A method for detecting pine wood nematode disease of a drone based on multi-modal features

Abstract

The present application relates to a kind of unmanned aerial vehicle pine wood nematode disease detection method based on multi-modal feature, by constructing image-text multi-modal dataset to provide prior semantic knowledge;Adopt vegetation index threshold mask filtering (VTMF) to enhance spectral difference and inhibit noise;Utilize the W-KAN module of fusion wavelet transform and KAN attention to carry out frequency domain denoising and feature reservation;Through four-domain diffusion graph convolution residual attention module (Q-DGCRAB) modeling long-distance spatial dependence relationship;Introduce epidemic semantic enhancement detection module (ESADM) to realize text-visual feature fusion;And adopt forest guardian collaborative optimization algorithm (FGA-CE) dynamic optimization training process;The present application solves the problem of low pine wood nematode disease detection precision caused by complex background, texture repetition, significant noise and insufficient semantic information in high-altitude unmanned aerial vehicle image, significantly improves the detection precision and robustness of pine wood nematode disease in complex stand environment, realizes early accurate, real-time monitoring of pine wood nematode disease.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forest pest monitoring, and particularly relates to a pine wood nematode disease detection method based on multi-modal features. BACKGROUND

[0002] Pine wilt disease (PWD) is a typical pathogen-dominated disease with short disease course and fast spread. Pine wilt disease causes rapid wilting and death of pine trees by blocking the water transport system of pine trees. The direct economic loss and biodiversity threat caused by pine wilt disease have become a major challenge to the sustainable development of forestry. Therefore, promoting the automation and intelligence of remote sensing monitoring of pine trees is an urgent need to be solved.

[0003] In the early stage, the detection of pine wilt disease mainly relies on the subjective experience of planters for positioning. This method not only consumes time and effort, but also easily leads to incorrect judgment for inexperienced personnel, significantly reducing the disease recognition efficiency and accuracy. In the field of intelligent forestry, forestry remote sensing technology helps forestry workers improve the analysis ability of pine tree diseases. This technology can accurately judge the damage degree caused by pine wilt disease, so as to specify the precise prevention and control scheme, and bring revolutionary changes to disease detection. However, researchers find that the method relying solely on image data is difficult to fully grasp the development law of pine wilt disease, and also lacks key detailed information to judge the severity of the disease, especially in the complex environment of mountain forests. The reliability of the traditional detection method will be greatly reduced.

[0004] With the development of computer vision technology and deep learning, object detection has become the core technology of pine wilt disease image analysis. Detection networks based on deep learning are generally divided into two categories: (1) two-stage detection models, represented by region-based two-stage convolutional neural network (R-CNN); (2) single-stage detection models, represented by single shot multibox detector (SSD) and you only look once (YOLO). The core difference between the two models lies in the detection mechanism: the two-stage detection model needs to generate candidate boxes that may contain diseased trees first, and then perform classification and positioning operations on the candidate boxes; the single-stage detection model can directly use the network feature map to predict the position and category information of the diseased area simultaneously. Due to the lower computational complexity of single-stage models, convolutional neural network (CNN)-based methods effectively improve the detection accuracy and efficiency of disease areas by analyzing image multi-scale features and feature fusion mechanisms, and occupy a dominant position in practical applications. Among them, the YOLO series of models enhance the ability to capture image semantic information through multi-level feature extraction mechanisms, and their network structure design significantly reduces computational complexity, thus meeting real-time detection requirements. YOLOv5 introduces multi-size models, automatic anchor point calculation, and mixed precision training, enhancing small target recognition and complex scene adaptability. The latest YOLOv8 abandons the traditional method of relying on predefined anchor boxes and uses an anchor-free detection mechanism to directly predict target center points and bounding boxes, simplifying the model structure, reducing computational complexity, and significantly improving small target detection accuracy. At the same time, the Transformer architecture has also begun to make a name for itself in the field of object detection. DETR (DEtection TRansformer) first introduced this architecture into object detection, using an encoder-decoder structure to directly predict targets, eliminating the region proposal and non-maximum suppression (NMS) steps, and simplifying the process. Subsequently, RT-DETR achieved real-time object detection with DETR through an efficient hybrid encoder and uncertainty-minimizing query selection mechanism. The continuous innovation of the YOLO series not only consolidates its own technological leadership but also drives the development of the entire single-stage detector field, making it a mainstream choice in scenarios such as autonomous driving and remote sensing that require fast response.

[0005] Despite the significant progress made by the single-object detection model YOLO in object detection, there are still deficiencies in four key dimensions:

[0006] In unmanned aerial vehicle pine wilt disease detection, the complex background formed by various vegetation types, terrain undulations, shadow areas, and seasonal changes in the forest environment seriously interferes with the accurate identification of diseased trees.

[0007] Although the forest remote sensing image obtained by the high-altitude unmanned aerial vehicle contains multi-scale texture features such as tree crown edge, shadow texture and terrain change, the conventional convolutional neural network produces high-frequency detail loss in the down-sampling process due to the repetitive texture pattern and random noise interference, so that the model faces multiple technical challenges such as texture feature preservation, target boundary accurate interpretation and noise robustness improvement;

[0008] The unmanned aerial vehicle aerial image usually covers several square kilometers to tens of square kilometers of geographical range, and the spatial layout of forest vegetation and the continuous distribution of vegetation patches in the ecological system present complex spatial correlation. However, due to the limitation of the receptive field, the traditional convolutional neural network cannot effectively capture the long-distance dependence in the image when extracting target features, especially for the relationship between the target and the surrounding environment in a large scene or complex background.

[0009] The unmanned aerial vehicle aerial image presents high instability in the target detection model training process due to the significant multi-scale feature distribution, dynamic background interference and spatial heterogeneity of target-scene, which is manifested as slow convergence speed, frequent performance index oscillation, and easy to fall into local suboptimal solution in the scene with large target size dynamic range and high background texture complexity. SUMMARY

[0010] Therefore, the present application provides a kind of unmanned aerial vehicle pine wood nematode disease detection method based on multi-modal feature, can solve the problem of low detection precision, large background interference, low accuracy, poor generalization ability and unstable training when traditional unmanned aerial vehicle remote sensing detects pine wood nematode disease in complex forest environment, realize early accurate, real-time monitoring of pine wood nematode disease.

[0011] To achieve the above purpose, the present application provides a kind of unmanned aerial vehicle pine wood nematode disease detection method based on multi-modal feature, comprising the following steps:

[0012] S1, obtain the visible light image of pine forest through the unmanned aerial vehicle high-altitude remote sensing platform, screen out the pine wood nematode disease pine image after cutting, and process the image to establish the multi-modal data set of image-text pairing, the multi-modal data set is randomly divided into training set, verification set and test set according to the ratio of 8:1:1;

[0013] S2, the visible light image is preprocessed by adopting vegetation index threshold mask filter VTMF, and the spectral feature of pine health state is obtained;

[0014] S3, the image enhanced by VTMF is input into the text-remote sensing collaborative network model TRSSL-Net based on YOLO architecture,

[0015] S301, constructing a W-KAN feature extraction module at a P3 layer of the text-remote sensing collaborative network model TRSSL-Net, performing frequency domain feature decomposition, nonlinear modulation and noise suppression on the image enhanced by the VTMF through discrete wavelet transform DWT, and extracting a frequency domain feature map of the pine wood nematode disease;

[0016] S302, establishing a regional graph structure relationship by using a Q-DGCRAB four-domain diffusion graph convolution residual attention module, dividing the frequency domain feature map of the pine wood nematode disease into four domains of upper left, upper right, lower left and lower right, and constructing a graph structure by taking each region as a node;

[0017] S303, introducing a pre-trained BERT model text encoder by using an epidemic semantic enhancement detection module ESADM, realizing visual-text cross-modal feature fusion, and providing enhanced information at a semantic level;

[0018] S4, adopting a forest guardian collaborative optimization strategy FGA-CE, dynamically optimizing the learning rate in the training process through a group cooperation mechanism, and accelerating the convergence of the text-remote sensing collaborative network model TRSSL-Net;

[0019] S5, outputting the positioning and classification results of the pine wood nematode disease through the detection head.

[0020] Preferably, the image processing comprises the following steps:

[0021] S101, adjusting the pine wood nematode disease pine image in jpg format to 800x800 pixels and inputting it into the labelmg software;

[0022] S102, an expert manually marks the pine wood nematode disease pine image on a visualization platform, performs multi-dimensional parameter text description on each pine wood nematode disease pine image, the multi-dimensional parameters include distribution characteristics, morphological characteristics and position information of the pine wood nematode disease, makes a label of irregular disease and saves it in txt format, and establishes a multi-modal data set of image-text pairs.

[0023] Preferably, the vegetation index threshold mask filter VTMF enhances the spectral feature difference between vegetation and non-vegetation areas by a green leaf index GLI, and the expression is:

[0024] ;

[0025] wherein, 、 、 Rg, Rh and Rb respectively represent the reflectivity of green, red and blue bands;

[0026] The green-red vegetation index GRVI, the visible light atmospheric impedance index VARI and the RGB vegetation index capture the spectral characteristics of the vegetation health state, and form a multi-dimensional feature space.

[0027] Preferably, the W-KAN feature extraction module adopts discrete wavelet transform (DWT) to decompose the image features enhanced by VTMF into low-frequency subbands LL and high-frequency subbands LH, HL and HH through haar wavelet basis functions, realizing frequency domain separation of signals and noises.

[0028] The image feature transformation is performed using a two-dimensional haar wavelet basis function, and the expression is:

[0029] ;

[0030] wherein, represents a mother wavelet function, represents a scale parameter, represents a translation parameter, represents an independent variable of a wavelet function, used to represent the time, spatial position of a signal or image;

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] wherein, , represent low-pass and high-pass filter coefficients, respectively, represent row and column index positions of an output image, respectively, represent row and column offset indexes of a filter when convolved on an input LL image;

[0036] The KAN attention mechanism is used to perform nonlinear modulation on the four subband coefficients to extract specific feature maps of the pine wood nematode disease in the frequency domain, and the expression is:

[0037] ;

[0038] wherein, represents a feature map of an input image in a specific frequency band band.

[0039] Preferably, the Q-DGCRAB four-domain diffusion graph convolution residual attention module establishes a regional graph structure relationship, which specifically includes the following steps:

[0040] The input frequency domain feature map of the pine wood nematode disease is proceeding group convolution dimension reduction wherein, denotes a real space, with four dimensions: denotes the number of samples of the input network, denotes the number of channels of the corresponding feature map, denotes the spatial height of the feature map, denotes the spatial width of the feature map;

[0041] extract the representations of the upper left, upper right, lower left and lower right four domains to construct the inter-regional connection relationship, and the regional representation set The expression is:

[0042] ;

[0043] wherein, , , , denote the regional representations obtained by average pooling of the upper left, upper right, lower left and lower right four regions respectively;

[0044] A diffusion factor based on spatial distance perception is introduced to transmit messages between regions, so that the interaction intensity between regions with farther distance is adaptively lower, and the expression of the diffusion factor is:

[0045] ;

[0046] wherein, , denote the region index respectively;

[0047] The message function of graph convolution is:

[0048] ;

[0049] wherein, denotes the attention weight, denotes the normalization factor, denotes the feature of the region;

[0050] A diffusion threshold mechanism is used for feature updating, when the aggregation operation is executed, otherwise the original value is kept, and the expression of the aggregation operation is:

[0051] ;

[0052] wherein, denotes the updated feature representation of the th neighbor region, denotes the neighbor set of node , denotes a pooling function;

[0053] application of the gated residual connection, the expression is:

[0054] ;

[0055] ;

[0056] wherein, denotes adaptive control information flow, denotes the weight matrix of the gating network, denotes the sigmoid activation function, denotes the original feature of the th region before feature update, denotes the final enhanced feature representation after the original feature and the neighborhood updated feature are fused through the gated residual mechanism;

[0057] fixed weight and channel attention mechanism are used for feature fusion, and the enhanced pine wilt disease feature map is output, and the expression is:

[0058] ;

[0059] wherein, denotes the batch normalization operation, denotes convolution operation, denotes the channel attention mechanism.

[0060] Preferably, the epidemic semantic enhancement detection module ESADM loads a pre-trained BERT model through a freezing parameter strategy, extracts deep semantic representations of the text, and performs dimension reduction and nonlinear transformation through a two-layer fully connected dimension reduction network, compresses high-dimensional text features output by the BERT model to a dimension space compatible with visual features, dynamically adjusts the contribution ratio of text and visual features through learnable parameters k, and performs text-visual fusion at P4 and P6 layers of the text-remote sensing collaborative network model TRSSL-Net, and the expression is:

[0061] ;

[0062] wherein, denotes an adaptive fusion weight, ensuring that text semantic information can enhance visual feature expression at different abstraction levels, denotes text-visual fusion, denotes visual features, denotes text features;

[0063] The epidemic semantic enhancement detection module ESADM further comprises a WorldDetect detection head, which realizes deep alignment of visual features and text semantics through a contrastive learning mechanism. The WorldDetect detection head calculates the semantic matching degree between text-visual features through a ContrastiveHead module using cosine similarity.

[0064] Preferably, the forest guardian cooperative optimization strategy FGA-CE optimizes the learning rate by simulating three behaviors of the unmanned aerial vehicle in forest monitoring: regional hovering behavior, high-altitude overlooking behavior, and remote surveying behavior.

[0065] The regional hovering behavior develops locally around the current optimal solution, and the position update expression is:

[0066] ;

[0067] ;

[0068] wherein, represents the optimal position found by the current unmanned aerial vehicle, represents the current search agent position of the unmanned aerial vehicle, , represents a coefficient vector, represents the element-by-element multiplication of two vectors;

[0069] The high-altitude overlooking strategy simulates the vertical lifting observation of the unmanned aerial vehicle, and the position update expression is:

[0070] ;

[0071] ;

[0072] wherein, represents the distance between the unmanned aerial vehicle and the best observation point, represents a spiral descent constant, represents a random number, represents the modulus of a vector, used to represent the Euclidean distance between the current position of the unmanned aerial vehicle and the optimal position;

[0073] The remote surveying strategy is triggered when , forcing the unmanned aerial vehicle to randomly select an unmonitored area for global exploration, by randomly selecting a reference point , the expression is:

[0074] ;

[0075] The position update expression is:

[0076] .

[0077] Preferably, the forest guardian cooperative optimization strategy FGA-CE introduces a multi-population cooperative evolution framework, divides the total unmanned aerial vehicle group N into multiple sub-group teams, and each sub-group team independently patrols;

[0078] An information sharing mechanism is established, and information exchange and elite migration are carried out regularly: information exchange is carried out between different sub-group teams every fixed iteration time, and high-quality detection information is transmitted from the optimal group team to other group teams according to the detection quality sorting; the dynamic path optimization strategy selects the worst detection point in the receiving group team for replacement, and adds a small disturbance to the transmission detection point, and the amplitude is The search range;

[0079] After each iteration, boundary checking is performed on all unmanned aerial vehicles to ensure that the updated position is within the monitoring space, and the fitness of all detection points is re-evaluated, and the optimal solution of the sub-group team and the global optimal solution are updated;

[0080] The target function is constructed through the historical verification index mAP50, and the learning rate is predicted and adjusted through momentum.

[0081] Compared with the prior art, the beneficial effects of the present application are:

[0082] The TRSSL-Net provided by the present application provides supplementary semantic knowledge by integrating text information input, and significantly enhances the disease tree feature recognition capability; meanwhile, a data enhancement framework integrating multiple vegetation index algorithms is constructed to capture the spectral features of the pine tree health status from different angles. The core of the model is the innovative W-KAN feature extraction module, which combines multi-scale decomposition of discrete wavelet transform (DWT) and KAN nonlinear activation mechanism to realize excellent noise suppression capability and complex nonlinear relationship capture; supplemented by the Q-DGCRAB four-domain diffusion graph convolution residual attention module to enhance the information transmission and integration between regions, effectively solving the problem of scattered and large scale change of disease tree targets under high-altitude perspective; in order to improve the training efficiency, a forest guardian cooperative optimization algorithm (FGA-CE) is designed, which accelerates the model convergence through group cooperation mechanism, overcomes the shortcomings of traditional convolutional neural network in frequency domain screening mechanism, and can effectively distinguish the weak signal of pine wilt disease symptoms in more than five hundred meters of aerial image from complex environmental noise such as light change, seasonal background difference, device noise and atmospheric scattering. The method provided by the present application significantly improves the detection accuracy of pine wilt disease, enhances the robustness of the model under complex background and multi-scale targets, and supports real-time detection, which is suitable for large-scale forestry monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0083] Fig. 1 is a flow framework diagram of the unmanned aerial vehicle pine wilt disease detection method based on multi-modal features of the present application;

[0084] Figure 2The image and text are combined to form a multi-modal instance graph of the present application;

[0085] Figure 3 In the present application, (a) is an image data processing flowchart, (b) is a structural diagram of ESADM, VTMF and Q-DGCRAB, (c) is a structural diagram of MRA-KAN, (d) is a structural diagram of FGA-CE, and (a)-(d) are combined into a structural diagram of a text-remote sensing collaborative network model TRSSL-Net of the present application;

[0086] Figure 4 The contrastive network visualization result contrastive graph of the present embodiment is shown in the following table:

[0087] Figure 5 The generalization dataset visualization result contrastive graph of the present embodiment is shown in the following table DETAILED DESCRIPTION

[0088] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purposes, the specific embodiments, structures, features and effects according to the present application are described in detail as follows in combination with the drawings and preferred embodiments.

[0089] The present embodiment provides a multi-modal feature-based unmanned aerial vehicle pine wood nematode disease detection method, which comprises the following steps:

[0090] S1, obtaining a visible light image of a pine forest through an unmanned aerial vehicle high-altitude remote sensing platform, screening out pine wood nematode disease pine tree images after cutting, and performing image processing to establish an image-text paired multi-modal dataset, and randomly dividing the multi-modal dataset into a training set, a validation set and a test set according to a ratio of 8:1:1;

[0091] The pine wood nematode disease image dataset used in the present embodiment is shot at a flight height of more than 500 meters using a vertical-horizontal CW-15, with a ground sampling distance (GSD) of 10 cm, a full-frame 61 million pixel sensor, and a 35mm focal length lens. This dataset covers 1174.5 square kilometers of forest images in the wild. Due to the large coverage of high-altitude unmanned aerial vehicle images, the high-altitude unmanned aerial vehicle images are cut and 785 image datasets containing multiple diseased pine tree pictures are screened out.

[0092] The image processing comprises the following steps:

[0093] S101, adjusting the pine wood nematode disease pine tree images in jpg format to 800x800 pixels and inputting them into the labelmg software, which is to prevent overfitting and speed up the training,

[0094] S102, the expert uses professional knowledge to manually mark the pine image of the pine wilt disease in the data set on the visualization platform, and performs multi-dimensional parameter text description on each pine image of the pine wilt disease, the multi-dimensional parameters include the distribution characteristics, morphological characteristics and position information of the pine wilt disease, a label of irregular disease is made and saved in txt format, and a multi-modal data set of image-text pairing is established, as shown in Figure 1 Each image is finely marked and text annotated, the system records the spatial distribution pattern and morphological characteristics of the diseased trees, and these text descriptions rich in semantic information provide additional learning references for the neural network, effectively promoting the accurate capture of the significant characteristics of the diseased trees by the model, and significantly improving the accuracy and reliability of disease recognition.

[0095] S2, the visible light image is preprocessed by using vegetation index threshold mask filtering VTMF to obtain the spectral characteristics of the pine health state, and a pine wilt disease pine candidate region is obtained;

[0096] The vegetation index threshold mask filtering VTMF does not need to modify the network structure, but embeds the vegetation index information as a multi-dimensional feature into the training data in the data preprocessing stage, so that the network can adaptively learn and use these biological spectral characteristics; in the unmanned aerial vehicle pine wilt disease target detection application, the green leaf index (GLI, Green Leaf Index) provides key technical support for solving the problem of complex background interference through its unique spectral response characteristics, therefore, the vegetation index threshold mask filtering VTMF enhances the spectral feature difference between the vegetation and non-vegetation regions through the green leaf index GLI, and the expression is:

[0097] ;

[0098] Among them, 、 、 respectively represent the reflectivity of green, red and blue bands, the GLI value of a healthy pine is significantly higher than that of surrounding vegetation, and the GLI value of a pine infected with pine wilt disease will sharply decrease to the range of 0.2-0.4 due to needle yellowing and wilting phenomenon, and this spectral response difference forms an obvious distinguishing feature on the RGB spectral curve, which is helpful for the model to quickly locate the potential disease tree area;

[0099] Other complementary vegetation indices are integrated, including green red vegetation index GRVI, visible light atmospheric impedance index VARI and RGB vegetation index, to capture the spectral characteristics of the vegetation health state and form a multi-dimensional feature space;

[0100] The expression of the green red vegetation index GRVI is:

[0101]

[0102] The expression of the visible light atmospheric impedance index VARI is:

[0103]

[0104] The expression of the RGB vegetation index is:

[0105]

[0106] These indices capture the spectral characteristics of vegetation health from different angles, forming a multi-dimensional feature space that significantly improves the sensitivity of the model to the early symptoms of pine wood nematode disease. The annotated data set covers a variety of environmental factors, including yellow sand, roads, marshes, grasslands, buildings, valleys, and rice fields. The diversity of these environmental conditions not only ensures the robustness and real-world applicability of the trained model, but also adds a lot of background noise interference. The low response characteristics of non-green objects of the vegetation index, especially the GLI index, can effectively help the model filter non-vegetation targets and achieve precise positioning of diseased trees in complex backgrounds, thereby reducing the false alarm rate.

[0107] S3, as shown in Figure 2 , the image enhanced by VTMF is input into the text-remote sensing collaborative network model TRSSL-Net based on the YOLO architecture;

[0108] S301, for the large amount of random noise appearing in the high-altitude unmanned aerial vehicle image, a multi-level denoising strategy based on wavelet transform is proposed, and a W-KAN feature extraction module (original resolution ) is constructed at the P3 layer of the text-remote sensing collaborative network model TRSSL-Net based on the analysis of signal-to-noise ratio. The P3 layer retains sufficient details while the noise propagation has not yet accumulated to an uncontrollable level.

[0109] The image enhanced by VTMF is decomposed in the frequency domain, nonlinearly modulated, and noise suppressed through discrete wavelet transform DWT to extract the frequency domain feature map of pine wood nematode disease;

[0110] This is because the characteristic signals of the early symptoms of pine wood nematode disease are mainly concentrated in a specific frequency band, while environmental noise shows a more dispersed frequency spectrum distribution, so the denoising advantage of wavelet transform is manifested. Discrete wavelet transform decomposes the signal into different frequency subbands to realize multi-resolution analysis. In the W-KAN module, the discrete wavelet transform DWT decomposes the image features enhanced by VTMF into low-frequency subbands LL and high-frequency subbands LH, HL, and HH through the haar wavelet basis function, realizing frequency domain separation of signal and noise.

[0111] ​​​The two-dimensional Haar wavelet basis function is used for image feature transformation, and the expression is:

[0112] ;

[0113] wherein, represents the mother wavelet function, represents the scale parameter, represents the translation parameter, represents the independent variable of the wavelet function, which is used to represent the time, spatial position of the signal or image;

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] wherein, , represent low-pass and high-pass filter coefficients respectively, represent the row and column index positions of the output image respectively, represent the row and column offset indexes of the filter when convolved on the input LL image, in this embodiment:

[0119] ;

[0120] ;

[0121] Simple threshold processing cannot meet the adaptive denoising requirements. Therefore, in this embodiment, the KAN attention mechanism is introduced to perform nonlinear modulation on the wavelet coefficients; the complex mapping relationship constructed based on the B-spline function enables the model to learn the optimal filtering strategy under different noise patterns, and the B-spline basis function is defined in a recursive form:

[0122] ;

[0123] wherein, represents the spline order, in this embodiment , represents the node vector, and the initial condition is:

[0124] ;

[0125] The expression of the mapping function used by KAN is:

[0126] ;

[0127] wherein, denotes a learnable coefficient;

[0128] The feature mapping is composed of a basic linear transformation and a spline transformation, and the expression is:

[0129]

[0130] wherein, denotes a basic activation function (SiLU), is a B-spline basis function vector, and are learnable weights, and the KAN attention mechanism is used to nonlinearly modulate four sub-band coefficients to extract specific feature maps of the pine wood nematode disease in the frequency domain, and the expression is:

[0131]

[0132] wherein, denotes a feature map of an input image in a specific frequency band band.

[0133] The W-KAN module realizes adaptive weighting of different frequency band information, and compared with traditional denoising methods, this data-driven adaptive denoising mechanism shows stronger environmental adaptability; in theory, the KAN mechanism can retain the high-frequency components related to disease characteristics while effectively suppressing background noise, and an excellent balance between maintaining feature integrity and noise suppression is achieved, the pine wood nematode disease symptoms have specific representation in the frequency domain, and the W-KAN module effectively extracts these key features through frequency domain analysis and nonlinear feature enhancement;

[0134] S302, a Q-DGCRAB four-domain diffusion graph convolution residual attention module is used to establish a regional graph structure relationship, and the frequency domain feature map of the pine wood nematode disease is divided into four domains of upper left, upper right, lower left and lower right, and each region is taken as a node to construct a graph structure;

[0135] ​​The feature expression in the target detection task faces the challenge of insufficient capture of spatial correlation. The traditional convolution operation is limited to a local receptive field and is difficult to establish a dependency relationship between long-distance features. Although the attention mechanism partially solves this problem, its secondary computational complexity is inefficient on high-resolution feature maps, especially in the dense small target scene. Local features cannot provide enough context information, leading to target missed detection and false detection. To solve this problem, the embodiment of the application proposes a Quad-Region Diffusion Graph Convolutional Residual Attention Block (Q-DGCRAB). The Q-DGCRAB is deployed at the front end of the detection head, which realizes global context modeling of features while maintaining computational efficiency. The Quad-Region Diffusion Graph Convolutional Residual Attention Block establishes a regional graph structure relationship, which includes the following steps:

[0136] The frequency domain feature map of the input pine wilt disease is performed group convolution dimension reduction to obtain , wherein represents a real space with four dimensions: represents the number of samples input into the network, represents the number of channels of the corresponding feature map, represents the spatial height of the feature map, represents the spatial width of the feature map;

[0137] The left upper, right upper, left lower, and right lower four regions are extracted to construct the inter-regional connection relationship, and the regional representation set is expressed as:

[0138] ;

[0139] wherein , , , respectively represent the regional representation obtained by average pooling of the left upper, right upper, left lower, and right lower four regions, which retains the spatial topological relationship of the features. Unlike traditional global attention, this module only interacts information between regional representations, greatly reducing the computational complexity, and at the same time establishing long-range dependencies through inter-regional message passing. It conforms to the spatial correlation decreasing law of visual features, and introduces a diffusion factor based on spatial distance perception to pass messages between regions, so that the interaction strength between regions with a farther distance is adaptively lower. The expression of the diffusion factor is:

[0140] ;

[0141] wherein , These represent the region indexes,

[0142] The message function for graph convolution is:

[0143] ;

[0144] in, Indicates attention weights, Represents the normalization factor. express Characteristics of the region;

[0145] Employing a diffusion threshold mechanism for feature updates ensures that information transfer is performed only when necessary, effectively preventing feature degradation caused by excessive smoothing. The aggregation operation is performed if the value is true, otherwise the original value is retained. The aggregation operation expression is:

[0146] ;

[0147] in, Indicates the updated number The characteristic representation of each neighboring region Represents a node The neighborhood group, Represents aggregate functions;

[0148] The expression for applying gated residual join is:

[0149] ;

[0150] ;

[0151] in, This indicates the flow of adaptive control information. This represents the weight matrix of the gated network. This represents the sigmoid activation function. Indicates the first The original features of each region before feature update. This represents the final enhanced feature representation after fusing the original features and the updated features from the neighborhood through a gated residual mechanism;

[0152] Using fixed weights Feature fusion with channel attention mechanism outputs an enhanced pine wilt disease feature map, expressed as:

[0153] ;

[0154] in, This indicates a batch normalization operation. express Convolution operation, representing a channel attention mechanism;

[0155] The gating residual structure in the DGCRAB module and the channel attention mechanism ensure the selectivity of information in the transmission process, enhancing the model's response to key features. Notably, Q-DGCRAB adopts a fixed fusion weight, avoiding the overfitting risk caused by overfitting and improving the model's generalization ability in complex scenarios.

[0156] In summary, Q-DGCRAB effectively solves the problem of long-range feature dependency modeling in target detection through a spatially aware graph convolution structure, while maintaining computational efficiency;

[0157] S303, using the epidemic semantic augmented detection module ESADM (Epidemic Semantics-Augmented Detection Module, ESADM) to introduce a pre-trained BERT model text encoder, realizing visual-text cross-modal feature fusion and providing semantic-level enhanced information;

[0158] Single visual modality in the target detection task faces the challenge of insufficient semantic understanding. Traditional convolutional neural networks rely solely on visual features for target recognition, lacking deep understanding of the semantic properties of target categories. In complex epidemic scenarios, false positives and false negatives are likely to occur, especially in epidemic-related target detection. The high visual similarity between different targets makes it difficult to accurately distinguish them based solely on visual features, limiting detection accuracy.

[0159] The epidemic semantic augmented detection module ESADM loads the pre-trained BERT model through a parameter freezing strategy to extract deep semantic representations of text, significantly improving the model's semantic perception ability while maintaining computational efficiency,

[0160] After two layers of fully connected dimension reduction network (768→256→128) for dimension reduction and nonlinear transformation, the high-dimensional text features output by the BERT model are compressed to a dimension space compatible with visual features. The contribution ratio of text and visual features is dynamically adjusted through learnable parameters k. Text-visual fusion is performed at the P4 and P6 layers of the text-remote sensing collaborative network model TRSSL-Net, with the expression being:

[0161] ;

[0162] where, represents an adaptive fusion weight, ensuring that text semantic information can enhance visual feature expression at different abstraction levels. Unlike traditional post-fusion methods, this strategy deeply embeds semantic information into the feature extraction process, allowing the model to obtain semantic guidance during feature learning, representing text-visual fusion, representing visual features, representing text features;

[0163] The processing procedure of the text features includes: firstly extracting deep semantic representation of the text through the frozen BERT encoder, then performing feature dimension reduction and nonlinear transformation through two fully connected networks, and finally obtaining text representation matching the dimension of the visual features; a linear weighting strategy is adopted in the fusion process, and the contribution ratio of the text and visual features is dynamically adjusted through a learnable parameter k, which avoids complex attention calculation, maintains the calculation efficiency while ensuring the fusion effect;

[0164] The epidemic semantic augmented detection module ESADM also includes a WorldDetect detection head specially designed for the fusion task of text embedding and target detection, which realizes deep alignment of visual features and text semantics through a contrastive learning mechanism. The WorldDetect detection head calculates the semantic matching degree between text-visual features through the ContrastiveHead module using cosine similarity, which enhances the model's ability to distinguish the categories of epidemic-related targets. ESADM adopts a fixed-level fusion strategy, avoiding overly complex fusion network design and improving the model's generalization ability in different epidemic scenarios. Through cross-modal feature fusion, ESADM effectively captures these key semantic features, providing important technical support for intelligent monitoring in epidemic prevention.

[0165] S4, a forest guardian collaborative optimization strategy FGA-CE (Forest Guardian Algorithm with Collaborative Evolution, FGA-CE) is used for dynamic optimization of the learning rate of the pine wood nematode disease unmanned aerial vehicle target detection network; combined with the global search ability of the forest cruise detection algorithm and the multi-population collaborative evolution strategy, it can adaptively adjust the learning rate parameter according to the verification index during training, effectively improving the model performance. The core idea of the forest guardian collaborative optimization algorithm comes from the three typical behaviors of unmanned aerial vehicles in forest monitoring: regional hovering, high-altitude overlooking, and remote surveying. We abstract these monitoring modes into mathematical models and introduce a multi-population collaborative evolution mechanism to enhance the global search ability and convergence speed of the algorithm;

[0166] FGA-CE dynamically optimizes the learning rate during the training process through a group collaboration mechanism and accelerates the convergence of the text-remote sensing collaborative network model TRSSL-Net;

[0167] The forest guardian collaborative optimization strategy FGA-CE optimizes the learning rate by simulating three behaviors of unmanned aerial vehicles in forest monitoring: regional hovering, high-altitude overlooking, and remote surveying;

[0168] The local spiral behavior develops around the current optimal solution, and the position update expression is:

[0169] ;

[0170] ;

[0171] where, represents the optimal position found by the current UAV, represents the current search agent position of the UAV, , represents the coefficient vector, represents the element-wise multiplication of the two vectors;

[0172] ;

[0173] linearly decreases from 2 to 0, and , where and are random numbers within the interval

[0174] The high-altitude overlooking strategy simulates the vertical lift observation of the UAV, approaching the optimal solution in a spiral path, and the position update expression is:

[0175] ;

[0176] ;

[0177] where, represents the distance between the UAV and the best observation point, represents the spiral descent constant, represents a random number within the interval (linearly decreasing from -1 to -2), represents the norm of the vector, used to represent the Euclidean distance between the current position of the UAV and the optimal position;

[0178] The remote survey strategy is triggered when , forcing the UAV to randomly select an unmonitored area for global exploration, by randomly selecting a reference point , the expression is:

[0179] ;

[0180] The position update expression is:

[0181] ; ​

[0182] The forest guardian cooperative optimization strategy FGA-CE introduces a multi-population cooperative evolution framework, divides the total unmanned aerial vehicle group N into multiple sub-group teams, in this embodiment, into 3 sub-group teams, and each sub-group team independently patrols;

[0183] An information sharing mechanism is established, and information exchange and elite migration are carried out regularly: information exchange is carried out every fixed number of iterations (5 times in this embodiment) between different sub-group teams, and the detection quality is sorted, and the best detection information is transmitted from the best team to other teams, and in this embodiment, the transmission rate is set to 0.2; the dynamic path optimization strategy selects the worst detection point in the receiving team for replacement, and adds a small disturbance to the transferred detection point, and the amplitude is The search range is effectively enhanced, and the monitoring diversity and global detection capability are enhanced;

[0184] After each iteration, boundary checking is performed on all unmanned aerial vehicles to ensure that the updated position is within the monitoring space, and the fitness of all detection points is re-evaluated, and the optimal solution of the sub-group team and the global optimal solution are updated;

[0185] A target function is constructed by a historical verification index mAP50, a momentum prediction and an adjustment of a learning rate, and the expression is:

[0186] ;

[0187] Wherein, represents the latest verification index, represents the relative change rate of the learning rate, represents a momentum trend factor calculated based on historical data; record the verification index of each epoch, and perform learning rate optimization every fixed interval (default 5 epochs), use the weighted historical trend, and the weight is , to ensure decay over time. The algorithm is triggered after verification, to balance the optimization effect and the calculation overhead. At the same time, the optimization range is limited to the preset minimum learning rate and the maximum learning rate , to prevent the learning rate from taking extreme values; the FGA-CE realizes self-adaptive adjustment of the learning rate in the pine wilt disease unmanned aerial vehicle YOLO11 detection network training, enhances the learning ability of the model to the features of infected pine trees, realizes the optimization process through CPU parallel computing, avoids occupying valuable GPU resources, and provides effective technical support for early detection of forest diseases.

[0188] S5, outputting the positioning and classification results of the pine wilt disease through the detection head.

[0189] Embodiment 2

[0190] To improve the accuracy of experimental results and minimize the interference of external variables, all experiments in this embodiment are completed in the same hardware and software environment; the hardware configuration used in the experiment is provided by the AutoDL platform, which ensures that all experiments are conducted on the same hardware basis and avoids the impact of hardware differences; in terms of software, the research adopts a standardized development environment to ensure consistency in the version of the operating system and related software tools, thereby minimizing the interference of software version differences on experimental results. The specific hardware specifications and software environment settings are shown in Table 1:

[0191] Table 1

[0192]

[0193] The effectiveness of each aspect in TRSSL-Net is comprehensively evaluated using five evaluation indicators; then, the effectiveness of each network on the internally planned dataset is verified to determine the progress; ablation experiments are further used to carefully examine the interdependence of individual networks to confirm their effectiveness; finally, experiments on two other pine wilt datasets are performed to verify the generalization ability of TRSSL-Net; in general, these experiments highlight the excellent detection performance of the TRSSL-Net high-altitude unmanned aerial vehicle pine wilt target detection framework on datasets of different sizes.

[0194] In this embodiment, the performance of the model is evaluated by recall (R), mAP50, mAP50-95, FPS, and GFLOPs.

[0195] Recall (R) represents the proportion of correct positive predictions to all positive actual, and the expression is:

[0196] ;

[0197] where TP represents the number of correctly detected positive samples, and FN represents the number of undetected positive samples;

[0198] Single-class mAP50 is calculated by calculating the average precision at an IoU threshold of 0.5, which comprehensively reflects the detection ability of the model in this class. Through the above steps and formulas, the performance of the model in the single-class target detection task can be clearly evaluated, and the expression is:

[0199] ;

[0200] mAP50-95 is a comprehensive indicator calculated by calculating the average precision at multiple thresholds ranging from 0.5 to 0.95 (step size 0.05) in IoU, and the expression is:

[0201] ;

[0202] wherein, denotes the IoU value, denotes the average precision at different IoU thresholds t;

[0203] The frame rate per second (FPS) is an important indicator of detection speed, which in this embodiment represents the average number of images detected per second, and the expression is:

[0204]

[0205] GFLOPs are used to measure model size and complexity.

[0206] Module effectiveness experiment:

[0207] 1. Vegetation index threshold mask filtering module effectiveness experiment;

[0208] In the embodiment, the enhancement effect of four vegetation indexes (GRVI, GLI, VARI and RGBVI) in the target detection task is evaluated; in order to comprehensively explore the influence of vegetation index on the performance of TRSSL-Net, a series of comparative experiments are designed to test the effect of single index and different combination strategies, and the experimental results are shown in Table 2:

[0209] Table 2:

[0210]

[0211] The experimental results show that among all the tested vegetation indexes, RGBVI (red-green-blue vegetation index) performs the most outstanding, significantly improves the target detection accuracy, and achieves the best results on all evaluation indicators; specifically, RGBVI achieves an mAP50 value of 0.875 and an mAP50-95 value of 0.416, and the R value also reaches 0.824, which comprehensively surpasses other vegetation indexes; other vegetation indexes such as GRVI, VARI and GLI also have certain enhancement effect, but the overall performance is not as good as RGBVI; this finding is consistent with the hypothesis of this embodiment, indicating that the effectiveness of vegetation index in a specific target detection task is highly dependent on its correlation with target features.

[0212] It is worth noting that multiple index combination experiments are also carried out, but the multiple index combination strategy fails to bring the expected performance improvement, which may be due to the existence of information redundancy or interference between different vegetation indexes, which instead weakens the expression of effective features; this shows that when designing feature enhancement modules, more attention should be paid to feature quality rather than quantity, and accurate selection of specific indexes highly related to target detection tasks may be more effective than blindly increasing feature dimensions.

[0213] ​Considering the model accuracy and efficiency, adding RGBVI alone provides the most cost-effective enhancement strategy, which can significantly improve the detection performance (mAP50 reaches 0.875, and mAP50-95 reaches 0.416) while maintaining a relatively low computational overhead; this finding has important practical value for high-precision target detection in resource-constrained scenarios, and also provides valuable experimental basis for subsequent research.

[0214] 2. Effectiveness of W-KAN

[0215] In this embodiment, the effectiveness of various feature enhancement methods in the WVKnet target detection framework is evaluated, and a new module integrating 'haar' wavelet transform and KAN attention mechanism is proposed, which is strategically placed at a specific stage of the model backbone network. The experimental results are shown in Table 3:

[0216] Table 3

[0217]

[0218] The experimental results show that the method provided in this embodiment improves by 2.2 percentage points compared to the best comparative method NonLocalMeans (mAP50 is 0.853); this performance gain is mainly due to the effective decomposition ability of wavelet transform on multi-scale features, combined with the adaptive feature fusion of KAN attention mechanism, which enables the model to capture both low-frequency global information and high-frequency detail information of the image, especially when dealing with complex scenes and small target detection, this advantage is more obvious.

[0219] It is worth noting that although BilateralFilter achieves a good result of 0.409 in the mAP50-95 indicator, close to the 0.416 of the method provided in this embodiment, its R value (0.734) and mAP50 value (0.843) are significantly lower than the method provided in this embodiment, indicating that the module proposed in this embodiment has more advantages in overall detection performance; although the traditional MedianFilter performs well in the R value (0.808), there is a significant gap in the accuracy indicator compared to the method provided in this embodiment.

[0220] In addition, the method provided in this embodiment is more efficient than the computationally complex NonLocalMeans operation, which performs well in mAP50 (0.853) but has a large computational overhead; SoftMaxPool achieves a result of 0.851 in the mAP50 indicator, but only 0.38 in the mAP50-95 indicator, indicating that its performance decreases significantly under more stringent evaluation criteria.

[0221] Through comparative experiments, it is found that different feature enhancement methods have their own advantages: GaussianFilter and BilateralFilter perform better in high-precision detection (mAP50-95), while NonLocalMeans and SoftMaxPool are more advantageous in medium-precision detection (mAP50); the method provided in the embodiment achieves balanced and excellent performance in all dimensions, proving the effectiveness of the combination of wavelet transform and KAN attention mechanism.

[0222] In summary, the method provided in the embodiment is superior to traditional filtering and feature enhancement techniques in all evaluation indicators, achieving the best performance of R value 0.824, mAP50 value 0.875, and mAP50-95 value 0.416; the feature enhancement module proposed in the embodiment provides significant performance improvement (R value 0.824, mAP50 value 0.875, and mAP50-95 value 0.416) for the TRSSL-Net model through efficient multi-scale feature extraction and enhancement capability, surpassing all traditional feature enhancement methods and being a cost-effective network enhancement strategy.

[0223] 3. Effectiveness experiment of four-domain diffusion graph convolution residual attention module

[0224] In the embodiment, the effectiveness of various graph convolutions and attention mechanisms in the target detection framework is evaluated, and a feature enhancement module based on spatial graph convolution, SpatialGraphConv, is proposed. The module is strategically placed in the feature map processing stage before the detection head, and the experimental results are shown in Table 4:

[0225] Table 4

[0226]

[0227] The experimental results show that the Q-DGCRAB proposed in the embodiment is superior to traditional graph convolution and attention mechanisms in all evaluation indicators, achieving the best performance of R value 0.824, MAP50 value 0.875, and MAP50-95 value 0.416; compared with the best comparative method Graph (MAP50 is 0.844), the method provided in the embodiment improves by 3.1 percentage points. This performance gain is mainly due to the effective modeling capability of Q-DGCRAB for the spatial relationship of feature maps. By constructing a spatial relationship graph of feature maps and applying graph convolution operations, the model can capture both local details and global context information, especially when dealing with complex scenes and dense target detection, this advantage is more obvious.

[0228] It is worth noting that although CBAM achieved a good result of 0.395 on the MAP50-95 indicator, close to 0.416 of the method provided in the embodiment, its R value (0.744) is significantly lower than the method provided in the embodiment, indicating that Q-DGCRAB has a significant advantage in detection recall rate; although the traditional ChebNetModule performs well in R value (0.791), there is a significant gap in the precision indicator compared with the method provided in the embodiment.

[0229] In addition, the method provided in the embodiment is more efficient than the computationally complex Non-Local Networks operation, which performs well in R value (0.79) but has high computational overhead. SelfAttention, although simple in structure, performs poorly in all indicators, with an R value of only 0.72 and a MAP50 of 0.826, indicating that a simple attention mechanism cannot fully capture the complex relationships between features.

[0230] Through comparative experiments, it is found that different feature enhancement methods have their own advantages: CBAM performs well in high-precision detection (MAP50-95), while Graph is more advantageous in medium-precision detection (MAP50); Q-DGCRAB provided in the embodiment balances and excels in all dimensions, proving the effectiveness of spatial graph convolution in feature enhancement.

[0231] In summary, the Q-DGCRAB module proposed in the embodiment provides significant performance improvement for target detection models through efficient spatial relationship modeling and graph convolution operations, surpassing all traditional graph convolution and attention mechanism methods, and is a cost-effective network enhancement strategy.

[0232] 4. Optimization algorithm effectiveness experiment

[0233] The optimization algorithm effectiveness experiment was carried out around the TRSSL-Net detection framework, and the two mainstream intelligent optimization algorithms PSO and WAM were compared with the method proposed in the embodiment; to ensure fair and reliable results, all algorithms were embedded in the same stage of the backbone network, maintaining consistent training dataset, data augmentation strategy, loss weight and hyperparameter settings, and using the same random seed for three independent reproductions and taking the average value. The experimental results are shown in Table 5:

[0234] Table 5

[0235]

[0236] From the results, it can be seen that the method proposed in the embodiment has significant advantages in all key indicators; specifically, compared with the best performing contrast algorithm PSO (mAP50 = 0.853), the method proposed in the embodiment achieves a 2.2 percent point improvement in mAP50; compared with WAM (mAP50-95 = 0.391), it improves by 2.5 percent points in the more stringent mAP50-95 indicator; in addition, the method proposed in the embodiment can still maintain higher precision in the high recall section of the PR curve, indicating that the positioning ability of the model for difficult-to-detect targets is also enhanced.

[0237] The fundamental reason for the performance improvement is that the method proposed in the embodiment simultaneously simulates and fuses three main search strategies: global exploration, local exploitation, and adaptive jumping; supplemented by a momentum update mechanism, the particles can make full use of the information of the historical optimal solution in the search space, avoid oscillation and accelerate convergence; at the same time, the multi-population cooperative evolution strategy introduces the ideas of "inter-population migration" and "elite exchange", which maintains diversity while facilitating rapid discovery of the global optimum; in summary, this mechanism effectively alleviates the pain points of traditional intelligent optimization, such as being easily trapped in local optimum and limited convergence speed, thereby significantly improving the overall precision and robustness of the target detection task.

[0238] 5. Ablation experiment

[0239] In order to verify the effectiveness of the target detection-based method proposed in the embodiment, an ablation experiment was conducted on the basis network, and the embodiment used the control variable method to add ESADM, W-KAN module, Q-DGCRAB, optimization algorithm (FGA-CE) and vegetation index threshold mask filter VTMF in turn, and then conducted 16 groups of ablation experiments on these 5 modules, and the experimental results are shown in Table 6:

[0240] Table 6

[0241]

[0242] Through the comparison of the first and second rows, it is proved that the integration of the text modal can effectively improve the detection performance, making the R value increase from 0.728 to 0.752 and the mAP50 increase from 0.815 to 0.831; when the W-KAN module is added, the mAP50 is further improved to 0.84745 and the MAP50-95 is improved to 0.38464, proving the importance of multi-scale feature decomposition; the introduction of Q-DGCRAB makes the recall rate R increase to 0.752 and the mAP50 reach 0.8395, indicating that the graph structure features are helpful for capturing the spatial relationship between targets;

[0243] Comparing row 13 and row 16, it can be found that when VTMF is added in the case of containing ESADM and Q-DGCRAB, the R value is improved from 0.76 to 0.778, and the mAP50 is improved from 0.849 to 0.858, which proves the importance of VTMF for improving the performance of the model, especially for detecting vegetation targets; when W-KAN module, Q-DGCRAB and FGA-CE are used in combination (row 13), the mAP50-95 reaches 0.404, which shows the effect of multi-module cooperation; after further adding VTMF (row 16), all indicators reach the optimal level.

[0244] In summary, compared with the base network, the mAP50 of the complete model (all modules are used) provided by the embodiment is improved from 0.815 to 0.875 (increased by 7.4%), and the R value is improved from 0.728 to 0.824 (increased by 13.2%), although the GFlops has increased (from 6.6 to 10.438), but the performance is significantly improved; the results of multiple experiments show that the text modal integration, W-KAN module, four-domain diffusion graph convolution residual attention module, forest guardian collaborative optimization algorithm and vegetation index threshold mask filtering can effectively improve the model precision, especially when the five modules are used together, the best detection effect can be achieved (R=0.824, mAP50=0.875, mAP50-95=0.416); therefore, the method proposed in the embodiment exhibits excellent performance in the target detection task.

[0245] 6. Comparative test

[0246] In the same test environment and test set, the embodiment and several traditional and current advanced target detection methods are compared, and the performance of the model provided by the embodiment is further analyzed, and the test results are shown in Table 7:

[0247] Table 7

[0248]

[0249] In the target detection model, the YOLO series has attracted widespread attention due to its excellent speed and accuracy balance. YOLOv3 initially introduced multi-scale feature prediction, but its recall rate was only 0.659, and MAP50 was 0.747, which performed poorly in detecting complex scenes. Subsequently, YOLOv5 and YOLOv8 achieved MAP50 of 0.751, but the recall rates were 0.688 and 0.656, respectively, which were still not ideal. YOLOv9t improved MAP50 to 0.760, but GFlops reached 8.5, with a large computational overhead. YOLOv10n optimized through attention mechanism, increasing MAP50-95 to 0.436 and MAP50 to 0.776, showing better generalization ability under different IoU thresholds. In the complex detector, RT-DETR adopted the Transformer architecture, achieving MAP50 of 0.802 and recall rate of 0.768, but its GFlops of 86.5 and FPS of 28.42 limited its application in resource-constrained scenarios. PSS-Net optimized through feature pyramid, achieving MAP50 of 0.803 and MAP50-95 of 0.437, achieving a good balance between accuracy and speed. Notably, YOLOv11n performed best in the MAP50-95 indicator, reaching 0.485, demonstrating its excellent detection quality.

[0250] According to the experimental results, the model proposed in this embodiment is superior to the existing detectors in most indicators, especially in recall rate, the model proposed in this embodiment reaches 0.824, which is 5.6 percentage points higher than the RT-DETR with the closest performance; in MAP50, the model proposed in this embodiment reaches 0.875, which is 6 percentage points higher than YOLOv11n; although in MAP50-95, the model proposed in this embodiment 0.416 is lower than 0.485 of YOLOv11n, but it is still superior to most comparative models; the reason why the model proposed in this embodiment can achieve such excellent performance is as follows: (a) the model proposed in this embodiment is improved based on the modern detection framework, and multiple advanced technologies are integrated to improve the precision and speed of the model; (b) the text modal fusion mechanism is introduced to enhance the ability of the model to understand semantic information; (c) the W-KAN module is used for multi-scale feature decomposition to effectively extract different frequency information; (d) the spatial graph convolution (SGC) effectively captures the spatial relationship between features and enhances the understanding ability of the model to complex scenes; (e) the FGA-CE optimization algorithm improves the model training process, so that the features are more discriminative; (f) the vegetation index enhancement (VGE) is specifically optimized for vegetation target detection, which improves the applicability of the model in agricultural scenes; although the model proposed in this embodiment is slightly higher than part of the models in GFlops (10.438), it still maintains 54.26 FPS, which fully meets the real-time detection requirements, especially compared with RT-DETR (86.5 GFlops, 28.42 FPS), the model proposed in this embodiment has a significant advantage in computing efficiency while maintaining higher accuracy; in summary, the experimental results show that the comprehensive feature enhancement framework proposed in this embodiment has superior performance in the target detection task, especially in application scenarios that require high recall rate and high precision.

[0251] 7. Visual analysis

[0252] Figure 3 The visualization comparison results of the model proposed in this embodiment and other 10 kinds of mainstream target detectors in aerial image detection are shown, from which it can be clearly seen that the performance difference of the model proposed in this embodiment and other detectors when processing complex scenes;

[0253] In the first column of images, the embodiment selects a scene with complex buildings and trees mixed to observe the detection results; compared with YOLOv11n and PSSNET, the model proposed in the embodiment not only can detect small targets more accurately, but also can better distinguish the boundaries of trees and buildings; especially in the shadow area, YOLOv3 and YOLOv5 often mistake the shadow as a target or completely ignore the target under the shadow, while the model proposed in the embodiment can correctly identify, which is mainly due to the spatial graph convolution (SGC) module in the model proposed in the embodiment, which can capture the spatial relationship between targets and make more accurate identification combined with text modal information.

[0254] In the third and fourth columns of images, the embodiment selects a building scene with complex roof structure; it can be seen that YOLOv8 and YOLOv10n often mistakenly identify the red roof as vegetation or trees, and the confidence is also low; in contrast, the model proposed in the embodiment can more accurately distinguish artificial structures from natural vegetation through the vegetation index enhancement (VGE) module, significantly reducing the false detection rate; even in the case of uneven lighting, the model proposed in the embodiment maintains high detection accuracy.

[0255] In the fifth and sixth columns, the embodiment shows the detection results of mixed scenes of water and buildings; RT-DETR performs well on large targets, but its detection accuracy at the junction of water and vegetation is not as good as the model proposed in the embodiment; the model proposed in the embodiment can better handle the transition between different texture regions through the W-KAN module for multi-scale feature decomposition, improving the completeness and accuracy of detection.

[0256] In addition, from the overall detection results, the model proposed in the embodiment has higher confidence and more accurate bounding box positioning in all test scenes; for example, in the small building detection in the seventh column, YOLOv6 and YOLOv9t have obvious missed detection phenomenon, while the model proposed in the embodiment detects almost all targets; this is due to the FGA-CE optimization algorithm that improves the feature expression ability of the model, enabling it to handle more complex scene changes.

[0257] Overall, the visualization results clearly show that the comprehensive feature enhancement framework proposed in the embodiment has significant advantages in the aerial image target detection task, especially in handling shadow areas, distinguishing similar textures (such as red roofs and vegetation), and detecting small targets in complex backgrounds. These improvements not only improve detection accuracy, but also optimize common missed and false detection problems in existing models.

[0258] 8、PDT public dataset generalization experiment

[0259] To verify the generalization ability of the model proposed in this embodiment, this embodiment selects the PDT (Pests and Diseases Tree) dataset jointly established by Shandong Computer Science Center, Peking University and other institutions in China for experiment; the PDT dataset is the first high-precision tree pest and disease target detection dataset based on unmanned aerial vehicle, aiming to solve the problem of lack of professional dataset in the development of pest and disease detection model in the field of agriculture; the dataset is composed of high-definition images collected by unmanned aerial vehicle at a height of about 200 meters, contains multiple tree pest and disease categories, has high-precision annotation and complex real scene characteristics, and at the same time, the dataset is described in text, which serves as the TRSSL-Net text dataset.

[0260] The complete model proposed in this embodiment is compared with the current mainstream target detector on the PDT dataset, and the experimental results are shown in Table 8.

[0261] Table 8

[0262]

[0263] From the experimental data, it can be seen that the model proposed in this embodiment achieves the best performance in all evaluation indicators, with R value reaching 0.845, MAP50 reaching 0.921, and MAP50-95 reaching 0.604; compared with the better YOLOv5 (R = 0.825, MAP50 = 0.905) and RT-DETR (R = 0.815, MAP50 = 0.909), the model proposed in this embodiment improves the MAP50 performance by 2% and 1.2% respectively, and under the more stringent MAP50-95 evaluation standard, the model proposed in this embodiment (0.604) improves the performance of YOLOv5 (0.578) and PSS-NET (0.584) by 2.6% and 2% respectively, showing stronger generalization ability and detection accuracy.

[0264] From the visualization results of the PDT dataset, as shown in the left three columns of Table 7, the model (OURS) proposed in this embodiment has obvious advantages in detection effect in the comparative experiments shown in the left three columns of Table 7; in the complex soil background scene in the first column, the model proposed in this embodiment can accurately identify and locate more pest targets, and the number of detection boxes is obviously more than that of other methods, and the positioning is accurate; in the dense green vegetation scenes in the second and third columns, the detection ability of the model proposed in this embodiment on the blue disease spot is particularly outstanding, and it can accurately identify the scattered disease and pest area in the high-density green background; compared with other mainstream detectors, such as YOLOv11, PSS-NET, RT-DETR, YOLOv10 and the like, the methods in the same scene have different degrees of missed detection phenomenon, especially the earlier versions of YOLOv8, YOLOv6, YOLOv5 and YOLOv3, the detection performance in the complex background is obviously insufficient.

[0265] Further analysis of the details of the visualization results, the detection boxes of the model proposed in this embodiment are not only more in number, but also more accurate in positioning, and the boundaries are more in line with the target contour; especially in dealing with small targets and edge targets, the model proposed in this embodiment shows excellent performance, which is particularly obvious in the green vegetation scenes in the second and third columns - other methods can only detect more obvious large disease spots, while the model proposed in this embodiment can identify more small disease and pest targets; the visualization results directly verify the advantages of the quantitative indicators and show the detection quality of the model proposed in this embodiment in the actual application scene.

[0266] It is worth noting that the performance of the model proposed in this embodiment on the PDT dataset is better than the results on the self-built dataset; this is mainly because the UAV shooting height of the PDT dataset is about 200 meters, while the shooting height of the self-built dataset is about 500 meters; the lower flight height makes the target size in the PDT dataset relatively larger, and the texture features are more obvious, which is conducive to feature extraction and target recognition of the model; this phenomenon is also clearly verified in the visualization results - in the detection results of the PDT dataset, the boundaries of the disease and pest targets are clearer, the color contrast is more obvious (such as blue disease spots in the green background), and the confidence of the detection box is generally higher, which reflects the excellent processing capability of the model for large size and high contrast targets, and this phenomenon also verifies the advantages of the wavelet KAN module and the spatial graph convolution in the model proposed in this embodiment in different scale feature processing - when the target features are more obvious, these modules can more fully play their functions and extract more rich feature information.

[0267] The above generalization experiment proves that the feature enhancement framework proposed in this embodiment has excellent stability and precision when applied to public data sets, especially when processing images taken by unmanned aerial vehicles of different heights. The visual comparison results further support this conclusion, and from Figure 4 As can be clearly seen, in different complexity scenarios (from soil background to dense vegetation), the model proposed in this embodiment can maintain stable detection performance, and the quality and quantity of the detection box are significantly better than the comparison method, especially when processing small targets in complex backgrounds. It is worth noting that this advantage is particularly evident in the MAP50-95 indicator, which has a significant improvement compared to similar models, proving that the model proposed in this embodiment can maintain high performance characteristics at different IoU thresholds.

[0268] 9. Pine wilt disease public data set generalization experiment

[0269] In order to verify the generalization ability of the model proposed in this embodiment, this embodiment selects the pine wilt disease public data set for experiment; this data set is a standardized data set specially established for pine wilt disease detection research, which contains images of pine trees in different regions and different infection stages, providing an important benchmark for the evaluation of pine wilt disease detection models. In addition, text descriptions are made for this data set as text modal input for TRSSL-Net.

[0270] The complete model of this embodiment and the current mainstream target detector are compared on the pine wilt disease public data set, and the experimental results are shown in Table 9:

[0271] Table 9

[0272]

[0273] The model proposed in this embodiment achieves excellent results of R=0.776, MAP50=0.857, and MAP50-95=0.395. Compared with the better performing YOLOv5 (MAP50=0.827) and RT-DETR (MAP50=0.831), the model proposed in this embodiment improves by about 3.6% and 3.1% in the MAP50 indicator, respectively. Meanwhile, under the more stringent MAP50-95 evaluation standard, the model proposed in this embodiment (0.395) also maintains a competitive advantage over advanced methods such as YOLOv9 (0.399) and YOLOv8 (0.392), demonstrating stronger generalization ability and detection accuracy.

[0274] From the visualization results of the pine wilt disease public data set, as shown in Figure 4In the contrast experiments shown in the three columns on the right, the model proposed in this embodiment performs outstandingly in the detection effect in complex forest environments: in the mixed vegetation scene in the first column, the model proposed in this embodiment can accurately identify the pine wood nematode disease area, the detection frame is accurately positioned and the number is moderate, effectively avoiding the problem of over-detection; under different light and vegetation density conditions in the second and third columns, the model proposed in this embodiment shows good environmental adaptability and can accurately identify disease characteristics in a green background, especially for early disease symptoms with subtle color changes, the detection effect is outstanding.

[0275] Compared with the visualization of other mainstream detectors, the methods such as YOLOv11, PSS-NET, and RT-DETR have different degrees of detection instability in the same scene; in particular, it is worth noting that when dealing with complex forest backgrounds, most of the comparison methods either have false detection (such as rt-detr, YOLOv11n, YOLOv8, YOLOv6, and YOLOv3) or the detection frame positioning is not accurate enough (such as pssnet); in contrast, the model proposed in this embodiment maintains stable detection performance in all test scenes, and the detection frame not only has a reasonable number, but also has high boundary fitting degree, fully embodying the professionalism and reliability of the model in the pine wood nematode disease detection task.

[0276] It is worth noting that the performance of the model proposed in this embodiment on the public dataset is different from that on the self-built dataset, which is mainly due to the differences in image acquisition conditions of the two datasets, including factors such as unmanned aerial vehicle flight height, shooting angle, and lighting environment; from the visualization results, it can be observed that the disease characteristics in the pine wood nematode disease dataset are relatively more subtle and have low color contrast, which puts higher requirements on the feature extraction capability of the model; these differences cause the size and clarity of the target features to be different, thereby affecting the detection performance of the model; however, through visualization comparison, it can be seen that the W-KAN module and the four-domain diffusion graph convolution residual attention module (Q-DGCRAB) proposed in this embodiment can effectively handle these changes and still maintain stable detection effect under different conditions, which is particularly evident in the low-contrast scene in the third column.

[0277] The above generalization experiment proves that the feature enhancement framework proposed in the embodiment has excellent stability and precision, and exhibits good adaptability when applied to public data sets; the visual comparison result further verifies this conclusion, whether in a complex forest environment or under different light conditions, the model proposed in the embodiment can maintain consistent detection quality, and the precision and recall rate of the detection box are significantly better than those of the comparison method; the significant advantage in the MAP50-95 index proves that the model proposed in the embodiment can maintain high performance under different IoU thresholds, providing reliable technical support for the early detection and prevention of pine wood nematode disease.

[0278] In the embodiment, five innovative modules are integrated: vegetation index threshold mask filtering (VTMF), W-KAN module, four-domain diffusion graph convolution residual attention module (Q-DGCRAB), forest guardian cooperative optimization algorithm (FGA-CE), and epidemic semantic enhancement detection module (ESADM); these modules are carefully designed and optimized to form a complete technical system, which can work cooperatively and complement each other, effectively improving the accuracy and stability of pine wood nematode disease detection.

[0279] The method provided in the embodiment particularly emphasizes multi-scale feature extraction, spatial correlation modeling, and training process stabilization, not only focusing on the improvement of single detection performance, but also constructing a comprehensive solution to cope with multiple technical challenges in pine wood nematode disease detection. Through rigorous validation experiments on self-built data sets and public data sets, the method provided in the embodiment exhibits excellent performance and generalization ability.

[0280] The advantages of the embodiment framework can be summarized as follows: (1) The vegetation index threshold mask filter (VTMF) effectively solves the problem of complex forest background interference. Through a specially designed vegetation feature extraction and enhancement mechanism, the ability to distinguish between diseased trees and healthy trees and other vegetation types is significantly improved, even in complex environments such as terrain undulations, shadow areas, and seasonal changes. (2) The W-KAN module effectively preserves the high-frequency texture details in high-altitude remote sensing images through wave frequency domain decomposition and reconstruction technology, solving the problem of high-frequency information loss in the down-sampling process of traditional convolutional networks. This allows the model to more accurately capture key features such as tree crown edges and shadow textures, improving the accuracy of diseased tree identification. (3) The four-domain diffusion graph convolution residual attention module (Q-DGCRAB) breaks through the limitations of traditional convolution receptive fields and establishes a correlation mechanism between distant pixels in the image. This allows the model to effectively grasp the spatial dependency in wide-area drone images, especially for understanding and analyzing the continuous distribution of vegetation patches and the spatial structure of the ecosystem. (4) The Forest Guardian Cooperative Optimization Algorithm (FGA-CE) addresses the instability of drone image detection model training by proposing an adaptive adjustment strategy. This effectively alleviates the convergence speed lag and performance index oscillation problems, allowing the model to avoid getting stuck in local suboptimal solutions in complex scenarios and achieving a more stable and efficient training process. (5) The Epidemic Semantic Enhancement Detection Module (ESADM) provides semantic-level supplementary information for image features, enriching the model's understanding of disease characteristics and further improving detection performance and model robustness.

[0281] Through a series of ablation experiments and comparative experiments, the effectiveness of each module and its synergistic effect are verified. The experimental results on the self-built dataset and the public pine wilt disease dataset show that the model provided in the embodiment significantly outperforms existing mainstream detection models in key indicators such as MAP50 and MAP50-95. It is particularly noteworthy that the method provided in the embodiment performs well on images taken at different altitudes, demonstrating good adaptability and generalization ability.

[0282] The multi-modal feature enhancement framework proposed in the embodiment provides a systematic solution for drone pine wilt disease detection, effectively addressing current technical challenges and providing new ideas and methods for further research in the field of forest pest monitoring. These advances will help improve the early detection capabilities of forest pests and provide important technical support for forest resource protection and sustainable management.

[0283] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed with the preferred embodiments as above, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes, as long as the changes or modifications do not deviate from the technical solution of the present application. Any modification, change, equivalent change and modification of the above embodiments made according to the technical essence of the present application still belong to the scope of the technical solution of the present application.

Claims

1. A method for detecting pine wilt disease of a UAV based on multi-modal features, characterized in that, The method comprises the following steps: S1, obtaining a visible light image of a pine forest through a UAV high-altitude remote sensing platform, screening a pine nematode disease pine tree image after cutting, and performing image processing to establish an image-text paired multi-modal data set, and dividing the multi-modal data set into a training set, a validation set and a test set according to a ratio of 8:1:1; S2, performing pretreatment on the visible light image by using a vegetation index threshold mask filter VTMF to obtain spectral characteristics of a pine tree health state; S3, inputting the image enhanced by the VTMF into a text-remote sensing collaborative network model TRSSL-Net based on a YOLO architecture; S301, constructing a W-KAN feature extraction module at a P3 layer of the text-remote sensing collaborative network model TRSSL-Net, performing frequency domain feature decomposition, nonlinear modulation and noise suppression on the image enhanced by the VTMF through discrete wavelet transform DWT, and extracting a frequency domain feature map of the pine nematode disease; S302, establishing a regional graph structure relationship by using a Q-DGCRAB four-domain diffusion graph convolution residual attention module, dividing the frequency domain feature map of the pine nematode disease into four domains of upper left, upper right, lower left and lower right, and constructing a graph structure by taking each region as a node; S303, introducing a pre-trained BERT model text encoder by using an epidemic semantic enhancement detection module ESADM to realize visual-text cross-modal feature fusion and provide enhanced information at a semantic level; S4, adopting a forest guardian collaborative optimization strategy FGA-CE to dynamically optimize a learning rate in a training process through a group collaboration mechanism, and accelerating convergence of the text-remote sensing collaborative network model TRSSL-Net; S5, outputting a positioning and classification result of the pine nematode disease through a detection head. 2.The method of claim 1, wherein, The image processing comprises the following steps: S101, adjusting the pine nematode disease pine tree image in a jpg format to 800*800 pixels and inputting the image into labelmg software; S102, marking the pine nematode disease pine tree image manually on a visual platform by an expert, performing multi-dimensional parameter text description on each pine nematode disease pine tree image, the multi-dimensional parameters including distribution characteristics, morphological characteristics and position information of the pine nematode disease, making a label of an irregular disease and saving the label in a txt format, and establishing an image-text paired multi-modal data set. 3.The method of claim 1, wherein, The vegetation index threshold mask filter VTMF enhances spectral feature differences between vegetation and non-vegetation areas by using a green leaf index GLI, and the expression is: ； wherein , , Rg, Rh, and Rb represent reflectance of green, red, and blue band, respectively; The green-red vegetation index GRVI, the visible light atmospheric impedance index VARI and the RGB vegetation index capture spectral characteristics of a vegetation health state, forming a multi-dimensional feature space. 4.The method of claim 1, wherein, The W-KAN feature extraction module adopts discrete wavelet transform DWT to decompose image features after VTMF enhancement into low-frequency subbands LL and high-frequency subbands LH, HL and HH through a haar wavelet basis function, realizing frequency domain separation of signals and noises; Image feature transformation is performed by using a two-dimensional haar wavelet basis function, and the expression is: ； wherein, denotes a mother wavelet function, denotes a scale parameter, denotes a translation parameter, denotes an argument of the wavelet function, for representing a time, spatial position of a signal or image; ； ； ； ； wherein, , respectively denote low-pass, high-pass filter coefficients, respectively denote row and column index positions of the output image, respectively denote row and column offset indices for the filter when convolved on the input LL image. The KAN attention mechanism is used to perform nonlinear modulation on four sub-band coefficients to extract the specific feature map of the pine wood nematode disease in the frequency domain, and the expression is as follows: ； wherein, represents a feature map of the input image under a specific frequency band band.

5. The method of claim 4, wherein the method comprises: The Q-DGCRAB four-domain diffusion graph convolution residual attention module establishes a regional graph structure relationship, and specifically includes the following steps: Frequency domain feature map of input pine wilt disease Conducting Group convolution dimension reduction obtains , wherein represents a real number space, having four dimensions: represents the number of samples of the first input network, represents the number of channels of the corresponding feature map, represents the spatial height of the feature map, represents the spatial width of the feature map; The region representation set is constructed by extracting the region representations of the upper left, upper right, lower left and lower right four domains to construct the inter-region connection relationship The expression is: ； wherein, , , , respectively represent the region representations obtained by average pooling of the top-left, top-right, bottom-left, and bottom-right four regions. A diffusion factor based on spatial distance perception is introduced to transfer messages between regions, so that the interaction intensity between regions with a farther distance is adaptively lower, and the expression of the diffusion factor is as follows: ； wherein , respectively represent a region index; The message function of the graph convolution is as follows: ； wherein, denotes an attention weight, denotes a normalization factor, denotes a feature of a region; The feature is updated by using a diffusion threshold mechanism, and when an aggregation operation is performed, otherwise the original value is kept, and the aggregation operation expression is: ； wherein, represents an updated feature representation of the neighbor region, represents a neighbor set of the node , represents an aggregation function; The gated residual connection is applied, and the expression is as follows: ； ； wherein, denotes adaptive control information flow, denotes a weight matrix of a gating network, denotes a sigmoid activation function, denotes the original feature of the th region before feature update, denotes the final enhanced feature representation after fusing the original feature and the neighborhood updated feature through the gating residual mechanism. Using fixed weights The feature fusion is performed with a channel attention mechanism, and an enhanced pine wood nematode disease feature map is output, and the expression is: ； wherein, denotes a batch normalization operation, denotes a convolution operation, denotes a channel attention mechanism. 6.The method of claim 5, wherein the method further comprises: The epidemic semantic enhancement detection module ESADM loads a pre-trained BERT model through a freezing parameter strategy, extracts deep semantic representations of the text, and performs dimension reduction and nonlinear transformation through a two-layer fully connected dimension reduction network, so as to compress high-dimensional text features output by the BERT model to a dimension space compatible with visual features. The contribution ratio of the text and the visual features is dynamically adjusted through a learnable parameter k, and the text-visual fusion is performed at the P4 layer and the P6 layer of the text-remote sensing collaborative network model TRSSL-Net, and the expression is as follows: ； wherein, denotes an adaptive fusion weight, ensuring that the text semantic information can enhance the visual feature expression at different abstraction levels, denotes text-visual fusion, denotes visual features, denotes text features; The epidemic semantic enhancement detection module ESADM further includes a WorldDetect detection head, which realizes deep alignment of visual features and text semantics through a contrastive learning mechanism. The WorldDetect detection head calculates the semantic matching degree between the text-visual features through a ContrastiveHead module by using cosine similarity.

7. The method of claim 1, wherein the method is characterized by, The forest guardian collaborative optimization strategy FGA-CE optimizes the learning rate by simulating three behaviors of the unmanned aerial vehicle in forest monitoring: regional hovering behavior, high-altitude overlooking behavior, and remote surveying behavior. The regional hovering behavior develops locally around the current optimal solution, and the position update expression is as follows: ； ； wherein, represents the optimal position found by the current drone, represents the current search agent position of the drone, , represents the coefficient vector, represents the element-wise multiplication of the two vectors. The high-altitude overlooking strategy simulates the vertical observation of the unmanned aerial vehicle, and the position update expression is as follows: ； ； wherein, represents the distance of the drone from the best observation point, represents the constant of the spiral descent, represents a random number, represents the modulus of the vector, used to represent the Euclidean distance between the current position of the drone and the optimal position; The remote survey strategy is triggered at The UAV is forced to randomly select an unmonitored area for global exploration by randomly selecting a reference point The expression is: ； The position update expression is as follows: 。 8.The method of claim 7, wherein, The forest guardian collaborative optimization strategy FGA-CE introduces a multi-population collaborative evolution framework, divides the total unmanned aerial vehicle group N into multiple sub-population teams, and each sub-population team independently patrols. Establish information sharing mechanism, regular information exchange and elite migration: every fixed iteration in different sub-group team information exchange, according to the detection quality sorting, from the best team to other team transfer high quality detection information; dynamic path optimization strategy selection receiving team in the worst detection point replacement, while adding a small perturbation to the transfer detection point, amplitude is Search range; After each iteration, boundary checks are performed on all unmanned aerial vehicles to ensure that the updated positions are within the monitoring space, and the fitness of all detection points is re-evaluated, and the optimal solution of the sub-population team and the global optimal solution are updated. A target function is constructed through a historical verification index mAP50, and the learning rate is predicted and adjusted through momentum.

Images