Cotton insect pest identification and classification method based on unmanned aerial vehicle accurate pesticide application

Abstract

The invention relates to the technical field of pest control, in particular to a cotton pest recognition and classification method based on unmanned aerial vehicle precise pesticide application, which comprises the following steps: S1, establishing a cotton pest leaf image data set, and after preprocessing and data enhancement are performed on cotton leaf images, performing label labeling to obtain image data; s2, establishing a detection model, and training the detection model by using the processed image data; s3, cutting the collected cotton unmanned aerial vehicle image to obtain a plurality of images, sequencing the images in sequence, inputting the sequenced image data into the trained cotton pest model in sequence, and performing cotton pest detection to obtain a cotton pest detection result; the method can effectively solve the problem that the model classification accuracy is low due to the fact that data cannot be migrated across time and space and defects exist in the aspect of the detection scale in an existing cotton pest recognition method.

Description

Technical Field

[0001] The present invention relates to the technical field of pest control, and in particular to a method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles. Background Art

[0002] Cotton is an important economic crop. During its growth, it will be attacked by a variety of insects, which has a significant impact on the quality and yield of cotton. Therefore, accurate identification of cotton pests is a key link in the prevention and control of cotton pests. It can also reduce the use of pesticides and promote the development of green agriculture. In recent years, with the rapid development of Agriculture 4.0, the use of advanced information systems and Internet technologies can realize the collection, analysis, processing and realization of a large amount of agricultural data, providing assistance for agricultural decision support systems. The pest and disease characteristics of cotton leaves are extracted using image processing methods, thereby realizing the classification of cotton pests and diseases.

[0003] At present, image classification methods can be roughly divided into two categories, namely, methods based on traditional artificial feature extraction and automatic feature learning methods based on deep learning. Traditional artificial feature extraction methods are usually based on features such as grayscale value, shape and texture in the image, and these features are manually designed using expert knowledge to achieve automatic segmentation of the target area, such as threshold classification, clustering classification, region-based classification and edge detection-based classification. These methods usually require a lot of prior knowledge to extract artificial features for classification. However, the designed artificial features are only effective for classification tasks of specific data sets, and the performance is unstable. The classification method based on deep learning adopts the idea of ​​pixel classification. Different from the traditional artificial feature extraction method, it can automatically extract semantic information from the image and learn the representation method of data, which effectively overcomes the limitations of traditional artificial feature classification. The Swin Transformer provides a new direction for research in the field of vision. Transformer-based vision research has developed rapidly, covering tasks such as static image classification, target detection and semantic segmentation, as well as temporal tasks such as video classification and target tracking.

[0004] However, the existing cotton pest identification methods still have shortcomings. Specifically: the existing models based on remote sensing methods can achieve large-scale detection of plant pests, but have difficulties in cross-temporal and spatial migration. The cotton pest detection models based on deep learning mostly use single cotton plants as the research object, and have limitations in detection scale. At the same time, due to the large number of parameters and high complexity of deep learning models, they are challenged to run on edge devices with limited computing power.

[0005] Therefore, a method for identifying and classifying cotton pests based on precise pesticide application by drones is needed to solve the problems raised in the above background technology. Summary of the invention

[0006] The purpose of the present invention is to provide a method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for identifying and classifying cotton pests based on drone precision pesticide application, comprising the following steps:

[0009] S1, establish a cotton insect leaf image dataset, perform label annotation on the cotton leaf images after preprocessing and data enhancement, and obtain image data;

[0010] S2, building a detection model, and using the processed image data to train the detection model;

[0011] S3, cropping the collected cotton drone images to obtain multiple images, and sequenced them in sequence, and inputting the sequenced image data into the trained cotton pest model in sequence to detect cotton pests, and marking the output results with blocks of equal size and different colors, and splicing the detection results together in the cropping order to finally generate a pest distribution map.

[0012] As a preferred solution of the present invention, the specific method of labeling in S1 is to correctly classify the collected cotton leaf images according to the leaf characteristics of different insect pests under the guidance of plant protection experts, rename the file name of each cotton image, and complete the classification labeling.

[0013] As a preferred solution of the present invention, the specific method of preprocessing and data enhancement in S1 is: the data file is cropped into a uniform size, and the cropped cotton image is rotated 90°, mirrored 180°, dimmed by 0.9, brightened by 1.5, and Gaussian filtered in turn to achieve image augmentation.

[0014] As a preferred solution of the present invention, the detection model in S2 is a deep learning network that uses Swin-Transformer as the backbone and integrates a multi-scale residual module and a jump connection. The detection model includes a block segmentation module, a linear sorting module, a multi-scale residual module, a network backbone module, and a classification module.

[0015] As a preferred solution of the present invention, the network backbone module is composed of multiple Transformer encoders and includes window attention and sliding window attention.

[0016] As a preferred solution of the present invention, the multi-scale residual module includes a multi-scale deep convolution layer (DWConv), a normalization layer and a ReLU activation function layer body, and the specific structural expression is: Where X is the input, σ is the ReLU activation function, BN is the normalization operation, and the output is y.

[0017] As a preferred solution of the present invention, the specific steps of training the detection model in S2 are: inputting the image file into the detection model, the image is an H*W*3 image, that is, a two-dimensional image of H*W pixels and 3 channels, the detection model extracts a feature map of size H*W*3 through the residual module, the block segmentation part divides the feature map into a series of image blocks of equal size, linear sorting performs linear transformation on the channel data of each pixel in the image block, and embeds the position information into the feature vector so that the detection model can perceive the relative position of the image block, the image feature map becomes H / 4*W / 4*3, Patch Merging performs down sampling, and uses SwinTransformerblock to extract classification features. The classification module classifies the extracted feature information.

[0018] As a preferred solution of the present invention, a skip connection is introduced between the linear sorting and the SwinTransformerblock to enhance the expression ability of the network.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] 1. In the present invention, by establishing a cotton insect pest leaf image data set, after preprocessing and data enhancement work on the cotton leaf image, labeling is performed to obtain image data, a detection model is established, and the processed image data is used to train the detection model. The collected cotton drone image is cropped to obtain multiple images, and the images are sequenced in sequence. The sorted image data are sequentially input into the trained cotton insect pest model to detect cotton insect pests, and the output results are labeled with types using blocks of equal size and different colors. The detection results are spliced ​​together according to the cropping order, and finally a pest distribution map is generated. The detection model is based on Swin Transformer, and a multi-scale residual module and jump connection are constructed and introduced to strengthen the extraction of small leaf features, and a classification model for cotton insect pests is constructed to realize the identification of cotton insect pests in a field environment. Finally, a pest distribution map is generated, providing technical support for precise pesticide application, and realizing cross-regional, rapid and accurate classification, solving the problems that the existing cotton insect pest identification method cannot migrate data across time and space and has insufficient detection scale, resulting in low model classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the structure of the detection model of the present invention;

[0022] Figure 2 Schematic diagram of the structure of the multi-scale residual module of the present invention;

[0023] Figure 3 A schematic diagram of a process for drawing a cotton pest distribution map of the present invention; DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0025] To facilitate understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings, and several embodiments of the present invention are given. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0026] It should be noted that when an element is referred to as being "fixed on" another element, it may be directly on the other element or there may also be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used in this article are for illustrative purposes only.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0028] For examples, see Figure 1-3 , the present invention provides a technical solution:

[0029] A method for identifying and classifying cotton pests based on drone precision pesticide application, comprising the following steps:

[0030] S1, establish a cotton insect leaf image dataset, perform label annotation on the cotton leaf images after preprocessing and data enhancement, and obtain image data;

[0031] S2, building a detection model, and using the processed image data to train the detection model;

[0032] S3, cropping the collected cotton drone images to obtain multiple images, and sequenced them in sequence, and inputting the sequenced image data into the trained cotton pest model in sequence to detect cotton pests, and marking the output results with blocks of equal size and different colors, and splicing the detection results together in the cropping order to finally generate a pest distribution map.

[0033] Furthermore, the specific method of labeling in S1 is to correctly classify the collected cotton leaf images according to the leaf features of different insect pests under the guidance of plant protection experts, rename the file name of each cotton image, and complete the classification labeling.

[0034] Furthermore, the specific method of preprocessing and data enhancement in S1 is: cropping the data file into a uniform size, rotating the cropped cotton image by 90°, mirroring it by 180°, dimming the brightness by 0.9, brightening it by 1.5, and performing Gaussian filtering to achieve image augmentation.

[0035] Furthermore, the detection model in S2 is a deep learning network that uses Swin-Transformer as the backbone and integrates a multi-scale residual module and a jump connection. The detection model includes a tile segmentation module, a linear sorting module, a multi-scale residual module, a network backbone module, and a classification module.

[0036] Furthermore, the network backbone module is composed of multiple Transformer encoders, and includes window attention and sliding window attention.

[0037] Furthermore, the multi-scale residual module includes a multi-scale deep convolution layer, a normalization layer and a ReLU activation function layer. The specific structural expression is: Where X is the input, σ is the ReLU activation function, BN is the normalization operation, and the output is y.

[0038] Furthermore, the specific steps of training the detection model in S2 are as follows: input the image file into the detection model, the image is an H*W*3 image, that is, a two-dimensional image of H*W pixels and 3 channels. The detection model extracts a feature map of size H*W*3 through the residual module. The block segmentation part divides the feature map into a series of image blocks of equal size. The linear sorting performs a linear transformation on the channel data of each pixel in the image block, and embeds the position information into the feature vector so that the detection model can perceive the relative position of the image block. The image feature map becomes H / 4*W / 4*3, PatchMerging is downsampled, and SwinTransformerblock is used to extract classification features. The classification module classifies the extracted feature information.

[0039] As a preferred solution of the present invention, a skip connection is introduced between the linear sorting and the SwinTransformerblock to enhance the expression ability of the network.

[0040] Specific implementation cases

[0041] Establish a cotton insect pest leaf image dataset, crop the data file into a uniform size, rotate the cropped cotton image by 90°, mirror it by 180°, dim the brightness by 0.9, brighten it by 1.5, and perform Gaussian filtering to achieve image augmentation, correctly classify the collected cotton leaf images according to the leaf features of different insect pests, rename the file name of each cotton image, complete the classification label, and obtain the image data;

[0042] A detection model is established. The detection model is a deep learning network that uses Swin-Transformer as the backbone and integrates multi-scale residual modules and jump connections. The detection model includes a tile segmentation module, a linear sorting module, a residual module, a network backbone module, and a classification module. The network backbone module consists of multiple Transformer encoders and contains window attention and sliding window attention. The residual module includes a convolutional layer, a normalization layer, and a ReLU activation function layer. The specific structural expression is: Where X is the input, σ is the ReLU activation function, BN is the normalization operation, and the output is y;

[0043] The image file is input into the detection model. The image is an H*W*3 image, that is, a two-dimensional image with H*W pixels and 3 channels. The detection model extracts a feature map of size H*W*3 through the residual module. The block segmentation part divides the feature map into a series of image blocks of equal size. The linear sorting performs a linear transformation on the channel data of each pixel in the image block and embeds the position information into the feature vector so that the detection model can perceive the relative position of the image block. The image feature map becomes H / 4*W / 4*3, and PatchMerging is used for downsampling. SwinTransformerblock is used to extract classification features. Skip connections are introduced between linear sorting and SwinTransformerblock to enhance the expression ability of the network. The classification module classifies the extracted feature information.

[0044] The collected cotton drone images are cropped to obtain multiple images, which are sequenced in sequence. The sorted image data are input into the trained cotton pest model in sequence to detect cotton pests. The output results are labeled with types using blocks of equal size and different colors. The detection results are spliced ​​together in the cropping order to finally generate a pest distribution map.

[0045] Comparison of data enhancement results

[0046] In order to prove the effectiveness of data augmentation, we trained and verified the model on the original dataset and the augmented dataset. The experimental results show that the accuracy rate increased from 94.62% to 95.09% when the original model was used for training and verification on the original dataset and the augmented dataset. The accuracy rate increased from 96.10% to 97.38% when the improved model was used for training and verification, an increase of 1.28%. The experimental data proves that using data augmentation datasets can achieve higher accuracy.

[0047] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles, characterized in that: The following steps are involved: S1, establish a cotton insect leaf image dataset, perform label annotation on the cotton leaf images after preprocessing and data enhancement, and obtain image data; S2, building a detection model, and using the processed image data to train the detection model; S3, cropping the collected cotton drone images to obtain multiple images, and sequenced them in sequence, and inputting the sequenced image data into the trained cotton pest model in sequence to detect cotton pests, and marking the output results with blocks of equal size and different colors, and splicing the detection results together in the cropping order to finally generate a pest distribution map.

2. The method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles according to claim 1, characterized in that: The specific method of labeling in S1 is to correctly classify the collected cotton leaf images according to the leaf features of different insect pests under the guidance of plant protection experts, rename the file name of each cotton image, and complete the classification labeling.

3. The method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles according to claim 1, characterized in that: The specific method of preprocessing and data enhancement in S1 is: cutting the data file into a uniform size, rotating the cropped cotton image by 90°, mirroring it by 180°, dimming the brightness by 0.9, brightening it by 1.5, and performing Gaussian filtering to achieve image augmentation.

4. The method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles according to claim 1, characterized in that: The detection model in S2 is a deep learning network that uses Swin-Transformer as the backbone and integrates residual modules and jump connections. The detection model includes a tile segmentation module, a linear sorting module, a residual module, a network backbone module, and a classification module.

5. The method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles according to claim 4, characterized in that: The network backbone module consists of multiple Transformer encoders, which also include window attention and sliding window attention.

6. The method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles according to claim 4, characterized in that: The multi-scale residual module includes a multi-scale deep convolution layer (DWConv), a normalization layer, and a ReLU activation function layer. The specific structural expression is: , where X is the input, is the ReLU activation function, BN is the normalization operation, and the output is y.

7. The method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles according to claim 1, characterized in that: The specific steps for training the detection model in S2 are as follows: input the image file into the detection model, the image is an H * W * 3 image, that is, a two-dimensional image with H * W pixels and 3 channels. The detection model extracts a feature map of size H * W * 3 through the residual module. The block segmentation part divides the feature map into a series of image blocks of equal size. The linear sorting performs a linear transformation on the channel data of each pixel in the image block, and embeds the position information into the feature vector so that the detection model can perceive the relative position of the image block. The image feature map becomes H / 4 * W / 4 * 3, Patch Merging is down-sampled, and the classification features are extracted using the Swin Transformer block. The classification module classifies the extracted feature information.

8. The method for identifying and classifying cotton pests based on precise pesticide application by unmanned aerial vehicles according to claim 7, characterized in that: A skip connection is introduced between the linear sorting and the Swin Transformer block to enhance the expression ability of the network.

Images