Intelligent detection method and system for forest pine wood nematode disease

Abstract

The invention discloses a forest pine wood nematode disease intelligent detection method and system, and belongs to the technical field of image recognition and target detection, and the method comprises the steps: obtaining forest pine wood historical images, and constructing a pine wood image sample training set; and constructing and training a pine wood image recognition network based on the pine wood image sample. A pine forest is divided into a plurality of areas, and pine forest images of different areas are acquired. And performing preliminary identification on the acquired pine forest image based on the first identification network, and outputting a suspected disease image. And performing deep identification on the suspected disease image based on a second identification network, and outputting a disease detection result. According to the method, characteristics such as growth potential are obtained through the CNN network, an accurate threshold value is set for preliminary screening, the preliminarily screened area is further judged through the improved MobileNet V3-small network, basic characteristics such as needle leaf color change, growth potential stop and resin secretion of initial symptoms of the pine wood nematode disease are fully considered, and the accuracy of the initial symptoms of the pine wood nematode disease is improved. And the detection capability and the recognition precision of sparse small disease spots in the early stage of pine wood nematode disease are enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image recognition and target detection, and particularly relates to a forest pine wood nematode disease intelligent detection method and system. BACKGROUND

[0002] Pine wilt disease (PWD) is a devastating forest disease caused by pine wood nematode (PWN), which is known as the 'cancer' of pine trees. PWD is characterized by fast spread, strong pathogenicity and great difficulty in prevention and control, posing a serious threat to pine forest resources, ecological security and forestry economy in China. The initial symptoms of PWD are extremely subtle, which are manifested as needle discoloration, growth stagnation and reduced resin secretion, and are difficult to be detected by the naked eye. When the needles turn red-brown and the whole tree dies, it is already in the late stage of the disease, which has lost its value for prevention and control. Therefore, developing efficient and accurate early detection technology for PWD is of great significance for 'early detection and early disposal' and effectively containing the spread of the disease. At present, the early diagnosis techniques for PWD mainly include the following categories: (1) manual ground survey method, which relies on the experience of forestry technicians to observe the color change of needle, growth potential, resin secretion and traces of insect vectors for judgment. This method is highly dependent on personal experience, low in efficiency and difficult to implement large-scale monitoring. (2) Punching and resin flow method, which indirectly judges by detecting the change of resin secretion capacity of trees. This method is a lossy detection and is easily affected by environmental factors. (3) Immunological diagnosis method, such as ELISA detection, which has high accuracy but needs sampling and laboratory analysis, resulting in high cost and long cycle, which is difficult to meet the demand of large-area and real-time monitoring. With the development of remote sensing and artificial intelligence technology, disease detection methods based on image recognition have become a research hotspot. This kind of method usually uses unmanned aerial vehicles or satellite platforms to obtain forest images, and uses computer vision algorithms to automatically identify suspected diseased trees, which has the advantages of wide coverage, high efficiency and non-contact. There are some existing patent solutions for PWD detection through image recognition. For example, the Chinese invention patent with publication number CN114140703A discloses an intelligent recognition method and system for forest PWD; the Chinese invention patent with publication number CN114595975A discloses an unmanned aerial vehicle remote sensing PWD monitoring method based on a deep learning model. Although these existing solutions use image recognition methods, they generally have the following defects: 1. Limited recognition accuracy: Most of the above solutions use general image recognition models or deep learning frameworks, which fail to closely combine the unique microscopic morphological characteristics of PWD that appear in the early stage for targeted modeling. The model lacks a deep understanding and learning of the unique pathological characteristics of the disease, resulting in low recognition accuracy and robustness for early and atypical symptoms, and easy to miss detection or false positives. 2. Single model structure and weak small target recognition ability: In the early stage of the disease, features such as disease spots appear as small targets in the image. The models used in existing solutions often have simple feature fusion methods, which are difficult to effectively capture and fuse multi-scale disease features, and are likely to lose key small target information in the feature extraction process or smooth it out in the feature fusion process, thereby failing to effectively recognize early subtle disease spots.3. The calculation efficiency and accuracy are difficult to balance: the massive image data of the entire forest area need to be calculated by complex deep learning model, which has large calculation load and low efficiency, and it is difficult to meet the demand of rapid and periodic survey of large area forest in practical application. SUMMARY

[0003] To solve the above problems, the present application provides a forest pine wood nematode disease intelligent detection method and system to solve the problems of low recognition accuracy of existing forest pine wood nematode disease, weak small target recognition ability, and difficulty in effectively capturing and fusing multi-scale disease characteristics, so as to effectively identify early fine disease spots.

[0004] A forest pine wood nematode disease intelligent detection method, comprising: obtaining forest pine wood historical images to construct a pine wood image sample training set; constructing and training a pine wood image recognition network based on the pine wood image sample, including a first recognition network and a second recognition network; dividing the pine forest into multiple regions and obtaining pine forest images of different regions; preliminarily identifying the obtained pine forest images based on the first recognition network, and outputting suspected disease images; deeply identifying the suspected disease images based on the second recognition network, and outputting disease detection results.

[0005] According to an embodiment of the present application, obtaining forest pine wood historical images to construct a pine wood image sample training set further comprises: selecting forest pine wood historical images of different growth periods to create a pine wood image database, the forest pine wood historical images including healthy pine wood images and pine wood nematode disease images; selecting multiple pine wood nematode disease images and forest pine wood images taken by a UAV from the pine wood image database to construct the pine wood image sample training set.

[0006] According to an embodiment of the present application, constructing and training a pine wood image recognition network based on the pine wood image sample further comprises: constructing the first recognition network using a CNN convolutional neural network, and inputting the pine wood image sample training set into the first recognition network for training; constructing the second recognition network using an improved MobileNet V3-small network, establishing small target enhanced samples, and inputting them into the second recognition network for training; The improved MobileNet V3-small network includes 11 inverted residual modules bneck, namely {bneck0, bneck1, ..., bneck10}. Among them, bneck0, bneck3, bneck4, bneck5, and bneck8 contain SE Blocks. Lightweight multi-head self-attention mechanism modules MHSA are inserted after bneck5 and bneck8 respectively.

[0007] According to a specific embodiment of the present invention, a first recognition network is constructed using a CNN convolutional neural network, and the training set of pine wood image samples is input into the first recognition network for training, which further includes: The first recognition network is constructed using a CNN convolutional neural network, and the loss function is determined. L MSE for:

[0008] In the formula, For the true value, The value is the model prediction, and N is the number of samples. The pine wood sample image is input into the first recognition network for feature extraction, resulting in a growth potential feature vector {ΔH, Sshape, CYSR, DSR}, where... Sshape = Area(C) / Area(A)

[0009] CYSR= /

[0010] DSR= /

[0011] Where ΔH is the canopy color, Sshape is the canopy integrity, CYSR is the proportion of new shoots, DSR is the proportion of dead shoots, Area() is the number of pixels, C is the original canopy, and A is the convex polygon formed by the points of the outer contour of the original canopy C, i.e., the convex hull. For yellowing pixels, Total pixels of the canopy For new shoot pixels, , Reddish-brown pixels; Input the growth potential feature vector {ΔH, Sshape, CYSR, DSR} into the loss function. L MSE The system is trained to output predicted values ​​for each feature in different tree age ranges.

[0012] According to a specific embodiment of the present invention, constructing a second recognition network using an improved MobileNet V3-small network, and further comprising establishing small target augmentation samples and inputting them into the second recognition network for training: A second recognition network is constructed using an improved MobileNet V3-small network, and the loss function is determined. L total for: L total = λ cls L cls + λ box L box + λ attn L attn In the formula, L cls For classification loss function, L box For box regression loss function, L attn To compare the distillation loss function, λ cls , λ box , λ attn These are the weighting coefficients for each loss function; Disease lesion images of 64×64~128×128 pixels were randomly cropped from the pine wood image sample training set and pasted onto healthy pine wood images with a transparency of 0.3~0.5 to obtain small target augmented samples; Small target augmented samples are input into a second recognition network for feature extraction and feature fusion to obtain multi-scale fused features; Multi-scale fused features are input into the loss function. L total The training process is performed, and the predicted values ​​of the multi-scale fused features are output.

[0013] According to a specific embodiment of the present invention, inputting small target enhancement samples into a second recognition network for feature extraction and feature fusion to obtain multi-scale fused features further includes: The small target enhancement sample is input into the second recognition network for feature extraction, resulting in detail features, structural features and semantic features. The detail features include detail edge information, the structural features include structural and color block information, and the semantic features include semantic information. Using semantic features as the query vector Q and structural features as the key vector K and value vector V, the semantic features and structural features are weighted and fused to obtain semantically enhanced structural features. The semantically enhanced structural features are then concatenated with the original structural features to obtain the mid-level fused features. Using the mid-level fusion feature as the query vector Q and the detail feature as the key vector K and value vector V, the structural feature and detail feature are weighted and fused to obtain the structural feature-enhanced detail feature. The structural feature-enhanced detail feature is then upsampled to the original resolution to obtain the multi-scale fusion feature.

[0014] According to a specific embodiment of the present invention, the preliminary identification of the acquired pine forest image based on the first identification network and the output of the suspected disease image further includes: Based on the first recognition network, feature extraction is performed on the acquired pine forest image to obtain the growth potential feature vector {ΔH,Sshape,CYSR,DSR}. Based on the feature threshold table, query the growth potential feature table vector {ΔH} t Sshape t ,CYSR t DSR t}, and concatenate it with the growth potential eigenvector {ΔH, Sshape, CYSR, DSR} to obtain an 8-dimensional eigenvector {ΔH, Sshape, CYSR, DSR, ΔH}. t Sshape t ,CYSR t DSR t}; The 8-dimensional feature vector is input into the first recognition network for convolution processing, and the first suspicion score is output. Determine whether the first suspicion score is greater than the first preset threshold. If so, output the corresponding suspected disease image.

[0015] According to a specific embodiment of the present invention, the method of performing deep recognition on suspected disease images based on a second recognition network and outputting disease detection results further includes: Based on the second recognition network, feature extraction is performed on suspected disease images to obtain detailed features, structural features and semantic features; By fusing detailed features, structural features, and semantic features, multi-scale fused features are obtained. The multi-scale fusion features are input into the second recognition network for calculation, and the second suspicion score is output. Determine whether the second suspicion score is greater than the second preset threshold. If so, output the disease detection result.

[0016] A smart detection system for forest pine wilt disease includes: The training set creation module is used to acquire historical images of pine timber in forests and construct a training set of pine timber image samples. An image recognition network creation module is used to construct and train a pine wood image recognition network based on pine wood image samples, including a first recognition network and a second recognition network; The image acquisition module is used to divide the pine forest into multiple areas and acquire images of the pine forest in different areas; The first identification module is used to perform preliminary identification on the acquired pine forest images based on the first identification network and output suspected disease images; The second recognition module is used to perform deep recognition on suspected disease images based on the second recognition network and output disease detection results.

[0017] According to a specific embodiment of the present invention, the training set creation module further includes: The database creation module is used to select historical images of forest pine wood at different growth stages to create a pine wood image database. The historical images of forest pine wood include images of healthy pine wood and images of pine wood nematode disease. The training set generation module is used to select multiple images of pine wilt disease from the pine wood image database and forest pine wood images taken by drones to construct a pine wood image sample training set.

[0018] According to a specific embodiment of the present invention, the image recognition network creation module further includes: The first recognition network creation module is used to construct the first recognition network using a CNN convolutional neural network and input the pine wood image sample training set into the first recognition network for training. The second recognition network creation module is used to construct a second recognition network using an improved MobileNet V3-small network, and to create small target augmentation samples and input them into the second recognition network for training. The improved MobileNet V3-small network includes 11 inverted residual modules bneck, namely {bneck0, bneck1, ..., bneck10}. Among them, bneck0, bneck3, bneck4, bneck5, and bneck8 contain SE Blocks. Lightweight multi-head self-attention mechanism modules MHSA are inserted after bneck5 and bneck8 respectively.

[0019] According to a specific embodiment of the present invention, the first identification module further includes: The first feature extraction module is used to extract features from the acquired pine forest image based on the first recognition network to obtain a growth potential feature vector; The feature concatenation module is used to query the growth potential feature vector based on the feature threshold table and concatenate it with the growth potential feature vector to obtain an 8-dimensional feature vector. The first calculation module is used to input the 8-dimensional feature vector into the first recognition network for convolution processing and output the first suspicion score. The first judgment module is used to determine whether the first suspicion score is greater than the first preset threshold. If so, the corresponding suspected disease image is output.

[0020] According to a specific embodiment of the present invention, the second identification module further includes: The second feature extraction module is used to extract features from suspected disease images based on the second recognition network to obtain detailed features, structural features and semantic features; The feature fusion module is used to fuse detailed features, structural features, and semantic features to obtain multi-scale fused features. The second calculation module is used to input multi-scale fused features into the second recognition network for calculation and output a second suspicion score. The second judgment module is used to determine whether the second suspicion score is greater than the second preset threshold. If so, the disease detection result is output.

[0021] An electronic device includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the above-mentioned intelligent detection method for forest pine wilt disease.

[0022] A computer-readable storage medium storing a computer program, which is loaded and executed by a processor to implement the above-mentioned intelligent detection method for forest pine wilt disease.

[0023] Compared with the prior art, this application has the following advantages: 1. Significantly improved the accuracy and reliability of disease identification.

[0024] This invention combines initial screening and fine screening. In the initial screening stage, a lightweight CNN network architecture is selected to filter out suspected disease images, significantly reducing the amount of data requiring detailed computation. In the fine screening stage, an improved MobileNet V3-small network is established by introducing a lightweight multi-head self-attention mechanism module after specific bottleneck layers (bneck-5 and bneck-8). This network uses an attention mechanism to guide the fusion of semantic features to structural features, and structural features to guide the fusion of detail features, achieving dynamic adjustment of the receptive field and dual spatial-channel attention. The fine screening stage performs deep discrimination on suspected disease images, avoiding missed detections or misjudgments caused by a single model's "one-size-fits-all" approach, thus improving the overall accuracy and reliability of identification. This network can more intelligently focus on key regions related to disease in the image (such as lesions of specific shapes or abnormally colored coniferous clusters) and effectively fuse multi-scale features (details, structure, semantics), significantly improving the ability to distinguish lesions of different morphologies against complex backgrounds. This overcomes the problem in existing technologies where simple feature vector fusion easily leads to the smoothing or misjudgment of small target features.

[0025] 2. It enhances the ability to detect early, sparse, small lesions, enabling early detection and early warning of diseases.

[0026] This invention employs a small-target enhancement strategy during the training phase (randomly pasting small lesion images onto healthy images), which specifically expands the quantity and quality of "sparse, small, and varied" early lesion samples in the training data. This forces the neural network to learn the characteristics of tiny lesions, greatly enhancing the model's sensitivity and robustness in identifying early diseases. By utilizing image recognition and spectral detection for mutual verification, it fully considers the basic characteristics of early symptoms of pine wilt disease, such as needle color changes, cessation of growth, and resin secretion, to identify pine wilt disease. This overcomes the shortcomings of traditional methods and some image recognition methods that are insensitive to early symptoms, gaining valuable time for timely control measures.

[0027] 3. While ensuring high precision, it also takes into account detection efficiency, making it suitable for widespread application.

[0028] In the fine screening stage, the MobileNetV3-small network selected in this invention is itself a lightweight network with low computational burden. Building upon this, this invention selectively inserts attention modules at higher levels (bneck-5 and bneck-8) rather than at lower high-resolution layers, cleverly avoiding an explosive increase in computational load while introducing a powerful attention mechanism. Combined with standardized drone photography specifications (uniform location, angle, altitude, and time), the consistency and optimality of input image quality are ensured, reducing model performance fluctuations and additional preprocessing overhead caused by differences in image quality, and significantly improving detection efficiency.

[0029] 4. It has achieved automated and standardized intelligent monitoring, reducing reliance on manual labor and costs.

[0030] This invention automates the entire process from image acquisition (automated drone inspection) and preprocessing to intelligent recognition, completely changing the past model that relied on manual on-site surveys—a model that was inefficient, subjective, and costly. Strict periodic inspections and standardized shooting protocols ensure strong temporal comparability of monitoring data. By analyzing the growth characteristics of images from different periods, dynamic tracking and trend analysis of the health status of individual trees or specific forest areas can be achieved, providing a more scientific and comprehensive basis for forestry management decisions. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of an intelligent detection method for forest pine wilt disease according to an embodiment of the present invention.

[0033] Figure 2 This is a flowchart of a method for constructing a training set of pine image samples by obtaining historical images of forest pine wood according to an embodiment of the present invention.

[0034] Figure 3 This is a flowchart of a method for constructing and training a pine wood image recognition network based on pine wood image samples according to an embodiment of the present invention.

[0035] Figure 4 This is a flowchart of a method according to an embodiment of the present invention, which uses a CNN convolutional neural network to construct a first recognition network and inputs a training set of pine wood image samples into the first recognition network for training.

[0036] Figure 5 This is a flowchart of a method according to an embodiment of the present invention, which uses an improved MobileNet V3-small network to construct a second recognition network, establishes small target enhancement samples, and inputs them into the second recognition network for training.

[0037] Figure 6 This is a flowchart of a method for inputting small target enhancement samples into a second recognition network for feature extraction and feature fusion according to an embodiment of the present invention.

[0038] Figure 7This is a flowchart of a method for preliminary identification of an acquired pine forest image based on a first identification network, according to an embodiment of the present invention.

[0039] Figure 8 This is a flowchart of a method for deep recognition of suspected disease images based on a second recognition network according to an embodiment of the present invention.

[0040] Figure 9 This is a structural diagram of an intelligent detection system for forest pine wilt disease provided according to an embodiment of the present invention.

[0041] Figure 10 This is a structural diagram of a training set creation module provided according to an embodiment of the present invention.

[0042] Figure 11 This is a structural diagram of an image recognition network creation module provided according to an embodiment of the present invention.

[0043] Figure 12 This is a structural diagram of the first identification module provided according to an embodiment of the present invention.

[0044] Figure 13 This is a structural diagram of the second identification module provided according to an embodiment of the present invention.

[0045] Figure 14 This is a schematic diagram of a computer device structure according to an embodiment of the present invention.

[0046] Figure label: 01-Training set creation module; 02-Image recognition network creation module; 03-Image acquisition module; 04-First recognition module; 05-Second recognition module; 011-Database creation module; 012-Training set generation module; 021 - First identification network creation module; 022 - Second identification network creation module; 041-First Feature Extraction Module; 042-Feature Concatenation Module; 043-First Calculation Module; 044-First Judgment Module; 051 - Second feature extraction module; 052 - Feature fusion module; 053 - Second calculation module; 054 - Second judgment module. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] Example 1 Additional aspects and advantages of embodiments of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of embodiments of the invention. Figures 1-8 This invention provides an intelligent detection method for forest pine wilt disease, comprising: S1: Obtain historical images of forest pine wood and construct a training set of pine wood image samples.

[0049] S2: Construct and train a pine wood image recognition network based on pine wood image samples, including a first recognition network and a second recognition network.

[0050] S3: Divide the pine forest into multiple regions and obtain images of the pine forest in different regions.

[0051] S4: Based on the first recognition network, perform preliminary recognition on the acquired pine forest images and output suspected disease images.

[0052] S5: Based on the second recognition network, perform deep recognition on suspected disease images and output disease detection results.

[0053] Specifically, step S1, acquiring historical images of forest pine wood and constructing a training set of pine wood image samples, further includes: S11: Select historical images of forest pine timber from different growth stages to create a pine timber image database.

[0054] S12: Select multiple images of pine wilt disease and forest pine images taken by drones from the pine image database to construct a pine image sample training set.

[0055] In a specific embodiment of this invention, considering that pine tree crowns are mostly conical or broad pyramidal, and the crown width often "first rapidly widens and then slowly increases in height" with age, drone aerial photography under side lighting can reveal a stepped outline, which displays the following characteristics in the image: a smooth curve at the top, small gaps between crowns, and consistent shadow direction. Therefore, it is easier to determine the growth vigor of pine forests using image recognition technology. This invention obtains samples by photographing pine forests with drones, and simultaneously obtains pine wilt disease samples by constructing a database, combining both to form a sample set for subsequent training of the pine wood image recognition network. Specifically, firstly, historical images of forest pine wood at different growth stages and with different health conditions are selected to establish a pine wood image database. The historical images of forest pine wood include images of healthy pine wood and images of pine wilt disease, each image carrying corresponding label information. Then, multiple high-quality images of pine wilt disease are selected from the established pine wood image database as positive samples. Meanwhile, images of healthy pine wood taken by drones in real forest environments were selected as negative samples to form the initial training sample set, providing reliable data support for subsequent training of the image recognition network.

[0056] Specifically, step S2, which involves constructing and training a pine wood image recognition network based on pine wood image samples, further includes: S21: A first recognition network is constructed using a CNN convolutional neural network, and the training set of pine wood image samples is input into the first recognition network for training, further including: S211: Construct the first recognition network using a CNN convolutional neural network and determine the loss function. L MSE for: (1) In the formula, For the true value, is the model's predicted value, and N is the number of samples.

[0057] In a specific embodiment of the present invention, a lightweight CNN network is used to construct a first recognition network to learn and recognize growth potential characteristics. By making a preliminary judgment on the image, areas that may be affected by pine wilt disease are screened out so that they can be input into the second recognition network to accurately identify the images of suspected pine wilt disease areas, further determine whether pine wilt disease has occurred, and output the specific plants that have been affected.

[0058] S212: Input the pine wood sample image into the first recognition network for feature extraction to obtain the growth potential feature vector {ΔH, Sshape, CYSR, DSR}, where... Sshape = Area(C) / Area(A) (2) (3) CYSR= / (4) DSR= / (5) Where ΔH is the canopy color, Sshape is the canopy integrity, CYSR is the proportion of new shoots, DSR is the proportion of dead shoots, Area() is the number of pixels, C is the original canopy, and A is the convex polygon formed by the points of the outer contour of the original canopy C, i.e., the convex hull. For yellowing pixels, Total pixels of the canopy For new shoot pixels, , The pixels are reddish-brown.

[0059] In one specific embodiment of the present invention, the canopy is a vertical layer composed of the collection of the canopies of all individual pine trees. H represents canopy color. It is calculated by randomly selecting N canopy pixels in the HSV color space and calculating the standard deviation of their H values ​​(i.e., hue standard deviation). The smaller the ΔH, the more consistent the hue, indicating a healthy and uniform canopy. The larger the ΔH, the more chaotic the hue, indicating yellowing, disease spots, or shadows, and decreased growth. Canopy width refers to the maximum horizontal width of a single pine tree's canopy. Canopy integrity (Sshape) is represented by notches: Canopy notch = 1 - (Actual projected area of ​​the canopy ÷ Area of ​​the circumscribed bulge). The larger the notch value, the more severe the "gap" caused by disease. New shoots are "candle-shaped" elongated branches that emerge from the terminal or lateral buds of pine trees. The proportion of new shoots (CYSR) = (number of light green / yellowish-green "candle-shaped" shoots visible to the naked eye ÷ total number of shoots in the canopy) × 100%. Dead shoots are branches that have died due to dehydration caused by disease or climate. The proportion of dead shoots (DSR) = (number of brown, silver-gray, bare, and leafless shoots visible to the naked eye ÷ total number of shoots in the canopy) × 100%.

[0060] Specifically, for Sshape (canopy integrity), the ExG (Excess Green Index) of each pixel in the sample image is first calculated in the RGB space. The Excess Green Index is a mathematical operation that enhances the green component in a color image to effectively separate green vegetation from the background. Its calculation formula is as follows: ExG = 2G - RB (6) In the formula, G is the pixel value in the green channel, R is the pixel value in the red channel, and B is the pixel value in the blue channel.

[0061] After obtaining the ExG image, the color ExG image is converted into a grayscale image, and the Otsu method is used to calculate the optimal threshold T on the grayscale image to divide the grayscale image into black and white parts. The white pixel region is the original tree crown C, and the convex polygon formed by the points of the outer contour of the original tree crown C is the convex hull A. The number of pixels of the original tree crown C and the convex hull A are calculated respectively, and then the crown integrity Sshape is calculated according to formula (2).

[0062] For ΔH (canopy color), CYSR (new shoot ratio), and DSR (withered shoot ratio), the original image is first converted from RGB space to HSV space. The growth potential characteristics ΔH, CYSR, and DSR are obtained by setting different thresholds for the three components H (hue), S (saturation), and V (brightness). For example: pixels with H∈[25,70] and S≥80 and V≥40 are set as "healthy" green; pixels with H∈[40,70] and V>120 and S>100 are set as "bright" green, i.e., new shoots; pixels with H∈[10,25] and S≥100 are set as yellow; pixels with H∈[0,15] or H>165 and S>90 are set as reddish-brown, i.e., old shoots. Therefore, ΔH, CYSR, and DSR are calculated by formulas (3), (4), and (5), respectively.

[0063] S213: Input the growth potential feature vector {ΔH,Sshape,CYSR,DSR} into the loss function. L MSE The system is trained to output predicted values ​​for each feature in different tree age ranges.

[0064] In a specific embodiment of the present invention, a first recognition network is first established using a lightweight CNN convolutional neural network. The lightweight CNN convolutional neural network can be a 1×1 convolution followed by a Sigmoid function, which outputs a confidence score between 0 and 1 to obtain the characteristics of the growth potential of the pine forest in the region. Then, the pine wood sample images in the sample set are preprocessed and input into the first recognition network for training to obtain the growth potential feature vector {ΔH,Sshape,CYSR,DSR}. The obtained growth potential feature vector {ΔH,Sshape,CYSR,DSR} is then input into the CNN network for 1×1 convolution processing to convert the 4 channels into a 1 channel image. Finally, the convolved image is sent into the Sigmoid function to output the "suspicion" score between 0 and 1. The present invention uses the MSE loss function shown in formula (1) for training, sets the label {healthy=0, diseased=1}, and obtains the predicted value of each feature in the 0-1 interval. Through the above training, the threshold of each feature in different tree age intervals can be obtained, as shown in Table 1: Table 1

[0065] S22: A second recognition network is constructed using an improved MobileNet V3-small network. Small target augmentation samples are generated and input into the second recognition network for training. Further steps include: S221: Construct a second recognition network using an improved MobileNet V3-small network and determine the loss function. L total for: L total = λ cls L cls + λ box L box + λ attn L attn (7) In the formula, L cls For classification loss function, L box For box regression loss function, L attn To compare the distillation loss function, λ cls , λ box , λ attn These are the weighting coefficients for each loss function.

[0066] In one specific embodiment of this invention, an improved MobileNet V3-small network is used to construct a second recognition network for further recognition of the initially determined image. The MobileNet V3-small network structure is a lightweight network structure with a built-in attention module (SE Block, Squeeze-and-Excitation Block), which can use the attention mechanism to fuse multi-scale features. The MobileNet V3-small network structure contains 11 bnecks (residual modules), namely {bneck0, bneck1,...bneck10}, among which bneck0, bneck3, bneck4, bneck5, and bneck8 contain SE Blocks, which improve representation capabilities through compression, excitation, and recalibration. However, the SE Block only performs global average pooling and channel weight calculation, and the receptive field is fixed at global average and only channel attention is provided, which makes structural features prone to confusion. This invention improves the MobileNet V3-small network structure by inserting a lightweight multi-head self-attention (MHSA) module after bneck-5 and bneck-8. This enables dynamically variable receptive fields and dual spatial-channel attention. The lightweight MHSA is only inserted after bneck-5 and bneck-8, avoiding the computational explosion caused by the high resolution of bneck-0. Specifically, in bneck0, bneck3, bneck4, bneck5, and bneck8, each residual module includes a 1×1 up-dimensional convolution, a 3×3 depthwise separable convolution, an SE channel attention module, and a 1×1 down-dimensional convolution. Each SE channel attention module includes a global average pooling layer and a fully connected layer. The 1×1 upscaling convolution upscales the input channels, providing sufficient capacity for subsequent depthwise convolutions. The 3×3 depthwise separable convolution performs sliding window processing on multiple channels separately, followed by pointwise linear combination to extract local spatial features (e.g., detail features, structural features, and semantic features). The SE channel attention module first performs global average pooling on the local spatial features, then passes them through two fully connected layers. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Hard-sigmoid activation function. The resulting value is then multiplied by each element of the original local spatial feature matrix. The 1×1 downscaling convolution reduces the dimensionality to the initial dimension through pointwise convolution. Then, two lightweight MHSA (lightweight multi-head self-attention mechanism modules) are used to fuse the extracted local features (detail features, structural features, and semantic features) to obtain multi-scale fused features. The role of each bneck (residual module) is shown in Table 2. Table 2

[0067] S222: Randomly crop diseased images of 64×64~128×128 from the pine wood image sample training set, and paste them onto healthy pine wood images with a transparency of 0.3~0.5 to obtain small target enhancement samples.

[0068] In one specific embodiment of the present invention, to enhance robustness, small target enhancement is used to improve the neural network's ability to identify early-stage "sparse, small, and scattered" lesions. In this embodiment, lesion images of 64×64~128×128 pixels are randomly cropped from the pine image sample training set and pasted onto the same or other healthy pine images with a transparency of 0.3~0.5. The pasting positions are randomly offset within the canopy area, ensuring at least three small targets in each image, thus obtaining small target enhancement samples. These enhanced samples are then input into the second recognition network for training.

[0069] S223: Input the enhanced small target samples into the second recognition network for feature extraction and feature fusion to obtain multi-scale fused features, which further include: S2231: Input the small target enhancement sample into the second recognition network for feature extraction to obtain detail features, structural features and semantic features. The detail features include detail edge information, the structural features include structural and color block information, and the semantic features include semantic information.

[0070] In a specific embodiment of the present invention, small target enhancement samples are input into a second recognition network for feature extraction, specifically through the inverted residual module bneck in the improved MobileNet V3-small network, to obtain multi-scale features, including detail features, structural features, and semantic features. Detail features have high resolution and are mainly used to capture detailed edge information of the image, such as the serrations at the tips of pine needles or the edges of withered branches. Structural features have medium resolution and are used to capture structural and color patch information, such as lesion patches. Semantic features have low resolution and are used to capture abstract semantic information, such as distinguishing the entire tree crown or main branch area. The feature information at each scale is shown in Table 3. Table 3

[0071] S2232: Using semantic features as the query vector Q and structural features as the key vector K and value vector V, the semantic features and structural features are weighted and fused to obtain semantically enhanced structural features. The semantically enhanced structural features are then concatenated with the original structural features to obtain the mid-level fused features.

[0072] S2233: Using the mid-level fusion feature as the query vector Q and the detail feature as the key vector K and value vector V, the structural feature and the detail feature are weighted and fused to obtain the structural feature-enhanced detail feature. The structural feature-enhanced detail feature is then upsampled to the original resolution to obtain the multi-scale fusion feature.

[0073] In a specific embodiment of this invention, after multi-scale feature extraction is completed, two lightweight MHSA modules are used to perform feature fusion on the extracted detail features, structural features, and semantic features. Specifically, this includes semantic feature-guided structural feature fusion and structural feature-guided detail feature fusion. 1) The semantic feature-guided structural feature fusion method is described below: Semantic features are upsampled to the resolution of structural features. The semantic features are used as the query vector Q, and the structural features as the key vector K and value vector V. A similarity matrix between Q and K is calculated to measure the correlation between semantics and structure. For example, to determine which regions in the structural features belong to the main body of the long structure, V is weighted according to the similarity between Q and K to strengthen semantically relevant regions and weaken irrelevant regions, such as strengthening the main structure of the long structure and weakening background noise. The resulting semantically enhanced structural features at the same resolution as Q are then upsampled to the original resolution of the structural features and concatenated with the original structural features to obtain the intermediate-level fusion features.

[0074] 2) The method for fusing structural features with detailed features is as follows: The mid-level fusion features are upsampled to the resolution of the detail features. The mid-level fusion features are used as the query vector Q, and the detail features are used as the key vector K and value vector V. The similarity matrix between Q and K is calculated to identify which details are important to the current structure, such as the precise pixel-level position of the long material outline. Then, V is weighted according to the similarity to enhance the structure-related details and suppress irrelevant textures, such as enhancing the true outline edges and suppressing rust spots and reflective noise. The structure-guided detail features with the same resolution as Q are output, and the output features are upsampled to the original image resolution as the final fusion result.

[0075] S224: Input the multi-scale fused features into the loss function L total The training process is performed, and the predicted values ​​of the multi-scale fused features are output.

[0076] In a specific embodiment of the present invention, the extracted multi-scale fusion features are input into the loss function shown in formula (7) for training, and the training samples are consistent with the training samples in the preliminary screening stage to ensure the recognition accuracy of the model.

[0077] Specifically, step S3 divides the pine forest into multiple regions and acquires images of the pine forest in different regions.

[0078] In one specific embodiment of the present invention, based on the geographical information, area, and terrain complexity of the target pine forest, the entire target pine forest is divided into multiple rectangular inspection areas of similar size and regular shape. The principle for dividing each inspection area is to ensure that a single drone flight can complete the image acquisition task for that area within its battery life, and to ensure a certain degree of overlap (e.g., 5%) between adjacent areas to avoid missed detections. After dividing the forest into multiple areas, the drone periodically acquires images of the target pine forest in multiple areas to obtain the images to be inspected.

[0079] In a specific embodiment of the present invention, to ensure the consistency, comparability, and high quality of the acquired images, strict photographic specifications are formulated and implemented, specifically including the following: (1) Shooting positions and angles: For each inspection area, a series of fixed shooting positions (e.g., GPS coordinates) are set in advance. At each shooting position, the drone must maintain a uniform shooting angle, that is, use a side lighting angle of 30°-45° to shoot, so as to use the light and shadow effect to enhance the texture features of the tree crown and branches in the image, which is convenient for the display of disease features.

[0080] (2) Shooting height: The flight height of the drone (i.e., the shooting height) should be greater than 30 meters to ensure that macroscopic tree canopy information is obtained, while ensuring that the image resolution is ≤3cm, so that the details of the lesions can be clearly distinguished.

[0081] (3) Shooting time: Image acquisition should be carried out during a period of stable lighting conditions each day, specifically from 10:00 to 14:00 local time, in order to minimize the interference to image quality caused by dim or strong light or color temperature differences at dawn and dusk.

[0082] (4) Shooting frequency: According to the preset inspection plan, the drone is driven to fly to the fixed shooting position of each inspection area in turn. The shooting frequency of each area is once a month, so as to realize the regular and dynamic monitoring of the growth and health status of the entire pine forest.

[0083] Specifically, step S4, based on the first recognition network, performs preliminary identification on the acquired pine forest image and outputs suspected disease images, further including: S41: Based on the first recognition network, feature extraction is performed on the acquired pine forest image to obtain the growth potential feature vector {ΔH,Sshape,CYSR,DSR}.

[0084] S42: Query the growth potential feature table vector {ΔH} based on the feature threshold table. t Sshape t ,CYSR t DSR t}, and concatenate it with the growth potential eigenvector {ΔH, Sshape, CYSR, DSR} to obtain an 8-dimensional eigenvector {ΔH, Sshape, CYSR, DSR, ΔH}. t Sshape t ,CYSR t DSR t}

[0085] S43: Input the 8-dimensional feature vector into the first recognition network for convolution processing and output the first suspicion score.

[0086] S44: Determine whether the first suspicion score is greater than the first preset threshold. If so, output the corresponding suspected disease image.

[0087] In a specific embodiment of the present invention, the collected pine forest images to be detected are identified by a trained pine wood image recognition network. The recognition process is similar to the training process, specifically including: firstly, the collected pine forest images are input into a first recognition network for initial screening; secondly, the pine forest images are preprocessed and the growth potential feature vector {ΔH, Sshape, CYSR, DSR} is extracted; and thirdly, the growth potential feature vector {ΔH} is obtained by looking up the feature thresholds in Table 1. t Sshape t ,CYSR t DSR t Then, the extracted growth potential feature vector {ΔH, Sshape, CYSR, DSR} is compared with the growth potential feature table vector {ΔH} obtained by looking up the table. t Sshape t ,CYSR t DSR t The features are concatenated to obtain an 8-dimensional feature vector {ΔH,Sshape,CYSR,DSR,ΔHt,Sshapet,CYSRt,DSRt}. The concatenated 8-dimensional feature vector is then input into a CNN network (the first recognition network) for convolution processing, converting the 8-channel image into a 1-channel image. The convolved feature image is then input into a Sigmoid function, which outputs a suspicion score between 0 and 1 (i.e., the first suspicion score). Based on the output first suspicion score and a first preset threshold, it is determined whether the image is a suspected disease image. When the first suspicion score is greater than the first preset threshold, the image is determined to be a suspected disease image, and the image is output to enter the fine screening stage.

[0088] Specifically, step S5, based on the second recognition network, performs deep recognition on the suspected disease image and outputs the disease detection result, further including: S51: Based on the second recognition network, feature extraction is performed on suspected disease images to obtain detailed features, structural features and semantic features.

[0089] S52: Feature fusion of detail features, structural features and semantic features to obtain multi-scale fused features.

[0090] S53: Input the multi-scale fusion features into the second recognition network for calculation and output the second suspicion score.

[0091] S54: Determine whether the second suspicion score is greater than the second preset threshold. If so, output the disease detection result.

[0092] In a specific embodiment of the present invention, suspected disease images output from the initial screening stage are input into an improved MobileNet V3-small network (second recognition network) for accurate identification. First, the second recognition network extracts features from the suspected disease images, obtaining multi-scale features (detail features, structural features, and semantic features). Then, utilizing the attention mechanism in the improved MobileNet V3-small network, semantic features guide the fusion of structural features, and structural features guide the fusion of detail features, ultimately obtaining multi-scale fused features, achieving dual spatial and channel attention. The fused features are then input into the second recognition network for calculation, outputting a second suspicion score. The output second suspicion score is compared with a second preset threshold. If the second suspicion score is greater than the second preset threshold, the image is determined to be a pine nematode disease image, and the disease detection result is output.

[0093] Example 2 Based on the above method, embodiments of the present invention also provide an intelligent detection system for forest pine wilt disease, such as... Figures 9-13 As shown, it includes: Training set creation module 01 is used to obtain historical images of forest pine wood and construct a training set of pine wood image samples.

[0094] Image recognition network creation module 02 is used to construct and train a pine wood image recognition network based on pine wood image samples, including a first recognition network and a second recognition network.

[0095] Image acquisition module 03 is used to divide the pine forest into multiple areas and acquire images of the pine forest in different areas.

[0096] The first identification module 04 is used to perform preliminary identification on the acquired pine forest image based on the first identification network and output suspected disease images.

[0097] The second recognition module 05 is used to perform deep recognition on suspected disease images based on the second recognition network and output disease detection results.

[0098] Specifically, the training set creation module 01 also includes: The database creation module 011 is used to select historical images of forest pine wood from different growth stages to create a pine wood image database. The historical images of forest pine wood include images of healthy pine wood and images of pine wood nematode disease.

[0099] The training set generation module 012 is used to select multiple images of pine wilt disease and forest pine images taken by drones from the pine image database to construct a pine image sample training set.

[0100] In a specific embodiment of the present invention, a pine wood image database is first established using a database creation module 011, which selects historical images of forest pine wood from different growth cycles and health conditions. These historical images include images of healthy pine wood and images of pine wilt disease, each image carrying corresponding label information. Then, a training set generation module 012 selects multiple high-quality images of pine wilt disease from the established pine wood image database as positive samples. Simultaneously, healthy pine wood images taken by drones in a real forest environment are selected as negative samples, together forming the initial training sample set, providing reliable data support for subsequent training of the image recognition network.

[0101] Specifically, the image recognition network creation module 02 also includes: The first recognition network creation module 021 is used to construct the first recognition network using a CNN convolutional neural network and input the pine wood image sample training set into the first recognition network for training.

[0102] The second recognition network creation module 022 is used to construct a second recognition network using an improved MobileNet V3-small network, and to create small target augmentation samples and input them into the second recognition network for training.

[0103] The improved MobileNet V3-small network includes 11 inverted residual modules bneck, namely {bneck0, bneck1, ..., bneck10}. Among them, bneck0, bneck3, bneck4, bneck5, and bneck8 contain SE Blocks. Lightweight multi-head self-attention mechanism modules MHSA are inserted after bneck5 and bneck8 respectively.

[0104] In a specific embodiment of the present invention, a first recognition network is constructed using a first recognition network creation module 021 to learn and recognize growth potential features. The present invention uses a lightweight CNN convolutional neural network to train the first recognition network. The lightweight CNN convolutional neural network can be a 1×1 convolution followed by a sigmoid function, outputting a confidence score between 0 and 1 to obtain features for judging the growth potential of pine forests in the region. By preprocessing pine wood sample images from the sample set and inputting them into the lightweight CNN convolutional neural network for training, a growth potential feature vector {ΔH, Sshape, CYSR, DSR} is obtained. The obtained growth potential feature vector {ΔH, Sshape, CYSR, DSR} is then input into the CNN network for 1×1 convolution processing, converting a 4-channel image into a 1-channel image. Finally, the convolved image is fed into the sigmoid function, outputting a "suspicion" score between 0 and 1. Through the above training, the first recognition network is obtained.

[0105] In a specific embodiment of the present invention, a second recognition network is constructed using a second recognition network creation module 022 to perform multi-scale feature extraction and fusion on preliminarily screened images suspected of being infected with pine nematodes, further determining whether pine nematode disease has occurred and outputting the specific plants with the disease. The present invention uses an improved MobileNet V3-small network to construct the second recognition network. Specifically, based on the MobileNet V3-small network structure, a lightweight multi-head self-attention mechanism module (MHSA) is inserted after bneck-5 and bneck-8, enabling dynamically variable receptive fields and dual spatial-channel attention. Specifically, in bneck0, bneck3, bneck4, bneck5, and bneck8, each residual module includes a 1×1 up-dimensional convolution, a 3×3 depthwise separable convolution, an SE channel attention module, and a 1×1 down-dimensional convolution. Each SE channel attention module includes a global average pooling layer and a fully connected layer. The 1×1 upscaling convolution increases the dimensionality of the input channels, providing sufficient capacity for subsequent depthwise convolutions. The 3×3 depthwise separable convolution performs sliding window processing on multiple channels separately, followed by pointwise linear combination to extract local spatial features (e.g., detail features, structural features, and semantic features). The SE channel attention module first performs global average pooling on the local spatial features, then passes them through two fully connected layers. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Hard-sigmoid activation function. The resulting value is then multiplied by each element of the original local spatial feature matrix. The 1×1 downscaling convolution reduces the dimensionality to the initial dimension through pointwise convolution. Finally, two lightweight MHSA (Multi-Head Self-Attention) modules are used to fuse the extracted local features (detail features, structural features, and semantic features) to obtain multi-scale fused features.

[0106] Specifically, the image acquisition module 03 is used to divide the entire target pine forest into multiple rectangular inspection areas of similar size and regular shape based on the target pine forest's geographical information, area, and terrain complexity. The principle for dividing each inspection area is to ensure that a single drone can complete the image acquisition task for that area within its battery life, and to ensure a certain degree of overlap between adjacent areas (e.g., 5%) to avoid missed detections. After dividing the forest into multiple areas, the drone periodically acquires images of the target pine forest in these areas to obtain the images to be inspected.

[0107] Specifically, the first identification module 04 also includes: The first feature extraction module 041 is used to extract features from the acquired pine forest image based on the first recognition network to obtain a growth potential feature vector.

[0108] The feature concatenation module 042 is used to query the growth potential feature vector based on the feature threshold table and concatenate it with the growth potential feature vector to obtain an 8-dimensional feature vector.

[0109] The first calculation module 043 is used to input the 8-dimensional feature vector into the first recognition network for convolution processing and output the first suspicion score.

[0110] The first judgment module 044 is used to determine whether the first suspicion score is greater than the first preset threshold. If so, the corresponding suspected disease image is output.

[0111] In a specific embodiment of the present invention, the collected pine forest images are first input into the first feature extraction module 041 for initial screening. The pine forest images are preprocessed and the growth potential feature vector {ΔH, Sshape, CYSR, DSR} is extracted. Then, the growth potential feature table vector {ΔH} is obtained through the feature stitching module 042. t Sshape t ,CYSR t DSR t Then, the extracted growth potential feature vector {ΔH, Sshape, CYSR, DSR} is compared with the growth potential feature table vector {ΔH} obtained by looking up the table. t Sshape t ,CYSR t DSR t The features are concatenated to obtain an 8-dimensional feature vector {ΔH,Sshape,CYSR,DSR,ΔHt,Sshapet,CYSRt,DSRt}. The concatenated 8-dimensional feature vector is then input into the first calculation module 043 for convolution processing, converting the 8-channel image into a 1-channel image. The convolved feature image is then input into the Sigmoid function, which outputs a suspicion score between 0 and 1 (i.e., the first suspicion score). Finally, the first judgment module 044 determines whether the image is a suspected disease image. When the first suspicion score is greater than the first preset threshold, the image is determined to be a suspected disease image, and the image is output to enter the fine screening stage.

[0112] Specifically, the second identification module 05 also includes: The second feature extraction module 051 is used to extract features from suspected disease images based on the second recognition network to obtain detailed features, structural features and semantic features.

[0113] Feature fusion module 052 is used to fuse detailed features, structural features and semantic features to obtain multi-scale fused features.

[0114] The second calculation module 053 is used to input multi-scale fusion features into the second recognition network for calculation and output the second suspicion score.

[0115] The second judgment module 054 is used to determine whether the second suspicion score is greater than the second preset threshold. If so, the disease detection result is output.

[0116] In a specific embodiment of the present invention, the suspected disease images output from the initial screening stage are first input into the second feature extraction module 051 to extract features from the suspected disease images, obtaining multi-scale features (detail features, structural features, and semantic features). Then, using the attention mechanism in the feature fusion module 052, the structural features are fused to guide the fusion of semantic features, and the detailed features are fused to guide the fusion of structural features, ultimately obtaining multi-scale fused features, achieving dual attention of space and channel. The fused features are then input into the second calculation module 053 for calculation, outputting a second suspicion score. Finally, the second judgment module 054 determines whether it is a pine nematode disease image. If the second suspicion score is greater than a second preset threshold, it is determined to be a pine nematode disease image, and the disease detection result is output.

[0117] Example 3 like Figure 14 As shown, this embodiment of the invention also provides an electronic device, including a processor and a memory. The memory stores a computer program, which is loaded and executed by the processor to implement the above-described intelligent detection method for forest pine wilt disease. The device in this invention can be a server, PC, PAD, mobile phone, etc.

[0118] Furthermore, this embodiment of the invention also provides a computer-readable storage medium storing a computer program, which is loaded and executed by a processor to implement the above-described intelligent detection method for forest pine wilt disease.

[0119] In summary, the intelligent detection method and system for forest pine wilt disease provided by this invention has the following advantages: 1. Significantly improved the accuracy and reliability of disease identification.

[0120] This invention combines initial screening and fine screening. In the initial screening stage, a lightweight CNN network architecture is selected to filter out suspected disease images, significantly reducing the amount of data requiring detailed computation. In the fine screening stage, an improved MobileNet V3-small network is established by introducing a lightweight multi-head self-attention mechanism module after specific bottleneck layers (bneck-5 and bneck-8). This network uses an attention mechanism to guide the fusion of semantic features to structural features, and structural features to guide the fusion of detail features, achieving dynamic adjustment of the receptive field and dual spatial-channel attention. The fine screening stage performs deep discrimination on suspected disease images, avoiding missed detections or misjudgments caused by a single model's "one-size-fits-all" approach, thus improving the overall accuracy and reliability of identification. This network can more intelligently focus on key regions related to disease in the image (such as lesions of specific shapes or abnormally colored coniferous clusters) and effectively fuse multi-scale features (details, structure, semantics), significantly improving the ability to distinguish lesions of different morphologies against complex backgrounds. This overcomes the problem in existing technologies where simple feature vector fusion easily leads to the smoothing or misjudgment of small target features.

[0121] 2. It enhances the ability to detect early, sparse, small lesions, enabling early detection and early warning of diseases.

[0122] This invention employs a small-target enhancement strategy during the training phase (randomly pasting small lesion images onto healthy images), which specifically expands the quantity and quality of "sparse, small, and varied" early lesion samples in the training data. This forces the neural network to learn the characteristics of tiny lesions, greatly enhancing the model's sensitivity and robustness in identifying early diseases. By utilizing image recognition and spectral detection for mutual verification, it fully considers the basic characteristics of early symptoms of pine wilt disease, such as needle color changes, cessation of growth, and resin secretion, to identify pine wilt disease. This overcomes the shortcomings of traditional methods and some image recognition methods that are insensitive to early symptoms, gaining valuable time for timely control measures.

[0123] 3. While ensuring high precision, it also takes into account detection efficiency, making it suitable for widespread application.

[0124] In the fine screening stage, the MobileNetV3-small network selected in this invention is itself a lightweight network with low computational burden. Building upon this, this invention selectively inserts attention modules at higher levels (bneck-5 and bneck-8) rather than at lower high-resolution layers, cleverly avoiding an explosive increase in computational load while introducing a powerful attention mechanism. Combined with standardized drone photography specifications (uniform location, angle, altitude, and time), the consistency and optimality of input image quality are ensured, reducing model performance fluctuations and additional preprocessing overhead caused by differences in image quality, and significantly improving detection efficiency.

[0125] 4. It has achieved automated and standardized intelligent monitoring, reducing reliance on manual labor and costs.

[0126] This invention automates the entire process from image acquisition (automated drone inspection) and preprocessing to intelligent recognition, completely changing the past model that relied on manual on-site surveys—a model that was inefficient, subjective, and costly. Strict periodic inspections and standardized shooting protocols ensure strong temporal comparability of monitoring data. By analyzing the growth characteristics of images from different periods, dynamic tracking and trend analysis of the health status of individual trees or specific forest areas can be achieved, providing a more scientific and comprehensive basis for forestry management decisions.

[0127] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A smart detection method for forest pine wilt disease, characterized in that, include: Obtain historical images of pine wood in forests to construct a training set of pine wood image samples; A pine wood image recognition network was constructed and trained based on pine wood image samples, including a first recognition network and a second recognition network; The pine forest was divided into multiple regions, and images of the pine forest in different regions were obtained; The first recognition network performs preliminary identification on the acquired pine forest images and outputs images of suspected diseases; The suspected disease images are subjected to deep recognition based on the second recognition network, and the disease detection results are output.

2. The intelligent detection method for forest pine wilt disease according to claim 1, characterized in that, The step of acquiring historical images of forest pine wood and constructing a training set of pine wood image samples further includes: A pine wood image database was created by selecting historical images of forest pine wood at different growth stages. The historical images of forest pine wood include images of healthy pine wood and images of pine wood nematode disease. A training set of pine wood image samples was constructed by selecting multiple images of pine wilt disease and forest pine wood images taken by drones from the pine wood image database.

3. The intelligent detection method for forest pine wilt disease according to claim 1, characterized in that, The construction and training of the pine wood image recognition network based on pine wood image samples further includes: A first recognition network is constructed using a CNN convolutional neural network, and the training set of pine wood image samples is input into the first recognition network for training. A second recognition network was constructed using an improved MobileNet V3-small network, and small target augmentation samples were generated and input into the second recognition network for training. The improved MobileNet V3-small network includes 11 inverted residual modules bneck, namely {bneck0, bneck1, ..., bneck10}. Among them, bneck0, bneck3, bneck4, bneck5, and bneck8 contain SE Blocks. Lightweight multi-head self-attention mechanism modules MHSA are inserted after bneck5 and bneck8 respectively.

4. The intelligent detection method for forest pine wilt disease according to claim 3, characterized in that, The step of constructing a first recognition network using a CNN convolutional neural network and inputting a training set of pine wood image samples into the first recognition network for training further includes: The first recognition network is constructed using a CNN convolutional neural network, and the loss function is determined. L MSE for: In the formula, For the true value, The value is the model prediction, and N is the number of samples. The pine wood sample image is input into the first recognition network for feature extraction, resulting in a growth potential feature vector {ΔH, Sshape, CYSR, DSR}, where... Sshape = Area(C) / Area(A) CYSR= / DSR= / Where ΔH is the canopy color, Sshape is the canopy integrity, CYSR is the proportion of new shoots, DSR is the proportion of dead shoots, Area() is the number of pixels, C is the original canopy, and A is the convex polygon formed by the points of the outer contour of the original canopy C, i.e., the convex hull. For yellowing pixels, Total pixels of the canopy For new shoot pixels, , Reddish-brown pixels; Input the growth potential feature vector {ΔH, Sshape, CYSR, DSR} into the loss function. L MSE The system is trained to output predicted values ​​for each feature in different tree age ranges.

5. The intelligent detection method for forest pine wilt disease according to claim 3, characterized in that, The construction of the second recognition network using the improved MobileNet V3-small network, and the creation of small target augmentation samples and their input into the second recognition network for training, further include: A second recognition network is constructed using an improved MobileNet V3-small network, and the loss function is determined. L total for: L total = λ cls L cls + λ box L box + λ attn L attn In the formula, L cls For classification loss function, L box For box regression loss function, L attn To compare the distillation loss function, λ cls , λ box , λ attn These are the weighting coefficients for each loss function; Disease lesion images of 64×64~128×128 pixels were randomly cropped from the pine wood image sample training set and pasted onto healthy pine wood images with a transparency of 0.3~0.5 to obtain small target augmented samples; Small target augmented samples are input into a second recognition network for feature extraction and feature fusion to obtain multi-scale fused features; Multi-scale fused features are input into the loss function. L total The training process is performed, and the predicted values ​​of the multi-scale fused features are output.

6. The intelligent detection method for forest pine wilt disease according to claim 5, characterized in that, The step of inputting the small target enhanced sample into the second recognition network for feature extraction and feature fusion to obtain multi-scale fused features further includes: The small target enhancement sample is input into the second recognition network for feature extraction to obtain detail features, structural features and semantic features. The detail features include detail edge information, the structural features include structural and color block information, and the semantic features include semantic information. Using semantic features as the query vector Q and structural features as the key vector K and value vector V, the semantic features and structural features are weighted and fused to obtain semantically enhanced structural features. The semantically enhanced structural features are then concatenated with the original structural features to obtain the mid-level fused features. Using the mid-level fusion feature as the query vector Q and the detail feature as the key vector K and value vector V, the structural feature and detail feature are weighted and fused to obtain the structural feature-enhanced detail feature. The structural feature-enhanced detail feature is then upsampled to the original resolution to obtain the multi-scale fusion feature.

7. The intelligent detection method for forest pine wilt disease according to claim 1, characterized in that, The preliminary identification of the acquired pine forest images based on the first identification network, and the output of suspected disease images, further includes: Based on the first recognition network, feature extraction is performed on the acquired pine forest image to obtain the growth potential feature vector {ΔH,Sshape,CYSR,DSR}. Based on the feature threshold table, query the growth potential feature table vector {ΔH} t Sshape t ,CYSR t DSR t }, and concatenate it with the growth potential eigenvector {ΔH, Sshape, CYSR, DSR} to obtain an 8-dimensional eigenvector {ΔH, Sshape, CYSR, DSR, ΔH}. t Sshape t ,CYSR t DSR t }; The 8-dimensional feature vector is input into the first recognition network for convolution processing, and the first suspicion score is output. Determine whether the first suspicion score is greater than the first preset threshold. If so, output the corresponding suspected disease image.

8. The intelligent detection method for forest pine wilt disease according to claim 1, characterized in that, The step of performing deep recognition on the suspected disease image based on the second recognition network and outputting the disease detection result further includes: Based on the second recognition network, feature extraction is performed on the suspected disease image to obtain detailed features, structural features and semantic features; By fusing detailed features, structural features, and semantic features, multi-scale fused features are obtained. The multi-scale fusion features are input into the second recognition network for calculation, and the second suspicion score is output. Determine whether the second suspicion score is greater than the second preset threshold. If so, output the disease detection result.

9. A smart detection system for forest pine wilt disease, characterized in that, include: The training set creation module is used to acquire historical images of pine timber in forests and construct a training set of pine timber image samples. An image recognition network creation module is used to construct and train a pine wood image recognition network based on pine wood image samples, including a first recognition network and a second recognition network; The image acquisition module is used to divide the pine forest into multiple areas and acquire images of the pine forest in different areas; The first identification module is used to perform preliminary identification on the acquired pine forest images based on the first identification network and output suspected disease images; The second identification module is used to perform deep identification on the suspected disease image based on the second identification network and output the disease detection result.

10. The intelligent detection system for forest pine wilt disease according to claim 9, characterized in that, The training set creation module also includes: The database creation module is used to select historical images of forest pine wood at different growth stages to create a pine wood image database. The historical images of forest pine wood include images of healthy pine wood and images of pine wood nematode disease. The training set generation module is used to select multiple images of pine wilt disease and forest pine images taken by drones from the pine image database to construct a pine image sample training set.

11. The intelligent detection system for forest pine wilt disease according to claim 9, characterized in that, The image recognition network creation module also includes: The first recognition network creation module is used to construct the first recognition network using a CNN convolutional neural network and input the pine wood image sample training set into the first recognition network for training. The second recognition network creation module is used to construct a second recognition network using an improved MobileNet V3-small network, and to create small target augmentation samples and input them into the second recognition network for training. The improved MobileNet V3-small network includes 11 inverted residual modules bneck, namely {bneck0, bneck1, ..., bneck10}. Among them, bneck0, bneck3, bneck4, bneck5, and bneck8 contain SE Blocks. Lightweight multi-head self-attention mechanism modules MHSA are inserted after bneck5 and bneck8 respectively.

12. The intelligent detection system for forest pine wilt disease according to claim 9, characterized in that, The first identification module further includes: The first feature extraction module is used to extract features from the acquired pine forest image based on the first recognition network to obtain a growth potential feature vector; The feature concatenation module is used to query the growth potential feature vector based on the feature threshold table and concatenate it with the growth potential feature vector to obtain an 8-dimensional feature vector. The first calculation module is used to input the 8-dimensional feature vector into the first recognition network for convolution processing and output the first suspicion score. The first judgment module is used to determine whether the first suspicion score is greater than the first preset threshold. If so, the corresponding suspected disease image is output.

13. The intelligent detection system for forest pine wilt disease according to claim 9, characterized in that, The second identification module also includes: The second feature extraction module is used to extract features from the suspected disease image based on the second recognition network to obtain detailed features, structural features and semantic features; The feature fusion module is used to fuse detailed features, structural features, and semantic features to obtain multi-scale fused features. The second calculation module is used to input multi-scale fused features into the second recognition network for calculation and output a second suspicion score. The second judgment module is used to determine whether the second suspicion score is greater than the second preset threshold. If so, the disease detection result is output.

14. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the intelligent detection method for forest pine wilt disease as described in any one of claims 1 to 8.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the intelligent detection method for forest pine wilt disease as described in any one of claims 1 to 8.

Images