Method, device and equipment for detecting tea green leafhopper and storage medium

Abstract

The present application relates to the technical field of pest detection, and particularly relates to a method and device for detecting tea green leafhoppers, equipment and a storage medium. The method first constructs a vegetation index feature vector based on an acquired five-band image, performs spectral feature coding processing on the vegetation index feature vector to obtain a spectral hint embedding vector, performs visual coding processing on an acquired RGB image to obtain a visual block feature sequence, performs multi-head cross-attention cross-modal comparison processing on the spectral hint embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map, determines a plurality of damage areas based on the cross-modal fusion feature map, acquires corresponding scan images, inputs each scan image into a pre-trained pest detection model to perform tea green leafhopper detection processing, obtains the bounding box coordinates of each pest individual, and determines the pest density value per unit area corresponding to each damage area based on all the bounding box coordinates, so as to improve the positioning accuracy of pests.

Description

Technical Field

[0001] This invention relates to the field of pest detection technology, and in particular to a method, apparatus, equipment and storage medium for detecting the tea green leafhopper. Background Technology

[0002] The tea green leafhopper is a prevalent pest in Chinese tea gardens. Both adults and nymphs suck sap from tender shoots, causing buds and leaves to scorch and curl, consistently hindering tea yield and quality. In major tea-producing areas across China, the annual affected area exceeds 20 million mu (approximately 1.3 million hectares), resulting in significant economic losses. Traditional methods combining manual inspections with chemical pesticide control suffer from low inspection efficiency, delayed pest response, and excessive pesticide residues, making them unsuitable for the precision plant protection needs of large-scale tea gardens. While drones and multispectral remote sensing technologies are increasingly being applied to crop pest monitoring, multispectral imaging can capture red-edge and near-infrared band information invisible to the human eye. By leveraging vegetation indices to deduce crop physiological stress states, it provides new insights for early pest identification. However, existing monitoring solutions still face multiple technical limitations when implemented in tea garden settings.

[0003] Current mainstream monitoring solutions either rely solely on multispectral data for macroscopic disease grading or simply on visible light images for target detection, resulting in a disconnect between spectral anomaly identification and precise spatial localization. Traditional target detection networks use sequential decoding logic to process multi-regional pest targets, with inference time increasing proportionally with the number of targets. This makes them unsuitable for the real-time computing requirements of UAV-based edge computing and insufficient for localizing sub-pixel-level small pest areas in tea gardens. Furthermore, existing monitoring systems largely depend on full-process manual intervention, with the data acquisition and analysis process operating in a one-way open-loop manner. UAVs cannot autonomously plan subsequent tasks such as fine scanning and retesting based on the identification results. They also lack a complete business loop encompassing pest detection, spatial localization, risk assessment, and data iteration optimization. This leads to poor identification stability in complex tea garden environments such as hilly terraces and variable lighting conditions. Both equipment and algorithms struggle to support the implementation of autonomous and intelligent plant protection in tea gardens. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, device, equipment and storage medium for detecting tea green leafhopper, which aims to solve the problems of spectral and visual modal fragmentation and low insect positioning accuracy, improve the pest positioning accuracy and quantify the regional pest density.

[0005] The first aspect of this invention provides a method for detecting the tea green leafhopper, comprising: acquiring a five-band image and an RGB image; constructing a vegetation index feature vector based on the five-band image; performing spectral feature encoding on the vegetation index feature vector to obtain a spectral cue embedding vector; performing visual encoding on the RGB image to obtain a visual block feature sequence; performing multi-head cross-attention cross-modal comparison processing on the spectral cue embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map; determining multiple infestation areas based on the cross-modal fusion feature map and acquiring a scanned image corresponding to each infestation area; inputting each scanned image into a pre-trained pest detection model for tea green leafhopper detection processing to obtain the bounding box coordinates of each individual pest, and determining the pest density per unit area corresponding to each infestation area based on all the bounding box coordinates.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the step of constructing a vegetation index feature vector based on the five-band image includes: obtaining the five-band reflectance of the five-band image; calculating a vegetation index based on the five-band reflectance; and concatenating the five-band reflectance with the vegetation index to obtain the vegetation index feature vector.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the step of performing spectral feature encoding processing on the vegetation index feature vector to obtain a spectral cue embedding vector includes: obtaining a pre-trained spectral cue encoder, the spectral cue encoder including a first fully connected layer, a second fully connected layer, a third fully connected layer, and a linear projection layer connected in sequence; inputting the vegetation index feature vector into the first fully connected layer for nonlinear feature extraction processing to obtain a first multidimensional intermediate feature vector; inputting the first multidimensional intermediate feature vector into the second fully connected layer for nonlinear feature extraction processing to obtain a second multidimensional intermediate feature vector; inputting the second multidimensional intermediate feature vector into the third fully connected layer for nonlinear feature extraction processing to obtain a spectral feature vector; and inputting the spectral feature vector into the linear projection layer for linear up-dimensional mapping processing to obtain the spectral cue embedding vector.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the step of visually encoding the RGB image to obtain a visual block feature sequence includes: segmenting the RGB image using a uniform block algorithm to obtain multiple image block units of the same size; inputting each image block unit into a pre-trained visual encoder for visual encoding to obtain multiple image feature vectors; obtaining a preset image spatial arrangement rule; and concatenating and integrating all the image feature vectors based on the image spatial arrangement rule to obtain the visual block feature sequence.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the step of performing multi-head cross-attention cross-modal comparison processing on the spectral cue embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map includes: using the spectral cue embedding vector as a query vector, and using each image feature vector in the visual block feature sequence as a key vector and a value vector, respectively; performing attention calculation on each key vector based on the query vector to obtain multiple attention scores, and normalizing all the attention scores to obtain an attention weight matrix; performing multi-branch weighted integration operation processing on the attention weight matrix and all the value vectors to obtain multiple single-head attention output features; obtaining a preset aggregation rule, and performing multi-head aggregation and splicing processing on the multiple single-head attention output features based on the aggregation rule to obtain the cross-modal fusion feature map.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, determining multiple hazardous regions based on the cross-modal fusion feature map includes: obtaining a vegetation spectral disease feature set for each pixel in the cross-modal fusion feature map; calculating an adaptive weight for each feature in the vegetation spectral disease feature set using a random forest algorithm; calculating a CHI value corresponding to each pixel based on the vegetation spectral disease feature set and the corresponding adaptive weight; obtaining a preset CHI threshold; comparing each CHI value with the CHI threshold; identifying pixels with CHI values ​​greater than the CHI threshold as hazardous pixels; and performing connected component aggregation processing on all hazardous pixels using a preset neighborhood determination rule to obtain multiple hazardous regions.

[0011] Optionally, in the sixth implementation of the first aspect of the present invention, after determining the unit area pest density value corresponding to each of the harmful areas based on all the bounding box coordinates, the method further includes: obtaining multiple historical CHI values, constructing a historical CHI index time series based on the multiple historical CHI values, inputting the historical CHI index time series into a pre-trained logistic growth model to predict future pest population changes, and obtaining a pest population change curve; calculating a comprehensive risk score based on the pest population change curve, and determining an early warning level based on the comprehensive risk score; and determining a prevention and control response strategy based on the early warning level.

[0012] A second aspect of the present invention provides a tea green leafhopper detection device, comprising: a feature vector construction module for acquiring five-band images and RGB images, and constructing a vegetation index feature vector based on the five-band images; a spectral feature encoding module for performing spectral feature encoding processing on the vegetation index feature vector to obtain a spectral cue embedding vector; a visual encoding module for performing visual encoding processing on the RGB images to obtain a visual block feature sequence; a cross-modal comparison module for performing multi-head cross-attention cross-modal comparison processing on the spectral cue embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map; a hazard area determination module for determining multiple hazard areas based on the cross-modal fusion feature map and acquiring a scanned image corresponding to each hazard area; and a pest detection module for inputting each scanned image into a pre-trained pest detection model for tea green leafhopper detection processing, obtaining the bounding box coordinates of each individual pest, and determining the pest density value per unit area corresponding to each hazard area based on all the bounding box coordinates.

[0013] A third aspect of the present invention provides a tea green leafhopper detection device, the tea green leafhopper detection device comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the tea green leafhopper detection device to perform the various steps of the tea green leafhopper detection method described in any of the preceding claims.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the tea green leafhopper detection method described in any of the preceding claims.

[0015] In the technical solution of this invention, five-band images and RGB images are first acquired. A vegetation index feature vector is constructed based on the five-band images. The vegetation index feature vector is then subjected to spectral feature encoding to obtain a spectral cue embedding vector. The RGB images are then subjected to visual encoding to obtain a visual block feature sequence. The spectral cue embedding vector and the visual block feature sequence are subjected to multi-head cross-attention cross-modal comparison processing to obtain a cross-modal fusion feature map. Based on the cross-modal fusion feature map, multiple hazard areas are determined, and a scan image corresponding to each hazard area is acquired. Each scan image is input into a pre-trained pest detection model for tea green leafhopper detection processing to obtain the bounding box coordinates of each individual pest. Based on all bounding box coordinates, the pest density value per unit area corresponding to each hazard area is determined. This approach aims to solve the problems of spectral and visual modal separation and low insect localization accuracy, thereby improving pest localization accuracy and quantifying regional pest density. Attached Figure Description

[0016] Figure 1A flowchart of the tea green leafhopper detection method provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the autonomous detection trajectory in a tea garden simulation environment provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the pest quantity change curve provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the tea green leafhopper detection device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the tea green leafhopper detection device provided in an embodiment of the present invention. Detailed Implementation

[0017] This invention provides a method, apparatus, device, and storage medium for detecting the tea green leafhopper. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the tea green leafhopper detection method of the present invention includes: 101. Acquire five-band images and RGB images, and construct vegetation index feature vectors based on the five-band images; In this embodiment, a drone equipped with a five-band imaging device (such as a multispectral camera) completes an airspace inspection of the tea garden along a preset route. The flight altitude can be set to 80 meters and the travel speed can be set to 5 m / s. The imaging device outputs five-band images at a preset acquisition frequency (such as 1 Hz). These images consist of five standard bands: blue, green, red, red-edge, and near-infrared. The drone simultaneously records geographic coordinates and flight attitude data using an onboard positioning and attitude sensing module. During the operation, multiple vegetation indices are generated by calculating the reflectance values ​​of each band. The RGB image is visible light three-channel imaging data, which can be synchronously acquired by the drone's visible light imaging unit (such as an RGB camera). The reflectance values ​​of each band corresponding to the five-band image are extracted. Based on the reflectance values ​​of each band, multiple sets of vegetation indices are calculated. The reflectance data of all bands and the calculated vegetation indices are input into a pre-trained spectral cue encoder to complete the feature dimension splicing, and finally a 9-dimensional vegetation index feature vector is generated. This vector integrates the original spectral reflectance information and vegetation physiological stress characterization indicators, and completely preserves the subtle spectral changes caused by pests affecting the tea tree canopy. Based on this, subsequent cross-modal feature fusion can give full play to the disease prior role of multi-band spectral data, make up for the deficiency that a single visible light image cannot capture the physiological damage inside the vegetation. After multi-dimensional spectral feature fusion, the recognition sensitivity of pest-stressed areas in tea gardens can be improved, and input features with complete dimensions and information can be provided for subsequent spectral and visual feature association matching.

[0019] 102. Perform spectral feature encoding on the vegetation index feature vector to obtain a spectral cue embedding vector; In this embodiment, spectral feature encoding processing relies on a pre-trained spectral cue encoder. This encoder consists of a first fully connected layer, a second fully connected layer, a third fully connected layer, and a linear projection layer, all connected in series. The input data is a 9-dimensional vegetation index feature vector formed by integrating five-band reflectance and four types of vegetation indices. The vector is first fed into the first fully connected layer for linear transformation and then activated by ReLU to introduce nonlinear expression, outputting a first multidimensional intermediate feature vector with a dimension of 64. This first multidimensional intermediate feature vector is then fed into the second fully connected layer for secondary nonlinear feature extraction, generating a second multidimensional intermediate feature vector with a dimension of 32. This second multidimensional intermediate feature vector is then fed into the third fully connected layer for deep feature compression and nonlinear mapping, resulting in a spectral feature vector with a dimension of 16. Finally, the spectral feature vector is... The data is fed into a linear projection layer to perform a linear upscaling operation without activation functions, outputting a spectral cue embedding vector of dimension 768. This vector is a unified dimensional representation of the spectral stress information of tea tree pests, and can maintain dimensional alignment with the visual features output by the visual encoding. The entire hierarchical encoding structure first mines the physiological stress correlation features hidden within the original 9-dimensional spectral data through a three-layer fully connected network, and then relies on the linear projection layer to complete the dimensional unification adaptation. This not only fully mines the subtle pest response signals within the low-dimensional spectral features, but also achieves dimensional matching between spectral modality and visual modality features, eliminating the dimensional barriers that exist when interacting with cross-modal features. This provides a spectral prior representation with higher adaptability and more complete feature information retention for subsequent multi-head cross-attention cross-modal comparison, effectively improving the matching accuracy between spectral anomaly regions and the spatial location of visual images.

[0020] 103. Perform visual encoding processing on the RGB image to obtain a visual block feature sequence; In this embodiment, the visual encoding process first uses a uniform block partitioning algorithm to spatially divide the RGB image, generating several image block units of uniform size. All image block units are sequentially fed into a pre-trained visual encoder to perform feature extraction operations. After encoding, a single image block outputs an image feature vector with a dimension of 768. All image feature vectors are sequentially concatenated and integrated according to a preset image spatial arrangement rule, ultimately forming a visual block feature sequence. The sequence contains multiple image feature vectors, each with a dimension of 768. This sequence completely preserves the global spatial texture and visual representation information of pests in the RGB image, and can maintain a unified feature dimension with the spectral cue embedding vector. The hierarchical block extraction encoding method decomposes the complete image into independent spatial units for feature mapping, reducing the size of a single image. Figure 1The computational cost of one-time encoding is offset by the fact that the serialized features generated after block division can be adapted to the multi-head cross-attention computing mechanism, enabling precise correlation and matching between region-by-region spectral priors and visual details. This fully opens up the interaction channel between spectral stress information and image spatial visual information, eliminates the data adaptation barrier between the two modalities from the feature level, and improves the spatial positioning accuracy of pest-damaged areas during subsequent cross-modal comparisons.

[0021] 104. Perform multi-head cross-attention cross-modal comparison processing on the spectral cue embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map; In this embodiment, the multi-head cross-attention cross-modal comparison processing uses the spectral cue embedding vector as the query vector. All image feature vectors within the visual block feature sequence simultaneously serve as key vectors and value vectors. In the attention calculation stage, the attention score is first calculated based on the query vector and each key vector. The calculation process introduces the square root of the dimension scaling factor 768 to constrain the numerical range. Then, an attention weight matrix is ​​generated through normalization. Each element in the matrix quantifies the correlation between the spectral semantic components and the features of each spatial block of the image. By using the weight matrix and all value vectors to perform a multi-branch weighted summation operation, multiple single-head attention output features can be obtained. According to the preset aggregation rules, all single-head output features are spliced ​​and linearly projected to fuse, finally generating cross-modal attention. The fused feature map achieves pixel-level alignment between spectral stress semantics and spatial visual features of the image. The spatial region within the map that matches the high-hazard comprehensive index has a higher feature response amplitude, which can provide accurate prior constraints on the target region for the parallel box decoder. The entire calculation process uses a multi-head mechanism to simultaneously mine multi-dimensional cross-modal correlation features, relies on spectral priors to enhance the feature response of the pest region, breaks the limitation of modal independence between spectral data and visible light image, and significantly reduces the interference of irrelevant background on pest localization. At the same time, it relies on block-sequential visual features to adapt to parallel computing logic, reduces the computing power overhead brought by global retrieval of a single large image, ensures the real-time inference of the airborne edge, and completely constructs a stable cross-modal mapping path from spectral anomalies to visual target regions.

[0022] 105. Based on the cross-modal fusion feature map, determine multiple hazardous areas and obtain a scan image corresponding to each hazardous area; In this embodiment, vegetation spectral disease feature sets corresponding to all pixels are extracted from the cross-modal fusion feature map. The adaptive weights of each indicator within the feature set are calculated using the random forest algorithm. CHI values ​​are calculated pixel-by-pixel based on the feature parameters and corresponding weights. The calculated CHI values ​​are compared and filtered against a preset CHI threshold. Pixels with CHI values ​​lower than or equal to the CHI threshold are identified as non-stressed blocks and maintain a coarse-scan operation mode. Pixels with values ​​greater than the CHI threshold are uniformly marked as hazardous pixels. Connected component aggregation is performed on spatially adjacent hazardous pixels to obtain multiple contiguous hazardous areas. A parallel box decoder architecture is used to perform a one-time parallel decoding of the cross-modal fusion feature map. A fixed-length sequence of 6 is used as the basic atomic unit of the bounding box. A maximum of 8 sets of hazardous area bounding boxes are output simultaneously in a single forward propagation, with three semantic labels: mild, moderate, and severe. After the UAV reads the geographic coordinates corresponding to each set of bounding boxes, it initiates an autonomous fine-scan process. The flight platform autonomously descends to a height of 5 meters above the target area, and the gimbal drives the imaging device to maintain a vertically downward shooting posture, in conjunction with RealSense. The D435 depth camera continuously detects obstacles such as tea trees, supports, and cables in the field, maintaining a safe distance of no less than 1.5 meters. The device travels at a speed of 1 m / s along a tic-tac-toe pattern across a 10m × 10m area, while simultaneously switching to a high-resolution RGB camera (such as a Sony A7R IV) to complete near-canopy imaging acquisition. Ultimately, it obtains scanned images of millimeter-level details matching each affected area. The entire area determination process relies on multi-feature weighted quantification of physiological stress, and a parallel decoding mechanism synchronously outputs multi-area positioning coordinates to avoid inference delays caused by sequential detection. The coarse and fine scanning layered operation modes take into account both large-scale surveys and local detail identification needs. Depth sensing and flight control ensure the safety of low-altitude operations. High-resolution fine scanning imaging can completely capture the distribution characteristics of tiny pest individuals, providing a clear and effective image data source for subsequent pest quantity statistics and population density quantification.

[0023] 106. Input each of the scanned images into the pre-trained pest detection model to perform tea green leafhopper detection processing, obtain the bounding box coordinates of each individual pest, and determine the pest density value per unit area corresponding to each of the infested areas based on all the bounding box coordinates.

[0024] In this embodiment, the pre-trained pest detection model adopts the YOLOv8-nano lightweight network architecture. The network is specifically optimized for the tiny target of the tea green leafhopper, which has a body length of 3 mm. It can stably complete the insect identification in images with a ground sampling distance of millimeters. The tea green leafhopper detection processing uses a 640×640 high-resolution RGB image acquired by a high-resolution RGB camera during the fine scanning stage as input. The model inference outputs the bounding box coordinates and confidence scores ranging from 0 to 1 for all individual pests. Valid bounding boxes with a confidence score of not less than 0.5 are selected to eliminate false detections. The bounding box coordinates are used to characterize the position and coverage of a single pest in the image plane. The total number of all valid bounding boxes is counted as the total number of pests in the region. Then, the pest density per unit area corresponding to each affected area is calculated. For example, if the affected area is 100 square meters, the pest density per unit area is obtained by dividing the total number of pests in the affected area by 100. The unit of the value is heads per square meter. After the density calculation is completed, the same... The coarse scanning stage involves bidirectional cross-validation of the CHI values ​​of corresponding plots. For example, CHI value ranges of 0.3 to 0.5, 0.5 to 0.7, and 0.7 and above correspond to mild, moderate, and severe grading standards, respectively. Pest densities below 50 per square meter, 50 to 150 per square meter, and above 150 per square meter are matched with the same grading standard. When the two types of indicators are consistent, the higher level is selected as the final hazard level of the plot. If the difference between the two types of indicators exceeds one level, a second fine scan or manual on-site verification is initiated. This eliminates spectral misjudgment problems caused by vegetation stress and light interference. The lightweight detection network is adapted to the low computing power operating environment of UAV-borne edge devices. The micro-target optimization improves the detection rate of small insects near the canopy at low altitudes. The two-dimensional judgment mechanism of cross-validation of spectral index and on-site insect density avoids the judgment bias of a single data source, making the plot hazard level classification results more objective and reliable. At the same time, the quantitative output density value can provide accurate quantitative basis for subsequent pest trend prediction and the formulation of graded plant protection plans.

[0025] like Figure 2 As shown, Figure 2This diagram illustrates the autonomous detection trajectory in a simulated tea garden environment, providing a complete and intuitive demonstration of the layered operational logic of the general survey followed by detailed local surveys. The diagram represents a 50m x 30m standardized tea garden from above, with horizontal gray-scale stripes corresponding to rows of tea bushes. The square marker in the lower left corner indicates the drone's operational starting point, while the triangular marker in the upper left corner indicates the RTK differential positioning base station, providing centimeter-level geolocation support for the drone throughout the process. The three circular blocks enclosed by dashed lines represent the light, moderate, and severe pest infestation areas identified in step 105, each with a bounding box generated through parallel positioning. The drone completes steps 101 to 105 along an 80m high-altitude coarse-scanning route, identifying multiple infestation areas. Then, it performs a 5m low-altitude fine scan on each infestation area. The drone autonomously descends based on the geographical coordinates of the boundary boxes of the infestation areas, acquiring millimeter-level high-definition scan images of each area, which are then fed into the pest detection model in step 106 for insect identification and density calculation. The entire flight path and timeline planning correspond one-to-one with the aforementioned steps. Relying on RTK high-precision positioning, it enables autonomous switching between coarse and fine sweep routes and precise revisiting of affected areas. The timeline decomposition and calculation are balanced with the flight mission to reduce the computing load on the airborne edge. While ensuring the efficiency of rapid survey of large tea gardens, it also takes into account the accuracy of identifying small targets of local pests. This provides a standardized observation dataset with complete georegistration and timeline alignment for subsequent pest density statistics, timeline population trend prediction, and the formulation of hierarchical plant protection and control strategies.

[0026] It is understood that the technical solution of this application can also be used for monitoring other crop diseases and pests, such as rice planthopper monitoring, citrus Huanglongbing monitoring, grape downy mildew monitoring, etc., only by adjusting the input band and vegetation index combination of the spectral prompt encoder.

[0027] In this embodiment of the invention, constructing a vegetation index feature vector based on the five-band image includes: obtaining the five-band reflectance of the five-band image; calculating a vegetation index based on the five-band reflectance; and concatenating the five-band reflectance with the vegetation index to obtain the vegetation index feature vector.

[0028] In this embodiment, tea garden inspection is carried out using a drone equipped with a five-band imaging device. The device includes blue, green, red, red-edge, and near-infrared imaging channels with center wavelengths of 475nm, 560nm, 668nm, 717nm, and 842nm, respectively. The raw DN values ​​obtained from the imaging are combined with irradiance data synchronously recorded by the airborne illumination sensor to complete radiometric calibration calculations, converting them into five-band reflectance. The five-band reflectance includes blue band surface reflectance, green band surface reflectance, red band surface reflectance, red-edge band surface reflectance, and near-infrared band surface reflectance. Based on four types of standardized formulas, the vegetation index is calculated by substituting the corresponding band reflectance. The specific calculation expression is as follows: , , , , in, Represents the normalized vegetation index. Indicates the normalized vegetation index in the green band. Indicates the red-edge normalized vegetation index. The vegetation index represents a simple ratio. Indicates the surface reflectance in the near-infrared band. Indicates the surface reflectance in the red band. Indicates the surface reflectance in the green band. This indicates the surface reflectance in the red-edge band.

[0029] In this embodiment, a pre-trained spectral cue encoder is used to concatenate five-band reflectance with vegetation indices. The pre-trained spectral cue encoder is trained under supervision using a dataset of combined band reflectance and vegetation indices corresponding to multiple sets of pest stress samples from tea gardens. A multi-layer fully connected network is used to fit the mapping relationship between spectral values ​​and the degree of damage to tea trees. During the training phase, the network weights and bias parameters are continuously updated until the model loss converges. The five sets of band reflectance and four types of vegetation indices are directly concatenated in a fixed-dimensional order to complete the feature concatenation operation and generate a vegetation index feature vector with a dimension of 9. The vector simultaneously carries the original spectral physical information and the quantitative indicators of vegetation physiological stress, completely preserving the subtle shift features of the canopy spectrum caused by the tea green leafhopper infestation. The standardized features after unified concatenation can eliminate the defects of insufficient representation ability of single bands or single vegetation indices, providing a complete bottom-level input for subsequent spectral feature encoding, fully enhancing the sensitivity of spectral modality to the representation of pest abnormal areas, and effectively improving the accuracy of pest stress area localization when matching spectral and visual cross-modal features.

[0030] In this embodiment of the invention, the step of performing spectral feature encoding processing on the vegetation index feature vector to obtain a spectral cue embedding vector includes: obtaining a pre-trained spectral cue encoder, the spectral cue encoder including a first fully connected layer, a second fully connected layer, a third fully connected layer, and a linear projection layer connected in sequence; inputting the vegetation index feature vector into the first fully connected layer for nonlinear feature extraction processing to obtain a first multidimensional intermediate feature vector; inputting the first multidimensional intermediate feature vector into the second fully connected layer for nonlinear feature extraction processing to obtain a second multidimensional intermediate feature vector; inputting the second multidimensional intermediate feature vector into the third fully connected layer for nonlinear feature extraction processing to obtain a spectral feature vector; and inputting the spectral feature vector into the linear projection layer for linear up-dimensional mapping processing to obtain the spectral cue embedding vector.

[0031] In this embodiment, the original training samples used by the pre-trained spectral hint encoder are five-band reflectance and four-category vegetation index observation data collected from tea garden plots with different levels of pest stress. All original samples are pre-assembled and integrated in a fixed dimensional order, uniformly transformed into standardized 9-dimensional vegetation index feature vectors, and then fed into supervised training. The complete training process uniformly assigns random initial parameters to all learnable weight matrices and bias vectors within the encoder. The standardized sample vectors flow through a three-layer cascaded fully connected network to complete a progressive nonlinear transformation. The network output representation is compared with the actual pest stress labels of the plots to calculate the loss value. The parameter values ​​of all weight matrices and bias vectors are continuously updated iteratively using the backpropagation mechanism to continuously compress the model prediction loss until the loss value tends to stabilize and converge. Finally, all network parameters are solidified to complete the encoder pre-training process. Nonlinear feature extraction relies on the combined linear operations within the fully connected layer and the ReLU activation function. The input vector is first multiplied by the corresponding weight matrix of the layer. The result is then superimposed on the bias vector of the current layer to obtain a linear transformation. The transformation values ​​are then truncated using the ReLU activation function, unifying all negative values ​​to zero. This nonlinear mapping mechanism is used to mine the pest stress correlation features hidden in the low-dimensional spectral data. The 9-dimensional vegetation index feature vector is fed into the first fully connected layer to complete the first round of nonlinear feature extraction, generating a first multidimensional intermediate feature vector with a dimension of 64. This vector carries the feature expression of the original spectral basic information after the initial nonlinear mapping. The first multidimensional intermediate feature vector is fed into the second fully connected layer to repeat the linear transformation and ReLU activation operation, resulting in a second multidimensional intermediate feature vector with a dimension of 32. This achieves secondary compression of the basic spectral features and extraction of deep correlation information. The second multidimensional intermediate feature vector is fed into the third fully connected layer to perform the same nonlinear processing, outputting a spectral feature vector with a dimension of 16. This vector is a condensed core spectral representation of tea tree pest stress obtained after three layers of progressive feature extraction. The linear dimensionality upscaling process is performed by an independent projection layer, executing only pure linear operations without activation function constraints. The 16-dimensional spectral feature vector is multiplied by the projection layer's dedicated weight matrix and superimposed with the corresponding bias vector, directly linearly expanding the feature dimension to 768 dimensions to generate a spectral cue embedding vector. This step does not introduce any new nonlinear mappings; it only aligns the feature dimensions of the spectral and visual modalities. The entire hierarchical coding structure relies on a three-layer progressive nonlinear network to fully mine the weak pest response signals in the original low-dimensional spectral data. Then, the linear projection dimensionality upscaling operation unifies the feature output dimensions of the two modalities, eliminating the dimensionality adaptation barrier that exists in multi-head cross-attention calculations. This provides a spectral prior representation of dimensionality matching and complete pest stress information for subsequent cross-modal comparison operations.

[0032] In this embodiment of the invention, the step of performing visual encoding processing on the RGB image to obtain a visual block feature sequence includes: segmenting the RGB image using a uniform block algorithm to obtain multiple image block units of the same size; inputting each image block unit into a pre-trained visual encoder for visual encoding processing to obtain multiple image feature vectors; obtaining a preset image spatial arrangement rule; and concatenating and integrating all the image feature vectors based on the image spatial arrangement rule to obtain the visual block feature sequence.

[0033] In this embodiment, the uniform block segmentation algorithm uses a fixed-step sliding window block segmentation algorithm to complete the RGB image segmentation process. The segmentation process follows a uniform pixel size and slides along the horizontal and vertical sides of the image at fixed intervals. The complete image is decomposed into image block units of completely uniform size. Each unit has the same pixel scale and uniformly covers the entire spatial area of ​​the image without any omissions or overlaps. Each group of independent image block units carries spatial detail information such as tea tree canopy texture and visual representation of pests in the corresponding local space. All image block units are sequentially fed into a pre-trained visual encoder for visual encoding processing. The visual encoder adopts the Moon-ViT architecture. During the pre-training stage of the encoder, it uses a multi-scale RGB image sample set of pests in the tea garden to perform supervised learning. The real annotation of the pest area in the image is used as the training label. The internal weight parameters of the network are iteratively updated by backpropagation until the feature extraction loss of the model for small pest targets converges and stabilizes. After a single image block is input into the encoder, a global spatial feature mapping operation is completed, and the output is an image feature vector with a uniform dimension of 768. This vector transforms the visible light visual information in the image block into a standardized one. Numerical representation fully preserves detailed features such as the morphology of pests and the outline of the canopy in local areas. The image spatial arrangement rules are set according to the row and column traversal order when the image is segmented and cropped, strictly matching the original two-dimensional spatial coordinate logic of the image. This ensures that the arrangement order of the feature sequence corresponds one-to-one with the real spatial position of the image. The serial integration operation reads all image feature vectors in the order of the arrangement rules and splices them sequentially according to the row and column traversal order to generate a visual block feature sequence. The sequence contains multiple image feature vectors, and all vectors maintain a fixed 768-dimensional dimension. This is aligned with the 768-dimensional spectral cue embedding vector output by the spectral cue encoder. The block-based encoding architecture decomposes the complete large image into lightweight units and extracts features step by step, effectively reducing the computational load of a single operation at the airborne edge. The standardized visual sequence of the same dimension can be directly adapted to the multi-head cross-attention calculation logic, realizing precise correlation between spectral stress priors and local visual features of the image on a unit-by-unit basis. This opens up the interaction path between the underlying features of the spectral and visual modalities, eliminating the feature fusion obstacles caused by the dimensional mismatch between modalities. At the same time, it relies on the independent feature representation of blocks to improve the spatial positioning accuracy of small pest targets.

[0034] In this embodiment of the invention, the step of performing multi-head cross-attention cross-modal comparison processing on the spectral cue embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map includes: using the spectral cue embedding vector as a query vector, and using each image feature vector in the visual block feature sequence as a key vector and a value vector, respectively; performing attention calculation on each key vector based on the query vector to obtain multiple attention scores, and normalizing all the attention scores to obtain an attention weight matrix; performing multi-branch weighted integration operation processing on the attention weight matrix and all the value vectors to obtain multiple single-head attention output features; obtaining a preset aggregation rule, and performing multi-head aggregation and splicing processing on the multiple single-head attention output features based on the aggregation rule to obtain the cross-modal fusion feature map.

[0035] In this embodiment, the query vector is a 768-dimensional spectral cue embedding vector output by the spectral cue encoder. The vector carries complete semantic information representing spectral anomaly regions. Both the key and value vectors are taken from image feature vectors with a uniform dimension of 768 within the visual block feature sequence, which contains multiple image feature vectors. The formula for calculating the attention score based on the query vector and each set of key vectors is as follows: , in, Indicates the first Attention score, Represents the query vector. Indicates the first A key vector, Represents the transpose of a vector. For fixed values, This is used to perform numerical scaling, avoiding the vanishing gradient problem caused by excessively large values ​​in vector inner product operations. After traversing all key vectors, multiple attention scores are obtained, each representing the degree of association between spectral semantics and the corresponding visual features of a single block. Softmax normalization is then performed on each set of independent attention scores, mapping all scores to the 0-1 interval. All normalized results are then integrated according to the original arrangement order of image blocks to generate an attention weight matrix. The weighted integration operation is performed using matrix multiplication, and the specific calculation formula is as follows: , in, Represents the attention weight matrix. This represents the feature matrix formed by stacking all value vectors in the spatial order of image blocks. Represents the query vector. The feature matrix represents the stacked key vectors. The weight values ​​in each row of the attention weight matrix are matched with the value vectors at their corresponding positions. The weighted summation of each visual space feature is completed by matrix multiplication. Image blocks with higher semantic matching with spectral anomalies are assigned larger weight coefficients. After the operation, the single-head visual features with fused spectral prior information are output. The weighted operation is performed by 8 independent computation branches to obtain multiple sets of differentiated single-head attention output features. The multiple sets of features are then spliced ​​and projected according to a preset aggregation rule to generate a cross-modal fusion feature map. The preset aggregation rule is set according to the standard architecture of multi-head attention networks. The core of the rule is to first perform a first-to-last concatenation splicing operation on all 8 sets of single-head attention output features along the feature dimension. The high-dimensional features after splicing are sent to the linear projection layer to complete the dimension adaptation fusion. The multi-head aggregation splicing process is fully executed according to this rule, and finally a cross-modal fusion feature map is generated. The feature map achieves pixel-level alignment between spectral stress semantics and image space visual features. Insect pest areas that match the high-hazard comprehensive index have higher feature response amplitudes, which can provide accurate target region prior constraints for parallel box decoders. The entire multi-head cross-attention computation process relies on 8 independent single-head computation branches to simultaneously mine multi-dimensional correlation features between the spectrum and vision. It introduces spectral prior information to amplify the feature response amplitude of the image region corresponding to the pest, effectively reducing the interference of irrelevant background textures in the tea garden on target localization. The visual block serialization features are naturally adapted to parallel computing logic, greatly reducing the computing power consumption generated by global feature retrieval of complete large images. It is suitable for low computing power real-time inference scenarios on UAV airborne edge devices, and fully builds a stable cross-modal mapping link from spectral anomaly recognition to accurate localization of pest regions in images, improving the accuracy of subsequent boundary box generation of the hazard area and the efficiency of parallel decoding.

[0036] In this embodiment of the invention, determining multiple hazardous regions based on the cross-modal fusion feature map includes: obtaining a vegetation spectral disease feature set for each pixel in the cross-modal fusion feature map; calculating an adaptive weight for each feature in the vegetation spectral disease feature set using a random forest algorithm; calculating a CHI value corresponding to each pixel based on the vegetation spectral disease feature set and the corresponding adaptive weight; obtaining a preset CHI threshold, comparing each CHI value with the CHI threshold; identifying pixels with CHI values ​​greater than the CHI threshold as hazardous pixels, and performing connected component aggregation processing on all hazardous pixels using a preset neighborhood determination rule to obtain multiple hazardous regions.

[0037] In this embodiment, the vegetation spectral disease feature set corresponding to a single pixel includes three types of standardized spectral indices: the decrease in NDVI relative to the healthy baseline after normalization, the shortwave offset of the red edge position after offset normalization, and the increase in MSI relative to the healthy baseline after normalization. These three indices correspond to the responses of tea leaves to chlorophyll attenuation, red edge spectral shift, and canopy water stress, respectively. The process of solving the adaptive weights for each feature using the random forest algorithm is completed using a tea garden multi-gradient pest stress sample dataset. During the training phase, the Gini importance score corresponding to the three types of spectral features is calculated using the Gini coefficient, and then the adaptive weight conversion is completed using the weight normalization formula. The specific formula expression is as follows: , in, Indicates the first Gini importance score of spectral features This represents the sum of the importance scores for all three feature classes. After conversion, three sets of weight values ​​are obtained, respectively. , and , The adaptive weights are the percentage decrease in NDVI relative to the healthy baseline. The adaptive weights are the shortwave offsets corresponding to the red edge positions. The adaptive weights corresponding to the increase in MSI relative to the health baseline are used to calculate the CHI value for each pixel based on the vegetation spectral disease feature set and the corresponding adaptive weights. , in, This indicates the degree of decline in NDVI relative to a healthy baseline. This indicates the shortwave offset at the red edge position. The increase in MSI relative to the health benchmark is represented by multiplying the three sets of normalized spectral indices of a single pixel with the matched adaptive weights one by one and then summing them to obtain the pixel-level comprehensive pest stress index (CHI) value.

[0038] In this embodiment, the preset CHI threshold can be set to 0.3. This threshold is determined based on years of field pest density and spectral stress data from tea gardens, and is divided into four levels of damage: healthy, mild, moderate, and severe. CHI values ​​less than 0.3 correspond to healthy plots without stress, 0.3 to 0.5 correspond to mild pest stress, 0.5 to 0.7 correspond to moderate pest stress, and values ​​greater than or equal to 0.7 correspond to severe pest stress. Pixel-by-pixel comparison and calculation of the CHI values ​​obtained through pixel-level screening and connected component aggregation are compared with the 0.3 baseline threshold. Pixels with CHI values ​​higher than 0.3 are uniformly marked as harmful pixels. After traversing the entire feature map, the coordinates of all harmful pixels are extracted. Connected component aggregation is performed on all the harmful pixels using a preset neighborhood determination rule. The neighborhood determination rule stipulates that only spatially adjacent harmful pixels are divided into the same connected block. Each group of consecutive pixels corresponds to an independent pest infestation area. After aggregation, multiple contiguous harmful areas are output. The entire pixel-level stress quantification process integrates three multi-dimensional spectral stress features: chlorophyll, red edge, and moisture. It utilizes the Gini importance of random forests to achieve adaptive allocation of feature weights, avoiding the bias in determining the degree of regional stress caused by fixed weights. CHI multi-feature weighted calculation can accurately quantify the comprehensive damage level of the tea tree canopy. Threshold grading combined with connected component aggregation can completely separate the scattered pest areas within the image. The subsequent parallel box decoder can generate multiple sets of synchronously output bounding boxes based on contiguous damage areas. Coarse scanning of the entire domain stress grading is combined with low-altitude fine layered scanning to take into account both rapid surveys of large-scale tea gardens and accurate identification of local pest details. Depth sensing and flight control mechanisms ensure the safety of low-altitude near-canopy imaging operations. Millimeter-level high-resolution fine scan images provide reliable data support for pest density quantification and cross-validation of damage levels.

[0039] In this embodiment of the invention, after determining the unit area pest density value corresponding to each of the harmful areas based on all the bounding box coordinates, the method further includes: obtaining multiple historical CHI values, constructing a historical CHI index time series based on the multiple historical CHI values, inputting the historical CHI index time series into a pre-trained logistic growth model to predict future pest population changes, and obtaining a pest population change curve; calculating a comprehensive risk score based on the pest population change curve, and determining an early warning level based on the comprehensive risk score; and determining a prevention and control response strategy based on the early warning level.

[0040] In this embodiment, the multi-period historical CHI values ​​are comprehensive pest stress indices calculated pixel-by-pixel and region-by-region during each UAV-based full-area coarse sweep operation. Each coarse sweep operation can completely collect CHI data for the entire area, and the data collection frequency is much higher than the insect population density values ​​that can only be obtained during the low-altitude fine sweep stage, thus forming a continuous and complete time-series observation sample. When constructing the historical CHI index time series, all multi-period CHI data are grouped and classified according to the unique identifier of the plot, and then the multiple sets of CHI values ​​corresponding to a single plot are arranged in order according to the time sequence of UAV inspection and collection, forming a one-dimensional time series array arranged in ascending time order. This array is the plot-specific historical CHI index time series.

[0041] In this embodiment, the logistic growth model uses the standard S-shaped growth function as its basic architecture, and its core expression is: , in, express The number of pest species at any given time. Indicates environmental carrying capacity. This indicates the inherent growth rate of the pest population. This indicates the inflection point in time when the population enters the exponential growth phase. This represents the natural exponential function. The model training process relies on supervised fitting of CHI time-series samples accumulated through long-term monitoring of tea garden plots and synchronous field insect population density calibration datasets. Historical CHI time series are used as input features, and the corresponding field-statistical insect population density is used as the true fit value. The model parameters are iteratively optimized using a nonlinear least squares algorithm to continuously reduce the error between the fitted curve and the actual insect population changes until the parameters converge and stabilize, completing the model pre-training.

[0042] In this embodiment, when predicting future pest population changes, the complete historical CHI index time series of a single plot is input into a converged logistic growth model. The model fits the pest population change curve based on the CHI value change trend within the time series. The curve completely reconstructs the three-stage evolution of the pest population from latent and slow growth, rapid exponential expansion to environmental carrying capacity saturation. Based on the curve, the pest population size change trend of the plot in the next 7 to 14 days can be deduced. Figure 3As shown, the horizontal axis represents time in days, and the vertical axis represents the number of pests per unit area of ​​the corresponding plot in heads per hundred shoots. The solid curve P(t) is the pest population change curve output by the logistic growth model, showing a typical S-shaped evolutionary pattern. The horizontal dashed line K=500 represents the pest carrying capacity of the tea garden plot, that is, the upper limit of the saturation of the pest population that the plot's vegetation can support. The horizontal dashed line K / 2=250 is the population growth inflection point, corresponding to the critical pest population size at which the pest reproduction rate reaches its peak. The time interval corresponding to this inflection point is the optimal control window. The curve can be divided into three stages along the time axis: during the incubation period, the pest population base is low and the population growth rate is slow; during the exponential growth period, the pest reproduction rate increases rapidly, and the pest population size rises sharply in a short period of time, corresponding to the marked control window period. During this period, chemical intervention can effectively suppress the population base and avoid large-scale outbreaks of pests; during the saturation period, due to the constraints of tea tree nutrients and living space, pest growth gradually slows down and approaches the upper limit of the carrying capacity K. After the model is derived from the historical CHI index time series, it can accurately locate the prevention and control window corresponding to the exponential growth stage, provide early prediction of pest outbreaks in the plot, and determine the application time nodes for the graded prevention and control strategy.

[0043] The comprehensive risk score calculated based on the pest population change curve relies on the Sigmoid mapping formula that integrates multiple dimensions of factors. The formula expression is as follows: , in, This represents the overall risk score. express Weighting coefficient (initially 0.4). This indicates the current CHI index value (immediate level of harm). This indicates the inherent growth rate of the pest population. express The weighting coefficient (initially 0.2). This represents the comprehensive score of meteorological impact factors. The factors representing the growth period of tea trees (spring tea picking period 1.5 > summer tea period 1.0 > autumn tea period 0.8 > overwintering period 0.3) are used to calculate the comprehensive score of meteorological influence factors. , in, express The corresponding adaptive weight coefficients, The score represents the temperature suitability score. 25℃ to 28℃ is the optimal temperature range for the reproduction of the tea green leafhopper. The corresponding score is calculated based on the deviation of the actual air temperature from the suitable range. express The corresponding adaptive weight coefficients, The score represents the humidity suitability score. 70% to 80% air humidity is conducive to pest proliferation. The score is calculated based on the difference between the measured humidity in the field and the optimal humidity range. express The corresponding adaptive weight coefficients, The score indicates the intensity of the rainfall suppression sub-item. Rainfall washes away insect bodies and inhibits the expansion of insect populations. The greater the rainfall, the lower the score for this sub-item, which indicates the strength of the rainfall's inhibitory effect on insect pests.

[0044] After calculation, the multi-dimensional fusion result is mapped to the 0-1 range using the Sigmoid function, and the output value is the comprehensive risk score of the land parcel. After the comprehensive risk score is completed, the warning level is divided according to the fixed range threshold: a score less than 0.3 is marked as a green safety level, a score between 0.3 and 0.6 is marked as a yellow attention level, a score between 0.6 and 0.8 is marked as an orange warning level, and a score greater than or equal to 0.8 is marked as a red emergency level. Different warning levels are matched with graded pest control response strategies for tea gardens. Green level maintains regular periodic inspection operations; yellow level increases the frequency of field inspections and conducts local manual pest sampling and verification; orange level delineates key control areas and arranges targeted low-concentration pesticide local spraying; red level initiates unified prevention and control operations for the entire contiguous tea garden. At the same time, a 7 to 14 day window period is set after the pesticide application is completed for re-flying and detection. Drones return to the field to carry out coarse sweeping combined with local fine sweeping operations. The CHI values ​​and pest density data collected during the re-flying are compared to evaluate the effectiveness of pesticide control. The new time-series observation data obtained from the re-flying are simultaneously added to the model training dataset to complete parameter iteration and update, and continuously optimize the population trend prediction accuracy of the logistic growth model. The entire time-series prediction and early warning system relies on high-frequency collected CHI time-series data to compensate for the shortcomings of excessively long sampling intervals for insect population density. The S-shaped logistic curve closely matches the natural population reproduction and evolution patterns of the tea green leafhopper. Multi-dimensional risk scoring comprehensively considers the real-time damage level of the plot, the insect reproduction rate, environmental and meteorological conditions, and crop growth status. The graded early warning system, combined with differentiated prevention and control strategies, enables precise zonal management of tea garden pests. The data feedback and iteration mechanism continuously optimizes the model's prediction performance, gradually reducing the bias in pest trend prediction, and enabling early intervention in high-risk plots to reduce tea yield losses caused by large-scale pest outbreaks.

[0045] The method for detecting the tea green leafhopper in the embodiments of the present invention has been described above. The detection device for the tea green leafhopper in the embodiments of the present invention is described below. Please refer to [link / reference needed]. Figure 4 One embodiment of the tea green leafhopper detection device of the present invention includes: Feature vector construction module 401: used to acquire five-band images and RGB images, and construct vegetation index feature vectors based on the five-band images; Spectral feature encoding module 402: used to perform spectral feature encoding processing on the vegetation index feature vector to obtain a spectral cue embedding vector; Visual encoding module 403: used to perform visual encoding processing on the RGB image to obtain a visual block feature sequence; Cross-modal comparison module 404: used to perform multi-head cross-attention cross-modal comparison processing on the spectral cue embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map; Hazardous area determination module 405: used to determine multiple hazardous areas based on the cross-modal fusion feature map, and to acquire a scan image corresponding to each hazardous area; Pest detection module 406: is used to input each of the scanned images into a pre-trained pest detection model for tea green leafhopper detection processing, obtain the bounding box coordinates of each individual pest, and determine the pest density value per unit area corresponding to each of the infested areas based on all the bounding box coordinates.

[0046] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0047] above Figure 4 The tea green leafhopper detection device in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The tea green leafhopper detection device in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0048] Figure 5 This is a schematic diagram of the structure of a tea green leafhopper detection device 500 provided in an embodiment of the present invention. The tea green leafhopper detection device 500 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 510 (e.g., one or more processors) and a memory 520, and one or more storage media 530 (e.g., one or more mass storage devices) storing application programs 533 or data 532. The memory 520 and storage media 530 can be temporary or persistent storage. The program stored in the storage media 530 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the tea green leafhopper detection device 500. Furthermore, the processor 510 may be configured to communicate with the storage media 530 and execute the series of instruction operations in the storage media 530 on the tea green leafhopper detection device 500 to implement the steps of the tea green leafhopper detection method provided in the above-described method embodiments.

[0049] The tea green leafhopper detection device 500 may also include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input / output interfaces 560, and / or one or more operating systems 531, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 5 The illustrated structure of the tea green leafhopper detection device does not constitute a limitation on the tea green leafhopper detection device. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0050] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the tea green leafhopper detection method.

[0051] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0052] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0053] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting the tea green leafhopper, characterized in that, include: Acquire five-band images and RGB images, and construct vegetation index feature vectors based on the five-band images; The vegetation index feature vector is subjected to spectral feature encoding to obtain a spectral cue embedding vector; The RGB image is subjected to visual encoding processing to obtain a visual block feature sequence; The spectral cue embedding vector and the visual block feature sequence are subjected to multi-head cross-attention cross-modal alignment to obtain a cross-modal fusion feature map; Based on the cross-modal fusion feature map, multiple hazard regions are determined, and a scan image corresponding to each hazard region is obtained; Each scanned image is input into a pre-trained pest detection model for tea green leafhopper detection processing to obtain the bounding box coordinates of each individual pest, and the pest density per unit area corresponding to each affected area is determined based on all the bounding box coordinates.

2. The method for detecting the tea green leafhopper according to claim 1, characterized in that, The construction of vegetation index feature vectors based on the five-band images includes: Obtain the five-band reflectance of the five-band image; The vegetation index is calculated based on the five-band reflectance. The five-band reflectance is concatenated with the vegetation index to obtain the vegetation index feature vector.

3. The method for detecting the tea green leafhopper according to claim 1, characterized in that, The step of performing spectral feature encoding on the vegetation index feature vector to obtain a spectral cue embedding vector includes: Obtain a pre-trained spectral cue encoder, the spectral cue encoder comprising a first fully connected layer, a second fully connected layer, a third fully connected layer and a linear projection layer connected in sequence; The vegetation index feature vector is input into the first fully connected layer for nonlinear feature extraction processing to obtain the first multidimensional intermediate feature vector. The first multidimensional intermediate feature vector is input into the second fully connected layer for nonlinear feature extraction processing to obtain the second multidimensional intermediate feature vector; The second multidimensional intermediate feature vector is input into the third fully connected layer for nonlinear feature extraction processing to obtain the spectral feature vector. The spectral feature vector is input into the linear projection layer for linear up-dimensional mapping to obtain the spectral cue embedding vector.

4. The method for detecting the tea green leafhopper according to claim 1, characterized in that, The step of performing visual encoding processing on the RGB image to obtain a visual block feature sequence includes: The RGB image is segmented using a uniform block division algorithm to obtain multiple image block units of the same size; Each image block unit is input into a pre-trained visual encoder for visual encoding processing to obtain multiple image feature vectors; A preset image spatial arrangement rule is obtained, and all the image feature vectors are concatenated and integrated based on the image spatial arrangement rule to obtain the visual block feature sequence.

5. The method for detecting the tea green leafhopper according to claim 1, characterized in that, The step of performing multi-head cross-attention cross-modal alignment processing on the spectral cue embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map includes: The spectral cue embedding vector is used as the query vector, and each image feature vector in the visual block feature sequence is used as the key vector and value vector, respectively. Attention is calculated for each key vector based on the query vector to obtain multiple attention scores, and all attention scores are normalized to obtain an attention weight matrix. The attention weight matrix and all the value vectors are subjected to multi-branch weighted integration operations to obtain multiple single-head attention output features; Obtain preset aggregation rules, and perform multi-head aggregation and splicing processing on multiple single-head attention output features based on the aggregation rules to obtain the cross-modal fusion feature map.

6. The method for detecting the tea green leafhopper according to claim 1, characterized in that, The determination of multiple hazard regions based on the cross-modal fusion feature map includes: Obtain the vegetation spectral disease feature set for each pixel in the cross-modal fusion feature map; The adaptive weight of each feature in the vegetation spectral disease feature set is calculated using the random forest algorithm; Based on the vegetation spectral disease feature set and the corresponding adaptive weights, calculate the CHI value corresponding to each pixel; Obtain a preset CHI threshold, and compare each CHI value with the CHI threshold respectively; Pixels with CHI values ​​greater than the CHI threshold are identified as hazardous pixels, and connected component aggregation is performed on all hazardous pixels using a preset neighborhood determination rule to obtain multiple hazardous regions.

7. The method for detecting the tea green leafhopper according to claim 1, characterized in that, After determining the pest density value per unit area corresponding to each of the harmful areas based on all the bounding box coordinates, the method further includes: Obtain historical CHI values ​​from multiple periods, construct a historical CHI index time series based on the historical CHI values ​​from multiple periods, input the historical CHI index time series into a pre-trained logistic growth model to predict future pest population changes, and obtain pest population change curves. A comprehensive risk score is calculated based on the pest population change curve, and an early warning level is determined based on the comprehensive risk score. The prevention and control response strategy is determined based on the aforementioned warning level.

8. A detection device for the tea green leafhopper, characterized in that, include: Feature vector construction module: used to acquire five-band images and RGB images, and construct vegetation index feature vectors based on the five-band images; Spectral feature encoding module: used to perform spectral feature encoding processing on the vegetation index feature vector to obtain a spectral cue embedding vector; Visual encoding module: used to perform visual encoding processing on the RGB image to obtain a visual block feature sequence; Cross-modal comparison module: used to perform multi-head cross-attention cross-modal comparison processing on the spectral cue embedding vector and the visual block feature sequence to obtain a cross-modal fusion feature map; Hazardous area determination module: used to determine multiple hazardous areas based on the cross-modal fusion feature map, and to acquire a scan image corresponding to each hazardous area; Pest detection module: used to input each of the scanned images into a pre-trained pest detection model for tea green leafhopper detection processing, obtain the bounding box coordinates of each individual pest, and determine the pest density value per unit area corresponding to each of the infested areas based on all the bounding box coordinates.

9. A tea green leafhopper detection device, characterized in that, The tea green leafhopper detection device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the tea green leafhopper detection device to perform the steps of the tea green leafhopper detection method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the tea green leafhopper detection method as described in any one of claims 1-7.

Images