A multi-stage pine wilt disease detection method

Abstract

This invention discloses a multi-stage pine wilt disease detection method, belonging to the technical field of pine wilt disease-infected wood detection. Based on the different stages of pine wood disease infection, this invention annotates multispectral image samples and trains a pine wilt disease detection model using a training set of annotated multispectral image samples. The trained pine wilt disease detection model is then used to identify the pine wilt disease stage in multispectral images of target pine forests. This method can accurately identify infected wood at different stages of pine wilt disease, solving the problem of low accuracy in identifying different stages of pine wilt disease in current pine wilt disease detection methods.

Description

Technical Field

[0001] This invention relates to a multi-stage pine wilt disease detection method, belonging to the technical field of pine wilt disease detection in infected wood. Background Technology

[0002] In the prevention and control of pine wilt disease, early detection and precise location of infected trees are crucial for effective management. However, traditional methods of monitoring infected trees have significant limitations. While manual ground patrols are intuitive, they are inefficient, costly, and limited by terrain, making rapid surveys of large forest areas difficult. Although satellite remote sensing technology has a wide coverage area, its spatial resolution is limited, revisit cycles are long, and it is easily affected by weather conditions such as clouds, making it difficult to meet the timeliness requirements for precise monitoring of infected trees.

[0003] The introduction of multispectral remote sensing technology has provided a new solution for early diagnosis. Studies have shown that pine wilt nematode infection can cause changes in tree physiological parameters, such as decreased chlorophyll content and water stress. These changes show higher sensitivity in specific spectral bands such as red edge and near-infrared, as well as vegetation indices such as NDVI. However, most methods only use the original spectral bands and fail to effectively mine vegetation index features that are more sensitive to early disease responses. Especially in complex habitats such as mixed coniferous and broad-leaved forests, the spectral features of different tree species interfere with each other, resulting in significant background noise, which further increases the difficulty of early identification of infected trees. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-stage pine wilt disease detection method that can accurately identify infected wood at different stages of pine wilt disease, thus solving the problem of low accuracy in identifying different stages of pine wilt disease in current pine wilt disease detection methods.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution: This invention provides a method for detecting pine wilt disease at multiple stages, including: Acquire multispectral images of the target pine forest; The multispectral images of the target pine forest are input into a pre-trained pine wilt disease detection model to identify targets in the target pine forest that have pine wilt disease and the stage of disease in the targets. The training of the pine wilt disease detection model includes: Obtain multispectral image samples of pine trees at different stages of pine wilt disease infection; Based on the different stages of lesion infection, multispectral image samples are labeled to form a training set of labeled multispectral image samples; A pine wilt disease detection model was trained using a training set to learn the labeled features corresponding to different stages of lesion infection. Based on the identification results of different disease infection stages of pine trees during the learning process of the pine wilt disease detection model, the model parameters are adjusted until the pine wilt disease detection model can accurately identify targets with pine wilt disease and the disease stage of the targets.

[0006] Furthermore, the multispectral image includes the red-edge band, near-infrared band, NDVI vegetation index, and RGB visible light band.

[0007] Furthermore, the pine wilt disease detection model includes: The feature extraction module adopts a dual-branch structure of visible light and multispectral light, performs multiple feature extractions on the RGB part and the multispectral part of the multispectral image, and then fuses the extracted features to obtain the first fused feature. The multi-scale attention aggregation module uses both spatial and channel paths to enhance the first fused feature, resulting in enhanced features. The fusion module is used to progressively fuse high-level semantic features and low-level detail features of the enhanced features to obtain a second fused feature; The detection module is used to detect targets based on fusion features, identify targets with pine wilt disease and the stage of the disease.

[0008] Furthermore, the method of training the pine wilt disease detection model using a training set to learn the labeled features corresponding to different stages of lesion infection includes: The first initial feature and the second initial feature are extracted multiple times from the multispectral image at both the visible light level and the multispectral level. The first and second initial features extracted multiple times are fused to obtain the first fused feature; The first fusion feature is enhanced in space and along the channel path to obtain the enhanced feature; The enhanced features are progressively fused with high-level semantic features and low-level detail features to obtain the second fused feature.

[0009] Furthermore, the extraction of first and second initial features from the multispectral image multiple times at both the visible light and multispectral levels includes: The first initial feature is extracted using the following formula: ; The second initial feature is extracted using the following formula: ; in, Indicates the first initial feature, Indicates the second initial feature; This represents a feature map of a multispectral image in the visible light plane. Represents the feature map of a multispectral image at the multispectral level; This indicates extended layer processing, with a size of [size missing]. To expand the multiplier Initial number of channels Extended to ; This indicates depthwise convolution processing, with a size of [size missing]. stride is To expand the multiplier Initial number of channels Expand to .

[0010] Furthermore, the process of fusing the first and second initial features extracted multiple times to obtain the first fused feature includes: The first initial feature is weighted using the following formula: ; The first initial feature is weighted using the following formula: ; in, This represents the first initial feature after weighted processing. This represents the second initial feature after weighted processing. Indicates feature weighting; The fusion process is represented by the following formula: ; in, Indicates the first fusion feature; Indicates fusion processing, This represents a feature map of a multispectral image in the visible light plane. Represents the feature map of a multispectral image at the multispectral level; This indicates projection layer processing, with a size of [size missing]. The number of channels is .

[0011] Furthermore, the enhancement processing of the first fused feature in space and channel path to obtain the enhanced feature includes: The following formula represents the enhancement process performed on the spatial path: ; in, Represents a spatial attention map; This represents the Sigmoid activation function; This indicates convolution processing, with a size of [size missing]. ; This indicates that the output is concatenated along the channel dimension; This indicates convolution processing, with a size of [size missing]. ; This indicates convolution processing, with a size of [size missing]. ; Indicates the first fusion feature; The following formula represents the enhancement process performed on the channel path: ; in, Represents the channel attention map; This represents the parameters of the first fully connected layer. This represents the parameters of the second fully connected layer. Represents the linear rectified function. This indicates global average pooling. This indicates max pooling. The enhancement process is performed using the following formula: ; in, Indicates enhanced features, This indicates multiplication by channel.

[0012] Furthermore, the progressive fusion of high-level semantic features and low-level detail features in the enhancement features to obtain a second fused feature includes: The enhanced features are convolved to obtain high-level and low-level features; Low-level features are used to guide high-level features in the channel and spatial dimensions, and a second fused feature is obtained through progressive fusion, represented as: ; ; ; ; ; in, Indicates high-level characteristics, Indicates low-level features, Indicates the height of low-level features. Indicates the width of the low-level feature; Indicates upsampling; This indicates convolution processing, with a size of [size missing]. ; Represents the ReLU activation function; This represents high-level features that are guided by low-level features in the channel dimension. This indicates multiplication by channel. Indicates channel weight, This represents higher-level features that are further guided by lower-level features in the spatial dimension; Representing a spatial attention map, This indicates that the output is concatenated along the channel dimension. This indicates convolution processing, with a size of [size missing]. ; This indicates the output splicing characteristics after channel compression. Represents a learnable scalar. This indicates the second fusion feature.

[0013] Furthermore, the step of adjusting the model parameters based on the identification results of different disease infection stages of pine trees during the learning process of the pine wilt disease detection model, until the pine wilt disease detection model can accurately identify targets with pine wilt disease and the disease stage of the targets, includes: Bounding box regression loss and classification loss are used as loss functions for model training. Backpropagation is used to calculate the total loss to update the model parameters. The total loss is continuously reduced through iterative optimization.

[0014] Furthermore, the total loss is expressed as: ; ; ; in, Represents the total loss function; This represents the bounding box regression loss function. Indicates intersection, union, and ratio. This indicates the calculation of the Euclidean distance between two points. Indicates the center point of the prediction box. Indicates the center point of the true bounding box. This represents the diagonal length of the smallest bounding rectangle that simultaneously contains both the predicted and ground truth bounding boxes. Indicates the weighting coefficient; This represents a consistency parameter used to measure the consistency of the aspect ratio; Represents the classification loss function. Indicates the number of samples. Indicates sample The true label, Indicates sample Predicted probability of infection with diseases.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: 1. This invention annotates multispectral image samples according to different stages of pine wood disease infection, trains a pine wilt disease detection model using a training set of annotated multispectral image samples, and uses the trained pine wilt disease detection model to identify the pine wilt disease stage in multispectral images of target pine forests. It can accurately identify infected trees at different stages of pine wilt disease, solving the problem of low accuracy in identifying different stages of pine wilt disease in current pine wilt disease detection methods.

[0016] 2. The pine wilt disease detection model established in this invention extracts first and second initial features multiple times from multispectral images at both the visible light and multispectral levels. Based on the first and second initial features, enhancement processing is performed in space and channel paths to obtain enhanced features. Finally, the enhanced features are progressively fused with high-level semantic features and low-level detail features to obtain a second fused feature that combines high-level semantics and low-level details, which significantly improves the ability to identify early lesions. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the pine wilt disease detection model provided in an embodiment of the present invention; Figure 2 This is a flowchart of the training process for the pine wilt disease detection model provided in this embodiment of the invention. Figure 3 This is a schematic diagram of the feature extraction module in the pine wilt disease detection model provided in this embodiment of the invention (feature extraction of LWABlock branch in the visible light layer of multispectral image). Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention. Example

[0019] A method for detecting pine wilt disease at multiple stages includes: Acquire multispectral images of the target pine forest; The multispectral images of the target pine forest are input into a pre-trained pine wilt disease detection model to identify targets in the target pine forest that have pine wilt disease and the stage of disease in the targets. It should be noted that the pine forest multispectral image in this embodiment includes the red edge band, near-infrared band, NDVI vegetation index, and RGB visible light band.

[0020] like Figure 1 As shown, the detection model for pine wilt disease includes: The feature extraction module adopts a dual-branch structure of visible light and multispectral light, performs multiple feature extractions on the RGB part and the multispectral part of the multispectral image, and then fuses the extracted features to obtain the first fused feature. The multi-scale attention aggregation module uses an attention mechanism (MSAA) to enhance the first fused feature through both spatial and channel paths, resulting in enhanced features; The fusion module includes a multi-layer fusion path (HLAF), which is used to progressively fuse high-level semantic features and low-level detail features of the enhanced features to obtain a second fused feature. The detection module includes multiple detection heads (Detect) used to detect targets based on fused features, identify targets with pine wilt disease and the stage of the disease.

[0021] like Figure 1 As shown, the training of the pine wilt disease detection model includes: Obtain multispectral image samples of pine trees at different stages of pine wilt disease infection; Based on the different stages of lesion infection, multispectral image samples are labeled to form a training set of labeled multispectral image samples; The pine wilt disease detection model was trained using a training set to learn the labeled features corresponding to different stages of lesion infection; specifically: like Figure 3 As shown, for the multispectral image, first initial features and second initial features are extracted multiple times at both the visible light level and the multispectral level, including: the first initial feature is extracted using the following formula: ; The second initial feature is extracted using the following formula: ; in, Indicates the first initial feature, Indicates the second initial feature; This represents a feature map of a multispectral image in the visible light plane. Represents the feature map of a multispectral image at the multispectral level; This indicates extended layer processing, with a size of [size missing]. To expand the multiplier Initial number of channels Expand to ; This indicates depthwise convolution processing, with a size of stride is To expand the multiplier Initial number of channels Expand to ; The first and second initial features extracted multiple times are fused to obtain the first fused feature, which includes: weighting the first initial feature using the following formula: ; The first initial feature is weighted using the following formula: ; in, This represents the first initial feature after weighted processing. This represents the second initial feature after weighted processing. Indicates feature weighting; The fusion process is represented by the following formula: ; in, Indicates the first fusion feature; Indicates fusion processing; This indicates projection layer processing, with a size of [size missing]. The number of channels is ; The first fusion feature is enhanced in space and along the channel path to obtain the enhanced feature, including: The following formula represents the enhancement process performed on the spatial path: ; in, Represents a spatial attention map; This represents the Sigmoid activation function; This indicates convolution processing, with a size of [size missing]. ; This indicates that the output is concatenated along the channel dimension; This indicates convolution processing, with a size of [size missing]. ; This indicates convolution processing, with a size of [size missing]. ; Indicates the first fusion feature; The following formula represents the enhancement process performed on the channel path: ; in, Represents the channel attention map; This represents the parameters of the first fully connected layer. This represents the parameters of the second fully connected layer. Represents the linear rectified function. This indicates global average pooling. This indicates max pooling. The enhancement process is performed using the following formula: ; in, Indicates enhanced features, This indicates multiplication by channel; The enhanced features are progressively fused with high-level semantic features and low-level detail features to obtain the second fused feature, which includes: The enhanced features are convolved to obtain high-level and low-level features; Low-level features are used to guide high-level features in the channel and spatial dimensions, and a second fused feature is obtained through progressive fusion, represented as: ; ; ; ; ; in, Indicates high-level characteristics, Indicates low-level features, Indicates the height of low-level features. Indicates the width of the low-level feature; Indicates upsampling; Represents the ReLU activation function; This represents high-level features that are guided by low-level features in the channel dimension. Indicates channel weight, This represents higher-level features that are further guided by lower-level features in the spatial dimension; Represents a spatial attention map; This indicates the output splicing characteristics after channel compression. Represents a learnable scalar. This indicates the second fusion feature.

[0022] Based on the pine wilt disease detection model's identification results of different disease infection stages in pine trees during the learning process, the model parameters were adjusted until the pine wilt disease detection model could accurately identify targets with pine wilt disease and the disease stage of the targets. Specifically: Bounding box regression loss and classification loss are used as loss functions for model training. Backpropagation is used to calculate the total loss to update model parameters, and iterative optimization is employed to continuously reduce the total loss. The total loss is expressed as: ; ; ; Represents the total loss function; This represents the bounding box regression loss function. Indicates intersection, union, and ratio. This indicates the calculation of the Euclidean distance between two points. Indicates the center point of the prediction box. Indicates the center point of the true bounding box. This represents the diagonal length of the smallest bounding rectangle that simultaneously contains both the predicted and ground truth bounding boxes. Indicates the weighting coefficient; This represents a consistency parameter used to measure the consistency of the aspect ratio; Represents the classification loss function. Indicates the number of samples. Indicates sample The true label, Indicates sample Predicted probability of infection with diseases.

[0023] To verify the superiority of the method proposed in this embodiment, a comparative experiment was conducted using the following existing target detection models: A classic two-stage object detection model (Faster Region-Based Convolutional Neural Network, Faster R-CNN); an advanced single-stage object detection model (You Only Look Once version 11 Small YOLO, YOLOv11s); an iterative cross-attention fusion object detection model (Iterative Cross-Attention Fusion, ICAfusion); and a super-resolution-based object detection model (Super-ResolutionYOLO, Super-YOLO). The test metrics include Average Percision (AP), Frames Per Second (FPS), Parameters, Floating Point Operations Per Second (FLOPs), and F1 score; in, ; Represents the probability density function; Indicates precision. ; Indicates recall rate, ; In this context, TP (True Positive) represents a correct positive detection, which is the bounding box of pine wood lesions correctly predicted by the model; FP (False Positive) represents a false positive detection, which is the model incorrectly detecting background or healthy pine wood as diseased wood; and FN (False Negative) represents a missed positive target, which is the diseased wood that actually exists in the image but was not detected by the model. ; This represents the time it takes for the model to process one image. ; Indicates the number of output channels of the model. This indicates the number of channels in the input feature map. Indicates the height of the convolution kernel. Indicates the width of the convolution kernel; ; Indicates the height of the input feature map. Input the width of the feature map; ; in, This represents the F1 score.

[0024] The experimental results are shown in the table below: Table 1. Experimental Results

[0025] It should be noted that mAP represents the average accuracy, and mAP@0.5(%) represents the average accuracy when IoU = 0.5; mAP-Early represents the average accuracy of the model in detecting early-stage diseased trees, and mAP-Early@0.5(%) represents the average accuracy of the model in detecting early-stage diseased trees when IoU = 0.5.

[0026] First, the method proposed in this embodiment demonstrates its superior early detection capability with an AP-Early@0.5 of 85.3%, primarily due to the effective fusion of multispectral information such as red-edge, near-infrared, and NDVI. Simultaneously, the method requires only 4.38M parameters to achieve an mAP@0.5 of 88.0%, revealing the advantage of its lightweight architecture in feature representation efficiency. Furthermore, the inference speed of 83.6 FPS ensures its potential for real-time application on edge devices such as drones. In summary, the method proposed in this embodiment establishes significant advantages in detection accuracy, model lightweighting, and operational efficiency.

[0027] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0028] This application is described with reference to flowchart illustrations of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each step in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 One or more processes or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0029] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0030] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

[0031] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for detecting pine wilt disease at multiple stages, characterized in that, include: Acquire multispectral images of the target pine forest; The multispectral images of the target pine forest are input into a pre-trained pine wilt disease detection model to identify targets in the target pine forest that have pine wilt disease and the stage of disease in the targets. The training of the pine wilt disease detection model includes: Obtain multispectral image samples of pine trees at different stages of pine wilt disease infection; Based on the different stages of lesion infection, multispectral image samples are labeled to form a training set of labeled multispectral image samples; A pine wilt disease detection model was trained using a training set to learn the labeled features corresponding to different stages of lesion infection, including: The first initial feature and the second initial feature are extracted multiple times from the multispectral image at both the visible light level and the multispectral level. The first and second initial features extracted multiple times are fused to obtain the first fused feature; The first fusion feature is enhanced in space and along the channel path to obtain the enhanced feature; The enhanced features are progressively fused with high-level semantic features and low-level detail features to obtain the second fused feature, which includes: The enhanced features are convolved to obtain high-level and low-level features; Low-level features are used to guide high-level features in the channel and spatial dimensions, and a second fused feature is obtained through progressive fusion, represented as: ； ； ； ； ； in, Indicates high-level characteristics, Indicates low-level features, Indicates the height of low-level features. Indicates the width of the low-level feature; Indicates upsampling; This indicates convolution processing, with a size of [size missing]. ; Represents the ReLU activation function; This represents high-level features that are guided by low-level features in the channel dimension. This indicates multiplication by channel. Indicates channel weight, This represents higher-level features that are further guided by lower-level features in the spatial dimension; Representing a spatial attention map, This indicates that the output is concatenated along the channel dimension. This indicates convolution processing, with a size of [size missing]. ; This indicates the output splicing characteristics after channel compression. Represents a learnable scalar. Indicates the second fusion feature; Based on the identification results of different disease infection stages of pine trees during the learning process of the pine wilt disease detection model, the model parameters are adjusted until the pine wilt disease detection model can accurately identify targets with pine wilt disease and the disease stage of the targets.

2. The method for detecting multi-stage pine wilt disease according to claim 1, characterized in that, The multispectral image includes the red-edge band, near-infrared band, NDVI vegetation index, and RGB visible light band.

3. The method for detecting multi-stage pine wilt disease according to claim 1, characterized in that, The pine wilt disease detection model includes: The feature extraction module adopts a dual-branch structure of visible light and multispectral light, performs multiple feature extractions on the RGB part and the multispectral part of the multispectral image, and then fuses the extracted features to obtain the first fused feature. The multi-scale attention aggregation module uses both spatial and channel paths to enhance the first fused feature, resulting in enhanced features. The fusion module is used to progressively fuse high-level semantic features and low-level detail features of the enhanced features to obtain a second fused feature; The detection module is used to detect targets based on fusion features, identify targets with pine wilt disease and the stage of the disease.

4. The method for detecting multi-stage pine wilt disease according to claim 1, characterized in that, The process of extracting first and second initial features multiple times from the multispectral image at both the visible light and multispectral levels includes: The first initial feature is extracted using the following formula: ； The second initial feature is extracted using the following formula: ； in, Indicates the first initial feature, Indicates the second initial feature; This represents the feature map of a multispectral image in the visible light layer. Represents the feature map of a multispectral image at the multispectral level; This indicates extended layer processing, with a size of [size missing]. To expand the multiplier Initial channel number Expand to ; This indicates depthwise convolution processing, with a size of [size missing]. stride is To expand the multiplier Initial channel number Expand to .

5. The method for detecting multi-stage pine wilt disease according to claim 1, characterized in that, The process of fusing the first and second initial features extracted multiple times to obtain the first fused feature includes: The first initial feature is weighted using the following formula: ； The first initial feature is weighted using the following formula: ； in, This represents the first initial feature after weighted processing. This represents the second initial feature after weighted processing. Indicates feature weighting; The fusion process is represented by the following formula: ； in, Indicates the first fusion feature; Indicates fusion processing, This represents the feature map of a multispectral image in the visible light layer. Represents the feature map of a multispectral image at the multispectral level; This indicates projection layer processing, with a size of [size missing]. The number of channels is .

6. The method for detecting multi-stage pine wilt disease according to claim 1, characterized in that, The enhancement processing of the first fused feature in space and channel path to obtain the enhanced feature includes: The following formula represents the enhancement process performed on the spatial path: ； in, Represents a spatial attention map; This represents the Sigmoid activation function; This indicates convolution processing, with a size of [size missing]. ; This indicates that the output is concatenated along the channel dimension; This indicates convolution processing, with a size of [size missing]. ; This indicates convolution processing, with a size of [size missing]. ; Indicates the first fusion feature; The following formula represents the enhancement process performed on the channel path: ； in, Represents the channel attention map; This represents the parameters of the first fully connected layer. This represents the parameters of the second fully connected layer. Represents the linear rectified function. This indicates global average pooling. This indicates max pooling. The following formula is used for enhancement: ； in, Indicates enhanced features, This indicates multiplication by channel.

7. The method for detecting multi-stage pine wilt disease according to claim 1, characterized in that, The process involves adjusting the model parameters based on the identification results of different disease infection stages in pine trees during the learning process of the pine wilt disease detection model, until the pine wilt disease detection model can accurately identify targets with pine wilt disease and the disease stage of the targets, including: Bounding box regression loss and classification loss are used as loss functions for model training. Backpropagation is used to calculate the total loss to update the model parameters. The total loss is continuously reduced through iterative optimization.

8. The method for detecting multi-stage pine wilt disease according to claim 7, characterized in that, The total loss is expressed as: ； ； in, Represents the total loss function; This represents the bounding box regression loss function; Indicates intersection, union, and ratio; (·) indicates the calculation of the Euclidean distance between two points; Indicates the center point of the prediction box. Indicates the center point of the true bounding box; This represents the diagonal length of the smallest bounding rectangle that simultaneously contains both the predicted and ground truth boxes. Indicates the weighting coefficient; This represents a consistency parameter used to measure the consistency of the aspect ratio; This represents the classification loss function.

Images