Unmanned aerial vehicle pine wood nematode disease detection method based on multi-modal characteristics

Abstract

The invention relates to an unmanned aerial vehicle pine wood nematode disease detection method based on multi-modal features. The method comprises the following steps: constructing an image-text multi-modal data set to provide prior semantic knowledge; using vegetation index threshold mask filtering (VTMF) to enhance spectral difference and suppress noise; frequency domain denoising and feature retention are carried out by using a W-KAN module fusing wavelet transform and KAN attention; modeling a long-distance spatial dependency relationship through a four-domain diffusion diagram convolution residual attention module (Q-DGCRAB); an epidemic situation semantic enhancement detection module (ESADM) is introduced to realize text-visual feature fusion; the training process is dynamically optimized by adopting a forest guardian collaborative optimization algorithm (FGA-CE); according to the method, the problem of low pine wood nematode disease detection precision caused by complex background, repeated texture, significant noise and insufficient semantic information in a high-altitude unmanned aerial vehicle image is solved, the pine wood nematode disease detection precision and robustness in a complex forest stand environment are significantly improved, and the pine wood nematode disease can be accurately monitored in early stage in real time.

Description

Technical Field

[0001] This invention relates to the field of forest pest and disease monitoring technology, and in particular to a method for detecting pine wilt disease using unmanned aerial vehicles (UAVs) based on multimodal characteristics. Background Technology

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

[0003] In the early days, the detection of pine wilt disease relied mainly on the subjective experience of growers for localization. This method was not only time-consuming and labor-intensive, but also prone to errors by inexperienced personnel, significantly reducing the efficiency and accuracy of disease identification. In the field of smart forestry, forestry remote sensing technology helps forestry workers improve their ability to analyze pine diseases. This technology can accurately determine the extent of damage caused by pine wilt disease, thereby specifying precise control measures and bringing about a revolutionary change in disease detection. However, researchers have found that relying solely on image data is insufficient to fully grasp the development pattern of pine wilt disease, and also lacks key details for judging the severity of the disease. Especially in complex mountain forest environments, the reliability of traditional detection methods decreases significantly. With the development of computer vision technology and deep learning, target detection has become the core technology for image analysis of pine wilt disease. 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 networks (R-CNN); (2) single-stage detection models, represented by single-shot multi-box detectors (SSD) and single-stage target detection models (YOLO). The core difference between these two types of models lies in their detection mechanisms: the two-stage detection model first needs to generate candidate boxes that may contain diseased tree regions, and then perform classification and localization operations on these candidate boxes; the single-stage detection model can directly use network feature maps to synchronously predict the location and category information of the diseased area; since the single-stage model has lower computational complexity, the convolutional neural network (CNN)-based method can analyze multiple images... Scale features and feature fusion mechanisms have effectively improved the detection accuracy and efficiency of diseased areas, and have become mainstream in practical applications. Among them, the YOLO series models enhance the ability to capture image semantic information through multi-level feature extraction mechanisms, and their network structure design significantly reduces computational complexity, thereby meeting the needs of real-time detection. YOLOv5 introduces multi-size models, automatic anchor point calculation, and mixed-precision training, which enhances the ability to recognize small targets and adapt to complex scenes. The latest YOLOv8 abandons the traditional method of relying on predefined anchor boxes and adopts an anchor-free detection mechanism to directly predict the target center point and bounding box, simplifying the model structure, reducing computational complexity, and significantly improving the detection accuracy of small targets. At the same time, the Transformer architecture has also begun to emerge in the field of target detection. DETR (DEtection TRansformer) was the first to introduce this architecture into object detection, directly predicting targets using an encoder-decoder structure, eliminating the region proposal and non-maximum suppression (NMS) steps, and simplifying the process. Subsequently, RT-DETR, through an efficient hybrid encoder and a query selection mechanism with minimal uncertainty, first realized the application of DETR in real-time object detection. The continuous innovation of the YOLO series has not only consolidated its own technological leadership, but also promoted the development of the entire single-stage detector field, making it the mainstream choice in scenarios requiring rapid response, such as autonomous driving and remote sensing.

[0004] Despite significant progress in object detection, the YOLO single-object detection model still has shortcomings in four key dimensions: In the detection of pine wilt disease by drones, the complex background formed by various vegetation types, topographic relief, shaded areas and seasonal changes in the forest environment seriously interferes with the accurate identification of diseased trees. Although forest remote sensing images acquired by high-altitude UAVs contain multi-scale texture features such as canopy edges, shadow textures and terrain changes, repetitive texture patterns and random noise interference cause high-frequency detail loss during downsampling of conventional convolutional neural networks, resulting in multiple technical challenges for the model, such as texture feature preservation, accurate interpretation of target boundaries and improvement of noise robustness. Because of its wide coverage, drone aerial images typically cover a geographical area of ​​several to tens of square kilometers. The spatial layout of forest vegetation and the continuous distribution of vegetation patches in the ecosystem exhibit complex spatial relationships. However, when extracting target features, traditional convolutional neural networks are limited by their receptive field and have difficulty effectively capturing long-distance dependencies in the image, especially in large scenes or complex backgrounds where they lack a good grasp of the relationship between the target and its surrounding environment. Due to the significant multi-scale feature distribution, dynamic background interference, and spatial heterogeneity between the target and the scene, the target detection model training process of drone aerial images is highly unstable, which manifests as slow convergence speed, frequent oscillation of performance indicators, and easy to get trapped in local suboptimal solutions in scenes with a large dynamic range of target scale and high background texture complexity. Summary of the Invention

[0005] In view of this, the present invention provides a method for detecting pine wilt disease by UAV based on multimodal features, which can solve the problems of low detection accuracy, large background interference, low accuracy rate, poor generalization ability and unstable training when traditional UAV remote sensing detects pine wilt disease in complex forest environments, and realize early accurate and real-time monitoring of pine wilt disease.

[0006] To achieve the above objectives, the present invention provides a method for detecting pine wilt disease using unmanned aerial vehicles (UAVs) based on multimodal features, comprising the following steps: S1. Visible light images of pine forests are acquired through a drone high-altitude remote sensing platform. After being cut and screened, images of pine trees affected by pine wilt disease are selected. Image processing is performed to establish an image-text paired multimodal dataset. The multimodal dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1. S2. The visible light image is preprocessed using Vegetation Index Threshold Mask Filter (VTMF) to obtain the spectral characteristics of the pine tree's health status. S3. Input the VTMF-enhanced image into TRSSL-Net, a text-remote sensing collaborative network model based on the YOLO architecture. S301. Construct a W-KAN feature extraction module in the P3 layer of the text-remote sensing collaborative network model TRSSL-Net. Perform frequency domain feature decomposition, nonlinear modulation and noise suppression on the image enhanced by VTMF through discrete wavelet transform (DWT) to extract the frequency domain feature map of pine wilt disease. S302. The Q-DGCRAB four-domain diffusion graph convolutional residual attention module is used to establish the regional graph structure relationship. The frequency domain feature map of the pine wilt disease is divided into four domains: upper left, upper right, lower left, and lower right. Each region is used as a node to construct the graph structure. S303. The ESADM module for enhancing the semantic detection of the epidemic is used to introduce a pre-trained BERT model text encoder to achieve cross-modal feature fusion between vision and text, providing semantic enhancement information. S4. The Forest Guardian Collaborative Optimization Strategy (FGA-CE) is adopted, and the learning rate during the training process is dynamically optimized through the group collaboration mechanism, thereby accelerating the convergence of the text-remote sensing collaborative network model TRSSL-Net. S5. Output the location and classification results of pine wilt disease through the detection head.

[0007] Preferably, the image processing includes the following steps: S101. Adjust the pine tree image of pine wilt disease in JPG format to 800×800 pixels and input it into the labelmg software; S102. Experts manually label the pine tree images of pine wilt disease on the visualization platform, and provide multi-dimensional parameter text descriptions for each pine tree image of pine wilt disease. The multi-dimensional parameters include the distribution characteristics, morphological characteristics and location information of pine wilt disease. Irregular disease labels are created and saved in txt format to establish a multimodal dataset with image-text pairing.

[0008] Preferably, the vegetation index threshold mask filter (VTMF) enhances the difference in spectral characteristics between vegetated and non-vegetated areas through the green leaf index (GLI), as expressed by: ; in, , , These represent the reflectance of the green, red, and blue bands, respectively. Then, the Green-Red Vegetation Index (GRVI), the Visible Light Atmospheric Impedance Index (VARI), and the RGB Vegetation Index are calculated to capture the spectral characteristics of vegetation health status, forming a multidimensional feature space.

[0009] Preferably, the W-KAN feature extraction module uses Discrete Wavelet Transform (DWT) to decompose the image features enhanced by VTMF into low-frequency subband LL and high-frequency subbands LH, HL, and HH through the Haar wavelet basis function, thereby achieving frequency domain separation of signal and noise. Image feature transformation is performed using two-dimensional Haar wavelet basis functions, expressed as follows: ; in, Represents the mother wavelet function. Indicates the scale parameter. Indicates the translation parameter. The independent variable of the wavelet function is used to represent the time and spatial location of a signal or image; ; ; ; ; in, , These represent the coefficients of the low-pass and high-pass filters, respectively. These represent the row and column index positions of the output image, respectively. The column offset index is used to represent the row and column offset indices when the filter is convolved on the input LL image; The KAN attention mechanism is used to nonlinearly modulate the coefficients of the four subbands to extract specific feature maps of pine wilt disease in the frequency domain. The expression is as follows: ; in, This represents the feature map of the input image within a specific frequency band.

[0010] Preferably, the Q-DGCRAB four-domain diffusion graph convolutional residual attention module establishes the region graph structure relationship by specifically including the following steps: Frequency domain feature map of the input pine wilt disease conduct Grouped convolution dimensionality reduction yields ,in, The space representing the real number space has four dimensions: This indicates the number of samples input into the network at one time. This indicates the number of channels in the corresponding feature map. Represents the spatial height of the feature map. Indicates the spatial width of the feature map; Extracting representations from the top-left, top-right, bottom-left, and bottom-right domains to construct inter-region connectivity relationships, resulting in a set of region representations. The expression is: ; in, , , , These represent the region representations obtained by average pooling of the four regions: top left, top right, bottom left, and bottom right, respectively. A spatial distance-aware diffusion factor is introduced to transmit messages between regions, so that the interaction strength between regions that are farther apart adaptively decreases. The expression for the diffusion factor is: ; in, , These represent region indexes; The message function for graph convolution is: ; in, Indicates attention weights, Represents the normalization factor. express Characteristics of the region; Feature updates are performed using a diffusion threshold mechanism when The aggregation operation is performed if the value is true, otherwise the original value is retained. The aggregation operation expression is: ; in, Indicates the updated number The characteristic representation of each neighboring region Represents a node The neighborhood group, Represents aggregate functions; The expression for applying gated residual join is: ; ; 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; Using fixed weights Feature fusion with channel attention mechanism outputs an enhanced pine wilt disease feature map, expressed as: ; in, This indicates a batch normalization operation. express Convolution operation, This indicates the channel attention mechanism.

[0011] Preferably, the epidemic semantic enhancement detection module ESADM loads a pre-trained BERT model using a parameter freezing strategy, extracts deep semantic representations of the text, performs dimensionality reduction and nonlinear transformation through a two-layer fully connected dimensionality reduction network, compresses the high-dimensional text features output by the BERT model into a visually compatible dimensionality space, and dynamically adjusts the contribution ratio of text and visual features through a learnable parameter k. Text-visual fusion is performed at layers P4 and P6 of the text-remote sensing collaborative network model TRSSL-Net, expressed as: ; in, This indicates adaptive fusion weights, ensuring that semantic information from the text can enhance the expression of visual features at different levels of abstraction. This indicates text-visual fusion. Indicates visual characteristics, Representing text features; The epidemic semantic enhancement detection module ESADM also includes the WorldDetect detection head, which achieves deep alignment between visual features and text semantics through a contrastive learning mechanism. The WorldDetect detection head uses the ContrastiveHead module to calculate the semantic matching degree between text and visual features using cosine similarity.

[0012] Preferably, the Forest Guardian Collaborative Optimization Strategy FGA-CE optimizes the learning rate by simulating three behaviors of UAVs in forest monitoring: regional hovering behavior, high-altitude overlooking behavior, and remote surveying behavior. The region hovering behavior involves local development around the current optimal solution, and the position update expression is: ; ; in, This indicates the optimal location found by the drone. Indicates the current location of the drone's search agent. , Represents the coefficient vector. This indicates that corresponding elements of two vectors are multiplied one by one. The high-altitude observation strategy simulates the vertical take-off and landing observation of a UAV, approaching the optimal solution via a spiral path. The position update expression is: ; ; in, This indicates the distance between the drone and the optimal observation point. Represents the spiral descent constant. Represents a random number. The magnitude of the vector is used to represent the Euclidean distance between the UAV's current position and its optimal position. Remote surveying strategies in Triggered by time, the drone is forced to randomly select an unmonitored area for global exploration, and a reference point is randomly selected. The expression is: ; The position update expression is: .

[0013] Preferably, the Forest Guardian Co-optimization Strategy FGA-CE introduces a multi-swarm co-evolutionary framework, dividing the total drone swarm N into multiple sub-swarms, each of which cruises independently. Establish an information sharing mechanism and conduct regular information exchange and elite migration: information is exchanged between different subgroups every fixed number of iterations, and high-quality detection information is passed from the best subgroup to other subgroups according to detection quality; a dynamic path optimization strategy selects the worst probe point in the receiving subgroup for replacement, while adding a small perturbation to the transmitted probe point, with an amplitude of [missing value]. Search scope; After each iteration, boundary checks are performed on all UAVs to ensure that the updated positions are within the monitoring space, and the fitness of all probe points is re-evaluated, updating the subgroup optimal solution and the global optimal solution. The objective function is constructed using the historical validation metric mAP50, and the learning rate is adjusted based on momentum prediction.

[0014] Compared with the prior art, the beneficial effects of the present invention are: The TRSSL-Net proposed in this invention provides supplementary semantic knowledge by integrating textual information input, significantly enhancing the ability to identify diseased tree features. Simultaneously, it constructs a data augmentation framework integrating multiple vegetation index algorithms to capture the spectral features of pine tree health status from different perspectives. The core of the model is the innovative W-KAN feature extraction module, which combines multi-scale decomposition of Discrete Wavelet Transform (DWT) with the KAN nonlinear activation mechanism to achieve excellent noise suppression and capture of complex nonlinear relationships. A Q-DGCRAB four-domain diffusion map convolutional residual attention module enhances information transfer and integration between regions, effectively solving the problems of dispersed diseased tree targets and large scale variations from high-altitude perspectives. To improve training efficiency, the Forest Guardian Collaborative Optimization Algorithm (FGA-CE) is designed, accelerating model convergence through a group collaboration mechanism. This overcomes the shortcomings of traditional convolutional neural networks in frequency domain filtering mechanisms and can effectively distinguish weak signals of pine wilt disease symptoms from complex environmental noise such as light changes, seasonal background differences, equipment noise, and atmospheric scattering in aerial images above 500 meters. The method provided in this application significantly improves the detection accuracy of pine wilt disease, enhances the robustness of the model under complex backgrounds and multi-scale targets, and supports real-time detection, making it suitable for large-scale forestry monitoring. Attached Figure Description

[0015] Figure 1 is a flowchart of the UAV-based pine wilt disease detection method based on multimodal features of the present invention. Figure 2 This is an example diagram illustrating the multimodal representation formed by combining images and text in this invention. Figure 3 In the diagram, (a) is the image data processing flowchart of the present invention, (b) is the structure diagram of ESADM, VTMF and Q-DGCRAB of the present invention, (c) is the structure diagram of MRA-KAN of the present invention, (d) is the structure diagram of FGA-CE of the present invention, and (a)-(d) are combined into the structure diagram of the text-remote sensing collaborative network model TRSSL-Net of the present invention. Figure 4 This is a comparison chart of the network visualization results in this embodiment; Figure 5 This is a comparison chart of the visualization results of the generalized dataset in this embodiment. Detailed Implementation

[0016] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0017] This embodiment provides a method for detecting pine wilt disease using a drone based on multimodal features, including the following steps: S1. Visible light images of pine forests are acquired through a drone high-altitude remote sensing platform. After being cut and screened, images of pine trees affected by pine wilt disease are selected. Image processing is performed to establish an image-text paired multimodal dataset. The multimodal dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The pine wilt disease image dataset used in this embodiment was taken by a CW-15 drone at a flight altitude of over 500 meters, with a ground sampling distance of 10cm (GSD). The sensor was a full-frame 61-megapixel sensor with a 35mm focal length lens. This dataset covers 1174.5 square kilometers of forest images in the wild. Due to the large coverage area of ​​the high-altitude drone images, the high-altitude drone images were segmented and filtered to select an image dataset of 785 images containing multiple diseased pine trees. Image processing includes the following steps: S101. Adjust the pine tree image of pine wilt disease in JPG format to 800×800 pixels and input it into the labelmg software. This is to prevent overfitting and speed up the training process. S102. Experts used their professional knowledge to manually label pine tree images of pine wilt disease in the dataset on a visualization platform. Each image was given a multidimensional parameter text description, including the distribution characteristics, morphological characteristics, and location information of the pine wilt disease. Irregular disease labels were created and saved in txt format, establishing an image-text paired multimodal dataset, such as... Figure 1 As shown, each image is finely labeled and annotated with text. The system records the spatial distribution patterns and morphological characteristics of diseased trees. These semantically rich text descriptions provide additional learning references for the neural network, effectively promoting the model's accurate capture of the significant features of diseased trees and significantly improving the accuracy and reliability of disease identification.

[0018] S2. The visible light image is preprocessed using Vegetation Index Threshold Mask Filter (VTMF) to obtain the spectral characteristics of the pine tree's health status and to generate candidate regions for pine trees affected by pine wilt disease. Vegetation index threshold mask filtering (VTMF) does not require modification of the network structure. Instead, it embeds vegetation index information as a multidimensional feature into the training data during the data preprocessing stage, enabling the network to adaptively learn and utilize these biospectral features. In the application of UAV-based pine wilt disease target detection, the Green Leaf Index (GLI) provides key technical support for solving the problem of complex background interference through its unique spectral response characteristics. Therefore, VTMF enhances the difference in spectral features between vegetated and non-vegetated areas through the Green Leaf Index (GLI), as expressed by: ; in, , , The values ​​represent the reflectance in the green, red, and blue bands, respectively. Healthy pine trees, due to their rich chlorophyll content, have significantly higher GLI values ​​than the surrounding vegetation. However, pine trees infected with pine wilt disease, due to the yellowing and wilting of their needles, have GLI values ​​that drop sharply to the range of 0.2-0.4. This difference in spectral response forms a clear distinguishing feature on the RGB spectral curve, which helps the model quickly locate potential diseased tree areas. It further integrates other complementary vegetation indices, including the Green-Red Vegetation Index (GRVI), the Visible Light Atmospheric Impedance Index (VARI), and the RGB vegetation index, to capture the spectral characteristics of vegetation health status and form a multidimensional feature space. The expression for the Green-Red Vegetation Index (GRVI) is: ; The expression for the visible light atmospheric impedance index (VARI) is: ; The expression for the RGB vegetation index is: ; These indices capture the spectral characteristics of vegetation health from different perspectives, forming a multidimensional feature space. This significantly improves the model's sensitivity to early symptoms of pine wilt disease. The labeled dataset covers diverse environmental factors, including yellow sand, roads, swamps, grasslands, buildings, valleys, and rice paddies. This diversity of environmental conditions not only ensures the robustness of the trained model and its applicability to the real world but also adds a lot of background noise interference. Furthermore, the low response of vegetation indices, especially the GLI index, to non-green objects can effectively help the model filter out non-vegetation targets, enabling accurate localization of diseased trees in complex backgrounds and thus reducing the false alarm rate.

[0019] S3, such as Figure 2 As shown, the image enhanced by VTMF is input into TRSSL-Net, a text-remote sensing collaborative network model based on the YOLO architecture; S301. To address the significant random noise in high-altitude UAV images, a multi-level denoising strategy based on wavelet transform is proposed. A W-KAN feature extraction module (based on the original resolution) is constructed at the P3 layer of the text-remote sensing collaborative network model TRSSL-Net. Based on signal-to-noise ratio analysis, the P3 layer retains sufficient detail while noise propagation has not yet accumulated to an uncontrollable level. Frequency domain feature maps of pine wilt disease are extracted by performing frequency domain feature decomposition, nonlinear modulation, and noise suppression on the VTMF-enhanced image using Discrete Wavelet Transform (DWT). This is because the characteristic signals of early symptoms of pine wilt disease are mainly concentrated in specific frequency bands, while environmental noise exhibits a more dispersed spectral distribution. Therefore, the denoising advantage of wavelet transform is demonstrated. Discrete wavelet transform achieves multi-resolution analysis by decomposing the signal into different frequency sub-bands. In the W-KAN module, the discrete wavelet transform (DWT) decomposes the image features enhanced by VTMF into low-frequency sub-band LL and high-frequency sub-bands LH, HL, and HH through the Haar wavelet basis function, thereby achieving frequency domain separation of signal and noise. Image feature transformation is performed using two-dimensional Haar wavelet basis functions, expressed as follows: ; in, Represents the mother wavelet function. Indicates the scale parameter. Indicates the translation parameter. The independent variable of the wavelet function is used to represent the time and spatial location of a signal or image; ; ; ; ; in, , These represent the coefficients of the low-pass and high-pass filters, respectively. These represent the row and column index positions of the output image, respectively. The term "alternate" represents the row and column offset indices of the filter during convolution on the input LL image. In this embodiment: ; ; Simple thresholding is insufficient for adaptive denoising. Therefore, this embodiment introduces the KAN attention mechanism to nonlinearly modulate the wavelet coefficients. KAN is based on a complex mapping relationship constructed using B-spline functions, enabling the model to learn the optimal filtering strategy under different noise modes. The B-spline basis functions are defined in a recursive form: ; in, In this embodiment, the spline order is represented. , Represents a node vector, with the following initial conditions: ; The mapping function expression used by KAN is: ; in, Represents the learnable coefficient; The feature map consists of basic linear transformations and spline transformations, and its expression is: ; in, This represents the basic activation function (SiLU). It is a B-spline basis function vector. and These are learnable weights. The KAN attention mechanism is used to nonlinearly modulate the coefficients of the four sub-bands to extract specific feature maps of pine wilt disease in the frequency domain. The expression is: ; in, This represents the feature map of the input image within a specific frequency band.

[0020] The W-KAN module achieves adaptive weighting of information in different frequency bands. Compared with traditional denoising methods, this data-driven adaptive denoising mechanism shows stronger environmental adaptability. Theoretically, the KAN mechanism can retain high-frequency components related to disease characteristics while effectively suppressing background noise, achieving an excellent balance between maintaining feature integrity and noise suppression. Pine wilt disease symptoms have specific characteristics in the frequency domain, and the W-KAN module effectively extracts these key features through frequency domain analysis and nonlinear feature enhancement. S302. The Q-DGCRAB four-domain diffusion graph convolutional residual attention module is used to establish the regional graph structure relationship. The frequency domain feature map of pine wilt disease is divided into four domains: upper left, upper right, lower left, and lower right. Each region is used as a node to construct the graph structure. In object detection tasks, feature representation faces the challenge of insufficient capture of spatial correlation. Traditional convolutional operations are limited by local receptive fields, making it difficult to establish dependencies between long-distance features. Although attention mechanisms partially solve this problem, their secondary computational complexity is inefficient on high-resolution feature maps, especially in dense small-object scenes, where local features cannot provide sufficient contextual information, leading to missed and false detections. To address this issue, this embodiment proposes a Quad-Region Diffusion Graph Convolutional ResidualAttention Block (Q-DGCRAB). Deployed at the front end of the detection head, Q-DGCRAB achieves global contextual modeling of features while maintaining computational efficiency. The Q-DGCRAB Quad-Region Diffusion Graph Convolutional ResidualAttention Block establishes the region graph structure relationship through the following steps: Frequency domain feature map of the input pine wilt disease conduct Grouped convolution dimensionality reduction yields ,in, The space representing the real number space has four dimensions: This indicates the number of samples input into the network at one time. This indicates the number of channels in the corresponding feature map. Represents the spatial height of the feature map. Indicates the spatial width of the feature map; Extracting representations from the top-left, top-right, bottom-left, and bottom-right domains to construct inter-region connectivity relationships, resulting in a set of region representations. The expression is: ; in, , , , These represent the region representations obtained by average pooling of the four regions: top left, top right, bottom left, and bottom right, respectively, preserving the spatial topological relationships of the features. Unlike traditional global attention, this module only interacts with the region representations, significantly reducing computational complexity. Simultaneously, it establishes long-range dependencies through message passing between regions. Following the decreasing spatial correlation law of visual features, a diffusion factor based on spatial distance is introduced to pass messages between regions, adaptively reducing the interaction strength between regions that are farther apart. The expression for the diffusion factor is: ; in, , These represent the region indexes, The message function for graph convolution is: ; in, Indicates attention weights, Represents the normalization factor. express Characteristics of the region; 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: ; in, Indicates the updated number The characteristic representation of each neighboring region Represents a node The neighborhood group, Represents aggregate functions; The expression for applying gated residual join is: ; ; 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; Using fixed weights Feature fusion with channel attention mechanism outputs an enhanced pine wilt disease feature map, expressed as: ; in, This indicates a batch normalization operation. express Convolution operation, This indicates the channel attention mechanism; The gated residual structure and channel attention mechanism in the DGCRAB module ensure the selectivity of information during transmission and enhance the model's response to key features. It is worth noting that Q-DGCRAB uses fixed fusion weights, which avoids the risk of overfitting caused by overfitting and improves the model's generalization ability in complex scenarios. In summary, Q-DGCRAB effectively solves the problem of long-range feature dependency modeling in object detection through a spatially aware graph convolutional structure, while maintaining computational efficiency. S303. The Epidemic Semantics-Augmented Detection Module (ESADM) is used to introduce a pre-trained BERT model text encoder to achieve cross-modal feature fusion between vision and text, providing semantic enhancement information. In object detection tasks, single visual modalities face the challenge of insufficient semantic understanding. Traditional convolutional neural networks rely solely on visual features for object recognition, lacking a deep understanding of the semantic attributes of object categories. In complex epidemic scenarios, false detections and false negatives are prone to occur. In particular, in the detection of epidemic-related objects, the visual similarity between different objects is high, and it is difficult to accurately distinguish them by relying solely on visual features, resulting in limited detection accuracy. The ESADM (Epidemic-Related Semantic Enhancement and Detection Module) loads a pre-trained BERT model using a parameter-freezing strategy to extract deep semantic representations from the text. This significantly improves the model's semantic perception capabilities while maintaining computational efficiency. After dimensionality reduction and nonlinear transformation using a two-layer fully connected dimensionality reduction network (768→256→128), the high-dimensional text features output by the BERT model are compressed into a visually compatible dimensional space. The contribution ratio of text and visual features is dynamically adjusted using a learnable parameter k. Text-visual fusion is then performed at layers P4 and P6 of the text-remote sensing collaborative network model TRSSL-Net, expressed as: ; in, This refers to adaptive fusion weights, ensuring that semantic information from the text enhances the visual feature representation at different levels of abstraction. Unlike traditional post-fusion methods, this strategy deeply embeds semantic information into the feature extraction process, enabling the model to obtain semantic guidance during the feature learning stage. This indicates text-visual fusion. Indicates visual characteristics, Representing text features; The text feature processing includes: first, extracting deep semantic representations of the text through a frozen BERT encoder, then performing feature dimensionality reduction and nonlinear transformation through two fully connected networks to finally obtain a text representation that matches the dimensions of the visual features; during the fusion process, a linear weighting strategy is adopted, and the contribution ratio of text and visual features is dynamically adjusted through a learnable parameter k. This design avoids complex attention calculations and maintains computational efficiency while ensuring the fusion effect. The Epidemic Semantic Enhancement Detection Module (ESADM) also includes the WorldDetect head, which is specifically designed to handle the fusion task of text embedding and object detection. It achieves deep alignment between visual features and text semantics through a contrastive learning mechanism. The WorldDetect head uses the ContrastiveHead module to calculate the semantic matching degree between text and visual features using cosine similarity, which enhances the model's ability to classify 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 and control.

[0021] S4. The Forest Guardian Algorithm with Collaborative Evolution (FGA-CE) is adopted for dynamic optimization of the learning rate of the UAV target detection network for pine wilt disease. Combining the global search capability of the forest patrol detection algorithm with a multi-population collaborative evolution strategy, the learning rate parameter can be adaptively adjusted according to validation metrics during training, effectively improving model performance. The core idea of ​​the Forest Guardian collaborative optimization algorithm originates from three typical behaviors of UAVs in forest monitoring: regional hovering, high-altitude observation, and remote surveying. We abstract these monitoring modes into mathematical models and introduce a multi-population collaborative evolution mechanism to enhance the algorithm's global search capability and convergence speed. 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. The Forest Guardian Collaborative Optimization Strategy FGA-CE optimizes the learning rate by simulating three behaviors of drones in forest monitoring: regional hovering behavior, high-altitude overlooking behavior, and remote surveying behavior. The region hovering behavior involves local development around the current optimal solution, and the position update expression is: ; ; in, This indicates the optimal location found by the drone. Indicates the current location of the drone's search agent. , Represents the coefficient vector. This indicates that corresponding elements of two vectors are multiplied one by one. ; Linearly decreasing from 2 to 0, and ,in and for Random numbers within the interval; The high-altitude observation strategy simulates the vertical take-off and landing observation of a UAV, approaching the optimal solution via a spiral path. The position update expression is: ; ; in, This indicates the distance between the drone and the optimal observation point. Represents the spiral descent constant. express Random numbers within the interval (Linearly decreasing from -1 to -2). The magnitude of the vector is used to represent the Euclidean distance between the UAV's current position and its optimal position. Remote surveying strategies in Triggered by time, the drone is forced to randomly select an unmonitored area for global exploration, and a reference point is randomly selected. The expression is: ; The position update expression is: ; The Forest Guardian Co-optimization Strategy FGA-CE introduces a multi-swarm co-evolutionary framework, dividing the total drone swarm N into multiple sub-swarms. In this embodiment, it is divided into 3 sub-swarms, each of which cruises independently. An information sharing mechanism is established to conduct regular information exchange and elite migration: information is exchanged between different subgroups every fixed number of iterations (5 times in this embodiment). Based on detection quality, high-quality detection information is transferred from the best subgroup to other subgroups; in this embodiment, the transfer rate is set to 0.2. A dynamic path optimization strategy selects the worst detection point in the receiving subgroup for replacement, while adding a small perturbation to the transferred detection points, with an amplitude of [missing value]. The search scope effectively enhances the diversity of monitoring and the overall detection capability; After each iteration, boundary checks are performed on all UAVs to ensure that the updated positions are within the monitoring space, and the fitness of all probe points is re-evaluated, updating the subgroup optimal solution and the global optimal solution. The objective function is constructed using the historical validation metric mAP50, and the learning rate is adjusted based on momentum prediction. The expression is as follows: ; in, This indicates the most recent verification metric. This represents the relative rate of change of the learning rate. This represents the momentum trend factor calculated based on historical data; it records the validation metric for each epoch, and performs learning rate optimization at fixed intervals (default 5 epochs), using a weighted historical trend with weights of [value missing]. This ensures that the learning rate decays over time. The algorithm is triggered after validation, balancing optimization effectiveness with computational cost. Simultaneously, the optimization range is limited to a preset minimum learning rate. and maximum learning rate To prevent extreme values ​​in the learning rate, FGA-CE achieved adaptive adjustment of the learning rate during the training of the YOLO11 detection network for pine wilt disease using a drone. This enhanced the model's ability to learn the characteristics of infected pine trees. The optimization process was achieved through parallel computing on the CPU side, avoiding the consumption of valuable GPU resources and providing effective technical support for the early detection of forest diseases.

[0022] S5. Output the location and classification results of pine wilt disease through the detection head.

[0023] Example 2 To improve the accuracy of experimental results and minimize the interference of external variables, all experiments in this embodiment were conducted in the same hardware and software environment. The hardware configuration used in the experiments was provided by the AutoDL platform, ensuring that all experiments were conducted on the same hardware basis and avoiding the impact of hardware differences. On the software side, a standardized development environment was adopted to ensure that the versions of the operating system and related software tools were consistent, thereby minimizing the interference of software version differences on the experimental results. The specific hardware specifications and software environment settings are shown in Table 1. Table 1

[0024] Five evaluation metrics were used to comprehensively evaluate the effectiveness of TRSSL-Net across various aspects. Subsequently, the effectiveness of each network was verified on an internally curated dataset to determine progress. Ablation experiments were further used to carefully examine the interdependencies of individual networks, thereby confirming their effectiveness. Finally, experiments were conducted on two additional pine wilt disease datasets to verify the generalization ability of TRSSL-Net. Overall, these experiments highlight the superior detection performance of the TRSSL-Net high-altitude UAV pine wilt disease target detection framework on datasets of varying sizes.

[0025] In this embodiment, the model's performance is evaluated using recall (R), mAP50, mAP50-95, FPS, and GFLOPs.

[0026] Recall (R) represents the proportion of correct positive predictions to all actual positives, expressed as: ; Where TP represents the number of positive samples that were correctly detected, and FN represents the number of positive samples that were not detected; The single-class mAP50 is calculated by taking the average accuracy at an IoU threshold of 0.5, comprehensively reflecting the model's detection capability for that class. Through the above steps and formulas, the model's performance in single-class object detection tasks can be clearly evaluated; the expression is: ; mAP50-95 is a comprehensive metric calculated by taking the average accuracy of multiple thresholds within an IoU threshold range of 0.5 to 0.95 (step size 0.05). The expression is: ; in, Indicates the IoU value. This represents the average accuracy for different IoU thresholds of t; Frames per second (FPS) is an important indicator of detection speed. In this embodiment, it represents the average number of images detected per second, and is expressed as: ; GFLOPs are used to measure model size and complexity.

[0027] Module validity experiment: 1. Experiment on the effectiveness of the vegetation index threshold mask filtering module; In this embodiment, the enhancement effects of four vegetation indices (GRVI, GLI, VARI, and RGBVI) on the object detection task were evaluated. To comprehensively explore the impact of vegetation indices on the performance of TRSSL-Net, a series of comparative experiments were designed to test the effects of single indices and different combination strategies. The experimental results are shown in Table 2. Table 2:

[0028] Experimental results show that among all tested vegetation indices, RGBVI (Red-Green-Blue Vegetation Index) performed the best, significantly improving target detection accuracy and achieving the best results across all evaluation metrics. Specifically, RGBVI achieved an mAP50 value of 0.875 and an mAP50-95 value of 0.416, while its R-value reached 0.824, comprehensively surpassing other vegetation indices. Although other vegetation indices such as GRVI, VARI, and GLI also showed some enhancement effects, their overall performance was not as good as RGBVI. This finding is consistent with the assumptions of this embodiment, indicating that in specific target detection tasks, the effectiveness of vegetation indices is highly dependent on their correlation with target features.

[0029] It is worth noting that multi-index combination experiments were also conducted, but the multi-index combination strategy failed to bring the expected performance improvement. This may be due to information redundancy or interference between different vegetation indices, which weakens the expression of effective features. This indicates that when designing feature enhancement modules, more attention should be paid to feature quality rather than quantity. Accurately selecting specific indices that are highly relevant to the target detection task may be more effective than blindly increasing feature dimensions.

[0030] Taking into account both model accuracy and efficiency, adding RGBVI alone provides the most cost-effective enhancement strategy, which can significantly improve detection performance (mAP50 reaches 0.875, mAP50-95 reaches 0.416) while maintaining relatively low computational overhead. This finding has important practical value for high-precision target detection in resource-constrained scenarios, and also provides valuable experimental evidence for subsequent research.

[0031] 2. The effectiveness of W-KAN In this embodiment, the effectiveness of various feature enhancement methods in the WVKnet object detection framework was evaluated, and a novel module integrating 'haar' wavelet transform and KAN attention mechanism was proposed and strategically placed in a specific stage of the model backbone network. The experimental results are shown in Table 3. Table 3

[0032] Experimental results show that, compared with the best-performing comparison method NonLocalMeans (mAP50 of 0.853), the method provided in this embodiment improves performance by 2.2 percentage points. This performance gain is mainly attributed to the effective decomposition capability of wavelet transform for multi-scale features. Combined with the adaptive feature fusion of the KAN attention mechanism, the model can simultaneously capture low-frequency global information and high-frequency detail information of the image. This advantage is particularly evident when dealing with complex scenes and small target detection.

[0033] It is worth noting that although the BilateralFilter achieved a good result of 0.409 in the mAP50-95 index, which is 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 those of the method provided in this embodiment, indicating that the module proposed in this embodiment has a greater advantage in overall detection performance. Although the traditional MedianFilter performs well in terms of R value (0.808), it has a significant gap with the method provided in this embodiment in terms of accuracy.

[0034] Furthermore, the method provided in this embodiment is more efficient than the computationally complex NonLocalMeans operation, which, although performing well on mAP50 (0.853), has a large computational overhead; SoftMaxPool, although achieving a result of 0.851 on mAP50, only achieves 0.38 on mAP50-95, indicating that its performance degrades significantly under more stringent evaluation criteria.

[0035] Comparative experiments revealed that different feature enhancement methods each 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 this embodiment achieves balanced and excellent performance in all dimensions, proving the effectiveness of combining wavelet transform with the KAN attention mechanism.

[0036] In summary, the method provided in this embodiment outperforms traditional filtering and feature enhancement techniques across all evaluation metrics, achieving the best performance with an R-value of 0.824, an mAP50 value of 0.875, and an mAP50-95 value of 0.416. The feature enhancement module proposed in this embodiment provides a significant performance improvement for the TRSSL-Net model (R-value 0.824, mAP50 value 0.875, mAP50-95 value 0.416) through efficient multi-scale feature extraction and enhancement capabilities, surpassing all traditional feature enhancement methods and representing a cost-effective network enhancement strategy.

[0037] 3. Experiment on the effectiveness of the four-domain diffusion graph convolutional residual attention module In this embodiment, the effectiveness of various graph convolution and attention mechanisms in the object detection framework was evaluated, and a feature enhancement module based on spatial graph convolution, SpatialGraphConv, was proposed. This module was strategically placed in the feature graph processing stage before the detection head. The experimental results are shown in Table 4. Table 4

[0038] Experimental results show that the Q-DGCRAB proposed in this embodiment outperforms traditional graph convolution and attention mechanisms on all evaluation metrics, achieving the best performance with an R-value of 0.824, a MAP50 value of 0.875, and a MAP50-95 value of 0.416. Compared with the best-performing comparative method Graph (MAP50 of 0.844), the method provided in this embodiment improves performance by 3.1 percentage points. This performance gain is mainly attributed to Q-DGCRAB's effective modeling ability of spatial relationships in feature maps. By constructing a spatial relationship graph of feature maps and applying graph convolution operations, the model can simultaneously capture local details and global contextual information. This advantage is particularly evident when dealing with complex scenes and dense object detection.

[0039] It is worth noting that although CBAM achieved a good result of 0.395 on the MAP50-95 index, close to the 0.416 of the method provided in this embodiment, its R value (0.744) is significantly lower than that of the method provided in this embodiment, indicating that Q-DGCRAB has a significant advantage in detection recall. Although the traditional ChebNetModule performed well on the R value (0.791), it had a significant gap with the method provided in this embodiment in terms of precision.

[0040] Furthermore, the method provided in this embodiment is more efficient than computationally complex Non-Local Networks operations, which, although performing well in terms of R value (0.79), have a large computational cost. Although SelfAttention has a simple structure, it performs poorly on all metrics, with an R value of only 0.72 and a MAP50 of 0.826, indicating that the simple attention mechanism cannot fully capture the complex relationships between features.

[0041] Comparative experiments revealed that different feature enhancement methods each have their own advantages: CBAM performs better in high-precision detection (MAP50-95), while Graph is more advantageous in medium-precision detection (MAP50); Q-DGCRAB provided in this embodiment achieves balanced and excellent performance in all dimensions, proving the effectiveness of spatial graph convolution in feature enhancement.

[0042] In summary, the Q-DGCRAB module proposed in this embodiment provides a significant performance improvement for object detection models through efficient spatial relationship modeling and graph convolution operations, surpassing all traditional graph convolution and attention mechanism methods. It is a cost-effective network enhancement strategy.

[0043] 4. Optimization algorithm effectiveness experiment Experiments were conducted to evaluate the effectiveness of the optimization algorithms within the TRSSL-Net detection framework. The proposed method was comprehensively compared with two mainstream intelligent optimization algorithms, PSO and WAM. To ensure the fairness and reliability of the results, all algorithms were embedded at the same stage of the backbone network, maintaining consistent training datasets, data augmentation strategies, loss weights, and hyperparameter settings. The same random seed was used for three independent replications, and the average value was taken. The experimental results are shown in Table 5. Table 5

[0044] The results show that the method proposed in this embodiment has achieved significant advantages in all key metrics. Specifically, compared with the best-performing comparison algorithm PSO (mAP50 = 0.853), the method proposed in this embodiment achieves a 2.2 percentage point improvement in mAP50; compared with WAM (mAP50-95 = 0.391), it improves by 2.5 percentage points in the more stringent mAP50-95 metric; in addition, the method proposed in this embodiment can still maintain higher accuracy in the high recall segment of the PR curve, indicating that the model's ability to locate difficult-to-detect targets is also enhanced.

[0045] The fundamental reason for the performance improvement lies in the fact that the method proposed in this embodiment simultaneously simulates and integrates three main search strategies: global exploration, local mining, and adaptive jumping. Furthermore, a momentum update mechanism is employed, enabling particles to fully utilize information from historical optimal solutions in the search space, avoiding oscillations and accelerating convergence. Meanwhile, a multi-population co-evolution strategy introduces the ideas of "cross-population migration" and "elite exchange," facilitating the rapid discovery of the global optimum while maintaining diversity. Overall, this mechanism effectively alleviates the pain points of traditional intelligent optimization, such as the tendency to get trapped in local optima and limited convergence speed, thereby significantly improving the overall accuracy and robustness of the target detection task.

[0046] 5. Ablation test To verify the effectiveness of the target detection-based method proposed in this embodiment, ablation experiments were conducted on the basic network. In this embodiment, the ESADM, W-KAN module, Q-DGCRAB, optimization algorithm (FGA-CE), and vegetation index threshold mask filter (VTMF) were added sequentially using the controlled variable method. Then, 16 ablation experiments were conducted on these 5 modules. The experimental results are shown in Table 6. Table 6

[0047] The comparison between rows 1 and 2 demonstrates that text modality integration effectively improves detection performance, increasing the R-value from 0.728 to 0.752 and the mAP50 from 0.815 to 0.831. Adding the W-KAN module further enhances the mAP50 to 0.84745 and the MAP50-95 to 0.38464, highlighting the importance of multi-scale feature decomposition. The introduction of Q-DGCRAB increases the recall R-value to 0.752 and the mAP50 to 0.8395, indicating that graph structure features are helpful in capturing spatial relationships between targets. Comparing lines 13 and 16, we can see that with ESADM and Q-DGCRAB included, adding VTMF improves the R-value from 0.76 to 0.778 and the mAP50 from 0.849 to 0.858, demonstrating the importance of VTMF in improving model performance, especially for vegetation target detection. When the W-KAN module, Q-DGCRAB, and FGA-CE are used in combination (line 13), the mAP50-95 reaches 0.404, showing the effect of multi-module collaborative work. After further adding VTMF (line 16), all indicators reach optimal levels.

[0048] In summary, compared with the basic network, the complete model (using all modules) provided in this embodiment improves the mAP50 from 0.815 to 0.875 (an improvement of 7.4%) and the R-value from 0.728 to 0.824 (an improvement of 13.2%). Although the GFlops increase (from 6.6 to 10.438), the performance improvement is significant. Multiple experimental results show that text modality integration, W-KAN module, four-domain diffusion graph convolutional residual attention module, forest guardian collaborative optimization algorithm, and vegetation index threshold mask filtering can all effectively improve the model accuracy. In particular, when these 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 this embodiment shows excellent performance in object detection tasks.

[0049] 6. Comparative Experiment On the same testing environment and test set, this embodiment was compared with several traditional and current state-of-the-art object detection methods. The performance of the model provided in this embodiment was further analyzed, and the test results are shown in Table 7. Table 7

[0050] In object detection models, the YOLO series has garnered significant attention due to its excellent balance between speed and accuracy. YOLOv3 initially introduced multi-scale feature prediction, but its recall was only 0.659 and MAP50 was 0.747, performing poorly in complex scenes. Subsequently, YOLOv5 and YOLOv8 achieved a MAP50 of 0.751, but their recalls of 0.688 and 0.656 respectively remained unsatisfactory. While YOLOv9t improved MAP50 to 0.760, its GFlops reached 8.5, resulting in substantial computational overhead. YOLOv10n optimized the model through an attention mechanism, increasing MAP50-95 to 0.436 and achieving a MAP50 of 0.776, demonstrating its effectiveness for different IoU levels. Better generalization ability below the threshold; In complex detectors, RT-DETR adopts a Transformer architecture and achieves a MAP50 of 0.802 and a recall of 0.768, but its GFlops of 86.5 and FPS of 28.42 limit its application in resource-constrained scenarios; PSS-Net achieves a MAP50 of 0.803 and a MAP50-95 of 0.437 through feature pyramid optimization, achieving a good balance between accuracy and speed; Notably, YOLOv11n performs best in the MAP50-95 metric, reaching 0.485, demonstrating its excellent detection quality.

[0051] Based on the experimental results, the model proposed in this embodiment outperforms existing detectors in most metrics, especially in recall, where it reaches 0.824, 5.6 percentage points higher than the closest performing RT-DETR; on MAP50, it reaches 0.875, 6 percentage points higher than YOLOv11n; although on MAP50-95, the model proposed in this embodiment (0.416) is lower than YOLOv11n's 0.485, it still outperforms most comparative models. The reasons why the model proposed in this embodiment achieves such excellent performance are as follows: (a) The model proposed in this embodiment is based on an improved modern detection framework, integrating multiple advanced technologies to enhance the model's accuracy and speed; (b) (c) Introducing a text modality integration mechanism enhances the model's ability to understand semantic information; (d) Using the W-KAN module for multi-scale feature decomposition effectively extracts information of different frequencies; (e) Spatial graph convolution (SGC) effectively captures the spatial relationships between features, enhancing the model's ability to understand complex scenes; (f) The FGA-CE optimization algorithm improves the model training process, making features more discriminative; (g) Vegetation index enhancement (VGE) is specifically optimized for vegetation target detection, improving the model's applicability in agricultural scenarios. Although the model proposed in this embodiment slightly outperforms some models (10.438) in GFlops, it still maintains 54.26 FPS, fully meeting the requirements of real-time detection. In particular, compared with RT-DETR (86.5 GFlops, 28.42 FPS), the model proposed in this embodiment has a significant advantage in computational efficiency while maintaining higher accuracy. In summary, experimental results show that the comprehensive feature enhancement framework proposed in this embodiment has superior performance in target detection tasks, especially in application scenarios requiring high recall and high accuracy.

[0052] 7. Visual Analysis Figure 3 The visualization comparison results of the model proposed in this embodiment with 10 other mainstream target detectors in aerial image detection are shown. The performance difference between the model in this embodiment and other detectors when handling complex scenes can be clearly seen from the figure. In the first column of images, this embodiment selects a scene with complex buildings and trees to observe the detection results. Compared with YOLOv11n and PSSNET, the model proposed in this embodiment can not only detect small targets more accurately, but also better distinguish the boundaries of trees and buildings. Especially in the shadow area, YOLOv3 and YOLOv5 often mistake shadows for targets or completely ignore targets under shadows, while the model proposed in this embodiment can correctly identify them. This is mainly due to the Spatial Graph Convolution (SGC) module in the model proposed in this embodiment, which can capture the spatial relationship between targets and combine text modal information for more accurate recognition.

[0053] In the third and fourth columns of images, this embodiment selected a building scene with a complex roof structure. It can be clearly seen that YOLOv8 and YOLOv10n often misidentify red roofs as vegetation or trees, and the confidence level is also low. In contrast, the model proposed in this embodiment, through the Vegetation Index Enhancement (VGE) module, can more accurately distinguish between man-made structures and natural vegetation, significantly reducing the false detection rate. Even under uneven lighting conditions, the model proposed in this embodiment maintains high detection accuracy.

[0054] In the fifth and sixth columns, this embodiment demonstrates the detection results for mixed water and building scenes. Although RT-DETR performs well on large targets, its detection accuracy at the boundary between water and vegetation is not as good as the model proposed in this embodiment. The model proposed in this embodiment uses the W-KAN module for multi-scale feature decomposition, which can better handle the transition between different texture regions and improve the completeness and accuracy of detection.

[0055] Furthermore, from the overall detection results, the model proposed in this embodiment achieved higher confidence and more accurate bounding box localization in all test scenarios. For example, in the detection of small buildings in the seventh column, YOLOv6 and YOLOv9t showed obvious missed detections, while the model proposed in this embodiment detected almost all targets. This is attributed to the FGA-CE optimization algorithm improving the model's feature representation ability, enabling it to handle more complex scene changes.

[0056] Overall, the visualization results clearly demonstrate that the comprehensive feature enhancement framework proposed in this embodiment has significant advantages in aerial image target detection tasks, particularly in handling shadowed regions, distinguishing similar textures (such as red roofs and vegetation), and detecting small targets against complex backgrounds. These improvements not only enhance detection accuracy but also optimize the false negative and false positive issues commonly found in existing models.

[0057] 8. Generalization experiment on PDT public dataset To verify the generalization ability of the model proposed in this embodiment, this embodiment selected the PDT (Pests and Diseases Tree) dataset, jointly established by the Shandong Computer Science Center and Peking University, for experiments. The PDT dataset is the first high-precision tree pest and disease target detection dataset based on UAVs, aiming to solve the problem of the lack of professional datasets in the development of pest and disease detection models in the agricultural field. The dataset consists of high-definition images collected by UAVs at an altitude of about 200 meters, containing various tree pest and disease categories, with high-precision annotation and complex real-world scene characteristics. At the same time, the dataset is described with text and serves as the TRSSL-Net text dataset.

[0058] The complete model proposed in this embodiment was comprehensively compared with current mainstream object detectors on the PDT dataset. The experimental results are shown in Table 8. Table 8

[0059] The experimental data show that the model proposed in this embodiment achieved the best performance across all evaluation metrics, with an R-value of 0.845, a MAP50 of 0.921, and a MAP50-95 of 0.604. Compared to the better-performing YOLOv5 (R=0.825, MAP50=0.905) and RT-DETR (R=0.815, MAP50=0.909), the model proposed in this embodiment improved the MAP50 performance by 2% and 1.2%, respectively. Furthermore, under the more stringent MAP50-95 evaluation standard, the model proposed in this embodiment (0.604) improved by 2.6% and 2% compared to YOLOv5 (0.578) and PSS-NET (0.584), respectively, demonstrating stronger generalization ability and detection accuracy.

[0060] From the visualization results of the PDT dataset, as shown in the comparative experiments in the three columns on the left of Table 7, the model (OURS) proposed in this embodiment shows a significant advantage in detection performance. In the complex soil background scene in the first column, the model proposed in this embodiment can accurately identify and locate more pest and disease targets, with a significantly larger number of detection boxes than other methods, and the localization is more accurate. In the dense green vegetation scene in the second and third columns, the model proposed in this embodiment is particularly outstanding in its ability to detect blue lesions, and can accurately identify scattered pest and disease areas against a high-density green background. Compared with other mainstream detectors, such as YOLOv11, PSS-NET, RT-DETR, and YOLOv10, different degrees of missed detection occurred in the same scene. In particular, earlier versions of methods such as YOLOv8, YOLOv6, YOLOv5, and YOLOv3 showed significantly insufficient detection performance in complex backgrounds.

[0061] Further analysis of the visualization results reveals that the model proposed in this embodiment not only has a larger number of detection boxes but also more accurate positioning and boundaries that better fit the target outline. Particularly when dealing with small and edge targets, the model proposed in this embodiment demonstrates superior performance, which is particularly evident in the green vegetation scenes in the second and third columns—other methods often only detect relatively obvious large lesions, while the model proposed in this embodiment can identify more small pests and diseases. The visualization results intuitively verify the advantages of quantitative indicators and demonstrate the detection quality of the model proposed in this embodiment in practical application scenarios.

[0062] It is worth noting that the model proposed in this embodiment outperforms the model on the PDT dataset. This is mainly because the drone shooting altitude of the PDT dataset is about 200 meters, while that of the self-built dataset is about 500 meters. The lower flight altitude makes the target size in the PDT dataset relatively larger and the texture features more obvious, which is beneficial for the model to perform feature extraction and target recognition. This phenomenon is also clearly confirmed in the visualization results. In the detection results of the PDT dataset, the boundaries of pests and diseases are clearer and the color contrast is more obvious (such as blue spots on a green background). The confidence of the detection boxes is generally higher, which reflects the model's excellent ability to handle large-sized, high-contrast targets. This phenomenon also verifies the advantages of the wavelet KAN module and spatial graph convolution in the model proposed in this embodiment in feature processing at different scales. When the target features are more obvious, these modules can perform their functions more fully and extract richer feature information.

[0063] The generalization experiments described above demonstrate that the feature enhancement framework proposed in this embodiment exhibits excellent stability and accuracy when applied to public datasets, particularly demonstrating good adaptability when processing images captured by UAVs at different altitudes. Visual comparison results further corroborate this conclusion. Figure 4 It is evident that the model proposed in this embodiment maintains stable detection performance in scenarios of varying complexity (from soil backgrounds to dense vegetation). The quality and quantity of detection boxes are significantly better than those of the comparative methods, especially when dealing with small targets in complex backgrounds. It is worth noting that this advantage is particularly evident in the MAP50-95 metric, which shows a significant improvement compared to similar models, proving that the model proposed in this embodiment can maintain high performance under different IoU thresholds.

[0064] 9. Generalization experiment of public dataset for pine wilt disease To verify the generalization ability of the model proposed in this embodiment, the public dataset of pine wilt disease was selected for the experiment. This dataset is a standardized dataset specifically established for the detection and research of pine wilt disease. It contains images of pine trees in different regions and at different infection stages, providing an important benchmark for the evaluation of pine wilt disease detection models. In addition, a text description of the dataset was made as the text modal input of TRSSL-Net.

[0065] The complete model of this embodiment was comprehensively compared with current mainstream target detectors on a public dataset of pine wilt disease. The experimental results are shown in Table 9. Table 9

[0066] The model proposed in this embodiment achieved excellent results with 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 improved the MAP50 index by approximately 3.6% and 3.1%, respectively. At the same time, under the more stringent MAP50-95 evaluation standard, the model proposed in this embodiment (0.395) also maintained a competitive advantage over advanced methods such as YOLOv9 (0.399) and YOLOv8 (0.392), demonstrating stronger generalization ability and detection accuracy.

[0067] Visualization results from the public dataset on pine wilt disease, such as Figure 4 In the comparative experiments shown in the three columns on the right, the model (OURS) proposed in this embodiment performed excellently in complex forest environments: In the mixed vegetation scene in the first column, the model proposed in this embodiment could accurately identify the pine wilt nematode disease area, with precise detection box positioning and an appropriate number of detection boxes, 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 showed good environmental adaptability and could accurately identify disease characteristics against a green background, especially for early disease symptoms with subtle color changes.

[0068] Visual comparisons with other mainstream detectors show that YOLOv11, PSS-NET, RT-DETR, and other methods exhibit varying degrees of detection instability in the same scenarios. Particularly noteworthy is that when dealing with complex forest backgrounds, most of the comparison methods either produce false detections (e.g., rt-detr, YOLOv11n, YOLOv8, YOLOv6, YOLOv3) or have insufficiently accurate bounding box localization (e.g., pssnet). In contrast, the model proposed in this embodiment maintains stable detection performance in all test scenarios. The bounding boxes are not only reasonably numerous but also exhibit high boundary fit, fully demonstrating the model's professionalism and reliability in the pine wilt disease detection task.

[0069] It is worth noting that the model proposed in this embodiment performs differently on public datasets than on self-built datasets. This is mainly due to the different image acquisition conditions of the two datasets, including factors such as drone flight altitude, shooting angle, and lighting environment. From the visualization results, it can be observed that the disease features in the pine wilt disease dataset are relatively more subtle and have lower color contrast, which places higher demands on the model's feature extraction capabilities. These differences result in variations in the size and clarity of target features, thus affecting the model's detection performance. However, through visualization comparison, it can be seen that the W-KAN module and the four-domain diffusion map convolutional residual attention module (Q-DGCRAB) proposed in this embodiment can effectively handle these variations and maintain stable detection performance under different conditions, which is particularly evident in the low-contrast scene of the third column.

[0070] The generalization experiments described above demonstrate that the feature enhancement framework proposed in this embodiment possesses excellent stability and accuracy, exhibiting good adaptability when applied to public datasets. The visualization comparison results further validate this conclusion. Regardless of complex forest environments or different lighting conditions, the model proposed in this embodiment maintains consistent detection quality, with the precision and recall of the detection boxes significantly outperforming the comparative methods. The significant advantage in the MAP50-95 metric proves that the model proposed in this embodiment maintains high performance under different IoU thresholds, providing reliable technical support for the early detection and control of pine wilt disease.

[0071] This embodiment integrates five innovative modules: Vegetation Index Threshold Mask Filtering (VTMF), W-KAN module, Four-Domain Diffusion Graph Convolutional Residual Attention Module (Q-DGCRAB), Forest Guardian Collaborative Optimization Algorithm (FGA-CE), and Epidemic Semantic Enhancement Detection Module (ESADM). These modules have been carefully designed and optimized to form a complete technical system that can work together and complement each other to effectively improve the accuracy and stability of pine wilt disease detection.

[0072] The method provided in this embodiment places particular emphasis on multi-scale feature extraction, spatial correlation modeling, and training process stabilization. It not only focuses on improving the performance of a single detection but also constructs a comprehensive solution to address multiple technical challenges in pine wilt disease detection. Through rigorous validation experiments on self-built datasets and public datasets, the method provided in this embodiment demonstrates excellent performance and generalization ability.

[0073] The advantages of this framework can be summarized as follows: (1) Vegetation index threshold mask filtering (VTMF) effectively solves the problem of interference from complex forest backgrounds. Through a specially designed vegetation feature extraction and enhancement mechanism, it significantly improves the ability to distinguish between diseased trees and healthy trees and other vegetation types. It can maintain stable performance even in complex environments such as terrain undulations, shadow areas and seasonal changes; (2) The W-KAN module effectively preserves high-frequency texture details in high-altitude remote sensing images through wave frequency domain decomposition and reconstruction technology. It solves the problem of high-frequency information loss in the downsampling process of traditional convolutional networks, enabling the model to capture key features such as tree crown edges and shadow textures more accurately, and improving the accuracy of diseased tree identification; (3) The four-domain diffusion map convolutional residual attention module (Q-DGCRAB) breaks through the limitations of the receptive field of traditional convolution and establishes a correlation mechanism between distant pixels in the image. This enables the model to effectively grasp the spatial dependencies in wide-area images of UAVs, especially to gain a more comprehensive understanding and analysis of the continuous distribution of vegetation patches and the spatial structure of the ecosystem; (4) The Forest Guardian Collaborative Optimization Algorithm (FGA-CE) proposes an adaptive adjustment strategy to address the instability problem in training UAV image detection models. This effectively alleviates the problems of slow convergence speed and oscillation of performance indicators, enabling the model to avoid getting trapped in local suboptimal solutions in complex scenarios and achieve a more stable and efficient training process. (5) The Epidemic Semantic Enhancement Detection Module (ESADM) provides semantic-level supplementary information for image features, enriches the model's understanding of disease features, and further improves detection performance and model robustness.

[0074] Through a series of ablation and comparative experiments, the effectiveness of each module and its synergistic effect were verified. Experimental results on the self-built dataset and the public pine wilt disease dataset show that the model provided in this embodiment is significantly better than the existing mainstream detection models in key indicators such as MAP50 and MAP50-95. It is particularly noteworthy that the method provided in this embodiment performs well on images taken at different altitudes, demonstrating good adaptability and generalization ability.

[0075] The multimodal feature enhancement framework proposed in this embodiment provides a systematic solution for UAV-based pine wilt disease detection. It not only effectively addresses current technical challenges but also provides new ideas and methods for further research in the field of forest pest and disease monitoring. These advancements will help improve the early detection capabilities of forest pests and diseases and provide important technical support for forest resource protection and sustainable management.

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

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 diffusion factor expression 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 gating 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 feature 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 sub-population team optimal solution 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