Crop disease and pest detection method and device based on RGB image integrated optimization

Abstract

The invention relates to a crop disease and insect pest detection method and device based on RGB image integrated optimization. The method comprises the following steps: calculating 24 visible light vegetation indexes according to an RGB image of a disease and insect pest monitoring area; carrying out importance sorting on the normalized visible light vegetation indexes by adopting a random forest algorithm, and carrying out index selection; according to a pre-selected visible light vegetation index image, processing by adopting a BWO and a plurality of threshold segmentation methods to obtain a plurality of new vegetation index images; and carrying out image segmentation on each new vegetation index image by adopting a corresponding threshold segmentation method, and processing all obtained segmentation results by adopting an integration strategy based on BWO to obtain a pest detection result of the crop. According to the method, subsets of all vegetation indexes are selected through the random forest algorithm, so that the calculation complexity can be effectively reduced while the detection precision is ensured; and an integration strategy of a BWO optimization algorithm is adopted, so that the pest and disease detection precision is improved.

Description

Technical Field

[0001] The present application relates to the technical field of pest and disease detection, and in particular to a crop pest and disease detection method and device based on RGB image integrated optimization. Background Art

[0002] The rapid progress of remote sensing technology has made it an important method to use a variety of aviation and aerospace remote sensing technologies for pest and disease monitoring. In the field of pest and disease monitoring, satellite remote sensing technology was used early and has achieved remarkable results. However, the acquisition cost of high-resolution satellite data is high, the update speed is slow, and its spatial resolution is still insufficient to meet the needs when it is required to be accurate to the tree species level.

[0003] UAV remote sensing can quickly obtain multi-temporal, multi-angle, multi-spectral, and hyperspectral ultra-high-resolution images. It can also collect data on target areas under complex terrain conditions, providing technical support for forest disease monitoring in the region. The physical basis for using UAV technology to monitor forest pests and diseases is that when vegetation is threatened by pests and diseases, different stress symptoms or vegetation damage often occur, causing changes in vegetation spectral reflectance, which are captured by UAV sensors. Existing research results include using multispectral or hyperspectral UAV sensors to obtain high-resolution multispectral UAV data, combining vegetation spectral characteristic differences and high-resolution images, and using artificial intelligence-related algorithms to conduct research on pest and disease monitoring, and have achieved good detection accuracy. However, since UAV sensors carrying dedicated multispectral / hyperspectral cameras are expensive and require high professionalism in operation, the cost of obtaining high-resolution multispectral / hyperspectral UAV image data is high. In addition, the payload weight of small low-altitude UAV remote sensing platforms is limited, and they often cannot carry high-precision professional instruments and equipment. Usually, visible light cameras are used as the main cameras, resulting in no near-infrared information in the images. To this end, a variety of visible light vegetation indices were constructed using the valley-peak-valley characteristics of green vegetation reflectance in the visible light band, and have been applied in studies such as vegetation information extraction and chlorophyll content estimation.

[0004] Although current forest disease monitoring has made certain progress, the improvement in image resolution represents increasingly rich surface information, which brings many challenges to the monitoring accuracy and calculation speed of the algorithm.

[0005] Therefore, it is necessary to propose a crop disease and pest detection method based on RGB image integration optimization to address the problems of expensive equipment and difficult data processing in traditional hyperspectral / multispectral sensors. Summary of the invention

[0006] Based on this, it is necessary to provide a crop disease and pest detection method and device based on RGB image integrated optimization to address the above technical problems.

[0007] A crop disease and insect pest detection method based on RGB image integrated optimization, the method comprising:

[0008] According to the RGB images of the pest and disease monitoring area taken by the drone, 24 visible light vegetation indices are calculated, all visible light vegetation indices are normalized, and the corresponding visible light vegetation index image is generated; the visible light vegetation index image is a grayscale image.

[0009] The random forest algorithm is used to sort the importance of all visible vegetation indices and select the indexes to obtain several pre-selected visible vegetation index images.

[0010] All pre-selected visible light vegetation index images are processed using the Beluga optimization algorithm and multiple threshold segmentation methods to obtain new vegetation index images corresponding to each threshold method.

[0011] For each new vegetation index image, the corresponding threshold segmentation method is used for image segmentation, and all the segmentation results are processed using an integrated strategy based on the Beluga optimization algorithm to obtain the pest and disease detection results of crops in the pest and disease monitoring area.

[0012] A crop pest detection device based on RGB image integrated optimization, the device comprising:

[0013] The visible light vegetation index image generation module is used to calculate 24 visible light vegetation indices based on the RGB images of the pest and disease monitoring area taken by the drone, normalize all visible light vegetation indices, and generate corresponding visible light vegetation index images; the visible light vegetation index image is a grayscale image.

[0014] The visible light vegetation index image selection module is used to use the random forest algorithm to sort the importance of all visible light vegetation indices and select the indexes to obtain a number of pre-selected visible light vegetation index images.

[0015] The new vegetation index image generation module is used to process all pre-selected visible light vegetation index images using the White Whale optimization algorithm and multiple threshold segmentation methods to obtain a new vegetation index image corresponding to each threshold method.

[0016] The pest and disease detection result detection module is used to perform image segmentation on each new vegetation index image using the corresponding threshold segmentation method, and process all the segmentation results using an integrated strategy based on the Beluga optimization algorithm to obtain the pest and disease detection results of crops in the pest and disease monitoring area.

[0017] The above-mentioned crop pest and disease detection method and device based on RGB image integrated optimization, the method includes: calculating 24 visible light vegetation indexes based on the RGB images of the pest and disease monitoring area taken by the drone, and normalizing them; using the random forest algorithm to sort the importance of all normalized visible light vegetation indexes, and select the index; using the Beluga optimization algorithm + multiple threshold segmentation methods to process all pre-selected visible light vegetation index images to obtain multiple new vegetation index images; using multiple threshold segmentation methods to segment each new vegetation index image, and using the integrated strategy based on the Beluga optimization algorithm to process all the segmentation results to obtain the pest and disease detection results of crops in the pest and disease monitoring area. This method can effectively reduce the computational complexity while ensuring the detection accuracy by selecting a subset of all vegetation indices through the random forest importance ranking method; using the integrated strategy of the Beluga optimization algorithm to improve the detection accuracy and detection precision of pests and diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a process of crop pest and disease detection method based on RGB image integration optimization in one embodiment;

[0019] Figure 2 It is an overall technical flow chart of a crop disease and insect pest detection method based on RGB image integration optimization in another embodiment;

[0020] Figure 3 A random forest importance ranking graph in another embodiment;

[0021] Figure 4 It is a schematic diagram of the result of segmenting the 10 indexes selected by random forest importance using the OTSU algorithm in another embodiment, wherein (a) is a true value map, (b) is a schematic diagram of the result of segmenting the ExGR index, (c) is a schematic diagram of the result of segmenting the CIVE index, (d) is a schematic diagram of the result of segmenting the VI index, (e) is a schematic diagram of the result of segmenting the VEG index, (f) is a schematic diagram of the result of segmenting the COM1 index, (g) is a schematic diagram of the result of segmenting the EGRBDI index, (h) is a schematic diagram of the result of segmenting the RGBRI index, (i) is a schematic diagram of the result of segmenting the VDVI index, (j) is a schematic diagram of the result of segmenting the MGRVI index, (k) is a schematic diagram of the result of segmenting the GLI index, and (l) is a schematic diagram of the result of segmenting the new vegetation index after BWO optimization;

[0022] Figure 5Schematic diagram of the results of segmenting 10 vegetation indices selected by random forest using multiple threshold segmentation methods in another embodiment; wherein (a) is an ExGR index map, (b) is the OTSU segmentation result of the ExGR index, (c) is the OTSU segmentation result of the new vegetation index after BWO optimization, (d) is a CIVE index map, (e) is the WAT segmentation result of the CIVE index, (f) is the WAT segmentation result of the new vegetation index after BWO optimization, (g) is an ExB index map, (h) is the BM segmentation result of ExB, and (i) is the BM segmentation result of the new vegetation index after BWO optimization;

[0023] Figure 6 Schematic diagram of the effect of the number of features selected by random forest importance on the detection results in another embodiment, where (a) is CSI, (b) is OA, and (c) is Kappa;

[0024] Figure 7 is the impact of different integration strategies in another embodiment; wherein, (a) is the CSI of different integration strategies of RFVI_Ths, (b) is the OA of different integration strategies of RFVI_Ths, (c) is the Kappa of different integration strategies of RFVI_Ths, (d) is the CSI of different integration strategies of RFVI_BWO_Ths, (e) is the OA of different integration strategies of RFVI_BWO_Ths, and (f) is the Kappa of different integration strategies of RFVI_BWO_Ths;

[0025] Figure 8 : is a detection result image of the model in a new scene in another embodiment, wherein (a) is the original RGB image of the new scene 1, (b) is the true value image of the new scene 1, (c) is the detection result image of the new scene 1, (d) is the original RGB image of the new scene 2, (e) is the true value image of the new scene 2, and (f) is the detection result image of the new scene 2;

[0026] Fig. 9 The structure block diagram of a crop disease and insect pest detection device based on RGB image integrated optimization in one embodiment. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0028] Whale Optimization Algorithm, abbreviated as BWO.

[0029] Trees in forests are usually planted in large areas in mountainous areas and are also susceptible to pests and diseases. Manual inspections are costly. As an advanced data collection carrier, drones can easily achieve large-scale data collection. In view of the problems of expensive equipment and difficult data processing of traditional hyperspectral / multispectral sensors, this application proposes a forest pest and disease detection method based on RGB images.

[0030] In one embodiment, Figure 1 , Figure 2 As shown, a crop pest and disease detection method based on RGB image integration optimization is provided, and the method comprises the following steps:

[0031] Step 100: Calculate 24 visible light vegetation indices based on the RGB image of the pest monitoring area taken by the drone, normalize all visible light vegetation indices, and generate a corresponding visible light vegetation index image; the visible light vegetation index image is a grayscale image.

[0032] Specifically, the existing vegetation indices are mainly based on visible light bands and near-infrared bands, such as the Normalized Difference Vegetation Index (NDVI) and the Ratio Vegetation Index (RVI). The remote sensing images required are generally expensive to obtain, have poor real-time performance, and low spatial resolution. This method calculates 24 visible light vegetation indices using visible light images taken by drones.

[0033] High-resolution image data of the monitoring area was collected by UAV, and necessary preprocessing was performed to ensure data quality. A variety of visible light vegetation indices were extracted from the preprocessed high-resolution image data.

[0034] Each visible vegetation index reflects the specific response of vegetation under specific environmental conditions. Considering that some indices use normalized RGB bands, resulting in inconsistent dimensions between indices, this method normalizes the 24 visible vegetation indices based on RGB images and generates corresponding grayscale images. All subsequent operations are performed based on this normalized image.

[0035] Step 102: using a random forest algorithm to sort the importance of all visible light vegetation indices, and perform index selection to obtain a plurality of pre-selected visible light vegetation index images.

[0036] Specifically, the importance of multiple vegetation indices was evaluated using the random forest importance ranking method, and the results showed that using a subset of all indices can effectively reduce the computational complexity while ensuring detection accuracy.

[0037] Random forest is an ensemble learning method that is widely used in machine learning. The algorithm solves the same problem by building multiple decision trees and uses a voting algorithm to determine the final classification result. Compared with other methods, random forest has the advantages of being simple to use, strong robustness, and not prone to overfitting. Feature selection is an important task of the random forest algorithm, which measures and ranks the importance of features.

[0038] By analyzing the pseudo-color image results of visible vegetation indices, it can be found that most visible vegetation indices, such as the green leaf index (GLI), the red-green ratio index (RGRI), and the super green red-blue difference index (EGRBDI), show higher sensitivity on plants affected by pests and diseases. In particular, the super green and super red difference index (ExGR) is more sensitive to the identification of healthy plants. However, indices such as the Kawashima index (IKAW) and the Woebbecke index (WI) may not have significant advantages in reflecting the condition of plants. By comprehensively considering the sensitive responses of these vegetation indices to pest-infested plants, the accuracy of pest and disease monitoring can be effectively improved. This comprehensive analysis method can make full use of the complementarity of different vegetation indices, thereby achieving more accurate identification and evaluation of the impact of pests and diseases in remote sensing monitoring.

[0039] Step 104: Process all pre-selected visible light vegetation index images using the Beluga optimization algorithm and multiple threshold segmentation methods to obtain new vegetation index images corresponding to each threshold method.

[0040] Specifically, according to the pre-selected vegetation index, a new vegetation index can be obtained by using the White Whale optimization algorithm and a threshold segmentation method; there are multiple threshold segmentation methods, and multiple new vegetation indexes can be obtained, and each new vegetation index corresponds to the adopted threshold segmentation method.

[0041] The Beluga optimization algorithm (BWO) was applied to a small area of ​​the image, and the CSI index after single threshold segmentation was used as the optimization target. A new vegetation index based on the optimization algorithm was constructed, and the optimized weight was applied to the entire image. The new vegetation index is defined as: the weighted sum of all pre-selected visible light vegetation indices, where the weight of each pre-selected visible light vegetation index is obtained by the Beluga optimization algorithm.

[0042] Beluga whale optimization (BWO) is a population-based metaheuristic algorithm proposed by Zhong in 2022 based on the behavioral logic of beluga whales. The behaviors of whale swimming, predation, and whale falling are simulated by establishing three stages: exploration, exploitation, and whale falling. Beluga whales are regarded as search agents and move in the search space by changing the position vector. In the exploration stage, beluga whales are randomly selected to ensure the global search capability in the scan space. In the exploitation stage, the control algorithm performs a local search in the scan space, and the levy flight method is used in this stage to improve convergence. In the optimization process, after each iteration completes the exploration and exploitation stages, it enters the whale descent stage.

[0043] Threshold segmentation is a region-based image segmentation technique, the principle of which is to divide image pixels into several categories. Image threshold segmentation is a traditional and most commonly used image segmentation method. It has become the most basic and widely used segmentation technique in image segmentation because of its simple implementation, small amount of calculation and relatively stable performance. The purpose of image thresholding is to divide the pixel set according to the grayscale, and each subset obtained forms a region corresponding to the real scene. Each region has consistent attributes, while the adjacent regions do not have such consistent attributes. Such division can be achieved by selecting one or more thresholds based on the grayscale.

[0044] The basic principle of the threshold method is to first determine a threshold, and then divide all pixels into two categories according to the relationship between their eigenvalues ​​and the threshold. When the eigenvalue is greater than the threshold, the pixel is classified as the target class; otherwise, it is classified as the background class. By selecting a suitable threshold, the effective separation of the image target and background can be achieved.

[0045] The BWO optimization algorithm is adopted, RGB images are used, and the information of multiple vegetation indices in different dimensions is integrated to form a new vegetation index, which can achieve higher detection accuracy than a single vegetation index.

[0046] It is worth noting that the White Whale optimization algorithm here can also be replaced by any other optimization algorithm.

[0047] Step 106: Perform image segmentation on each new vegetation index image using a corresponding threshold segmentation method, and process all obtained segmentation results using an integrated strategy based on the Beluga optimization algorithm to obtain crop disease and insect pest detection results in the disease and insect pest monitoring area.

[0048] Specifically, each new vegetation index corresponds to a threshold segmentation method, so when performing segmentation later, the corresponding threshold segmentation method must also be used.

[0049] Ensemble learning trains multiple basic classifiers, combines them in series or parallel, and then uses a combination strategy to complete the learning task in order to reduce variance, bias and improve prediction. The main problem of ensemble learning is the selection of ensemble strategies. The more commonly used ensemble strategies are averaging, voting and learning. The averaging method is suitable for numerical regression prediction problems, while voting and learning are often used for classification tasks. Voting is a commonly used ensemble strategy in ensemble learning recognition tasks. It is also a method that follows the principle of majority, including absolute majority voting, relative majority voting and weighted voting. Given a task, construct an ensemble model F(x i ), assuming that the ensemble contains M base classification results {f1,f2,...,f M}, the classification category with the highest number of votes for each pixel in the result is selected as the final category. This classification strategy is the majority voting strategy.

[0050] For this method, the base classification result is the detection result map obtained by the threshold segmentation method. Since the performance of the base classification results varies, this method introduces a weighted voting method, which uses BWO to weight each base classification result, optimizes the weight based on the fitness function, and then uses the majority voting strategy to obtain the final voting result.

[0051] Based on the new vegetation index, a method of integrating multiple threshold segmentation methods for pest and disease detection is proposed. Through a simple integration strategy of majority voting, better detection performance can be achieved than single threshold segmentation and single vegetation index.

[0052] In the above-mentioned crop pest and disease detection method based on RGB image integrated optimization, the method includes: calculating 24 visible light vegetation indexes based on the RGB images of the pest and disease monitoring area taken by the drone, and normalizing them; using the random forest algorithm to sort the importance of all normalized visible light vegetation indexes, and select the index; using the Beluga optimization algorithm + multiple threshold segmentation methods to process all pre-selected visible light vegetation index images to obtain multiple new vegetation index images; using multiple threshold segmentation methods to segment each new vegetation index image, and using the integrated strategy based on the Beluga optimization algorithm to process all the segmentation results to obtain the pest and disease detection results of crops in the pest and disease monitoring area. This method can effectively reduce the computational complexity while ensuring the detection accuracy by selecting a subset of all vegetation indices through the random forest importance ranking method; using the integrated strategy of the BWO optimization algorithm to improve the detection accuracy and detection precision of pests and diseases.

[0053] In one embodiment, the 24 visible light vegetation indices in step 100 include but are not limited to: super green vegetation index, triangular greenness index, blue-green difference vegetation index, modified super green vegetation index, super red index, super green and super red difference index, vegetation factor, visible light difference vegetation index, vegetation color index, normalized green-red difference index, red-green ratio vegetation index, visible light atmospheric impedance index, improved green-red vegetation index, super green red and blue difference index, red-green-blue ratio vegetation index, enhanced normalized green-blue difference index, green leaf index, over-blue index, kawashima index, comprehensive vegetation index 1, comprehensive vegetation index 2, red-blue ratio vegetation index, Woebbecke index and vegetation index.

[0054] Specifically, R, G, B and 24 vegetation indices were selected for BWO weighting, and different vegetation indices were used to obtain important information about the physiological state and growth status of plants. The reasonable weighting of the borneol camphor vegetation index can improve the accuracy and efficiency of crop (for example: borneol camphor vegetation) disease and insect pest monitoring, and provide scientific basis and technical support for timely detection and prevention of borneol camphor vegetation diseases and insect pests.

[0055] In one embodiment, step 102 includes: in the RGB image, a preset number of pixels in a small area are selected to record the values ​​corresponding to 27 indexes, the 27 values ​​are used as independent variables, the standard disease map division result is used as the dependent variable, the importance value of each feature is obtained by using a random forest algorithm, and the 27 features are sorted from high to low according to the importance value, and the index is selected according to the sorting result to obtain a pre-selected number of visible light vegetation index images. The 27 indexes include R band, G band, B band and 24 visible light vegetation indexes.

[0056] Specifically, in this embodiment, in a 1000*1000 picture, a small area of ​​300*300 points is selected to record the values ​​corresponding to 27 indexes (wherein the 27 indexes include R band, G band, B band and 24 visible light vegetation indices), the 27 indexes are used as independent variables, the standard disease map division result is used as the dependent variable (expressed by 0, 1), and the random forest algorithm is used to obtain the importance value of each feature, and the 27 features are sorted from high to low according to the importance value.

[0057] In one embodiment, step 104 includes: setting the population size to PS, and the initial weight of each whale to <w1,w2,…,w n >; Select the small image of the training area and the corresponding true value area, use the White Whale optimization algorithm, weight the n pre-selected light-sensitive vegetation indices, and obtain PS new vegetation indices; the new vegetation index is:

[0058]

[0059] Among them, VI_new is the new vegetation index; <VI1,VI2,…,VI n > are n preselected visible light vegetation indices, and n is the number of preselected visible light vegetation index images.

[0060] Select a certain threshold segmentation algorithm to segment the new vegetation index, and calculate the accuracy evaluation index CSI by comparing the segmentation result with the corresponding true value image to obtain PS CSI values.

[0061] Taking the precision evaluation index CSI as the fitness function, according to the value of the balance factor Bf, entering the exploration or development stage, updating the behavior of the beluga whale, selecting the optimal individual according to the fall probability Wf, iterating the above process until the number of iterations is reached, and then selecting the optimal individual as the final weight, and applying the weight to the complete image of the n pre-selected vegetation indices to obtain a new vegetation index image corresponding to each threshold method.

[0062] In one embodiment, the multiple threshold segmentation methods in step 104 include but are not limited to: OTSU threshold segmentation algorithm, block OTSU threshold segmentation algorithm, local dynamic threshold method, Wellner adaptive threshold segmentation algorithm, maximum entropy threshold segmentation algorithm, bimodal method, minimum error method, iterative threshold segmentation method and global and local threshold joint segmentation algorithm.

[0063] Specifically, after random forest importance ranking, a certain number of vegetation indices are selected as the input of the next BWO. Assume that the selected n vegetation indices are <VI1,VI2,...,VI n >, the initial weight of each visible light vegetation index is <w1,w2,...,w n >, select the small image of the training area and the corresponding true value area, use the BWO algorithm to weight the selected n vegetation indices, set the population size to PS, and obtain PS new vegetation indices; the new vegetation index is defined as:

[0064]

[0065] Select a threshold segmentation algorithm, such as OTSU, to segment the new vegetation index, and calculate the accuracy evaluation index CSI by comparing the segmentation results with the corresponding true value image. PS CSI values ​​can be obtained. The CSI value is used as the fitness function. According to the value of the balance factor Bf, the behavior of the beluga whale is updated, and the exploration or development stage is entered. The behavior of the beluga whale is updated, and the optimal individual is selected according to the fall probability Wf. The above process is iterated until the number of iterations is reached, so as to select the optimal individual as the final weight, and apply the weight to the complete image of the selected n vegetation indices. The final new vegetation index is obtained according to the new vegetation index formula. It should be pointed out that different threshold segmentation algorithms will result in different weights and new vegetation indices, which also provides basic variables for subsequent integrated learning.

[0066] In this embodiment, the number of beluga whales is set to 5, 10, 15, 20, and 27, the population size is set to 50, the search space range is set to (-10, 10), and the number of iterations is set to 50. Subsequently, 9 threshold segmentation methods are used to segment and integrate the n selected new vegetation indices after BWO optimization to achieve accurate detection of pest-infested trees.

[0067] In one embodiment, step 106 includes: performing image segmentation on each new vegetation index image using a corresponding threshold segmentation method, weighting the detection result images of all new vegetation index threshold segmentation methods using the Beluga optimization algorithm, optimizing the weights based on the fitness function, and then using the majority voting strategy to obtain the final pest and disease detection results of crops in the pest and disease monitoring area.

[0068] Specifically, each threshold segmentation method generates a new vegetation index, and the corresponding vegetation index is segmented using this threshold method to obtain a segmentation result. There are 9 threshold segmentation methods, so there are 9 results, and then the White Whale algorithm is used to perform weighted optimization on the 9 results to obtain the final result.

[0069] Based on a variety of threshold segmentation methods and new vegetation indices, an integrated strategy based on the BWO optimization algorithm was proposed. By adaptively re-setting the weights of the single threshold method, the accuracy of pest and disease detection can be further improved steadily.

[0070] In one of the embodiments, multiple threshold segmentation methods are used to segment each new vegetation index image, the Beluga optimization algorithm is used to weight the detection result map of each threshold segmentation method, the weight is optimized based on the fitness function, and then the majority voting strategy is used to obtain the final pest and disease detection results of crops in the pest and disease monitoring area, including: using 9 threshold segmentation methods to segment the corresponding new vegetation index to obtain 9 base classification results, and selecting 9 base classification results in the sub-area with the same new vegetation index as the input of the Beluga optimization algorithm; setting the balance factor Bf and the fall probability Wf, setting the population size and the number of iterations, using CSI as the fitness function, and updating the weight of each base classification result in each iteration process; for a certain Beluga individual, according to the weights of the 9 base classification results <w1, w2, ..., w9>, each base classification result is copied, and after each base classification result is copied, the image containing The set of classification results is finally formed by adopting the majority voting strategy. Each pixel in the set is voted by the majority to obtain the final integrated result. The CSI between the integrated result and the true value is calculated as the fitness of the individual. The exploration or development stage is decided according to the balance factor Bf. After the falling process, the maximum number of iterations is reached and the optimal individual is selected. The corresponding weight is the final weight. The final weight is applied to the large image of the 9 base classification results, and each pixel is voted by the majority to obtain the final pest and disease detection result of crops in the pest and disease monitoring area.

[0071] It should be understood that although Figure 1 , Figure 2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in the application, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 , Figure 2 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0072] In a verification embodiment, a confusion matrix (also known as a sample-based error matrix) is used as a verification and evaluation method. According to the error matrix table, the situation of the misclassification of pests and diseases is comprehensively analyzed, and the following evaluation indicators are used: hit rate (POD), false alarm rate (FAR) and critical success index (CSI), overall accuracy (Overall Accuracy), Kappa coefficient and other indicators are used to comprehensively evaluate the pest and disease detection performance. The pest and disease detection of this application is a binary classification, so its confusion matrix is ​​shown in Table 1.

[0073] Table 1 Confusion matrix

[0074] Prediction is positive Prediction is negative True value is positive True Class (TP) False Negative (FN) True value is negative False Positive (FP) True Negative (TN)

[0075] 1) CSI (Critical Success Index): Critical success index, also known as accuracy, is used to evaluate the accuracy of identifying positive examples. CSI represents the ratio of correctly classified positive examples to all positive examples.

[0076]

[0077] 2) POD (Probability of Detection): Detection probability, also known as sensitivity or true positive rate (TPR), is used to evaluate the ability to correctly identify positive examples. POD represents the proportion of samples correctly identified as positive examples to all positive examples.

[0078]

[0079] 3) FAR (False Alarm Rate): False alarm rate, also known as false alarm rate, is used to evaluate the error rate of identifying negative examples. FAR represents the proportion of negative samples that are mistakenly identified as positive examples to all negative examples.

[0080]

[0081] 4) Overall Accuracy is a numerical value that expresses the degree to which all objects are correctly classified.

[0082]

[0083] 5) The Kappa coefficient expresses the accuracy of the entire image classification result. The maximum value of the Kappa coefficient is 1. The larger the Kappa coefficient, the higher the degree of consistency between the classification result and the real object.

[0084]

[0085] Where: P0 = OA,

[0086] (1) Feature Importance Analysis

[0087] In order to evaluate the ability of 27 variables (24 vegetation indices and three bands of R, G, and B) to monitor pests and diseases of Camphor camphor, a data set was constructed using a small area of ​​300*300 in the monitoring area, and the 27 variables were ranked by importance using a random forest model. First, 27 independent variable images of the small area were generated, and the true value of the disease map was used as the dependent variable (represented by 0, 1). The importance value of each variable was obtained using the random forest algorithm, and the 27 features were ranked from high to low according to the importance value. The results are as follows Figure 3 As shown:

[0088] From above Figure 3 It can be seen that ExGR and COM1 rank the top two in importance, which is much higher than other indices, which means that these two features play a more important role in classification, followed by 8 indices such as CIVE and VEG, while 17 indices such as B and G are less important, indicating that these indices have limited effects on accurate classification.

[0089] (2) Effectiveness of BWO for single OTSU algorithm

[0090] Random forests were used to select the 10 most important indexes, and the binary classification images of the 10 indexes were obtained based on the OTSU algorithm, which was also the pest and disease detection map. Then the BWO strategy was used to adaptively weight the 10 features, and the OTSU algorithm was used to adaptively segment the new index image to obtain the final pest and disease detection map. Figure 4 Schematic diagrams of the results of segmenting the 10 indices selected by random forest importance using the OTSU algorithm are given respectively, where (a) is the true value map, (b) is the segmentation result of the ExGR index, (c) is the segmentation result of the CIVE index, (d) is the segmentation result of the VI index, (e) is the segmentation result of the VEG index, (f) is the segmentation result of the COM1 index, (g) is the segmentation result of the EGRBDI index, (h) is the segmentation result of the RGBRI index, (i) is the segmentation result of the VDVI index, (j) is the segmentation result of the MGRVI index, (k) is the segmentation result of the GLI index, and (l) is the segmentation result of the new vegetation index after BWO optimization. Figure 4The black area in the middle is the detected diseased and insect-infested plants. It can be seen that compared with the true value map, the original single vegetation index direct segmentation effect is not good overall, and there are very serious misclassifications and false positives. The segmentation method after using 10 indexes BWO weighted is significantly better than the single vegetation index threshold segmentation, which is specifically manifested in a significant reduction in false positives, and the distribution and location of diseased trees are closer to the true value map, such as Figure 4 As shown in (l).

[0091] A variety of accuracy evaluation indicators are calculated for the above detection results, and the results are shown in Table 2 below. It can be seen that the effect of directly performing OTSU segmentation on the index is not very good, and the CSI score is very low, but the POD and FAR are very high, indicating that the segmentation results tend to segment more pixels into pests and diseases, and the low Kappa coefficient also illustrates this point. The OTSU threshold segmentation results based on BWO are significantly better than the results of OTSU segmentation of a single vegetation index in each indicator, indicating that BWO can effectively concentrate the valuable information in each index, so that the generated new index can also distinguish between pest and disease areas and normal areas. The difference in the index is a very important reason for this result. Each index only reflects the characteristics of a certain aspect, so a single index OTSU will have a large number of false alarms, and the results also show the advantages of BWO weighted segmentation.

[0092] Table 2: OTSU classification accuracy table of single index based on random forest importance ranking and new index optimized by BWO

[0093]

[0094]

[0095] (3) Effectiveness of BWO for various threshold segmentation methods

[0096] In order to test the diseased tree detection accuracy of various threshold segmentation algorithms under various index conditions, this embodiment uses 9 threshold segmentation methods to automatically segment the 10 indexes selected based on the random forest importance ranking, and obtains the diseased tree detection results of each index under each threshold segmentation method. Then, for each threshold segmentation method, the average value of the detection accuracy of the 10 indexes under this method is calculated as the average detection accuracy of the threshold segmentation method. At the same time, based on the 10 indexes, the detection accuracy of the comprehensive index after BWO weighting of the 10 indexes is calculated for each threshold segmentation method.

[0097] Figure 5The diagram shows the results of segmenting the 10 vegetation indices selected by random forest using multiple threshold segmentation methods; (a) is the ExGR index map, (b) is the OTSU segmentation result of the ExGR index, (c) is the OTSU segmentation result of the new vegetation index after BWO optimization, (d) is the CIVE index map, (e) is the WAT segmentation result of the CIVE index, (f) is the WAT segmentation result of the new vegetation index after BWO optimization, (g) is the ExB index map, (h) is the BM segmentation result of ExB, and (i) is the BM segmentation result of the new vegetation index after BWO optimization. The individual indexes come from the 10 indices selected by random forest importance ranking. Figure 5 It can be seen that directly using any threshold segmentation method will result in serious noise and serious false alarms, which is mainly reflected in the excessive black areas (that is, diseased trees), which are significantly different from the true value. However, the new index obtained by weighting the 10 indexes BWO can better segment the location of diseased trees using any threshold segmentation method, and is similar to the true value. Although there is a certain amount of noise, the noise intensity is low.

[0098] The quantitative evaluation results are shown in Table 3 below. The average detection accuracy of the nine threshold segmentation methods on the 10 indices selected by random forest importance ranking is calculated, denoted as Th_Avg, and the detection results on the BWO weighted new index of the 10 indices are denoted as Th_BWO. Figure 5 The subjective evaluation results are consistent with those of Th_Avg. For the 10 indices selected by random forest importance ranking, the average precision of each threshold method on all indices is similar. The CSI index value is distributed around 0.30. The lowest is the LDT method, which is only 0.2579, and the highest is WAT, which reaches 0.3974. Except for WAT, the POD values ​​of all other methods exceed 0.95, and even reach 0.99, and the FAR also exceeds 0.6, indicating that there are a large number of false positives in the detection results. The results of Th_BWO are significantly better than those of Th_Avg, with CSI values ​​exceeding 0.55, OA exceeding 0.85, and FAR maintained at around 0.3, indicating that there are still some false positives, but the degree is much lighter than that of Th_Avg.

[0099] Table 3 Quantitative evaluation results

[0100]

[0101] (4) Impact of the number of features on detection results

[0102] In this embodiment, the number of indexes is set to 5, 10, 15, 20, and all features, and 9 threshold segmentation methods are used respectively to calculate the average value of detection accuracy under each number index, recorded as Avg, and then under the same number index, BWO weighted calculation is used to calculate a new index, and 9 threshold segmentation methods are used to calculate the average value of detection accuracy for the new index, recorded as BWO. The influence of the number of features selected by random forest importance on the detection results is shown in the figure. Figure 6 As shown, (a) is CSI, (b) is OA, and (c) is Kappa.

[0103] Depend on Figure 6 It can be seen that different numbers of indexes have some impact on the importance ranking of random forests, but it is relatively stable. In terms of trend, for BWO, the importance ranking of random forests reaches the highest value when the index is 10, and it shows a downward trend as the number of indexes increases, indicating that in this data, the information contained in 10 indexes can better reflect the characteristics of pests and diseases. More indexes may bring certain redundancy and conflicts, resulting in a decrease in detection accuracy. For Avg, in the index of random forest importance ranking, as the index increases, the detection accuracy will show a trend of first increasing and then decreasing. However, regardless of the upward or downward trend, the detection accuracy of Avg always remains at a low level, which is much worse than BWO, which also shows that BWO has obvious advantages.

[0104] (5) Impact of integrated voting strategy

[0105] Ensemble learning can integrate the information of each component to obtain better results. In the previous experiments, this embodiment proves that the new index weighted by BWO can achieve significantly better detection accuracy than a single index on any adaptive threshold segmentation method. In this embodiment, the performance difference between the conventional majority voting strategy and the BWO weighted voting strategy is examined. For the various numbers of indices selected by the random forest importance assessment, 9 threshold segmentation methods are first used to segment each index to obtain a series of segmentation results. For example, 10 indices can obtain 90 segmentation results, suffixed as _Ths, and then the majority voting strategy and BWO weighted voting integration strategy are used to integrate this series of segmentation results to obtain the final detection result. For the new index weighted by BWO, a new index will be obtained for a certain number of indices. 9 threshold segmentation methods can obtain 9 segmentation results, suffixed as _BWO_Ths, and then the majority voting strategy and BWO weighted voting integration strategy are used to obtain the final detection result. CSI, OA and Kappa are used to evaluate the accuracy of the results of multiple strategies. The impact of different integration strategies is shown in the figure. Figure 7As shown; wherein, (a) is the CSI of different integration strategies of RFVI_Ths, (b) is the OA of different integration strategies of RFVI_Ths, (c) is the Kappa of different integration strategies of RFVI_Ths, (d) is the CSI of different integration strategies of RFVI_BWO_Ths, (e) is the OA of different integration strategies of RFVI_BWO_Ths, and (f) is the Kappa of different integration strategies of RFVI_BWO_Ths.

[0106] from Figure 7 It can be seen that, overall, the BWO weighted majority voting strategy has higher accuracy than the direct majority voting in various index numbers. Different numbers of indexes have a great impact on the final accuracy. The highest accuracy of direct index segmentation occurs in the first 5 indexes, while the highest accuracy of BWO weighted new index segmentation occurs in the first 10 indexes, which means that selecting a few indices with the greatest importance can often achieve higher detection accuracy than using all indexes. The use of BWO weighted new indexes can achieve significantly higher accuracy than direct index segmentation, which shows that the BWO weighted new index proposed in this application has obvious advantages. In terms of integrated strategies, the BWO weighted majority voting strategy is superior to the direct majority voting strategy in most index selection methods and quantities, showing certain advantages. It should be pointed out that the optimization target of BWO in this application is the CSI indicator. The definition of the CSI indicator does not focus on the situation where both the true value and the predicted value are TN at the same time, while OA and Kappa will take into account that when CSI reaches the maximum, OA and Kappa may not be optimal. Therefore, the value of direct majority vote in Figures (e) and (f) is higher than the value of BWO weighted majority vote. However, in general, the three accuracy indicators still show a high degree of consistency.

[0107] (6) More application scenarios

[0108] In order to verify the generalization performance of this method, the BWO weighted new index of the 10 indices of RFVI obtained from the previous training and the BWO weighted integration strategy method are applied to the new scene. The detection results of the model in the new scene are shown in the figure. Figure 8 As shown, (a) is the original RGB image of the new scene 1, (b) is the true value image of the new scene 1, (c) is the detection result image of the new scene 1, (d) is the original RGB image of the new scene 2, (e) is the true value image of the new scene 2, and (f) is the detection result image of the new scene 2. Figure 8 It can be seen that, in general, the detection results are highly similar to the true value map, especially the details of the diseased trees. Although there is still some noise, the diseased trees can basically be accurately detected.

[0109] According to the confusion matrix, the five accuracy indicators of CSI, POD, FAR, OA and Kappa of the two scenarios are calculated. The results are shown in Table 7 below:

[0110] Table 7 Detection accuracy of more scenarios

[0111] CSI POD FAR OA Kappa Scenario 1 0.5779 0.7233 0.2580 0.9594 0.7105 Scenario 2 0.5964 0.6901 0.1855 0.9759 0.7346

[0112] Comprehensive analysis of the data from two different scenarios shows that the model trained with the previous dataset shows high-precision recognition capabilities in other scenarios, which is not much different from the accuracy of the previous experiment, and can effectively identify pest and disease areas in RGB images. It is worth noting that the OA (Overall Accuracy) indicator not only evaluates the recognition accuracy of pest and disease areas, but also considers non-pest and disease areas, and the latter is usually more extensive. Therefore, the value of the OA indicator is very high. Overall, despite the challenges of new scenarios, the model still performs well and shows good generalization ability.

[0113] In one embodiment, Fig. 9 As shown, a crop pest and disease detection device based on RGB image integrated optimization is provided, comprising: a visible light vegetation index image generation module, a visible light vegetation index image selection module, a new vegetation index image generation module and a pest and disease detection result detection module, wherein:

[0114] The visible light vegetation index image generation module is used to calculate 24 visible light vegetation indices based on the RGB images of the pest and disease monitoring area taken by the drone, normalize all visible light vegetation indices, and generate corresponding visible light vegetation index images; the visible light vegetation index image is a grayscale image.

[0115] The visible light vegetation index image selection module is used to use the random forest algorithm to sort the importance of all visible light vegetation indices and select the indexes to obtain a number of pre-selected visible light vegetation index images.

[0116] The new vegetation index image generation module is used to process all pre-selected visible light vegetation index images using the White Whale optimization algorithm and multiple threshold segmentation methods to obtain a new vegetation index image corresponding to each threshold method.

[0117] The pest and disease detection result detection module is used to perform image segmentation on each new vegetation index image using the corresponding threshold segmentation method, and process all the segmentation results using an integrated strategy based on the Beluga optimization algorithm to obtain the pest and disease detection results of crops in the pest and disease monitoring area.

[0118] In one embodiment, the 24 visible light vegetation indices in the visible light vegetation index image generation module include but are not limited to: super green vegetation index, triangular greenness index, blue-green difference vegetation index, modified super green vegetation index, super red index, super green and super red difference index, vegetation factor, visible light difference vegetation index, vegetation color index, normalized green-red difference index, red-green ratio vegetation index, visible light atmospheric impedance index, improved green-red vegetation index, super green red and blue difference index, red-green and blue ratio vegetation index, enhanced normalized green-blue difference index, green leaf index, over-blue index, kawashima index, comprehensive vegetation index 1, comprehensive vegetation index 2, red-blue ratio vegetation index, Woebbecke index and vegetation index.

[0119] In one embodiment, the visible light vegetation index image selection module is also used to select a preset number of pixel points in a small area in the RGB image to record the values ​​corresponding to 27 indexes, use the 27 values ​​as independent variables, and the standard disease map division result as the dependent variable, use the random forest algorithm to obtain the importance value of each feature, and sort the 27 features from high to low according to the importance value, select the index according to the sorting result, and obtain a pre-selected number of visible light vegetation index images. The 27 indexes include R band, G band, B band and 24 visible light vegetation indexes.

[0120] In one embodiment, the new vegetation index image generation module is further used to set the population size to PS, and the initial weight of each whale is <w1,w2,…,w n >; Select the small image of the training area and the corresponding true value area, use the White Whale optimization algorithm, weight the n pre-selected light-sensitive vegetation indices, and obtain PS new vegetation indices; the new vegetation index is:

[0121]

[0122] Among them, VI_new is the new vegetation index; <VI1,VI2,…,VI n > are n preselected visible light vegetation indices, and n is the number of preselected visible light vegetation index images.

[0123] Select a certain threshold segmentation algorithm to segment the new vegetation index, calculate the accuracy evaluation index CSI of the segmentation result and the corresponding true value image, and obtain PS CSI values; use the accuracy evaluation index CSI as the fitness function, enter the exploration or development stage according to the value of the balance factor Bf, update the behavior of the beluga whale, select the optimal individual according to the fall probability Wf, iterate the above process until the number of iterations is reached, and then select the optimal individual as the final weight, and apply the weight to the complete image of n vegetation indices to obtain the new vegetation index image corresponding to each threshold method.

[0124] In one embodiment, the new vegetation index image generation module includes multiple threshold segmentation methods including but not limited to: OTSU threshold segmentation algorithm, block OTSU threshold segmentation algorithm, local dynamic threshold method, Wellner adaptive threshold segmentation algorithm, maximum entropy threshold segmentation algorithm, bimodal method, minimum error method, iterative threshold segmentation method and global and local threshold joint segmentation algorithm.

[0125] In one of the embodiments, the pest and disease detection result detection module is also used to perform image segmentation on each new vegetation index image using a corresponding threshold segmentation method, weight the detection result images of all new vegetation index threshold segmentation methods using the Beluga optimization algorithm, optimize the weights based on the fitness function, and then use the majority voting strategy to obtain the final pest and disease detection results of crops in the pest and disease monitoring area.

[0126] In one embodiment, the pest detection result detection module is also used to

[0127] Use 9 threshold segmentation methods to segment the corresponding new vegetation index to obtain 9 basic classification results. Select the 9 basic classification results in the sub-area with the same new vegetation index as the input of the white whale optimization algorithm; set the balance factor Bf and the fall probability Wf, set the population size and the number of iterations, use CSI as the fitness function, and update the weight of each basic classification result in each iteration; for a certain white whale individual, copy each basic classification result according to the weight of the 9 basic classification results <w1,w2,…,w9>, and after copying each base classification result, you can get The set of classification results is finally formed by adopting the majority voting strategy. Each pixel in the set is voted by the majority to obtain the final integrated result. The CSI between the integrated result and the true value is calculated as the fitness of the individual. The exploration or development stage is decided according to the balance factor Bf. After the falling process, the maximum number of iterations is reached and the optimal individual is selected. The corresponding weight is the final weight. The final weight is applied to the large image of the 9 base classification results, and each pixel is voted by the majority to obtain the final pest and disease detection result of crops in the pest and disease monitoring area.

[0128] For the specific definition of the crop disease and pest detection device based on RGB image integrated optimization, please refer to the definition of the crop disease and pest detection method based on RGB image integrated optimization above, which will not be repeated here. Each module in the above-mentioned crop disease and pest detection device based on RGB image integrated optimization can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0129] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A crop disease and insect pest detection method based on RGB image integration optimization, characterized in that: The method comprises: Calculate 24 visible light vegetation indexes based on the RGB image of the pest monitoring area taken by the drone, normalize all the visible light vegetation indexes, and generate corresponding visible light vegetation index images; the visible light vegetation index images are grayscale images; The random forest algorithm is used to sort the importance of all visible light vegetation indices and select the indexes to obtain several pre-selected visible light vegetation index images; According to all pre-selected visible light vegetation index images, the Beluga optimization algorithm and multiple threshold segmentation methods are used to process them, and a new vegetation index image corresponding to each threshold method is obtained; For each new vegetation index image, the corresponding threshold segmentation method is used for image segmentation, and all the segmentation results are processed using an integrated strategy based on the Beluga optimization algorithm to obtain the pest and disease detection results of crops in the pest and disease monitoring area.

2. The crop disease and insect pest detection method based on RGB image integrated optimization according to claim 1, characterized in that: The 24 visible light vegetation indices include but are not limited to: super green vegetation index, triangular greenness index, blue-green difference vegetation index, modified super green vegetation index, super red index, super green and super red difference index, vegetation factor, visible light difference vegetation index, vegetation color index, normalized green-red difference index, red-green ratio vegetation index, visible light atmospheric impedance index, improved green-red vegetation index, super green red and blue difference index, red-green-blue ratio vegetation index, enhanced normalized green-blue difference index, green leaf index, over-blue index, kawashima index, comprehensive vegetation index 1, comprehensive vegetation index 2, red-blue ratio vegetation index, Woebbecke index and vegetation index.

3. The crop disease and insect pest detection method based on RGB image integrated optimization according to claim 1, characterized in that: The random forest algorithm is used to sort the importance of all visible light vegetation indices and select the indexes to obtain several pre-selected visible light vegetation index images, including: In the RGB image, a preset number of pixels in a small area are selected to record the values ​​corresponding to 27 indices. The 27 values ​​are used as independent variables, and the standard disease map division result is used as the dependent variable. The random forest algorithm is used to obtain the importance value of each feature, and the 27 features are sorted from high to low according to the importance value. According to the sorting result, the index is selected to obtain several pre-selected visible light vegetation index images; the 27 indices include R band, G band, B band and 24 visible light vegetation indices.

4. The crop disease and insect pest detection method based on RGB image integrated optimization according to claim 1, characterized in that: According to all pre-selected visible light vegetation index images, the Beluga optimization algorithm and multiple threshold segmentation methods are used to process them, and new vegetation index images corresponding to each threshold method are obtained, including: Set the population size to PS, and the initial weight of each whale to <w1,w2,…,w n >; Select the small image of the training area and the corresponding true value area, use the White Whale optimization algorithm, weight the n pre-selected light-sensitive vegetation indices, and obtain PS new vegetation indices; the new vegetation index is: Among them, VI_new is the new vegetation index; <VI1,VI2,…,VI n > is n pre-selected visible vegetation indexes, n is the number of pre-selected visible vegetation index images; Select a certain threshold segmentation algorithm to segment the new vegetation index, calculate the accuracy evaluation index CSI by comparing the obtained segmentation result with the corresponding true value image, and obtain PS CSI values; Taking the precision evaluation index CSI as the fitness function, according to the value of the balance factor Bf, entering the exploration or development stage, updating the behavior of the beluga whale, selecting the optimal individual according to the fall probability Wf, iterating the above process until the number of iterations is reached, and then selecting the optimal individual as the final weight, and applying the weight to the complete image of the n pre-selected vegetation indices to obtain a new vegetation index image corresponding to each threshold method.

5. The crop disease and insect pest detection method based on RGB image integrated optimization according to claim 1, characterized in that: According to all pre-selected visible light vegetation index images, the Beluga optimization algorithm and multiple threshold segmentation methods are used for processing to obtain a new vegetation index image corresponding to each threshold method. The multiple threshold segmentation methods in the step include but are not limited to: OTSU threshold segmentation algorithm, block OTSU threshold segmentation algorithm, local dynamic threshold method, Wellner adaptive threshold segmentation algorithm, maximum entropy threshold segmentation algorithm, bimodal method, minimum error method, iterative threshold segmentation method and global and local threshold joint segmentation algorithm.

6. The crop disease and insect pest detection method based on RGB image integrated optimization according to claim 1, characterized in that: Each new vegetation index image is segmented using the corresponding threshold segmentation method. The segmentation results are processed using an integrated strategy based on the Beluga optimization algorithm to obtain the pest and disease detection results of crops in the pest and disease monitoring area, including: For each new vegetation index image, the corresponding threshold segmentation method is used for image segmentation. The Beluga optimization algorithm is used to weight the detection result images of all new vegetation index threshold segmentation methods. The weights are optimized based on the fitness function, and the majority voting strategy is used to obtain the final pest and disease detection results of crops in the pest and disease monitoring area.

7. The crop disease and insect pest detection method based on RGB image integrated optimization according to claim 6, characterized in that: Each new vegetation index image is segmented using the corresponding threshold segmentation method. The Beluga optimization algorithm is used to weight the detection result images of all new vegetation index threshold segmentation methods. The weights are optimized based on the fitness function. The majority voting strategy is then used to obtain the final pest and disease detection results of crops in the pest and disease monitoring area, including: Use 9 threshold segmentation methods to segment the corresponding new vegetation index, obtain 9 basic classification results, and select the 9 basic classification results in the sub-area with the same new vegetation index as the input of the Beluga optimization algorithm; Set the balance factor Bf and the fall probability Wf, set the population size and the number of iterations, use CSI as the fitness function, and update the weight of each base classification result in each iteration; For a certain beluga individual, according to the weights of the 9 basic classification results <w1,w2,…,w9>, each basic classification result is copied, and after each basic classification result is copied, the result containing The set of classification results is finally formed by using the majority voting strategy to perform the majority vote on each pixel in the set to obtain the final integrated result. Calculate the CSI between the integrated result and the true value as the fitness of the individual; According to the balance factor Bf, it is decided to enter the exploration or development stage. After the falling process, the maximum number of iterations is reached, and the optimal individual is selected. The corresponding weight is the final weight; The final weights are applied to the large image of the 9 base classification results, and the majority vote is performed on each pixel to obtain the final pest and disease detection results of crops in the pest and disease monitoring area.

8. A crop pest detection device based on RGB image integration optimization, characterized in that: The device comprises: A visible light vegetation index image generation module is used to calculate 24 visible light vegetation indices based on the RGB image of the pest monitoring area taken by the drone, normalize all the visible light vegetation indices, and generate a corresponding visible light vegetation index image; the visible light vegetation index image is a grayscale image; A visible light vegetation index image selection module is used to use a random forest algorithm to sort the importance of all visible light vegetation indices and select the indexes to obtain a number of pre-selected visible light vegetation index images; A new vegetation index image generation module is used to process all pre-selected visible light vegetation index images using the White Whale optimization algorithm and multiple threshold segmentation methods to obtain a new vegetation index image corresponding to each threshold method; The pest and disease detection result detection module is used to perform image segmentation on each new vegetation index image using the corresponding threshold segmentation method, and process all the segmentation results using an integrated strategy based on the Beluga optimization algorithm to obtain the pest and disease detection results of crops in the pest and disease monitoring area.

Images