Burwood nematode disease tree grading detection method based on unmanned aerial vehicle remote sensing and improved PWD-YOLOv10n algorithm

Abstract

The invention discloses a pine wood nematode disease tree grading detection method based on unmanned aerial vehicle remote sensing and an improved PWD-YOLOv10n algorithm, and relates to the technical field of forestry harmful diseased tree detection, and the method comprises the steps: (1) diseased tree remote sensing image collection; (2) constructing a data set; (3) network model improvement; (4) selecting evaluation indexes; (5) result evaluation: in the step (3), by designing a PWD-YOLOv10 model with high precision and low leak detection, introducing a CAA attention mechanism module into a backbone network of the YOLOv10n model so as to enhance the feature extraction capability; c2fEMSCP modules are introduced into a backbone part and a neck part respectively, the recognition precision of the model is further improved, the omission ratio is effectively reduced, and rapid and accurate grading detection on the pine wood nematode disease tree in a complex environment is realized by training the high-performance PWD-YOLOv10n model.

Description

Technical Field

[0001] The present invention relates to the technical field of plant disease detection and identification, and in particular to a pine wood nematode disease tree classification detection method based on unmanned aerial vehicle remote sensing and an improved PWD-YOLOv10n algorithm. Background Art

[0002] In recent years, global warming has accelerated the spread of pine wilt disease, posing a serious threat to my country's forest resources. Timely and accurate identification of pine wilt diseased trees is crucial to preventing the further spread of the disease, protecting forest resources and promoting ecological sustainable development. Therefore, a rapid and efficient detection method is urgently needed to curb the spread of pine wilt disease.

[0003] At present, the detection of pine wood nematode diseased trees mainly relies on manual inspections and satellite remote sensing, which are not only time-consuming and labor-intensive, but also prone to missed detections and misjudgments. With the development of drone remote sensing and deep learning technology, plant disease detection has ushered in new technological opportunities. However, when faced with complex forest environments, existing deep learning models are still insufficient in identifying the characteristics of diseased trees, with low detection accuracy, high missed detection rate, and poor real-time performance, resulting in limited detection accuracy and efficiency. Summary of the invention

[0004] The present invention aims to solve the problems of low accuracy and high missed detection rate in the existing models in the detection of pine wilt diseased trees. To this end, the present invention proposes a pine wilt diseased tree classification detection method based on UAV remote sensing and improved PWD-YOLOv10n algorithm, which solves the problems of insufficient attention paid by the model to diseased tree targets and difficulty in accurately identifying small target early diseased trees. It can distinguish and identify different disease stages of pine wilt diseased trees, improve detection accuracy and effectively reduce missed detection rate.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] The present invention provides a method for grading and detecting pine wood nematode diseased trees based on unmanned aerial vehicle remote sensing, the steps of which include: (1) remote sensing image acquisition of diseased trees; (2) data set construction; (3) network model improvement; (4) selection of evaluation indicators; (5) result evaluation;

[0007] (1) Remote sensing image collection of diseased trees: Plan a drone aerial photography route in the selected research area, use drones to collect characteristic information of pine wood nematode disease, and save remote sensing image data;

[0008] First, we selected the area affected by pine wilt disease as the study area and conducted a preliminary field survey to determine the affected area. Then, we used a DJI drone equipped with a visible light camera to set the aerial photography route and flight parameters to collect remote sensing images of the study area.

[0009] (2) Dataset construction: The original images collected by the drone were cropped and enhanced, divided into training and validation sets in proportion, and the diseased trees were accurately labeled using annotation tools;

[0010] First, the original images collected by the drone were cropped into multiple 1280×1280 images to increase the visual scale of the diseased tree targets. Then, the parts containing the diseased tree targets were screened out from the cropped images and divided into training and validation sets at a ratio of 7:3; then, the divided data sets were subjected to image enhancement processing, including random rotation, flipping, saturation adjustment, etc., to increase data diversity; finally, according to the different stages of diseased trees, Labelimg tools were used for labeling, marking early diseased trees, mid-stage diseased trees, late diseased trees and dead trees respectively.

[0011] (3) Network model improvement: A high-precision, low-miss detection PWD-YOLOv10 model is designed, and the CAA attention mechanism module is introduced into the backbone network of the YOLOv10n model to enhance the feature extraction capability; subsequently, the C2f_EMSCP module is introduced into the backbone and neck parts respectively to further improve the recognition accuracy of the model and effectively reduce the missed detection rate.

[0012] The model is based on the YOLOv10n network, and the structure of the model includes:

[0013] Backbone

[0014] Layer 1: Use 3x3 convolution to downsample the input image and extract preliminary features.

[0015] Layer 2: 3x3 convolution is used again for downsampling to further extract features.

[0016] Layer 3: Multi-scale features are integrated through the C2f module to improve the model's perception of multi-scale information.

[0017] Layer 4: Introduce the CAA attention mechanism module to enhance the focus on key features.

[0018] Layer 5: Use 3x3 convolution to downsample the feature map to 1 / 8 of the original size.

[0019] Layer 6: Use the C2f module to further fuse features and enhance feature expression.

[0020] Layer 7: Apply the CAA module again to strengthen the model’s recognition of key features.

[0021] Layer 8: Down-sampling is performed through the SCDown module to obtain more in-depth information.

[0022] Layer 9: Use the C2f_EMSCP module and combine it with multi-scale convolution to improve the expressiveness of features.

[0023] Layer 10: Optimize feature channels through the CAA module to improve model accuracy.

[0024] Layer 11: Use the SCDown module to continue downsampling the feature map and extract deeper information.

[0025] Layer 12: Introduce the C2f_EMSCP module to perform multi-scale fusion and enhance feature extraction capabilities.

[0026] Layer 13: Apply the CAA module to weight important feature channels to further improve feature extraction.

[0027] Layers 14-15: SPPF and PSA modules are added respectively to enhance the model’s adaptability to features of different scales.

[0028] Neck (Neck Network)

[0029] Layer 16-17: Feature maps at different levels are fused through upsampling and feature concatenation.

[0030] Layer 18: The fused features are further processed through the C2f_EMSCP module to extract more effective information.

[0031] Layers 19-20: Upsampling and feature concatenation are used again to fuse multi-level feature maps.

[0032] Layer 21: Apply the C2f module to process the concatenated feature map and generate features suitable for small target detection.

[0033] Layer 22: Use 3x3 convolution to downsample the feature map to P4 scale.

[0034] Layer 23: Multi-scale information fusion is completed through feature splicing.

[0035] Layer 24: Generates a feature map suitable for medium-sized object detection through the C2f_EMSCP module.

[0036] Layer 25: Use the SCDown module to downsample the feature map to P5 scale.

[0037] Layer 26: Feature splicing module, further integrating multi-scale features.

[0038] Layer 27: Use the C2fCIB module to generate feature maps suitable for large object detection.

[0039] Head

[0040] Layer 28: Use the v10Detect target detection module to detect small, medium, and large targets on the feature maps of layers 21, 24, and 26, respectively, and output the final detection results.

[0041] Improvements to the model include:

[0042] ① Enhance model feature extraction through CAA (Context Anchor Attention) attention mechanism, highlight important contextual features, and suppress irrelevant background information.

[0043] The process is as follows:

[0044] First, the input feature map is globally averaged pooled to extract the global context information of each channel, thereby capturing the overall semantics of the image. Then, information fusion between channels is performed through 1×1 convolution to generate preliminary context anchors. Subsequently, 1×11 and 11×1 depthwise separable convolutions are used to extract spatial context features in the horizontal and vertical directions, respectively, to enhance the model's perception of horizontal and vertical structures. 1×1 convolution is then used to further integrate features and extract key context anchor information. Finally, the attention feature map M is generated through the Sigmoid activation function. CAA ,Right now

[0045] F pool =Conv(P avg (F)) (1)

[0046] F W =DWConv 1×11 (F pool ) (2)

[0047] F H =DWConv 11×1 (F W ) (3)

[0048] M CAA =Sigmoid(Conv 1×1 (F H )) (4)

[0049] Where F is the input feature map, DWConv 1×11 and DWConv 11×1 They are depth-wise separable convolutions with kernel sizes of 1×11 and 11×1 respectively;

[0050] ② Enhance the perception and expression capabilities of the model through C2f_EMSCP (Efficient Multi-Scale-Conv plus), which is particularly suitable for processing images with different scale features.

[0051] The working process of EMSCP is as follows:

[0052] First, the input feature map X∈RC×H×W is divided into four subgroups, and the number of channels in each subgroup is C / 4; assuming that the sub-feature map after grouping is X i (i=1,2,3,4), then:

[0053] X=[X1,X2,X3,X4] (5)

[0054] Where X i ∈R (C / 4)×H×W ;

[0055] Then, for each subgroup X i Convolution operations are performed using convolution kernels of different sizes (1×1, 3×3, 5×5, and 7×7) to generate multi-scale features Y i (i=1,2,3,4):

[0056] Y i =Convk i (X i ),k∈{1,3,5,7} (6)

[0057] where Y i ∈R (C / 4)×H×W ;

[0058] Next, the convolution results of each subgroup are stacked to form a multi-scale feature map Y:

[0059] Y=[Y1,Y2,Y3,Y4] (7)

[0060] where Y∈R C×H×W ;

[0061] Finally, a 1×1 convolution is used to perform channel fusion on the stacked multi-scale feature map Y to obtain the final output feature map M p :

[0062] M p =Conv 1×1 (Y) (8)

[0063] Among them, M p ∈R C×H×W .

[0064] (4) Selection of evaluation indicators: The present invention uses the following evaluation indicators to compare the performance of different pine wood nematode disease tree detection models, including: mean average precision (mAP), parameters, detection speed, GFLOPs, and missed detection rate. The calculation expressions of various indicators are:

[0065]

[0066]

[0067]

[0068]

[0069] In the formula, TP (True Positive) represents the positive samples correctly classified by the model, FP (False Positive) represents the negative samples incorrectly classified by the model, FN (False Negative) represents the positive samples incorrectly classified by the model, TN (True Negative) represents the negative samples correctly classified by the model, and AP average precision N represents the number of categories of diseased tree targets.

[0070] (5) Result evaluation: In order to verify the improved recognition effect of the improved model on diseased trees, the present invention conducts an ablation experiment to compare the improvement of each module on the model performance, and selects the current mainstream target detection model for comparative experiments.

[0071] First, the contribution of each module to the model performance was evaluated through ablation experiments, and modules were gradually removed or added to observe their impact on detection accuracy, parameter quantity, and inference speed, to determine the actual role of each module in improving the identification of diseased trees. Then, the improved model was compared with the current mainstream target detection model to comprehensively evaluate its performance from multiple indicators. This comparative experiment further verified the detection advantage of the improved model for pine wood nematode diseased trees under complex backgrounds.

[0072] The beneficial effects of the present invention are:

[0073] The method disclosed in the present invention is a pine wood nematode diseased tree grading detection method based on unmanned aerial vehicle remote sensing and improved PWD-YOLOv10n algorithm. By improving the model recognition accuracy, recognition missed detection rate and other aspects, a high-performance PWD-YOLOv10n model is trained, thereby realizing rapid and accurate detection of pine wood nematode diseased trees in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 The flowchart of the pine wood nematode diseased tree classification detection method of the present invention is as follows;

[0075] Figure 2 : The network structure diagram of the model of the present invention, the red box is the improved module. DETAILED DESCRIPTION

[0076] The present invention will be further described below in conjunction with the accompanying drawings:

[0077] 1) Remote sensing image acquisition of diseased trees; 2) Dataset construction; 3) Network model improvement; 4) Evaluation index selection; 5) Result evaluation.

[0078] The specific implementation process of this method is as follows:

[0079] Step 1: Remote sensing image acquisition of diseased trees. First, the area affected by pine wilt disease was selected as the study area, and a preliminary field survey was conducted to determine the affected area. Then, a DJI M300 RTK drone equipped with a Zenmuse H20 visible light camera was used to collect characteristic information of pine wilt disease, set the aerial route and flight parameters, and collect remote sensing images of the study area.

[0080] Step 2: Dataset construction. First, the original images collected by the drone were cropped into multiple 1280×1280 images to increase the visual scale of the diseased tree targets. Then, the parts containing the diseased tree targets were screened out from the cropped images and divided into training and validation sets in a ratio of 7:3. Subsequently, the divided datasets were subjected to image enhancement processing, including random rotation, flipping, saturation adjustment, etc., to increase data diversity. Finally, according to the different stages of infection of the diseased trees, the Labelimg tool was used for labeling, marking early diseased trees, mid-stage diseased trees, late diseased trees, and dead trees respectively.

[0081] Step 3: Network model improvement. To address the shortcomings of the existing model, the present invention introduces the CAA attention mechanism module into the backbone network of the YOLOv10n model to enhance the feature extraction capability. Subsequently, the C2f_EMSCP module is introduced into the backbone and neck parts respectively to further improve the recognition accuracy of the model and effectively reduce the missed detection rate.

[0082] The structure of the model includes:

[0083] Backbone

[0084] Layer 1: Use 3x3 convolution to downsample the input image and extract preliminary features;

[0085] Layer 2: Use 3x3 convolution again for downsampling to further extract features;

[0086] Layer 3: Multi-scale features are integrated through the C2f module to improve the model's perception of multi-scale information;

[0087] Layer 4: Introduce the CAA attention mechanism module to enhance the focus on key features;

[0088] Layer 5: Use 3x3 convolution to downsample the feature map to 1 / 8 of the original size;

[0089] Layer 6: Use the C2f module to further fuse features and enhance feature expression effects;

[0090] Layer 7: Apply the CAA module again to strengthen the model's recognition of key features;

[0091] Layer 8: Down-sampling is performed through the SCDown module to obtain more in-depth information;

[0092] Layer 9: Use the C2f_EMSCP module and combine multi-scale convolution to improve the expressiveness of features;

[0093] Layer 10: Optimize feature channels through CAA module to improve model accuracy;

[0094] Layer 11: Use the SCDown module to continue downsampling the feature map and extract deeper information;

[0095] Layer 12: Introduce the C2f_EMSCP module to perform multi-scale fusion and enhance feature extraction capabilities;

[0096] Layer 13: Apply CAA module to weight important feature channels to further improve feature extraction effect;

[0097] Layers 14-15: SPPF and PSA modules are added to enhance the model's adaptability to features of different scales.

[0098] Neck (Neck Network)

[0099] Layer 16-17: Feature maps at different levels are fused through upsampling and feature concatenation;

[0100] Layer 18: The fused features are further processed through the C2f_EMSCP module to extract more effective information;

[0101] Layers 19-20: Upsampling and feature concatenation are used again to fuse multi-level feature maps;

[0102] Layer 21: Apply the C2f module to process the concatenated feature maps and generate features suitable for small target detection;

[0103] Layer 22: Use 3x3 convolution to downsample the feature map to P4 scale;

[0104] Layer 23: Multi-scale information fusion is completed through feature splicing;

[0105] Layer 24: Generates a feature map suitable for medium target detection through the C2f_EMSCP module;

[0106] Layer 25: Use the SCDown module to downsample the feature map to P5 scale;

[0107] Layer 26: Feature splicing module, further integrating multi-scale features;

[0108] Layer 27: Use the C2fCIB module to generate feature maps suitable for large object detection;

[0109] Head

[0110] Layer 28: Use the v10Detect target detection module to detect small, medium, and large targets on the feature maps of layers 21, 24, and 26, respectively, and output the final detection results.

[0111] The improved PWD-YOLOv10n model is as follows Figure 2 shown.

[0112] ① Enhance model feature extraction through CAA (Context Anchor Attention) attention mechanism, highlight important context features, and suppress irrelevant background information. Its working process is as follows:

[0113] First, the input feature map is globally averaged pooled to extract the global context information of each channel, thereby capturing the overall semantics of the image. Next, a 1×1 convolution is used to fuse information between channels and generate preliminary context anchors. Subsequently, 1×11 and 11×1 depthwise separable convolutions are used to extract spatial context features in the horizontal and vertical directions, respectively, to enhance the model's perception of horizontal and vertical structures. The features are further integrated through 1×1 convolution to extract key context anchor information. Finally, the attention feature map M is generated through the Sigmoid activation function. CAA ,Right now

[0114] F pool =Conv(P avg (F)) (1)

[0115] F W =DWConv 1×11 (F pool ) (2)

[0116] F H =DWConv 11×1 (F W ) (3)

[0117] M CAA =Sigmoid(Conv 1×1 (F H )) (4)

[0118] Where F is the input feature map, DWConv1×11 and DWConv11×1 are depth-wise separable convolutions with kernel sizes of 1×11 and 11×1, respectively.

[0119] ② Enhance the perception and expression capabilities of the model through C2f_EMSCP (Efficient Multi-Scale-Conv plus), which is particularly suitable for processing images with different scale features. The working process of EMSCP is as follows:

[0120] First, the input feature map is divided into four subgroups, and the number of channels in each subgroup is C / 4. Assuming that the sub-feature maps after grouping are Xi (i = 1, 2, 3, 4), then:

[0121] X=[X1,X2,X3,X4] (5)

[0122] Where X i ∈R (C / 4)×H×W .

[0123] Then, convolution operations are performed on each subgroup Xi using convolution kernels of different sizes (1×1, 3×3, 5×5, and 7×7) to generate multi-scale features Yi (i=1, 2, 3, 4):

[0124] Y i =Convk i (X i ),k∈{1,3,5,7} (6)

[0125] where Y i ∈R (C / 4)×H×W .

[0126] Next, the convolution results of each subgroup are stacked to form a multi-scale feature map Y:

[0127] Y=[Y1,Y2,Y3,Y4] (7)

[0128] where Y∈R C×H×W .

[0129] Finally, a 1×1 convolution is used to perform channel fusion on the stacked multi-scale feature map Y to obtain the final output feature map M p :

[0130] M p =Conv 1×1 (Y) (8)

[0131] Among them, M p ∈R C×H×W .

[0132] Step 4: Selection of evaluation indicators. The present invention uses the following evaluation indicators to compare the performance of different pine wood nematode disease tree detection models, including: mean average precision (mAP), parameters, and missed detection rate. The calculation expressions of various indicators are:

[0133]

[0134]

[0135]

[0136]

[0137] In the formula, TP (True Positive) represents the positive samples correctly classified by the model, FP (False Positive) represents the negative samples incorrectly classified by the model, FN (False Negative) represents the positive samples incorrectly classified by the model, TN (True Negative) represents the negative samples correctly classified by the model, and AP average precision N represents the number of categories of diseased tree targets.

[0138] Step 5: Result evaluation: In order to verify the improved recognition effect of the improved model on diseased trees, the present invention conducts an ablation experiment to compare the improvement of each module on the model performance, and selects the current mainstream target detection model for comparative experiments.

[0139] ① Evaluate the contribution of each module to the model performance through ablation experiments, gradually remove or add modules to observe their impact on detection accuracy, parameter quantity and inference speed, and determine the actual role of each module in improving the diseased tree identification effect. The improved modules and evaluation indicators are as follows:

[0140]

[0141] The data in the above table shows that after adding the CAA attention mechanism module to the YOLOv10n model, the model's feature extraction ability has been improved, and mAP@0.5 has increased by 1.1%, but the missed detection rate has increased slightly by 0.2%. The C2f_EMSCP module is further introduced to enhance feature fusion and target detection accuracy, and mAP@0.5 has increased by another 1.0%, and the missed detection rate has decreased by 0.5%, showing the detection effect of C2f_EMSCP on diseased trees in complex backgrounds. At the same time, the number of parameters has only increased slightly, and the overall model performance has been effectively improved.

[0142] Since most missed detections are concentrated in the early and middle stages of diseased trees, in order to further verify the effect of the improved model on diseased tree classification and identification, the identification of different diseased tree types was analyzed to evaluate the contribution of each module to the improvement of the model classification performance. The improved modules and evaluation indicators are as follows:

[0143]

[0144]

[0145] After the CAA attention mechanism module was introduced into the YOLOv10n model, the model's feature extraction capability was significantly enhanced, and the AP of early and mid-stage diseased trees increased by 1.0% and 0.9% respectively. Although the missed detection rate increased slightly, the effect on the identification of diseased tree types was significant. After further adding the C2f_EMSCP module, the model's feature fusion capability was enhanced, and the AP of early diseased trees increased by 1.7%, and the AP of dead diseased trees increased by 0.8%, verifying the advantages of this module in complex background and diseased tree classification detection. At the same time, the increase in the number of parameters was small, indicating that the improved model effectively controlled the computational cost while improving the recognition accuracy.

[0146] ② Through comparative experiments between the improved model and the current mainstream target detection model, its performance was comprehensively evaluated from multiple indicators. This comparative experiment further verified the detection advantage of the improved model for pine wood nematode diseased trees under complex backgrounds. The evaluation indicators of each model recognition are as follows:

[0147]

[0148] As can be seen from the data in the table, different models have significant differences in mAP@0.5 and parameter size. Traditional models such as Faster R-CNN and SSD have accuracies of 0.759 and 0.790 respectively, and relatively large parameter sizes of 108MB and 92.1MB respectively, which are difficult to meet lightweight requirements. In contrast, the YOLO series of models show higher accuracy and significant lightweight advantages, especially YOLOv5s, YOLOv7-tiny and YOLOv8n, with mAP@0.5 of 0.897, 0.880 and 0.91 respectively, and a significant reduction in parameters. YOLOv10n optimizes the parameter size to 5.51MB while achieving mAP@0.5 of 0.922, showing high accuracy and lightness. The improved model PWD-YOLOv10n further improved mAP@0.5 to 0.943 based on YOLOv10n, while only slightly increasing the parameter size to 5.77MB, showing the best balance between accuracy and lightweight. It is suitable for resource-constrained but high-precision diseased tree detection tasks, and can distinguish and identify different disease stages (such as early, middle, late, and dieback) of pine wood nematode diseased trees.

[0149] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in the technical field, several improvements can be made without departing from the principles of the present invention. These improvements should also be regarded as the scope of protection of the present invention without paying creative labor.

Claims

1. A pine wood nematode disease tree classification detection method based on UAV remote sensing and improved PWD-YOLOv10n algorithm, characterized by: The steps include: (1) remote sensing image acquisition of diseased trees; (2) data set construction; (3) network model improvement; (4) selection of evaluation indicators; (5) result evaluation; (1) Remote sensing image collection of diseased trees: Plan a drone aerial photography route in the selected research area, use the drone to collect characteristic information of pine wood nematode disease, and save the remote sensing image data; (2) Dataset construction: The original images collected by the drone were cropped and enhanced, divided into training and validation sets in proportion, and the diseased trees were accurately labeled using annotation tools; (3) Network model improvement: A high-precision, low-miss detection PWD-YOLOv10 model is designed, and the CAA attention mechanism module is introduced into the backbone network of the YOLOv10n model to enhance the feature extraction capability; then, the C2f_EMSCP module is introduced into the backbone and neck parts respectively to further improve the recognition accuracy of the model and effectively reduce the missed detection rate; (4) Selection of evaluation indicators: The present invention uses the following evaluation indicators to compare the performance of different pine wood nematode disease tree detection models, including: average accuracy value, parameter quantity, detection speed, GFLOPs, and missed detection rate; (5) Result evaluation: In order to verify the improved model’s recognition effect on diseased trees, an ablation experiment was conducted to compare the improvement of each module on model performance, and the current mainstream target detection model was selected for comparative experiments.

2. The method for grading and detecting pine wood nematode diseased trees based on UAV remote sensing and improved PWD-YOLOv10n algorithm according to claim 1, characterized in that: The model is based on the YOLOv10n network, and the structure of the model includes: Backbone Layer 1: Use 3x3 convolution to downsample the input image and extract preliminary features; Layer 2: Use 3x3 convolution again for downsampling to further extract features; Layer 3: Multi-scale features are integrated through the C2f module to improve the model's perception of multi-scale information; Layer 4: Introduce the CAA attention mechanism module to enhance the focus on key features; Layer 5: Use 3x3 convolution to downsample the feature map to 1 / 8 of the original size; Layer 6: Use the C2f module to further fuse features and enhance feature expression effects; Layer 7: Apply the CAA module again to strengthen the model's recognition of key features; Layer 8: Down-sampling is performed through the SCDown module to obtain more in-depth information; Layer 9: Use the C2f_EMSCP module and combine multi-scale convolution to improve the expressiveness of features; Layer 10: Optimize feature channels through CAA module to improve model accuracy; Layer 11: Use the SCDown module to continue downsampling the feature map and extract deeper information; Layer 12: Introduce the C2f_EMSCP module to perform multi-scale fusion and enhance feature extraction capabilities; Layer 13: Apply CAA module to weight important feature channels to further improve feature extraction effect; Layers 14-15: SPPF and PSA modules are added to enhance the model's adaptability to features of different scales. Neck (Neck Network) Layer 16-17: Feature maps at different levels are fused through upsampling and feature concatenation; Layer 18: The fused features are further processed through the C2f_EMSCP module to extract more effective information; Layers 19-20: Upsampling and feature concatenation are used again to fuse multi-level feature maps; Layer 21: Apply the C2f module to process the concatenated feature maps and generate features suitable for small target detection; Layer 22: Use 3x3 convolution to downsample the feature map to P4 scale; Layer 23: Multi-scale information fusion is completed through feature splicing; Layer 24: Generates a feature map suitable for medium target detection through the C2f_EMSCP module; Layer 25: Use the SCDown module to downsample the feature map to P5 scale; Layer 26: Feature splicing module, further integrating multi-scale features; Layer 27: Use the C2fCIB module to generate feature maps suitable for large object detection; Head Layer 28: Use the v10Detect target detection module to detect small, medium, and large targets on the feature maps of layers 21, 24, and 26, respectively, and output the final detection results.

3. The method for grading and detecting pine wood nematode diseased trees based on UAV remote sensing and improved PWD-YOLOv10n algorithm according to claim 1, characterized in that: In step (1), the area affected by pine wilt disease was selected as the study area, and a preliminary field survey was conducted to determine the affected area. Then, a DJI drone equipped with a visible light camera was used to set the aerial photography route and flight parameters to collect remote sensing images of the study area.

4. The method for grading and detecting pine wood nematode diseased trees based on UAV remote sensing and improved PWD-YOLOv10n algorithm according to claim 1, characterized in that: In step (2), the original image collected by the drone is cropped into multiple 1280×1280 images to increase the visual scale of the diseased tree targets; the part containing the diseased tree targets is screened out from the cropped images and divided into a training set and a validation set in a ratio of 7:3; the divided data set is subjected to image enhancement processing, including random rotation, flipping, saturation adjustment, etc., to increase data diversity; according to the different infection stages of the diseased trees, the Labelimg tool is used for labeling, marking early diseased trees, mid-stage diseased trees, late diseased trees and dead trees respectively.

5. The method for grading and detecting pine wood nematode diseased trees based on UAV remote sensing and improved PWD-YOLOv10n algorithm according to claim 4, characterized in that: In step (3), the CAA attention mechanism is used to enhance model feature extraction, highlight important contextual features, and suppress irrelevant background information; The process is as follows: First, the input feature map is globally averaged pooled to extract the global context information of each channel, thereby capturing the overall semantics of the image. Then, information fusion between channels is performed through 1×1 convolution to generate preliminary context anchors. Subsequently, 1×11 and 11×1 depthwise separable convolutions are used to extract spatial context features in the horizontal and vertical directions, respectively, to enhance the model’s perception of horizontal and vertical structures. The features are further integrated through 1×1 convolution to extract key contextual anchor information; finally, the attention feature map M is generated through the Sigmoid activation function. CAA ,Right now F pool =Conv(P avg (F)) (1) F W =DWConv 1×11 (F pool ) (2) F H =DWConv 11×1 (F W ) (3) M CAA =Sigmoid(Conv 1×1 (F H )) (4) Where F is the input feature map, DWConv 1×11 and DWConv 11×1 They are depth-wise separable convolutions with kernel sizes of 1×11 and 11×1, respectively.

6. The method for grading and detecting pine wood nematode diseased trees based on UAV remote sensing and improved PWD-YOLOv10n algorithm according to claim 1, characterized in that: In step (3), the perception and expression capabilities of the model are enhanced by C2f_EMSCP, which is particularly suitable for processing images with different scale features; The working process of EMSCP is as follows: First, the input feature map X∈R C×H×W Divided into four subgroups, the number of channels in each subgroup is C / 4; assuming that the sub-feature map after grouping is X i (i=1,2,3,4), then: X=[X1,X2,X3,X4] (5) Where X i ∈R (C / 4)×H×W ; Then, for each subgroup X i Convolution operations are performed using convolution kernels of different sizes (1×1, 3×3, 5×5, and 7×7) to generate multi-scale features Y i (i=1,2,3,4): AND i =Convk i (X i ),k∈{1,3,5,7} (6) where Y i ∈R (C / 4)×H×W ; Next, the convolution results of each subgroup are stacked to form a multi-scale feature map Y: Y=[Y1,Y2,Y3,Y4] (7) where Y∈R C×H×W ; Finally, a 1×1 convolution is used to perform channel fusion on the stacked multi-scale feature map Y to obtain the final output feature map M p : M p =Conv 1×1 (Y) (8) Among them, M p ∈R C×H×W .

7. The method for grading and detecting pine wood nematode diseased trees based on UAV remote sensing and improved PWD-YOLOv10n algorithm according to claim 1, characterized in that: In step (4), the calculation expressions of various indicators are: In the formula, TP represents the positive samples correctly classified by the model, FP represents the negative samples incorrectly classified by the model, FN represents the positive samples incorrectly classified by the model, TN represents the negative samples correctly classified by the model, and AP average precision N represents the number of categories of diseased tree targets.

Images