Bursaphelenchus xylophilus tree detection method based on MASFNet framework

By adopting the multi-level attention and spatial sampling fusion network of the MASFNet framework in the detection of pine nematode disease trees, the problems of misdetecting disease trees and complex background information in the existing technology are solved, and higher detection accuracy and errors are achieved The detection rate is reduced.

CN119068331AActive Publication Date: 2024-12-03CHINA THREE GORGES UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411104615.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-12-03
Estimated Expiration
2044-08-13

AI Technical Summary

Technical Problem

In the existing tree detection technology for pine nematode disease, CNN and multi-level feature map detection methods are prone to misdetection of diseased trees, and the background information of land objects in the drone images is complex, making it difficult to distinguish land objects in diseased trees and non-forest areas. , resulting in a high false detection rate.

Method used

Using the remote sensing image detection method based on the MASFNet framework, the multi-level attention module and spatial sampling fusion module are built to enhance feature extraction capabilities, combine global and local attention to reduce the error detection rate.

Benefits of technology

It effectively reduces the false detection rate of diseased trees in non-forest areas, improves the ability to distinguish diseased trees from their background environment, and improves the accuracy and interpretability of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068331A_ABST
    Figure CN119068331A_ABST
Patent Text Reader

Abstract

A pine wood nematode disease tree detection method based on a MASFNet framework comprises the following steps: step 1, acquiring a remote sensing image in a target range by using an unmanned aerial vehicle, then preprocessing data, analyzing and training data quality and scale, and generating a good pine wood nematode disease tree data set; 2, performing an experiment by using a detection algorithm of classical deep learning, verifying effects of different models, comparing and selecting an optimal model structure and model complexity, and designing and testing an appropriate disease tree detection model so as to ensure that the model keeps a stable prediction capability for a disease tree detection problem; and step 3, constructing a MASFNet model, and carrying out gradual improvement and ablation contrast experiments to explain reasons of model prediction result differences and enhance the interpretability and generalization ability of the model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method for detecting pine wood nematode diseased trees using remote sensing images based on a MASFNet framework. Background Art

[0002] Pine wilt (PWD) is caused by pine wood nematodes (Bergdahl, 1988). It is a devastating disease with a wide range of habitats, multiple transmission pathways, rapid lethality, and high prevention and control costs. Once infected, pine trees will wither and die within 40 days, so they are called the "cancer" of pine trees. The spread of pine wilt disease will cause serious damage to pine forest resources, natural landscapes, and the ecological environment. According to statistics, from 1998 to 2017, the average annual total economic losses caused by pine wilt disease in China reached RMB 71.7 billion. Timely and accurate positioning and removal of infected pine trees to destroy the transmission pathways of the disease is currently the most effective way to reduce and prevent new infections. Therefore, accurate and efficient detection of pine wilt disease is of great significance in helping governments and forestry departments to take timely prevention and control measures, reduce economic losses, and protect the ecological environment.

[0003] The traditional method of monitoring pine wood nematodes mainly relies on technical personnel to conduct in-depth field surveys in forest areas, which is not only time-consuming, labor-intensive and unsafe, but also difficult to achieve full coverage. In recent years, unmanned aerial vehicle (UAV) remote sensing has been widely used in low-altitude remote sensing research in agriculture and forestry due to its advantages of real-time, low cost and high spatial resolution. Compared with traditional manual methods, pine wood nematode diseased tree detection based on artificial intelligence technology and UAV RGB (red, green, blue) image acquisition can capture ultra-fine information in centimeter-level images, and has the advantages of being faster, more efficient and less expensive.

[0004] Deep learning networks mostly rely on convolutional layers and pooling layers for feature extraction. They have strong local feature extraction capabilities for images, but lack the ability to capture global information of images, resulting in a large number of false detections of diseased trees. UAV images contain a variety of environmental background information, such as forests, farmlands, and towns. Feature extraction methods that rely solely on local information are not sufficient to consider the environmental background information of objects, and objects in non-forest areas are easily misjudged as diseased trees. Using a larger range of pixel information can help the model better capture global image information and use a wide range of contextual information to avoid incorrect results caused by detectors with limited receptive fields. Introducing this idea into remote sensing detection of pine wilt diseased trees is expected to help the model distinguish the background environment of the detected objects, so as to reduce the model's false detection of diseased trees in non-forest areas.

[0005] In addition, in the actual work of detecting trees with pine wilt disease, due to the large variety of objects in UAV remote sensing images, existing algorithms are difficult to distinguish objects with a high similarity to diseased trees. In order to improve the detection accuracy of algorithms for objects of different scales, existing studies usually use multi-level feature maps to detect objects of different scales, assigning larger targets to the deepest feature map and assigning smaller targets to shallower feature maps for identification. However, the above studies regard objects as indivisible wholes, which is a coarse-grained instance allocation. Because UAVs are affected by objective conditions during flight, many features of objects in the image have a high similarity between classes, such as some bare land in the forest. This coarse-grained feature representation makes it difficult to distinguish objects with a high similarity to diseased trees.

[0006] Therefore, the applicant proposed a method for detecting pine wood nematode diseased trees using remote sensing images based on the MASFNet framework. Summary of the invention

[0007] The purpose of the present invention is to solve the technical problem that the common CNN and multi-level feature map detection methods in the existing detection technology for pine wood nematode diseased trees are prone to misdetection of diseased trees, and to provide a pine wood nematode diseased tree detection method based on remote sensing images of the MASFNet framework.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] A method for detecting pine wood nematode diseased trees based on the MASFNet framework comprises the following steps:

[0010] Step 1: Use drones to acquire remote sensing images within the target range, then preprocess the data, analyze the quality and scale of training data, and generate a good pine wood nematode disease tree dataset;

[0011] Step 2: Conduct experiments using classic deep learning detection algorithms to verify the effects of different models, compare and select the optimal model structure and model complexity, and design and test a suitable diseased tree detection model to ensure that the model maintains a stable predictive ability for diseased tree detection problems;

[0012] Step 3: Build the MASFNet model and conduct gradual improvements and ablation comparison experiments to explain the reasons for the differences in the model prediction results and enhance the interpretability and generalization ability of the model;

[0013] The constructed MASFNet model is used to detect pine wood nematode diseased trees in remote sensing data.

[0014] In step 1, the following steps are specifically included:

[0015] Step 1) using a drone to obtain remote sensing image data of a target area;

[0016] Step 2) Filter clear and high-quality images from all acquired images, and use labelImg tool to annotate the data to create a dataset of pine wood nematode diseased trees;

[0017] In step 3, the following steps are included:

[0018] Step 1) In terms of model structure, a large number of preliminary comparative experiments were conducted using classic detection models, including Faster R-CNN, Cascade R-CNN, CenterNet, Yolox, Fcos, and Yolov7; on this basis, a suitable model structure was selected to avoid overfitting or underfitting, and to obtain the best accuracy performance in pine wood nematode disease tree detection;

[0019] Step 2) Use convolutional neural network to build the backbone network, extract relevant features, and generate feature moment graph. The calculation formula of the backbone network is as follows:

[0020] x t =σ CNN (W CNN ·x t-1 +b CNN )

[0021] x t-1 represents the input feature map of the previous layer, x t Represents the feature map after calculation, w CNN represents the weight of the filter, b CNN represents the deviation parameter, σ CNN is the activation function;

[0022] Step 3) Extract features at different levels to form a multi-level feature map for target detection tasks:

[0023] Extract the features of x3, x4, and x5 layers to form a multi-level feature map;

[0024] Step 4) Construct a multi-level attention module, which integrates global attention, channel attention and spatial attention; for each layer of input features, perform calculations on the multi-level attention module, and the formula process is as follows:

[0025] After full attention calculation, g g (x i )=x i +bn(f(x i ·softmax(f(x i )))),x i The task represents the input feature, f(·) represents 1x1 convolution, and bn represents batchNorm;

[0026] Then after channel attention calculation, x i Represents input features;

[0027] Step 5) Construct a spatial sampling fusion module. For each layer of features, use spatial attention to sample the most important pixel area, and then fuse and superimpose them with each other. In this way, the pixel features of each layer will be conditionally fused to other feature layers. The formula process is as follows:

[0028] After spatial attention calculation, x i represents the input feature, P n Indicates the position of the calculated pixel, △P n Indicates the spatial offset;

[0029] Then perform fusion calculations at different layers

[0030] The MASFNet network model is established through the above steps.

[0031] In step 3, the following steps are specifically included:

[0032] Step 1) After synthesizing all the research results, the MASFNet model showed the most outstanding overall performance and achieved good results in the problem of false detection of pine wood nematode diseased trees;

[0033] Step 2) Further ablation experiments are performed to analyze the importance of features in the MASFNet model.

[0034] The MASFNet model used in the present invention is as follows:

[0035] The feature extraction backbone network includes a first feature extraction structure CBS, a second feature extraction structure CBS, a first cross-stage local network CSP, a third feature extraction structure CBS, a second cross-stage local network CSP, a fourth feature extraction structure CBS, a third cross-stage local network CSP, and a fifth feature extraction structure CBS;

[0036] Constructing a feature pyramid network includes a sixth feature extraction structure CBS, a first upsampling structure UpSampling, a first merging structure Concat, a seventh feature extraction structure CBS, a second upsampling structure UpSampling, a second merging structure Concat, and a first spatial pyramid pooling structure SPPF;

[0037] Constructing a detection result output module including a first multi-level attention structure MLAttention, a second multi-level attention structure MLAttention, a third multi-level attention structure MLAttention, a first spatial attention structure SpatialAttention, a second spatial attention structure SpatialAttention, a third spatial attention structure SpatialAttention, and a prediction head structure Predict;

[0038] The input image is connected to the first feature extraction structure CBS, the output of the first feature extraction structure CBS is connected to the second feature extraction structure CBS, the output of the second feature extraction structure CBS is connected to the first cross-stage local network CSP, the output of the first cross-stage local network CSP is connected to the third feature extraction structure CBS, the output of the third feature extraction structure CBS is connected to the second cross-stage local network CSP, the output of the second cross-stage local network CSP is connected to the fourth feature extraction structure CBS, the output of the fourth feature extraction structure CBS is connected to the third cross-stage local network CSP, and the output of the third cross-stage local network CSP is connected to the fifth feature extraction structure CBS to obtain a calculated feature map;

[0039] The output of the fifth feature extraction structure CBS is connected to the sixth feature extraction structure CBS, the output of the sixth feature extraction structure CBS is connected to the first upsampling structure UpSampling, the output of the first upsampling structure UpSampling and the output of the second cross-stage local network CSP are merged through the first merging structure Concat, the output of the first merging structure Concat is connected to the seventh feature extraction structure CBS, the output of the seventh feature extraction structure CBS is connected to the second upsampling structure UpSampling, the output of the second upsampling structure UpSampling and the first cross-stage local network CSP are merged through the second merging structure Concat, and the output of the fifth feature extraction structure CBS is connected to the first spatial pyramid pooling structure SPPF to obtain multi-level feature structures of different sizes;

[0040] The output of the first spatial pyramid pooling structure SPPF is connected to the first multi-level attention structure MLAttention, the output of the seventh feature extraction structure CBS is connected to the second multi-level attention structure MLAttention, the output of the second merging structure Concat is connected to the third multi-level attention structure MLAttention, the output of the first multi-level attention structure MLAttention is connected to the first spatial attention structure Spatial Attention, the output of the second multi-level attention structure MLAttention is connected to the second spatial attention structure Spatial Attention, the output of the third multi-level attention structure MLAttention is connected to the third spatial attention structure Spatial Attention, finally, the outputs of the first spatial attention structure SpatialAttention, the second spatial attention structure Spatial Attention and the third spatial attention structure SpatialAttention are connected to the prediction head structure Predict to obtain the final prediction result.

[0041] Compared with the prior art, the present invention has the following technical effects:

[0042] 1) In order to deal with the problem of misdetection of pine wood nematode diseased trees in non-forest areas caused by complex background information of remote sensing images, the present invention proposes a multi-level attention structure. Specifically, the present invention designs a multi-attention feature activation structure, which combines the advantages of various attention calculations, increases the diversity of mathematical operations and further strengthens the extraction of features; and strengthens the modeling of feature correlations between pixels in a larger range in the image from different dimensions, so that the network pays more attention to the connection between the ground objects and their environmental background; in non-forest areas such as human activity areas, the information of some ground objects that are easily confused by human activities will be easier to distinguish, thereby reducing the misdetection rate of diseased trees in non-forest areas.

[0043] 2) In order to deal with the problem of false detection of diseased trees due to the difficulty in distinguishing some highly similar objects in remote sensing images, the present invention designs a spatial sampling fusion structure. For each layer of features, the most important pixel area is sampled using spatial attention, and then fused and superimposed with each other, so that the pixel features of each layer can be conditionally fused to other feature layers. This merging or fusing of features from different levels or sources in a more sophisticated way can improve the model's ability to perceive and represent detailed information. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0045] Figure 1 This is a model architecture diagram of the present invention;

[0046] Figure 2 for Figure 1 Multi-level attention maps in ;

[0047] Figure 3 for Figure 1 The spatial sampling fusion structure diagram in ;

[0048] Figure 4 for Figure 3 Figure 2. The spatial attention structure in .

[0049] Figure 5 It is a comparative visualization diagram of the present invention;

[0050] Figure 6 This is a multi-level attention improvement effect diagram of the present invention;

[0051] Figure 7 This is a diagram showing the improvement effect of the spatial sampling fusion structure of the present invention. DETAILED DESCRIPTION

[0052] A method for detecting pine wood nematode diseased trees using remote sensing data based on the MASFNet framework comprises the following steps:

[0053] Step 1: Use drones to obtain remote sensing images of Yichang, then preprocess the data, analyze the quality and scale of training data, and generate a good pine wood nematode disease tree dataset;

[0054] Step 2: Conduct experiments using a wide range of classical deep learning detection algorithms to verify the effects of different models, compare and select the optimal model structure and model complexity, and design and test a suitable diseased tree detection model to ensure that the model maintains a stable predictive ability for diseased tree detection problems;

[0055] Step 3: Based on a large number of experiments, summarize the current detection problems, and make improvements and designs for specific problems. Introduce appropriate mathematical operation mechanisms, design MASFNet, and conduct gradual improvements and ablation comparison experiments to explain the reasons for the differences in model prediction results and enhance the interpretability and generalization ability of the model.

[0056] In step 1, the following steps are specifically included:

[0057] Step 1) Use drones to obtain remote sensing image data of the Dengcun Township area in Yichang, covering a total area of ​​300 square kilometers and obtaining more than 200GB of image data.

[0058] Step 2) Filter clear and high-quality images from all acquired images, and use labelImg tool to annotate the data to create a dataset of pine wood nematode diseased trees;

[0059] In step 2, if Figure 1 , Figure 2 , Figure 3 As shown in the figure, when building the MASFNet network model, the following steps are included:

[0060] Step 1) In terms of model structure, we used the classic detection models previously studied and conducted a large number of preliminary comparative experiments, including Faster R-CNN, Cascade R-CNN, CenterNet, Yolox, Fcos, and Yolov7. On this basis, we innovatively designed the MASFNet network model (such as Figure 1 A total of six different models (Table 1) were compared to select the appropriate model structure and complexity to avoid overfitting or underfitting and to obtain the best accuracy performance in pine wilt disease tree detection.

[0061] Step 2) Use convolutional neural network to build the backbone network, extract relevant features, and generate feature moment graph. The calculation formula of the backbone network is as follows:

[0062] x t =σ CNN (W CNN ·x t-1 +b CNN )

[0063] x t-1 represents the input feature map of the previous layer, x t Represents the feature map after calculation, w CNN represents the weight of the filter, b CNN represents the deviation parameter, σ CNN is the activation function;

[0064] Step 3) Extract features at different levels to form a multi-level feature map for target detection tasks:

[0065] The features of x3, x4, and x5 layers are extracted to form a multi-level feature map.

[0066] Step 4) Construct a multi-level attention module that integrates global attention, channel attention, and spatial attention. This diverse computational structure enhances feature extraction and models global context feature information from different dimensions. When faced with complex background information in remote sensing image tasks, this method helps to model more comprehensive feature correlations. For each layer of input features, a multi-level attention module is calculated. The formula process is as follows:

[0067] After full attention calculation, g g (x i )=x i +bn(f(x i ·softmax(f(xi )))),x i The task represents the input feature, f(·) represents 1x1 convolution, and bn represents batchNorm.

[0068] Then after channel attention calculation, x i Represents the input features.

[0069] Step 5) Construct a spatial sampling fusion module. For each layer of features, use spatial attention to sample the most important pixel area, and then fuse and superimpose them with each other. In this way, the pixel features of each layer will be conditionally fused to other feature layers. The formula process is as follows:

[0070] After spatial attention calculation, x i represents the input feature, P n Indicates the position of the calculated pixel, △P n Indicates a spatial offset.

[0071] Then perform fusion calculations at different layers

[0072] The MASFNet network model is established by the above steps;

[0073] In step 3, the following steps are specifically included:

[0074] Step 1) After synthesizing all the research results, the MASFNet model showed the most outstanding overall performance and achieved good results in the problem of false detection of pine wood nematode diseased trees;

[0075] Step 2) Further ablation experiments are performed to analyze the importance of features in the MASFNet model.

[0076] The MASFNet model used in the present invention is as follows:

[0077] The feature extraction backbone network includes a first feature extraction structure CBS2, a second feature extraction structure CBS3, a first cross-stage local network CSP4, a third feature extraction structure CBS5, a second cross-stage local network CSP6, a fourth feature extraction structure CBS7, a third cross-stage local network CSP8, and a fifth feature extraction structure CBS9;

[0078] Constructing a feature pyramid network includes a sixth feature extraction structure CBS10, a first upsampling structure UpSampling11, a first merging structure Concat12, a seventh feature extraction structure CBS13, a second upsampling structure UpSampling14, a second merging structure Concat15, and a first spatial pyramid pooling structure SPPF16;

[0079] Constructing a detection result output module including a first multi-level attention structure MLAttention17, a second multi-level attention structure MLAttention18, a third multi-level attention structure MLAttention19, a first empty sampling fusion structure SSF20, and a prediction head structure Predict21;

[0080] The input image 1 is connected to the first feature extraction structure CBS2, the output of the first feature extraction structure CBS2 is connected to the second feature extraction structure CBS3, the output of the second feature extraction structure CBS3 is connected to the first cross-stage local network CSP4, the output of the first cross-stage local network CSP4 is connected to the third feature extraction structure CBS5, the output of the third feature extraction structure CBS5 is connected to the second cross-stage local network CSP6, the output of the second cross-stage local network CSP6 is connected to the fourth feature extraction structure CBS7, the output of the fourth feature extraction structure CBS7 is connected to the third cross-stage local network CSP8, and the output of the third cross-stage local network CSP8 is connected to the fifth feature extraction structure CBS9 to obtain the calculated feature map;

[0081] The output of the fifth feature extraction structure CBS9 is connected to the sixth feature extraction structure CBS10, the output of the sixth feature extraction structure CBS10 is connected to the first upsampling structure UpSampling11, the output of the first upsampling structure UpSampling11 is merged with the output of the second cross-stage local network CSP6 through the first merging structure Concat12, the output of the first merging structure Concat12 is connected to the seventh feature extraction structure CBS13, the output of the seventh feature extraction structure CBS13 is connected to the second upsampling structure UpSampling14, the output of the second upsampling structure UpSampling14 and the first cross-stage local network CSP4 are merged through the second merging structure Concat15, the output of the fifth feature extraction structure CBS9 is connected to the first spatial pyramid pooling structure SPPF16, and multi-level feature structures of different sizes are obtained;

[0082] The output of the first spatial pyramid pooling structure SPPF16 is connected to the first multi-level attention structure MLAttention17, the output of the seventh feature extraction structure CBS13 is connected to the second multi-level attention structure MLAttention18, the output of the second merging structure Concat15 is connected to the third multi-level attention structure MLAttention19, the outputs of the first multi-level attention structure MLAttention17, the second multi-level attention structure MLAttention18 and the third multi-level attention structure MLAttention19 are connected to the first empty sampling fusion structure SSF20, and the output of the first empty sampling fusion structure SSF20 is connected to the prediction head structure Predict21 to obtain the final prediction result.

[0083] Example:

[0084] 1. Dataset Information

[0085] With the help of the Forest Pest Control and Quarantine Station, the areas with more serious infection were identified, and flight plans and routes were specified, and drone image data was collected. The flight platform used was the MD-25 drone, with a wingspan of 1.8m and a maximum cruising speed of 19m / s. The camera was Sony a7r, Zeiss 35mm fixed-focus lens, 36 million pixels, and a camera image resolution of 5,472×3,648 pixels (JPEG format). The drone flew at an altitude of 1,000 meters, with a lateral overlap rate of 70% and a forward overlap rate of 80%, to obtain RGB band orthophoto images with a spatial resolution greater than 0.6m. The drone was used to collect orthophoto remote sensing images and related data of ground vegetation of approximately 300km2.

[0086] The characteristics of pinewood nematode trees include reddish-brown crowns and needle-like leaves. After infection, trees mainly show characteristics such as pine needle color changes, wilting, and browning of bark. We selected high-quality and clear images. In order to unify the standards, we cropped the images to 1000*1000 pixels and annotated the wood nematode dataset using the python-based labelImg tool. In the end, there were 1652 images for training and 708 images for verification. These datasets will be used to train and evaluate the object detection model.

[0087] 2. Parameter settings

[0088] The experiments were conducted on an Ubuntu 18.04 system equipped with an NVIDIA RTX 3090 GPU. MASFNet was implemented based on the Pytorch framework. Pytorch is a Python framework for developing and evaluating deep learning models that can perform efficient mathematical operations. All images were cropped to 1000x1000 pixels. The mosic data augmentation method was used, the training batch size was set to 8, the initial learning rate was set to 0.01, and all models were optimized using synchronous SGD, with a weight decay of 0.0005, a momentum of 0.937, and 100 rounds of iterations.

[0089] 3. Comparative Experiment

[0090] A variety of classic algorithms were selected for comparison, including the most representative single-stage and two-stage detection networks. Most target detection network structures consist of a backbone network, a neck part, and a detection head. The main models include: Faster R-CNN, Cascade R-CNN, CenterNet, Fcos, Yolox, and Yolov7.

[0091] Table 1: Comparative test

[0092]

[0093]

[0094] 4. Comparison of ablation improvements:

[0095] Based on the classic algorithm, the advantages of multi-level attention and spatial sampling fusion network are combined, and the upper ablation experiment is carried out on the pine wood nematode dataset. The experiments of multi-level attention and spatial sampling modules are carried out independently, and the metrics are evaluated respectively.

[0096] For the detection of diseased trees, this method has improved the recognition accuracy and recall rate to a certain extent, and achieved an absolute gain of 2.5% in the mAP50 indicator.

[0097] Table 2: Improved experimental table

[0098]

[0099] 5. Summary:

[0100] This study used UAV technology for data collection and confirmed the effectiveness of using UAV remote sensing images and target monitoring technology to monitor pine wilt diseased trees. In order to address the problem of false detection of diseased trees caused by the easily confused background environment and the difficulty in distinguishing highly similar objects in remote sensing images, a multi-level attention and spatial sampling fusion network (MASFNet) was designed to mitigate the false detection problem. Our method effectively reduced the false detection rate of pine wilt diseased trees. In general, our method has made some progress in improving the accuracy of pine wilt diseased tree detection. We provide some new reference dimensions for forest pest control. The management of infected forests has great ecological and economic significance. We hope that these dimensions can provide useful insights for future forest pest control work.

Claims

1. A method for detecting pine wood nematode diseased trees based on the MASFNet framework. It is characterized in that The following steps are involved: Step 1: Use drones to acquire remote sensing images within the target range, then preprocess the data, analyze the quality and scale of training data, and generate a good pine wood nematode disease tree dataset; Step 2: Conduct experiments using classic deep learning detection algorithms to verify the effects of different models, compare and select the optimal model structure and model complexity, and design and test a suitable diseased tree detection model to ensure that the model maintains a stable predictive ability for diseased tree detection problems; Step 3: Build the MASFNet model and conduct gradual improvements and ablation comparison experiments to explain the reasons for the differences in the model prediction results and enhance the interpretability and generalization ability of the model; The constructed MASFNet model is used to detect pine wood nematode diseased trees in remote sensing data.

2. The method according to claim 1, It is characterized in that In step 1, the following steps are specifically included: Step 1) using a drone to obtain remote sensing image data of a target area; Step 2) Filter clear and high-quality images from all acquired images, and use labelImg tool to annotate the data to create a pine wood nematode disease tree dataset.

3. The method according to claim 1, It is characterized in that In step 3, the following steps are included: Step 1) In terms of model structure, a large number of preliminary comparative experiments were conducted using classic detection models, including FasterR-CNN, Cascade R-CNN, CenterNet, Yolox, Fcos, and Yolov7; on this basis, a suitable model structure was selected to avoid overfitting or underfitting, and to obtain the best accuracy performance in pine wood nematode disease tree detection; Step 2) Use convolutional neural network to build the backbone network, extract relevant features, and generate feature moment graph. The calculation formula of the backbone network is as follows: x t =σ CNN (W CNN ·x t-1 +b CNN ) x t-1 represents the input feature map of the previous layer, x t Represents the feature map after calculation, W CNN represents the weight of the filter, b CNN represents the deviation parameter, σ CNN is the activation function; Step 3) Extract features at different levels to form a multi-level feature map for target detection tasks: Extract x 3 , x 4 , x 5 The features of the layers form a multi-level feature map; Step 4) Construct a multi-level attention module, which integrates global attention, channel attention and spatial attention; for each layer of input features, perform calculations on the multi-level attention module, and the formula process is as follows: After full attention calculation, g g (x i )=x i +bn(f(x i ·softmax(f(x i )))),x i The task represents the input feature, f(.) represents 1x1 convolution, and bn represents batchNorm; Then after channel attention calculation, x i Represents input features; Step 5) Construct a spatial sampling fusion module. For each layer of features, use spatial attention to sample the most important pixel area, and then fuse and superimpose them with each other. In this way, the pixel features of each layer will be conditionally fused to other feature layers. The formula process is as follows: After spatial attention calculation, x i represents the input feature, P n Indicates the position of the calculated pixel, △P n Indicates the spatial offset; Then perform fusion calculations at different layers The MASFNet network model is established through the above steps.

4. The method according to claim 1, It is characterized in that In step 3, the following steps are specifically included: Step 1) After synthesizing all the research results, the MASFNet model showed the most outstanding overall performance and achieved good results in the problem of false detection of pine wood nematode diseased trees; Step 2) Further ablation experiments are performed to analyze the importance of features in the MASFNet model.

5. The method according to any one of claims 1 to 4, It is characterized in that The MASFNet model is as follows: The feature extraction backbone network includes a first feature extraction structure CBS (2), a second feature extraction structure CBS (3), a first cross-stage local network CSP (4), a third feature extraction structure CBS (5), a second cross-stage local network CSP (6), a fourth feature extraction structure CBS (7), a third cross-stage local network CSP (8), and a fifth feature extraction structure CBS (9); Constructing a feature pyramid network includes a sixth feature extraction structure CBS (10), a first upsampling structure UpSampling (11), a first merging structure Concat (12), a seventh feature extraction structure CBS (13), a second upsampling structure UpSampling (14), a second merging structure Concat (15), and a first spatial pyramid pooling structure SPPF (16); Constructing a detection result output module including a first multi-level attention structure MLAttention (17), a second multi-level attention structure MLAttention (18), a third multi-level attention structure MLAttention (19), a first empty sampling fusion structure SSF (20), and a prediction head structure Predict (21); The input image (1) is connected to the first feature extraction structure CBS (2), the output of the first feature extraction structure CBS (2) is connected to the second feature extraction structure CBS (3), the output of the second feature extraction structure CBS (3) is connected to the first cross-stage local network CSP (4), the output of the first cross-stage local network CSP (4) is connected to the third feature extraction structure CBS (5), the output of the third feature extraction structure CBS (5) is connected to the second cross-stage local network CSP (6), the output of the second cross-stage local network CSP (6) is connected to the fourth feature extraction structure CBS (7), the output of the fourth feature extraction structure CBS (7) is connected to the third cross-stage local network CSP (8), and the output of the third cross-stage local network CSP (8) is connected to the fifth feature extraction structure CBS (9), to obtain a calculated feature map; The output of the fifth feature extraction structure CBS (9) is connected to the sixth feature extraction structure CBS (10), the output of the sixth feature extraction structure CBS (10) is connected to the first upsampling structure UpSampling (11), the output of the first upsampling structure UpSampling (11) and the output of the second cross-stage local network CSP (6) are merged through the first merging structure Concat (12), the output of the first merging structure Concat (12) is connected to the seventh feature extraction structure CBS (13), the output of the seventh feature extraction structure CBS (13) is connected to the second upsampling structure UpSampling (14), the output of the second upsampling structure UpSampling (14) and the first cross-stage local network CSP (4) are merged through the second merging structure Concat (15), the output of the fifth feature extraction structure CBS (9) is connected to the first spatial pyramid pooling structure SPPF (16), and multi-level feature structures of different sizes are obtained; The output of the first spatial pyramid pooling structure SPPF (16) is connected to the first multi-level attention structure MLAttention (17), the output of the seventh feature extraction structure CBS (13) is connected to the second multi-level attention structure MLAttention (18), the output of the second merging structure Concat (15) is connected to the third multi-level attention structure MLAttention (19), the outputs of the first multi-level attention structure MLAttention (17), the second multi-level attention structure MLAttention (18) and the third multi-level attention structure MLAttention (19) are connected to the first empty sampling fusion structure SSF (20), and the output of the first empty sampling fusion structure SSF (20) is connected to the prediction head structure Predict (21) to obtain the final prediction result.

Citation Information

Patent Citations

  • Intelligent pine wood nematode disease tree identification method based on sample imbalance

    CN115019200A

  • Bursaphelenchus xylophilus tree detection method based on attention mechanism and center circle sampling

    CN115761477A

  • Bursaphelenchus xylophilus tree detection method based on small target sample expansion and pooling weighting

    CN115909066A

  • Gawky (GW) nucleic acid molecules to control insect pests

    US20170218391A1