in the solution space through a series of genetic operations such as selection, crossover, and mutation. This paper uses Python and combines GA with Recursive Feature Elimination Cross-Validation (RFECV) for feature selection. Genetic Algorithms are optimization algorithms that simulate natural selection and genetic mechanisms, finding the optimal feature subset through iterative evolution. The cross-validation accuracy of Support Vector Machines is used as the fitness function, and the performance of each subset is evaluated using RFECV to ultimately obtain the optimal feature subset.

[0033] To reduce the randomness of results that may result from single feature selection methods, and to comprehensively evaluate the effectiveness of different algorithms in screening sensitive features for wheat yellow mosaic disease severity, Spearman correlation analysis was introduced. Spearman correlation analysis is a non-parametric statistical method used to measure the degree of monotonic correlation between two variables. Unlike Pearson correlation analysis, Spearman correlation analysis does not rely on the linear relationship between variables, but rather on the rank (or order) of the variables rather than the original data values, thus making it suitable for non-linear data. In high-dimensional data analysis, Spearman correlation analysis can be used to calculate the correlation between different features, thereby identifying a subset of features related to the target feature. The Spearman correlation coefficient is defined as follows:

[0034] Where di represents the rank difference between two variables, and n represents the sample size. The Spearman correlation coefficient ranges from -1 to 1. A value closer to 1 indicates a stronger monotonically positive correlation between the two variables; a value closer to -1 indicates a stronger monotonically negative correlation; and a value closer to 0 indicates no monotonically correlated relationship. In spectral data analysis, Spearman correlation analysis can be used to calculate the correlation between different vegetation indices and target features (such as crop categories). If certain vegetation indices show a significant monotonically correlated relationship with the target feature, these indices are likely to be of significant value to the classification task. By calculating the Spearman correlation coefficient, these key vegetation indices can be selected, thereby improving the accuracy and explanatory power of the classification model.

[0035] Step 2.2, regarding the selection of classification algorithms, including: In this invention, a modeling dataset was constructed using 16 color indices from 216 plots across 5 collection times in an experimental breeding population. Based on a Python programming environment, the dataset was divided into training and test sets in a 7:3 ratio. The test set was strictly independent of the model training process and used solely to evaluate the generalization performance of the classification model. Three classic machine learning classification algorithms—Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost)—were employed to construct a wheat yellow mosaic disease severity monitoring model. By systematically comparing the classification performance of each algorithm on the training and test sets, the aim was to select the model method with the best prediction accuracy and generalization ability for wheat yellow mosaic disease severity identification.

[0036] Support Vector Machine (SVM) is a widely used supervised learning algorithm for classification and regression problems. Its core idea is to find an optimal hyperplane A in the feature space that separates sample points of different classes as much as possible and maximizes the margin between the two classes. This optimal hyperplane is called the decision boundary. Specifically, SVM attempts to find a hyperplane such that the margin of the sample points closest to this hyperplane is maximized. These sample points closest to the hyperplane are called support vectors, and they play a decisive role in determining the classification hyperplane. The decision boundary of SVM is entirely determined by these support vectors; therefore, SVM is also called a support vector classifier. Classification includes binary classification and multi-class classification. For binary classification problems, the goal of SVM is to find a hyperplane that separates the two classes of samples and maximizes the margin between the classes. Support Vector Machine is a powerful and flexible classification algorithm, especially suitable for high-dimensional data and nonlinear problems. Combined with the use of kernel functions, SVM can effectively solve complex classification tasks. This invention employs an "OvR" (one-to-many) approach, training multiple binary classifiers and separating each class from the rest to achieve multi-class classification.

[0037] Random Forest (RF) is an ensemble learning algorithm based on decision trees. It constructs multiple decision trees and combines their results to make classification predictions. Its basic principle is "integrated decision-making." It constructs multiple different subsets from the original training dataset through sampling with replacement, then builds multiple decision trees based on each subset, and finally combines the predictions of these decision trees to determine the final classification result. Typically, a voting method is used, selecting the class that appears most frequently as the final prediction. RF has high accuracy because it combines the results of multiple decision trees, effectively reducing model variance and mitigating the risk of overfitting.

[0038] XGBoost is a high-efficiency machine learning algorithm based on the gradient boosting framework, widely used in data mining and machine learning competitions. It's based on the Gradient Boosting Decision Tree (GBDT) framework, iteratively optimizing the objective function by continuously adding new trees. Each time a new tree is generated, it calculates the gradient of the loss function with respect to the predicted value based on the difference between the predictions of previous trees and the true values, and then fits a new tree to minimize this gradient, thereby gradually improving the model's accuracy. By considering second-order derivative information, XGBoost can fit data more accurately, achieving high prediction accuracy in many machine learning tasks; it also supports large-scale distributed training, adapting well to datasets of different sizes and computational resources.

[0039] Step 2.3: The selection of model evaluation metrics includes: This invention constructs a wheat disease classification model based on the Python 3.11.4 platform and systematically evaluates the predictive performance of three machine learning algorithms using a confusion matrix. During the evaluation, precision, recall, and F1 (F1-score) were first calculated for each category. Due to differences in the number of disease grades among different wheat varieties during disease investigation, class imbalance exists in the dataset. To address this, this invention comprehensively employs multi-dimensional evaluation metrics, including accuracy, weighted F1 (WF1), and micro-average F1 (MF1), to comprehensively and objectively assess the model's classification performance. This multi-metric comprehensive evaluation method not only reflects the model's classification effectiveness across categories but also effectively balances the impact of uneven class distribution on the evaluation results, thus providing a reliable basis for accurate model performance evaluation.

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048] In the formula, TP (TruePositive): the number of samples correctly classified as positive by the model; FN (FalseNegative): the number of positive samples misclassified as negative by the model (i.e., missed detection); FP (FalsePositive): the number of negative samples misclassified as positive by the model; TN (TrueNegative): the number of samples correctly classified as negative by the model; F1 i : F1 score of class i; Precision i Precision of the i-th class; Recall i : Recall rate of class i; w i The weight of class i, typically the proportion of samples of that class in the total sample; TP i : The number of true instances in class i; FP i : The number of false positives in class i; FN i : The number of false counterexamples of class i.

[0049] Step 3. Analyze the model selection and the results of evaluating the constructed model using evaluation metrics, including: Step 3.1: Feature selection for evaluation: Step 3.1.1 Feature selection based on genetic algorithm: Feature selection was performed on the color index at each time point of wheat yellow mosaic disease collection, and the selected features were used as input variables for the subsequent construction of the wheat yellow mosaic disease classification model.

[0050] Two graphs are used at each data collection time to illustrate the color index in the feature selection process of the genetic algorithm. The first graph is a feature selection trend graph: it shows the trend of the probability of each feature being selected as the number of generations increases during the feature selection process of the genetic algorithm. Each curve in the graph corresponds to a feature, and the trend of the curve reflects the change in the probability of that feature being selected in each generation. If a curve shows an upward trend, it means that the probability of that feature being selected in subsequent generations gradually increases, that is, it is becoming more and more favored by the genetic algorithm; if the curve is relatively flat or shows a downward trend, it means that the probability of that feature being selected is relatively stable or gradually decreasing, which may be a relatively less important feature, and it is likely to be eliminated. The second graph is a feature importance histogram: RFECV is used to evaluate the selected features, calculate the importance score of the features, and draw a histogram to clarify the importance of each feature.

[0051] Feature selection results based on the GA algorithm show that, during the data collection period on March 2nd, features NGBDI, NGRDI, and ExG were selected most frequently, while WI and EGRBDI were selected with extremely low probabilities, showing a trend of gradual elimination. Evaluation of the optimal feature subset using RFECV yielded the following importance ranking for the 10 features: RGBVI, NGBDI, NDI, ExGR, NGRDI, IKAW, GLA, ExR, ExG, and ExB.

[0052] Feature selection results based on the GA algorithm show that, during the data collection period on March 8th, features ExR, IKAW, and NDI were selected most frequently, while MGRVI, WI, and CIVE were selected with extremely low probabilities, showing a trend of gradual elimination. Evaluation of the optimal feature subset using RFECV yielded the following importance ranking for the seven features: EGRBDI, GLA, RGBVI, ExGR, NGRDI, IKAW, and ExR.

[0053] Feature selection results based on the GA algorithm show that, during the data collection period on March 15th, features MGRVI, RGBVI, NDI, and EGRBDI were selected most frequently, while ExB, ExR, and GLA were selected with extremely low probabilities, showing a trend of gradual elimination. Evaluation of the optimal feature subset using RFECV yielded the following importance ranking for the 10 features: ExGR, DEVI, ExG, NDI, EGRBDI, NGRDI, VARI, RGBVI, MGRVI, and KAW.

[0054] Feature selection results based on the GA algorithm show that, during the data collection period on March 22nd, features GLA, ExGR, and NDI were selected most frequently, while NGBDI and IKAW were selected with extremely low probability, showing a trend of gradual elimination. Evaluation of the optimal feature subset using RFECV yielded the following importance ranking for the 10 features: ExGR, ExG, GLA, NDI, WI, RGBVI, NGRDI, MGRVI, ExR, and ExB.

[0055] Feature selection results based on the GA algorithm show that, during the data collection period on March 29th, features WI and EGRBDI were selected most frequently, while GLA and ExGR were selected with extremely low probability, showing a trend of gradual elimination. Evaluation of the optimal feature subset using RFECV yielded the following importance ranking for the six features: CIVE, EGRBDI, MGRVI, RGBVI, WI, and ExR.

[0056] Finally, the feature selection color index results based on the GA algorithm at each acquisition time are shown in Table 3.

[0057] Table 3

[0058] Step 3.1.2: Feature selection based on Spearman correlation analysis: Spearman correlation analysis was performed between the color index and the disease severity (DL) of wheat yellow mosaic. On March 2nd, except for ExB, RGBVI, and NGBDI, the absolute values ​​of the Spearman correlation coefficients (ρ) of the other color indices were between 0.4 and 0.5, with VARI showing the highest correlation at -0.456. On March 8th, except for ExB, RGBVI, NGBDI, and EGRBDI, the absolute values ​​of the other color indices (ρ) were between 0.3 and 0.4, with WI showing the highest correlation at -0.37. On March 15th, only IKAW, NGBDI, WI, and ExB had absolute values ​​greater than 0.3. On March 22nd, IKAW, NGBDI, and ExB had absolute values ​​greater than 0.5, with IKAW having the largest absolute value at 0.594. On March 29th, seven color vegetation indices had absolute values ​​greater than 0.3, with WI showing the highest correlation at -0.39.

[0059] To compare with the features selected by the GA algorithm, six color indices with the highest Spearman correlation were selected for each acquisition time. The results are shown in Table 4.

[0060] Table 4

[0061] Step 3.2: Construction and evaluation of disease severity classification model Step 3.2.1: Setting model hyperparameters based on different search methods: This invention employs different search methods to optimize the hyperparameters of three machine learning algorithms: SVM, RF, and XGBoost. Particle Swarm Optimization (PSO) is used to optimize the hyperparameters of the SVM algorithm. PSO is an algorithm that solves optimization problems by simulating the foraging behavior of bird flocks, and it can efficiently search for the global optimum. For SVM, the PSO algorithm is mainly used to optimize the penalty coefficient C and the gamma value when the kernel function is a radial basis function (RBF).

[0062] Grid search is used for hyperparameter optimization in both RF and XGBoost algorithms. Grid search iterates through all combinations of hyperparameter variations in the model, determining the optimal parameter combination based on the test results of models trained with different parameters. This is highly effective on small datasets. For the RF algorithm, the main parameters optimized are the number of decision trees (n_estimators) and the maximum depth (max_depth). For the XGBoost algorithm, the main parameters optimized include the number of decision trees (n_estimators), the maximum depth (max_depth), and the learning rate (learning_rate).

[0063] By using the different search methods described above, the hyperparameters of each algorithm were optimized in different groups of selected features, and the optimal hyperparameters of SVM, RF and XGBoost algorithms were finally determined, as shown in Table 5.

[0064] Table 5

[0065] Step 3.2.2, Evaluation of the Disease Severity Classification Model: Using features selected by GA and Spearman correlation coefficients as inputs, and combining datasets from different collection times, we constructed classification models for wheat yellow mosaic disease severity in breeding populations using SVM, RF, and XGBoost machine learning algorithms. A total of 30 models were built (2 feature selection methods × 5 dates × 3 algorithms). The established models were tested using the test set of the aforementioned datasets, and five-fold cross-validation was performed. Evaluation metrics such as accuracy, Weighted-F1, Micro-F1, Precision, and Recall were recorded.

[0066] The results show that the classification models for wheat yellow mosaic disease severity levels constructed using three classification algorithms, based on features selected by GA and Spearman and combined with data collected at different times, all achieved good results. This demonstrates that using RGB imaging technology to collect wheat canopy images and extract color indices, combined with feature selection methods, can effectively achieve accurate classification of wheat yellow mosaic disease severity levels. Furthermore, optimizing the model's hyperparameters using grid search and PSO methods further improved the model's performance, validating the reliability and practicality of RGB imaging technology and feature selection methods in disease severity classification.

[0067] The test results of the 30 constructed models are summarized in Table 6.

[0068] Table 6

[0069] As can be observed from the data in Table 6, the Micro-F1 value is the same as the accuracy, indicating that the model performs relatively evenly across all classes, without excessive bias or neglect of any particular class. The Recall value is the same as the accuracy, indicating that the model can identify positive samples well. Overall, the model performance varies under different combinations of feature selection methods (GA and Spearman), data acquisition time, and machine learning algorithms (SVM, RF, XGBoost). The accuracy ranges approximately between 0.6 and 0.908, the Weight-F1 ranges approximately between 0.551 and 0.877, and the precision ranges approximately between 0.509 and 0.853, indicating that the model's performance has significant room for improvement and fluctuates significantly under different conditions.

[0070] Regarding feature selection methods, there are some differences in model performance between GA and Spearman feature selection methods. At different data collection times, Spearman sometimes performs better; at the 03:22 data collection time, the accuracy of SVM using Spearman reached 0.908, while the accuracy of SVM using GA was 0.892. However, Spearman does not outperform GA at all times, indicating that the impact of different feature selection methods on model performance is not absolute and may be related to the specific distribution and characteristics of the data.

[0071] Regarding the impact of data acquisition time, the performance metrics of each model showed a certain trend as the acquisition time progressed. Overall, from March 22nd to March 22nd, most models showed an upward trend in metrics such as accuracy, Weight-F1, and Micro-F1. The accuracy of SVM under the GA selection method increased from 0.646 to 0.892. This may indicate that with the passage of time, the accumulation of data or changes in data characteristics enabled the model to learn and fit better, thus improving performance. However, at the March 29th acquisition time, some metrics declined. The accuracy of SVM under the GA selection method dropped to 0.662, possibly due to the introduction of some interference factors from the new data or changes in the data distribution.

[0072] Different machine learning algorithms exhibit varying performance under the same feature selection method and data acquisition time. Among SVM, RF, and XGBoost, SVM achieves higher accuracy in some cases, reaching 0.892 with the GA selection method and a data acquisition time of 03:22. RF also performs well at certain time points, achieving an accuracy of 0.723 with a data acquisition time of 03:08. XGBoost's performance is relatively stable, but fluctuates across different time points and feature selection methods. This indicates that different algorithms differ in their adaptability to data and learning ability, making the selection of an appropriate algorithm crucial for model performance.

[0073] The above analysis shows that the wheat yellow mosaic disease classification model constructed based on RGB imaging technology and feature selection method has high practicality and reliability. In particular, the model can achieve good classification results when combined with Spearman feature selection method and SVM algorithm during the mid-term collection period (0315-0322). The model performance is particularly outstanding on the collection time of March 22 (as shown in Table 7), with an accuracy of 0.908.

[0074] Table 7

[0075] As shown in Table 7, disease level 0 has the largest number of samples, reaching 54, with a precision of 0.92, a recall of 1.00, and an F1-Score of 0.96. This indicates that the model can accurately identify all disease level 0 samples with almost no false negatives or missed positives. Disease level 1 has the smallest number of samples, only 4, with a precision, recall, and F1-Score of 0, indicating that the model cannot effectively identify disease level 1 samples, possibly due to the small sample size, making it difficult for the model to learn the features of this level. Disease level 2 also has 4 samples, with a precision, recall, and F1-Score of 0.75, showing moderate performance, but with some false positives and false negatives. This may be due to the limited sample size, resulting in insufficient learning of the features of this level. Disease level 3 has the smallest number of samples, with 3. The model achieved a precision of 1.00, a recall of 0.67, and an F1-Score of 0.80, indicating that the model can accurately predict samples of disease grade 3, but there are some missed detections. Overall, the model performs best on disease grade 0, which has a larger sample size, while it performs poorly on disease grades 1, 2, and 3, which have a smaller sample size, especially disease grade 1.

[0076] Due to the inherent subjectivity of manual identification, the assigned classification labels may deviate from the actual severity of wheat yellow mosaic disease. To more intuitively illustrate the differences between the SVM classification model for wheat yellow mosaic disease severity in the breeding population constructed using Spearman-selected features (collected on March 2nd) and manual identification, the disease severity levels identified manually and those predicted by the SVM model, as well as the differences between the two, were visualized. The constructed SVM classification model for wheat yellow mosaic disease severity in the breeding population can basically distinguish different severity levels of wheat yellow mosaic disease. Furthermore, the disease severity levels predicted by the SVM model differ from those identified manually by only one level (24), accounting for 92.31% of all incorrect model predictions (26 samples). This indicates that the classification model for wheat yellow mosaic disease severity constructed using the six color indices selected by Spearman can basically reflect the severity of wheat yellow mosaic disease in different varieties within the experimental breeding population.

[0077] In summary, this invention extracts the average DN values ​​of the R, G, and B bands from RGB images, calculates the corresponding color indices based on the spectral index formula, selects features using the genetic algorithm GA and Sperman correlation analysis, and constructs a wheat yellow mosaic disease classification model by combining three machine learning algorithms: SVM, RF, and XGBoost. The hyperparameter settings for different search methods were optimized, and the effectiveness was verified through model evaluation. By comparing different combinations of feature selection methods, acquisition time, and machine learning algorithms, the changing trends of model performance were analyzed, and the optimal model combination under different conditions was determined. This invention achieves the following results: (1) Regarding the impact of feature selection methods, the effects of GA and Spearman feature selection methods on model performance differ. At certain time points, the Spearman method can achieve higher accuracy and other index values. However, Spearman is not superior to GA at all time points, which indicates that the effectiveness of feature selection methods may be related to the specific distribution and characteristics of the data.

[0078] (2) As the data collection time progressed, the model performance indicators showed a certain trend. From March 22nd to March 22nd, the accuracy, Weight-F1, and Micro-F1 of most models showed an upward trend. This may mean that as time went on, the accumulation of data or changes in data characteristics enabled the model to learn and fit better, thereby improving performance. However, at the data collection time of March 29th, some indicators decreased. This may be because the new data introduced some interference factors or the data distribution changed.

[0079] (3) Under the same feature selection method and data acquisition time, different algorithms also showed different performance. Among the three algorithms SVM, RF, and XGBoost, SVM achieved higher accuracy in some cases, RF performed well at some time points, and XGBoost had relatively stable performance, but fluctuated under different time points and feature selection methods. This indicates that different algorithms have different adaptability and learning abilities to data, and choosing the appropriate algorithm is crucial for model performance.

[0080] (4) Based on RGB imaging technology, from the overall data perspective, the accuracy range of different feature selection methods (GA and Spearman), acquisition time, and machine learning algorithms (SVM, RF, XGBoost) is approximately between 0.6 and 0.908, the weight-F1 range is approximately between 0.551 and 0.877, and the precision range is approximately between 0.509 and 0.853. At the acquisition time of March 22, the Spearman feature selection method was used to obtain the best 6 color indices (IKAW, NGBDI, ExB, WI, RGBVI, EGRBDI) as input features. The optimal parameter combination (C=24.04, gamma=0.40) was found through PSO. The constructed SVM algorithm performed best in the wheat yellow mosaic disease classification task, with an accuracy of 0.908.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0083] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for unmanned aerial vehicle remote sensing monitoring of wheat yellow mosaic disease based on RGB images, characterized in that, The method comprises the following steps: A breeding population naturally infected with wheat yellow leaf virus is taken as the object, RGB image data is collected by a UAV, color indexes associated with wheat yellow leaf virus are extracted, and a data set is constructed; After feature extraction is performed on the data set by using a correlation analysis method, modeling is performed by using a machine learning algorithm, and model parameter optimization is performed by using an optimization algorithm, a monitoring model is constructed, and the wheat yellow leaf virus is identified by using UAV remote sensing technology.

2. The UAV remote sensing monitoring method for wheat yellow leaf virus based on RGB images according to claim 1, wherein: When the color indexes are extracted, the selected color indexes include an excess blue index ExB, an excess green index ExG, an excess red index ExR, a green leaf area index GLA, an IKAW index, an improved green-red vegetation index MGRVI, a normalized green-red difference index NGRDI, a red-green-blue vegetation index RGBVI, a visible atmosphere resistance index VARI, a Weibek index WI, an excess green minus excess red index ExGR, a normalized difference index NDI, a normalized green-blue difference index NGBDI, a vegetation extraction color index CIVE, a deep green color index DEVI, and an excess green-red-blue difference index EGRBDI.

3. The UAV remote sensing monitoring method for wheat yellow leaf virus based on RGB images according to claim 2, wherein: When the feature extraction is performed, the correlation analysis method includes a genetic algorithm or a Spearman.

4. The UAV remote sensing monitoring method for wheat yellow leaf virus based on RGB images according to claim 3, wherein: When the feature extraction is performed, the Spearman is selected for feature extraction.

5. The UAV remote sensing monitoring method for wheat yellow leaf virus based on RGB images according to claim 4, wherein: When the modeling is performed, the machine learning algorithm includes a support vector machine, a random forest, and a distributed gradient boosting.

6. The UAV remote sensing monitoring method for wheat yellow leaf virus based on RGB images according to claim 5, wherein: When the modeling is performed, the support vector machine is selected for model construction, and the model parameters are determined by using a particle swarm optimization.

7. A wheat yellow mosaic disease unmanned aerial vehicle remote sensing monitoring system based on RGB images, used to realize the wheat yellow mosaic disease unmanned aerial vehicle remote sensing monitoring method based on RGB images according to claim 1, characterized in that, It comprises: A data acquisition module is configured to take a breeding population naturally infected with wheat yellow leaf virus as the object, collect RGB image data by a UAV, extract color indexes associated with wheat yellow leaf virus, and construct a data set; An identification module is configured to perform feature extraction on the data set by using a correlation analysis method, perform modeling by using a machine learning algorithm, and perform model parameter optimization by using an optimization algorithm, and construct a monitoring model for identifying wheat yellow leaf virus by using UAV remote sensing technology.

8. The UAV remote sensing monitoring system for wheat yellow leaf virus based on RGB images according to claim 7, wherein: The data acquisition module, the selected color index includes: super blue index ExB, super green index ExG, super red index ExR, green leaf index GLA, Kawabata index IKAW, improved green red vegetation index MGRVI, normalized green red difference index NGRDI, red green blue vegetation index RGBVI, visible light atmospheric impedance index VARI, Wobbe index WI, super green minus super red index ExGR, normalized difference index NDI, normalized green blue difference index NGBDI, vegetation extraction color index CIVE, deep green color index DEVI and super green red blue difference index EGRBDI.

9. The wheat yellow mosaic disease unmanned aerial vehicle remote sensing monitoring system based on RGB images according to claim 8, characterized in that: The identification module selects Spearman for feature extraction, selects a support vector machine for model construction, and determines model parameters by using a particle swarm optimization.

Images