Cotton aphid infringement intelligent monitoring method based on multi-source remote sensing fusion data

Abstract

The invention provides a cotton aphid infringement intelligent monitoring method based on multi-source remote sensing fusion data, and relates to the field of pest control, and the method comprises the steps: obtaining the multispectral and panchromatic image data of a cotton field through an unmanned aerial vehicle; performing field investigation and collection on the cotton aphids; preprocessing and fusing the remote sensing image data of the unmanned aerial vehicle; constructing a spectrum index of cotton aphid analysis, and obtaining an absolute spectrum index in the fused image through correlation analysis; constructing a monitoring model of the cotton aphids; and establishing a deep learning model, and carrying out auxiliary verification analysis by using the deep learning model. According to the method, the instability of a single data source is made up by combining the spectral features of the multispectral image and the spatial details of the panchromatic image, the recognition capability and monitoring precision of the cotton aphid are improved, and important support is provided for precision agriculture and green agriculture.

Description

Technical Field

[0001] The present invention relates to the technical field of pest control, and in particular to an intelligent monitoring method for cotton aphid infestation based on multi-source remote sensing fusion data. Background Art

[0002] Cotton is an important economic crop in my country and is widely used in the textile, chemical and other fields. During the cotton cultivation process, pests and diseases are one of the main factors hindering the improvement of the quality and efficiency of the industry. Among them, aphids are the main pests and diseases that restrict the development of the cotton industry. The yield losses and infestation areas caused by them far exceed those of other pests. Aphids have rapid outbreaks, strong migration, and are easily affected by the environment. They suck plant sap through their piercing-sucking mouthparts, which leads to crop yield reduction and virus transmission. Currently, the monitoring of aphids in the cotton cultivation process mainly relies on manual labor, which not only leads to high costs and waste of resources, but is also easily interfered with by human subjective factors. The monitoring accuracy is poor, the errors are large, and the efficiency is low.

[0003] In recent years, monitoring technologies including satellite remote sensing, UAV remote sensing and ground vision have been widely used in the agricultural field. Among them, satellite remote sensing has the advantages of large range, real-time, objectivity and high efficiency. It can provide continuous farmland data and support the monitoring of pests and diseases, crop growth, etc. However, its spatial resolution limitations, weather and light influences, high cost and difficulty in ground data verification have seriously restricted its application in precise monitoring and early warning. UAV remote sensing can achieve accurate monitoring of pests and diseases, but its single data source has weak generalization ability in monitoring different environments and aphid damage, which leads to unstable monitoring results. Ground vision monitoring relies on manual or fixed equipment to collect images. Although it has high precision and convenience, its monitoring range is limited and efficiency is low. It is difficult to cover large areas of farmland and has strong limitations for large-scale cotton planting. Summary of the Invention

[0004] In response to the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide an intelligent monitoring method for cotton aphid infestation based on multi-source remote sensing fusion data. This method combines the spectral characteristics of multispectral images and the spatial details of panchromatic images to compensate for the instability of a single data source, thereby constructing a more accurate aphid infestation monitoring model. At the same time, it combines the bottom measured data for auxiliary verification and analysis, effectively improving the recognition ability and monitoring accuracy of pests and diseases.

[0005] The purpose of the present invention is achieved through the following technical solutions: An intelligent monitoring method for cotton aphid infestation based on multi-source remote sensing fusion data, comprising: Step S1: acquiring multispectral image data and panchromatic image data of the cotton field through a UAV; Step S2: Conducting field surveys and collecting data on the degree of damage caused by cotton aphids in cotton fields; Step S3: pre-processing the UAV remote sensing image data and fusing the multispectral image data with the panchromatic image data to obtain fused image data; Step S4: constructing a spectral index for cotton aphid analysis and obtaining the absolute spectral index in the fused image through correlation analysis; Step S5: constructing a monitoring model for cotton aphids and predicting the aphid damage level based on the absolute spectral index; Step S6: Establish a deep learning model, use the deep learning model to predict and identify ground measured samples, and compare them with the monitoring model in step S5.

[0006] Based on further optimization of the above scheme, a dual-camera ten-channel imaging system with multispectral and high-definition visible light sensors is set on the drone in step S1 to simultaneously obtain impact data and panchromatic image data of ten bands including visible light and near-infrared.

[0007] Based on the further optimization of the above scheme, in step S2, a 50×50 cm marking frame was used to survey the severity of aphids at each sampling point. At least 5 cotton plants were surveyed in each sampling point, and the aphid disease index in each sampling point was calculated. :

[0008] Where: n Indicates the level of each aphid disease; f Indicates the number of plants at each level; N Indicates the highest aphid damage level; The severity of the disease is divided according to the aphid disease index: Normal (Level 0): ; Mild (Grade 1): ; Moderate (Grade 2): ; Severe (Grade 3): ; Extremely severe (Grade 4): .

[0009] Based on the further optimization of the above solution, the preprocessing of the UAV remote sensing image data in step S3 includes camera calibration and image alignment; After importing the spectral image, calibrate the camera first: First, before the drone takes off, place a high-reflectivity multispectral calibration plate horizontally on flat ground. Use the drone's onboard multispectral camera to capture images of the calibration plate from different heights (e.g., 5m, 10m, and 15m), collecting at least two sets of images for each band (e.g., red, green, and NIR). Then, local feature points in the image are extracted, feature descriptors are used for feature matching, and the basic matrix or homography matrix is fitted using the RANSAC algorithm. The internal points are screened and the overlapping areas are determined to obtain the coefficient matching points. Then, based on the sparse matching points, a multi-view stereo matching algorithm is used to calculate the disparity map through epipolar geometric constraints, combined with camera parameters and triangulated to generate a dense 3D point cloud; Finally, the dense 3D point cloud data is processed to generate a digital surface model and orthophoto. Based on the elevation data of the digital surface model, the image pixels are resampled and combined with the geographic coordinate mapping of the orthophoto to achieve pixel-level geometric correction to ensure that each pixel is consistent with the actual geographic coordinates. Export the image for further processing: First, before the flight of the drone, a diffuse reflectance reference plate is photographed to eliminate the influence of environmental factors (atmosphere, sensor, and other) on the image radiation value and complete amplitude correction; Then, geometric correction is performed to accurately align the image with the ground control points. The control points with the same name are manually marked in the image and the actual geographic coordinate system to obtain their pixel coordinates and geographic coordinates. The image is resampled using the solved transformation parameters, and the pixels are mapped to the geographic coordinate system. After removing outliers, the parameters are iteratively optimized to ensure accurate alignment between the image and the ground coordinate system. After that, the image is subjected to pre-processing operations such as noise removal and filtering; Finally, the pre-processed image is cropped to obtain the image data of the experimental area, and the ground measurement area is marked by drawing the region of interest (ROI) to provide accurate data for subsequent spectral index calculation and other analyses.

[0010] Based on further optimization of the above solution, the digital surface model is generated as follows: For dense point cloud data, the grid resolution is defined as d , then for each grid center point ( X g ,Y g ), select its k Nearest neighbor point cloud data ( X i ,Y i ,Z i ), get its elevation value Z g :

[0011] Where: , p represents the power parameter (usually 2); Thus a digital surface model with regular grid is generated; Based on the elevation data of the digital surface model, the image pixels are resampled as follows:

[0012] Where: X 0 ,Y 0 ,Z 0) represents the projection center coordinate; f Indicates the focal length of the camera; ( x 0 ,y 0) represents the coordinates of the principal point; R =[ r ij ] represents the rotation matrix, obtained through the attitude angle; ( x,y ) represents the original image coordinates; like( x,y ) in the original image coordinates, get its pixel value V :

[0013] Where: dx 、 dy represent the normalized distances to neighboring pixels, V ij Represents the values of the four surrounding pixels; Map the resampled orthophoto to the target coordinate system:

[0014] Where: u,v ) represents the pixel coordinates of the orthophoto; ( , ) represents geographic coordinates; parameter a 0~ a 5. b 0~ b 5 Solve by ground control points.

[0015] Based on the further optimization of the above solution, the iterative optimization parameters after removing the outliers are specifically: First, perform outlier detection and calculate the predicted geographic coordinates for each ground control point. With actual coordinates The residual between:

[0016] Preset residual threshold e 0, if the residual e i If it is greater than the residual threshold, it is marked as an outlier; Calculate the residuals of all ground control points, count the number of outliers, remove outliers, and iterate the process of calculating and removing outliers. Q The model with the most inliers is retained.

[0017] Based on further optimization of the above solution, the specific steps for completing the fusion of multispectral image data and panchromatic image data in step S3 are: Principal component analysis is first used to extract principal components from multispectral images, eliminate redundant information, and ensure the independence of principal components through the Gram-Schmidt orthogonalization process; Extracting principal components from multispectral images: Assume that the multispectral image is a matrix M∈R h×w , where: h is the number of bands and w is the number of pixels;

[0018] Where: V represents the orthogonal eigenvector matrix, 1 w express w ×1 all-1 vector; Orthogonalize the principal components:

[0019] Where: represents the magnitude of a vector; represents the inner product of vectors; Then, the high spatial resolution information of the panchromatic image is introduced into the first principal component of the multispectral image. The degree of fusion is controlled by the fusion coefficient, and the panchromatic image and the first principal component of the multispectral image are fused by linear weighting:

[0020] Where: represents the fusion coefficient, ; Pan norm Indicates the normalized value of the panchromatic image;

[0021] Where: 、 represent the mean and standard deviation of the panchromatic image, respectively; 、 They represent the mean and standard deviation of the first principal component after Gram-Schmidt processing; Pan Represents the pixel value vector of the panchromatic image; Afterwards, the adjusted first principal component is orthogonalized with the other principal components using Gram-Schmidt:

[0022] Where: HR sim represents high-resolution simulation data; LR k Indicates the k An estimate or projection vector of low-resolution data; GS k Indicates the k orthogonal basis vectors generated by the Gram-Schmidt orthogonalization process; Then orthogonalize multiple GS n Reconstruct the multispectral image through inverse PCA transformation:

[0023] Finally, the inverse PCA transform is used to restore the fused principal components to the original image space to obtain the final high-resolution fused image.

[0024] Based on further optimization of the above scheme, the spectral indices for cotton aphid analysis include ten indices: ARI (Atmospheric Resistance Index), GLI (Green Leaf Index), GBI (Green Biome Index), RVI (Ratio Vegetation Index), DVI (Difference Vegetation Index), ARVI (Atmospherically Resistant Vegetation Index), GNDVI (Green Normalized Difference Vegetation Index), SAVI (Soil Adjusted Vegetation Index), SIPI (Structure Insensitive Pigment Index), and TCARI (Transformed Chlorophyll Absorption in Reflectance Index).

[0025] Based on the further optimization of the above scheme, the absolute spectral index in the fused image is obtained by correlation analysis as follows: the ten spectral indices of the target area in the remote sensing image and the aphid damage degree measured on the spot are subjected to Pearson correlation analysis using R language:

[0026] Where: x 、 y represent two variables, namely, the ten spectral indices of the target area and the degree of aphid damage measured on the spot; n represents the sample size; 、 Represent the means of the two variables respectively; The significance threshold and the correlation strength threshold are set, and the four indices with the highest absolute values of the correlation coefficients are obtained by plotting the correlation heat map, which are the spectrum indices.

[0027] Based on further optimization of the above scheme, the significance threshold is 0.05. In the Pearson correlation analysis performed by R language, the significance is less than the significance threshold; the correlation strength threshold is 0.3. In the Pearson correlation analysis performed by R language, the correlation strength is not less than the correlation strength threshold.

[0028] Based on further optimization of the above solution, step S5 is specifically as follows: Step S51: normalize the data according to the absolute spectral index obtained in step S4:

[0029] Where: B represents the input feature value; represents the feature mean; represents the standard deviation; The normalized dataset is divided into a training set and a test set (divided in a ratio of 8:2); Step S52: Build a monitoring model: First, initialize the model, and the initial prediction value is the target variable y The mean of:

[0030] Then, get the monitor model:

[0031] Where: R jm Represents the division of the regression tree into J m Leaf nodes of a region; m Indicates the number of model iterations; mx Indicates the maximum number of iterations; I represents the exponential function, when x Belong to the region R jm hour, I is 1, otherwise 0; Represents the model learning rate.

[0032] Based on further optimization of the above solution, the maximum number of iterations mx is 500 times, and the model learning rate is 0.1, and the regression tree depth is 3.

[0033] Based on further optimization of the above scheme, the deep learning model adopts the pre-trained YOLOv8 model.

[0034] The following are the technical effects of the technical solution of the present invention: This method uses the fusion of multispectral images and panchromatic images to conduct remote sensing monitoring of aphids in cotton fields. By combining the high spatial resolution of panchromatic images with the rich spectral characteristics of multispectral images, it achieves accurate identification of cotton aphid infestation areas and a comprehensive assessment of plant health status, effectively avoiding the limitations, instability and lag of a single image source, significantly improving monitoring accuracy and reliability, and completing accurate, real-time precision measurement and early warning of aphids in cotton fields. At the same time, this method further optimizes the monitoring performance of aphid severity and further improves the accuracy of cotton aphid monitoring by constructing a monitoring model based on fused images and combining it with the auxiliary verification of the YOLOv8 deep learning model.

[0035] This method provides a more efficient and accurate technical means for remote sensing monitoring of agricultural pests and diseases, and can effectively guide plant protection equipment to achieve precise prevention and control, thus providing important support for precision agriculture and green agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is the cotton field image acquisition system and ground data acquisition process of the present invention; wherein, Figure 1 (a) DJI M300 RTK flight monitoring platform; Figure 1 (b) Manual sampling and Qianxun RTK GNSS positioning.

[0037] Figure 2 These are cotton leaves under different insect pest levels; Figure 2 (a) is the 0-level blade, Figure 2 (b) is the first-stage blade, Figure 2 (c) is the second-stage blade, Figure 2 (d) is the 3rd level blade, Figure 2 (e) is a 4-stage blade.

[0038] Figure 3 Schematic diagram of the fusion of multispectral image data and panchromatic image data in the present invention.

[0039] Figure 4 It is the correlation index graph of the fused image of the present invention.

[0040] Figure 5 Schematic diagram of the structure evaluation of the machine learning model for fusion-affected image data of the present invention.

[0041] Figure 6 This is a graph showing the prediction results of the cotton sub-sea index of the present invention.

[0042] Figure 7 Flowchart of the monitoring method of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention. In the following description, for the purpose of illustration rather than limitation, specific details such as specific system structures and technologies are provided to facilitate a thorough understanding of the embodiments of the present invention.

[0044] Example 1: An intelligent monitoring method for cotton aphid infestation based on multi-source remote sensing fusion data, comprising: Step S1: Obtain multispectral image data and panchromatic image data of the cotton field through a drone: The drone is equipped with a dual-camera ten-channel imaging system with multispectral and high-definition visible light sensors to simultaneously obtain impact data and panchromatic image data in ten bands including visible light and near-infrared.

[0045] In this embodiment, a data acquisition platform consisting of a DJI M300RTK industrial application-grade drone, a RedEdge-MX Dual10 multispectral camera, and a DJI ZENMUSE P1 high-definition visible light camera is used. Figure 1 As shown in (a), the multispectral camera can simultaneously acquire data from 10 wavelengths, including visible light and near-infrared (444nm, 475nm, 531nm, 560nm, 650nm, 668nm, 705nm, 717nm, 740nm, and 842nm). To account for the impact of weather on data collection, data collection was performed at noon on a clear day with a wind speed of no more than 4m / s. Multispectral data was collected at an altitude of 50m, and the camera was calibrated before takeoff by placing it 1m above a calibration reflector. Panchromatic data was collected at an altitude of 20m, with both heading and lateral overlap rates of 80%.

[0046] Step S2: Conduct field surveys and collect data on the severity of cotton aphid damage in cotton fields: Use a 50×50 cm marking frame to survey the severity of aphid damage at sample points (e.g. Figure 1 (b) In each sampling point, at least 5 cotton plants were surveyed. The severity of aphids was classified according to the national standard "Technical Specifications for Cotton Aphid Monitoring and Forecasting" (GB / T 15799-2011, classification standards in Table 1 below). The aphid disease index in each sampling point was calculated. :

[0047] Where: n Indicates the level of each aphid disease; f Indicates the number of plants at each level; N Indicates the highest aphid damage level; The severity of the disease is divided according to the aphid disease index: Normal (Level 0): ; Mild (Grade 1): ; Moderate (Grade 2): ; Severe (Grade 3): ; Extremely severe (Grade 4): .

[0048] Table 1. Classification criteria for the degree of damage caused by cotton aphids:

[0049] Step S3: preprocessing the UAV remote sensing image data, including camera calibration and image alignment; Step S31: After importing the spectral image, calibrate the camera first: First, before the drone takes off, place a high-reflectivity multispectral calibration plate horizontally on flat ground. Use the drone's onboard multispectral camera to capture images of the calibration plate from different heights (e.g., 5m, 10m, and 15m), collecting at least two sets of images for each band (e.g., red, green, and NIR). Then, local feature points in the image are extracted, feature descriptors are used for feature matching, and the basic matrix or homography matrix is fitted using the RANSAC algorithm. The inliers are screened and the overlapping areas are determined to obtain the coefficient matching points. Specifically: Construct a Gaussian difference pyramid on the image and obtain a Gaussian blurred image:

[0050] Where: k represents the scale factor of adjacent scales; represents the scale parameter; I(x,y) represents the input image; For each pixel in the Gaussian blurred image, in space (x,y) and scale Fit on to get the extreme points:

[0051] Where: H Represents the second-order derivative Hessian matrix; represents the first-order derivative vector; Remove points whose response values are lower than the preset threshold and retain local extreme points as candidate points; For each key point area, get its gradient magnitude and direction :

[0052] Where: L(x,y) Gaussian blurred image representing the size of the key points; Centered on the key point, the neighborhood is divided into multiple sub-regions. The 8-directional gradient histogram is calculated for each sub-region to obtain a multi-dimensional descriptor vector, and the descriptor is normalized (to enhance illumination invariance). Feature matching is achieved through Euclidean distance.

[0053] Then, based on the sparse matching points, a multi-view stereo matching algorithm is used to calculate the disparity map through epipolar geometric constraints, combined with camera parameters and triangulated to generate a dense 3D point cloud; Finally, the dense 3D point cloud data is processed to generate a digital surface model and orthophoto. The specific steps of generating the digital surface model are as follows: For dense point cloud data, the grid resolution is defined as d , then for each grid center point ( X g ,Y g ), select its k Nearest neighbor point cloud data ( X i ,Y i ,Z i ), get its elevation value Z g :

[0054] Where: , p represents the power parameter (usually 2); Thus a digital surface model with regular grid is generated; Based on the elevation data of the digital surface model, the image pixels are resampled as follows:

[0055] Where: X 0 ,Y 0 ,Z 0) represents the projection center coordinate; f Indicates the focal length of the camera; ( x 0 ,y 0) represents the coordinates of the principal point; R =[r ij ] represents the rotation matrix, obtained through the attitude angle; ( x,y ) represents the original image coordinates; like( x,y ) in the original image coordinates, get its pixel value V :

[0056] Where: dx 、 dy represent the normalized distances to neighboring pixels, V ij Represents the values of the four surrounding pixels; Combined with the geographic coordinate mapping of the orthophoto, pixel-level geometric correction is achieved to ensure that each pixel is consistent with the actual geographic coordinates; that is, the resampled orthophoto is mapped to the target coordinate system:

[0057] Where: u,v ) represents the pixel coordinates of the orthophoto; ( , ) represents geographic coordinates; parameter a 0~ a 5. b 0~ b 5 Solve by ground control points.

[0058] Step S32: The exported image is further processed: First, before the flight of the drone, a diffuse reflectance reference plate is photographed to eliminate the influence of environmental factors (atmosphere, sensor, and other) on the image radiation value and complete amplitude correction; Then, geometric correction is performed to accurately align the image with the ground control points. The same-name control points are manually marked in the image and the actual geographic coordinate system to obtain their pixel coordinates and geographic coordinates. The image is resampled using the solved transformation parameters, and the pixels are mapped to the geographic coordinate system. After removing outliers, the parameters are iteratively optimized to ensure accurate alignment between the image and the ground coordinate system. The specific parameters for iterative optimization after removing outliers are as follows: Perform outlier detection and calculate the predicted geographic coordinates for each ground control point With actual coordinates The residual between:

[0059] Preset residual threshold e 0 (set according to actual situation), if the residual e i If it is greater than the residual threshold, it is marked as an outlier; Calculate the residuals of all ground control points, count the number of outliers, remove outliers, and iterate the process of calculating and removing outliers. Q Second, keep the model with the most inliers; Iterations:

[0060] Where: p represents the confidence level (in this embodiment, p =0.99); Indicates the predicted value of the abnormal point ratio; s Indicates the subset size (for example: s =6); After that, the image is subjected to pre-processing operations such as noise removal and filtering; Finally, the pre-processed image is cropped to obtain the image data of the experimental area, and the ground measurement area is marked by drawing the region of interest (ROI) to provide accurate data for subsequent spectral index calculation and other analyses.

[0061] Because panchromatic and multispectral data are collected from different sensors, the coordinate information of the panchromatic and multispectral images deviates. Therefore, coordinate alignment and image registration are required. The high-resolution panchromatic image is then fused with the multispectral image to obtain high-precision multispectral information. First, 127 feature point pairs are identified in each panchromatic and multispectral image for image registration, with the RMSE error kept within 3 mm. Second, a cubic convolution interpolation algorithm is used for image registration. Multispectral image pixel positions are mapped to the panchromatic image coordinate system based on affine parameters. A cubic convolution kernel function is used to calculate weights for the 4×4 pixel neighborhood of the target location. Neighboring pixel values are then superimposed according to the weights to generate a geometrically aligned image. Finally, the coordinate-aligned panchromatic and MSI datasets are obtained for image fusion processing.

[0062] Step S33: Complete the fusion of multispectral image data and panchromatic image data to obtain fused image data; the specific steps are: Principal component analysis is first used to extract principal components from multispectral images, eliminate redundant information, and ensure the independence of principal components through the Gram-Schmidt orthogonalization process; Extracting principal components from multispectral images: Assume that the multispectral image is a matrix M∈R h×w , where: h is the number of bands and w is the number of pixels;

[0063] Where: V represents the orthogonal eigenvector matrix, 1 w express w×1 all-1 vector; Orthogonalize the principal components:

[0064] Where: represents the magnitude of a vector; represents the inner product of vectors; Then, the high spatial resolution information of the panchromatic image is introduced into the first principal component of the multispectral image. The degree of fusion is controlled by the fusion coefficient, and the panchromatic image and the first principal component of the multispectral image are fused by linear weighting:

[0065] Where: represents the fusion coefficient, ; Pan norm Indicates the normalized value of the panchromatic image;

[0066] Where: 、 represent the mean and standard deviation of the panchromatic image, respectively; 、 They represent the mean and standard deviation of the first principal component after Gram-Schmidt processing; Pan Represents the pixel value vector of the panchromatic image; Afterwards, the adjusted first principal component is orthogonalized with the other principal components using Gram-Schmidt:

[0067] Where: HR sim represents high-resolution simulation data; LR k Indicates the k An estimate or projection vector of low-resolution data; GS k Indicates the k orthogonal basis vectors generated by the Gram-Schmidt orthogonalization process; Then orthogonalize multiple GS n Reconstruct the multispectral image through inverse PCA transformation:

[0068] Finally, the inverse PCA transform is used to restore the fused principal components to the original image space to obtain the final high-resolution fused image, as shown in Figure 3 shown.

[0069] Step S4, constructing spectral indices for cotton aphid analysis, including ten indices: ARI (Atmospheric Resistance Index), GLI (Green Leaf Index), GBI (Green Biome Index), RVI (Ratio Vegetation Index), DVI (Difference Vegetation Index), ARVI (Atmospherically Resistant Vegetation Index), GNDVI (Green Normalized Difference Vegetation Index), SAVI (Soil Adjusted Vegetation Index), SIPI (Structure Insensitive Pigment Index), and TCARI (Transformed Chlorophyll Absorption in Reflectance Index);

[0070] The absolute spectral index in the fused image was obtained through correlation analysis: the ten spectral indices of the target area in the remote sensing image and the aphid damage degree measured on the spot were subjected to Pearson correlation analysis using R language:

[0071] Where: x 、 y represent two variables, namely, the ten spectral indices of the target area and the degree of aphid damage measured on the spot; n represents the sample size; 、 Represent the means of the two variables respectively; The significance threshold and correlation strength threshold are set. The significance threshold is 0.05. In the Pearson correlation analysis using R language, the significance is less than the significance threshold. The correlation strength threshold is 0.3. In the Pearson correlation analysis using R language, the correlation strength is not less than the correlation strength threshold. By plotting the correlation heat map, the four indices with the highest absolute values of the correlation coefficients (such as Figure 4 As shown in the figure, the top four indices with the highest absolute values of correlation coefficients are GLI (0.89), RVI (0.83), DVI (0.79) and SAVI (0.78), which are the spectral indices.

[0072] Step S5: constructing a monitoring model for cotton aphids and predicting the aphid damage level based on the absolute spectral index; specifically: Step S51: normalize the data according to the absolute spectral index obtained in step S4:

[0073] Where: B represents the input feature value; represents the feature mean; represents the standard deviation; The normalized dataset is divided into a training set and a test set (divided in a ratio of 8:2); Step S52: Build a monitoring model: First, initialize the model, and the initial prediction value is the target variable y The mean of:

[0074] Then, get the monitor model:

[0075] Where: R jm Represents the division of the regression tree into J m leaf nodes of the region (in this embodiment, the regression tree depth is 3); m Indicates the number of model iterations; mx Indicates the maximum number of iterations (in this embodiment, the maximum number of iterations mx 500 times); I represents the exponential function, when x Belong to the region R jm hour, I is 1, otherwise 0; Indicates the model learning rate (in this embodiment, the model learning rate is 0.1).

[0076] The final prediction results of the monitoring model are as follows: Figure 5 As shown, the color gradually deepens from yellow to red, which intuitively reflects the changing trend of the degree of cotton aphid damage in different spatial locations and clearly shows the spatial distribution characteristics of aphid damage in the region.

[0077] Step S6: Establish a deep learning model. The deep learning model uses a pre-trained YOLOv8 model (a conventional YOLOv8 model can be used, which is not limited in this embodiment). The deep learning model is used to predict and identify ground-measured samples and compare them with the monitoring model in step S5 (the results show that the model monitoring accuracy is over 90%; in most aphid-aggregated areas, the aphid distribution predicted by the drone image is basically consistent with the actual results of the YOLOv8 ground monitoring, such as Figure 6 shown).

[0078] Example 2: A cotton aphid infestation intelligent monitoring system includes an industrial drone equipped with a multispectral camera and a high-definition visible light camera, and adopts the cotton aphid infestation intelligent monitoring method based on multi-source remote sensing fusion data as described in Example 1.

Claims

1. An intelligent monitoring method for cotton aphid infestation based on multi-source remote sensing fusion data, characterized by: include: Step S1: acquiring multispectral image data and panchromatic image data of the cotton field through a UAV; Step S2: Conducting field surveys and collecting data on the degree of damage caused by cotton aphids in cotton fields; Step S3: pre-processing the UAV remote sensing image data and fusing the multispectral image data with the panchromatic image data to obtain fused image data; Step S4: constructing a spectral index for cotton aphid analysis and obtaining the absolute spectral index in the fused image through correlation analysis; Step S5: constructing a monitoring model for cotton aphids and predicting the aphid damage level based on the absolute spectral index; Step S6: Establish a deep learning model, use the deep learning model to predict and identify ground measured samples, and compare them with the monitoring model in step S5.

2. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 1, characterized in that: In step S1, the drone is provided with a dual-camera ten-channel imaging system, comprising multispectral and high-definition visible light sensors, for simultaneously acquiring impact data and panchromatic image data in ten bands including visible light and near-infrared.

3. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 1 or 2, characterized in that: In step S2, a 50×50 cm marking frame is used to survey the severity of aphids at each sampling point. At least 5 cotton plants are surveyed in each sampling point, and the aphid disease index in each sampling point is calculated. : Where: n Indicates the level of each aphid disease; f Indicates the number of plants at each level; N Indicates the highest aphid damage level; The severity of the disease is divided according to the aphid disease index: normal: Mild: Moderate: ;serious: ; Extremely severe: .

4. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 1 or 3, characterized in that: In step S3, the preprocessing of the UAV remote sensing image data includes camera calibration and image alignment; After importing the spectral image, calibrate the camera first: First, before the drone takes off, a high-reflectivity multispectral calibration plate is placed horizontally on flat ground. The multispectral camera onboard the drone captures images of the calibration plate from different heights, collecting at least two sets of images for each band. Then, local feature points in the image are extracted, feature descriptors are used for feature matching, and the basic matrix or homography matrix is fitted using the RANSAC algorithm. The internal points are screened and the overlapping areas are determined to obtain the coefficient matching points. Then, based on the sparse matching points, a multi-view stereo matching algorithm is used to calculate the disparity map through epipolar geometric constraints, combined with camera parameters and triangulated to generate a dense 3D point cloud; Finally, the dense 3D point cloud data is processed to generate a digital surface model and orthophoto. Based on the elevation data of the digital surface model, the image pixels are resampled and combined with the geographic coordinate mapping of the orthophoto to achieve pixel-level geometric correction to ensure that each pixel is consistent with the actual geographic coordinates. Export the image for further processing: First, a diffuse reflectance reference plate is photographed before the drone takes flight to eliminate the impact of environmental factors on the image radiation value and complete amplitude correction; Then, geometric correction is performed to accurately align the image with the ground control points. The control points with the same name are manually marked in the image and the actual geographic coordinate system to obtain their pixel coordinates and geographic coordinates. The image is resampled using the solved transformation parameters, and the pixels are mapped to the geographic coordinate system. After removing outliers, the parameters are iteratively optimized to ensure accurate alignment between the image and the ground coordinate system. After that, the image is subjected to pre-processing operations such as noise removal and filtering; Finally, the pre-processed image is cropped to obtain the image data of the experimental area, and the ground measurement area is marked by drawing the region of interest to provide accurate data for subsequent spectral index calculation and other analyses.

5. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 4, characterized in that: The generating of the digital surface model is specifically as follows: For dense point cloud data, the grid resolution is defined as d , then for each grid center point ( X g ,Y g ), select its k Nearest neighbor point cloud data ( X i ,Y i ,Z i ), get its elevation value Z g : Where: , p represents the power parameter; Thus a digital surface model with regular grid is generated; Based on the elevation data of the digital surface model, the image pixels are resampled as follows: Where: X 0 ,Y 0 ,Z 0) represents the projection center coordinate; f Indicates the focal length of the camera; ( x 0 ,y 0) represents the coordinates of the principal point; R =[ r ij ] represents the rotation matrix, obtained through the attitude angle; ( x,y ) represents the original image coordinates; like( x,y ) in the original image coordinates, get its pixel value V : Where: dx 、 dy represent the normalized distances to neighboring pixels, V ij Represents the values of the four surrounding pixels; Map the resampled orthophoto to the target coordinate system: Where: u,v ) represents the pixel coordinates of the orthophoto; ( , ) represents geographic coordinates; parameter a 0~ a 5. b 0~ b 5 Solve by ground control points.

6. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 4, characterized in that: The iterative optimization parameters after removing outliers are specifically: First, perform outlier detection and calculate the predicted geographic coordinates for each ground control point. With actual coordinates The residual between: Preset residual threshold e 0, if the residual e i If it is greater than the residual threshold, it is marked as an outlier; Calculate the residuals of all ground control points, count the number of outliers, remove outliers, and iterate the process of calculating and removing outliers. Q The model with the most inliers is retained.

7. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 4, characterized in that: The specific steps for completing the fusion of multispectral image data and panchromatic image data in step S3 are: Principal component analysis is first used to extract principal components from multispectral images, eliminate redundant information, and ensure the independence of principal components through the Gram-Schmidt orthogonalization process; Extracting principal components from multispectral images: Assume that the multispectral image is a matrix M∈R h×w , where: h is the number of bands and w is the number of pixels; Where: V represents the orthogonal eigenvector matrix, 1 w express w ×1 all-1 vector; Orthogonalize the principal components: Where: represents the magnitude of a vector; represents the inner product of vectors; Then, the high spatial resolution information of the panchromatic image is introduced into the first principal component of the multispectral image. The degree of fusion is controlled by the fusion coefficient, and the panchromatic image and the first principal component of the multispectral image are fused by linear weighting: Where: represents the fusion coefficient, ; Pan norm Indicates the normalized value of the panchromatic image; Where: 、 represent the mean and standard deviation of the panchromatic image, respectively; 、 They represent the mean and standard deviation of the first principal component after Gram-Schmidt processing; Pan Represents the pixel value vector of the panchromatic image; Afterwards, the adjusted first principal component is orthogonalized with the other principal components using Gram-Schmidt: Where: HR sim represents high-resolution simulation data; LR k Indicates the k An estimate or projection vector of low-resolution data; GS k Indicates the k orthogonal basis vectors generated by the Gram-Schmidt orthogonalization process; Then orthogonalize multiple GS n Reconstruct the multispectral image through inverse PCA transformation: Finally, the inverse PCA transform is used to restore the fused principal components to the original image space to obtain the final high-resolution fused image.

8. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 1, characterized in that: The spectral indices for cotton aphid analysis include ten indices: ARI, GLI, GBI, RVI, DVI, ARVI, GNDVI, SAVI, SIPI, and TCARI.

9. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 1, characterized in that: The step S5 is specifically as follows: Step S51: normalize the data according to the absolute spectral index obtained in step S4: Where: B represents the input feature value; represents the feature mean; represents the standard deviation; Divide the normalized dataset into training set and test set; Step S52: Build a monitoring model: First, initialize the model, and the initial prediction value is the target variable y The mean of: Then, get the monitor model: Where: R jm Represents the division of the regression tree into J m Leaf nodes of a region; m Indicates the number of model iterations; mx Indicates the maximum number of iterations; I represents the exponential function, when x Belong to the region R jm hour, I is 1, otherwise 0; Represents the model learning rate.

10. The method for intelligent monitoring of cotton aphid infestation based on multi-source remote sensing fusion data according to claim 1, characterized in that: The deep learning model uses a pre-trained YOLOv8 model.

Images