YOLOv11-based beet armyworm prediction method and system

Abstract

The invention provides a YOLOv11-based beet armyworm prediction method and system, and the method comprises the steps: constructing an image data set through employing YOLOv11 in combination with the day and night activity characteristics of beet armyworms; and constructing an anti-interference feature library. And constructing a space-time attention enhanced LSTM-GRU hybrid network, introducing an insect pest migration diffusion matrix and a growth period dynamic weight mechanism, and outputting a beet armyworm occurrence prediction result. According to the method, the image data set is constructed, feature alignment and fusion are realized through processing, and the target detection accuracy is improved. And an insect pest migration diffusion matrix and a growth period dynamic weight mechanism are introduced, so that cross-region migration path quantitative prediction is realized, and prediction is more accurate. Interference targets are eliminated by means of an anti-interference feature library and a filtering algorithm, weighting processing is performed on features and information in combination with a network, a three-dimensional prediction result is output, a full-chain solution is formed, and the accuracy, dynamic performance and anti-interference capability of beet armyworm prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of green pest control technology in agriculture, in particular to a YOLOv11-based forecasting method and system for lepidopteran pests. BACKGROUND

[0002] Current forecasting research on lepidopteran pests mainly relies on computer vision and machine learning techniques. Traditional methods mostly use single modal data (such as visible light images) combined with YOLO series models for target detection, or use time series models (such as LSTM) to analyze pest occurrence trends. For example, some studies use visible light cameras to collect field images, identify lepidopteran pests using YOLOv5 / v8 models, and combine environmental temperature, humidity, rainfall, and other data to build a regression model to predict the occurrence degree. In addition, infrared thermal imaging technology has also been applied to night pest monitoring, which locates active pests through temperature anomalies. However, existing technologies generally have the following characteristics: data sources are mainly single or dual modal, lacking the coordinated use of multi-scale and multi-modal data; time series analysis mostly uses traditional recurrent neural networks, without fully considering the diurnal activity characteristics of lepidopteran pests and dynamic changes in their growth stages; and migration and diffusion path prediction mostly relies on empirical models or simple wind field simulation, lacking quantitative matrix support based on multi-modal data.

[0003] Due to the use of single modal data and the lack of pest migration and diffusion prediction mechanisms, most studies only focus on local area pest density, without establishing a migration and diffusion matrix based on wind speed, atmospheric vertical velocity, and crop continuity, making it difficult to achieve accurate prediction of cross-regional migration paths. Moreover, the prediction results are not linked to prevention. SUMMARY

[0004] The present application aims to at least solve the technical problem in the prior art that it is difficult to achieve accurate prediction of cross-regional migration paths, and particularly innovatively proposes a YOLOv11-based forecasting method and system for lepidopteran pests.

[0005] To achieve the above-mentioned purposes of the present application, the present application provides a YOLOv11-based forecasting method for lepidopteran pests, which comprises: S1, using YOLOv11 to identify visible light feature maps and infrared feature maps of lepidopteran pests under infrared light and visible light environments respectively, and combining the diurnal activity characteristics of lepidopteran pests to construct a lepidopteran pest image dataset; S2, using a cross-scale data alignment algorithm to preprocess the constructed lepidopteran pest image dataset, aligning the data obtained under infrared light and visible light environments at different scales to obtain an aligned image dataset; S3, constructing an anti-interference feature library; S4, filtering out interference targets in the aligned image data set by using the anti-interference feature library; S5, constructing a spatio-temporal attention enhanced LSTM-GRU hybrid network and introducing a pest migration and diffusion matrix and a dynamic weight mechanism of the growth period into the network; S6, inputting the filtered aligned image data set into the spatio-temporal attention enhanced LSTM-GRU hybrid network, and outputting a prediction result of the beet armyworm, the prediction result including the occurrence degree of the beet armyworm, the migration path of the beet armyworm, and the damage peak of the beet armyworm.

[0006] In another aspect, the present application also provides a YOLOv11-based beet armyworm prediction system, which is used to implement the above-mentioned YOLOv11-based beet armyworm prediction method; the system comprises: a multi-modal perception module for collecting multi-modal data; an image recognition module connected with the multi-modal perception module, for recognizing a visible light feature map and an infrared feature map of the beet armyworm by using YOLOv11 under infrared light and visible light environments respectively and combining the diurnal activity characteristics of the beet armyworm, and constructing a beet armyworm image data set; a dynamic preprocessing module connected with the image recognition module, for preprocessing the constructed beet armyworm image data set by using a cross-scale data alignment algorithm, aligning the data obtained under the infrared light and visible light environments at different scales, and obtaining an aligned image data set; a spatio-temporal prediction model module connected with the dynamic preprocessing module, for outputting a prediction result of the beet armyworm based on the aligned image data set, the prediction result including the occurrence degree of the beet armyworm, the migration path of the beet armyworm, and the damage peak of the beet armyworm; a prevention and control decision module connected with the spatio-temporal prediction model module, for recommending a corresponding prevention and control method based on the occurrence degree of the beet armyworm, and recommending a pesticide application time based on the migration path and the damage peak of the beet armyworm; a result output module connected with the prevention and control decision module, for outputting the prevention and control method and the pesticide application time.

[0007] The application has the beneficial effects that: the application effectively solves the problems of feature information loss and scale mismatch caused by traditional single modal data through multi-modal data fusion and cross-scale alignment technology; after the cross-scale data alignment algorithm processing, the high-precision feature alignment and fusion are realized, and the accuracy of target detection in complex field environment is significantly improved; in view of the pain point of missing pest migration and diffusion prediction, the pest migration and diffusion matrix based on wind speed, effective migration time, crop continuity and atmospheric vertical velocity and other factors is innovatively introduced, combined with the dynamic weight mechanism of growth period, the quantitative prediction of cross-regional migration path and the adaptive adjustment of growth period are realized, so that the prediction result is expanded from local pest density to accurate judgment of migration path and damage peak in the whole space scale; at the same time, through the anti-interference feature library constructed by collecting interference image data, combined with the Euclidean distance similarity filtering algorithm, the interference targets such as weeds and fallen leaves are effectively eliminated, combined with the weighted processing of long-term dependent features and spatio-temporal dynamic information by the spatio-temporal attention enhanced LSTM-GRU hybrid network, finally the three-dimensional prediction results including occurrence degree, migration path and damage peak are output, forming a whole-chain solution from data acquisition, preprocessing, interference filtering to accurate prediction, which significantly improves the precision, dynamic and anti-interference ability of the prediction of the beet armyworm.

[0008] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0009] The above and / or additional aspects and advantages of the application will become apparent and be readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 is a flow chart of a beet armyworm prediction method based on YOLOv11 in embodiment 1 of the application; Figure 2 is a structure diagram of a beet armyworm prediction system based on YOLOv11 in embodiment 2 of the application. DETAILED DESCRIPTION

[0010] The embodiments of the application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the application, and cannot be understood as a limitation of the application.

[0011] Embodiment 1 As shown in Figure 1 , a beet armyworm prediction method based on YOLOv11, the method comprises: S1, YOLOv11 is used to identify the visible light feature map and infrared feature map of the beet armyworm under infrared light environment and visible light environment respectively, and to construct a beet armyworm image dataset according to the diurnal activity characteristics of the beet armyworm; S2, the constructed beet armyworm image dataset is preprocessed by using a cross-scale data alignment algorithm, and the data obtained under infrared light environment and visible light environment are aligned on different scales to obtain an aligned image dataset; in step S2, it needs to be specified that the preprocessing first carries out data quality evaluation, and the quality of the hyperspectral, LiDAR and RGB image data is evaluated by indicators such as signal-to-noise ratio, point cloud density and definition, and the corresponding method is used for processing, and then fusion and occlusion compensation are carried out, such as using wavelet threshold denoising algorithm to process the waveband with SNR<30dB in hyperspectral.

[0012] The above fusion logic: through the convolution attention module (CAM), the visible light feature map (channels 1-3) and the infrared feature map (channel 4) are weighted and distributed, wherein the visible light weight is 0.7 and the infrared weight is 0.3 in the daytime (light intensity>5000lux), the infrared weight is 0.8 and the visible light weight is 0.2 at night (light intensity<1000lux), and the identification accuracy of small larvae is improved from the original 91.2% to 96.8%; The occlusion compensation implementation: when the occlusion rate of the insect body is greater than 40%, the contour features of the insect body of the adjacent 3 frames of images are called, and the complete insect body is restored through the "contour completion algorithm" (based on generative adversarial network), and the identification accuracy of the occlusion scene is improved by 22%. In this embodiment, the cross-scale data alignment is mainly realized through the "GPS timestamp+SIFT feature matching" double-checking mechanism.

[0013] Preliminary alignment based on GPS timestamp: the unmanned aerial vehicle data and the ground node data are preliminarily aligned through the GPS timestamp, so as to ensure that the time error is less than or equal to 1ms. Assuming that the timestamp of the unmanned aerial vehicle data is tuav, and the timestamp of the ground node data is tground, the error Δt of time alignment is |tuav-tground|, and through technical means such as 5G edge computing module, it is ensured that Δt≤1ms.

[0014] S3, collect interference image data to construct an anti-interference feature library; S4, use the anti-interference feature library to filter out the interference targets in the aligned image dataset; S5, construct a spatio-temporal attention enhanced LSTM-GRU hybrid network, and introduce a pest migration and diffusion matrix and a dynamic weight mechanism of the growth period in the network; S6, input the filtered alignment image dataset into a spatio-temporal attention enhanced LSTM-GRU hybrid network, and output a prediction result of the beet armyworm, the prediction result of the beet armyworm including a beet armyworm occurrence degree, a beet armyworm migration path and a beet armyworm damage peak value.

[0015] In the embodiment, the principle of a beet armyworm prediction method based on YOLOv11 is as follows: The method relies on a YOLOv11 target detection framework, fuses infrared and visible light dual-mode data, and combines the diurnal activity characteristics of the beet armyworm to construct a multi-modal image dataset. Through a cross-scale data alignment algorithm, a "GPS time stamp + SIFT feature matching" double-checking mechanism is used to realize high-precision alignment of different modal data in the space-time dimension, so that the spatial deviation is less than or equal to 0.1 m and the time error is less than or equal to 1 ms. On this basis, an anti-interference feature library containing multi-source data such as hyperspectral, LiDAR and environmental perception is constructed, and a Euclidean distance similarity algorithm is used to filter interference targets such as weeds and fallen leaves. The spatio-temporal attention enhanced LSTM-GRU hybrid network extracts time series long-term dependence features through the LSTM layer, and the GRU layer combines the spatio-temporal attention mechanism to weight the features, and finally outputs a three-dimensional prediction result containing the occurrence degree probability, the migration path vector and the damage peak value intensity. The particularly introduced pest migration and diffusion matrix integrates environmental factors such as wind speed and effective migration time, and cooperates with the dynamic weight mechanism of the growth period to realize the quantitative prediction of the cross-regional migration path and the adaptive adjustment of the growth period.

[0016] As an optional embodiment of the present application, optionally, in step S1, the beet armyworm image dataset includes: S101, performing convolution operation on the visible light feature map and the infrared feature map to obtain feature information of the visible light feature map and feature information of the infrared feature map respectively; In step S101, it needs to be explained in detail that different scale convolution kernels are used when performing convolution operation on the visible light feature map and the infrared feature map. For the visible light feature map, small scale convolution kernels of 3x3 and 5x5 are used to capture local detail features such as leaf edges and vein directions; for the infrared feature map, a medium scale convolution kernel of 7x7 is used to extract macro regional features of temperature distribution. Through parallel processing of multi-scale convolution kernels, micro morphological features and macro thermal features of the beet armyworm activity area can be obtained at the same time. After convolution operation, the ReLU activation function is used to enhance the sparsity of the features, and the Batch Normalization layer is used to accelerate the model convergence. Finally, the obtained visible light feature information contains 256-dimensional feature vectors, and the infrared feature information contains 128-dimensional feature vectors, which will be used as the input of the subsequent attention mechanism calculation.

[0017] S102, calculate the attention weights of the visible light feature map and the infrared feature map at different channels and spatial positions based on the feature information of the visible light feature map and the feature information of the infrared feature map respectively; The calculation formula of the attention weights of the visible light feature map and the infrared feature map at different channels and spatial positions is: attention score = (insect situation data correlation x growth period risk coefficient / (meteorological data volatility + 1) (cabbage rosette growth period risk coefficient = 1.3, 30% higher than seedling stage, at this time the weight of insect situation data is increased); In step S102, the calculation of the attention score comprehensively considers multiple factors such as insect situation data, growth period, and meteorological data. The insect situation data correlation reflects the correlation between the current area's activity of the beet armyworm and the historical data. The higher the correlation, the greater the continuity of the current insect situation development, and the greater the contribution to the attention score. The growth period risk coefficient is set according to different growth stages of soybeans. For example, the cabbage rosette growth period risk coefficient is 1.3, which is 30% higher than the seedling stage, meaning that the risk of damage to soybeans by the beet armyworm increases during this growth stage, so the weight of insect situation data in calculating the attention score is correspondingly increased. The meteorological data volatility reflects the influence of environmental factors on the activity of the beet armyworm. The greater the volatility, the more severe the environmental changes, which may interfere with the migration and reproduction of the beet armyworm. Therefore, it needs to be considered in calculating the attention score. By adding 1 to the denominator, the situation of zero denominator is avoided. Through this calculation method, reasonable attention weights can be allocated to the visible light feature map and the infrared feature map at different channels and spatial positions, so that the model pays more attention to important feature regions related to the activity of the beet armyworm, thereby improving the accuracy of subsequent feature fusion and prediction. After obtaining the attention weights, they are weighted and fused with the corresponding feature information to further highlight the key features.

[0018] S103, weighted fusion of the visible light feature map and the infrared feature map based on the attention weights, obtaining the fused feature map, and all the fused feature maps are combined to form a beet armyworm image dataset.

[0019] It needs to be explained in step S103 that the weighted fusion process is not a simple numerical superposition, but a dynamic adjustment of the visible light feature map and the infrared feature map through attention weights. Specifically, for each feature channel and spatial position in the visible light feature map, it is weighted according to its corresponding attention weight; similarly, the same weighting operation is performed on the infrared feature map. Then, the weighted visible light feature map and infrared feature map are spliced in the channel dimension to form a fused feature map. This fusion method not only retains the morphological detail information in the visible light feature map, but also integrates the physiological feature information such as temperature distribution in the infrared feature map, so that the fused feature map can more comprehensively reflect the activity of the beet armyworm. Finally, all the fused feature maps are arranged according to the time sequence or the spatial position to form a complete beet armyworm image data set.

[0020] As an optional embodiment of the present application, optionally, constructing the anti-interference feature library in step S3 includes: S301, under visible light and infrared light environment, collecting interference object image data, and labeling the category of the interference object in the interference object image data; In step S301, it needs to be explained in detail that when collecting interference object image data under visible light and infrared light environment, different areas of the field and different time periods are photographed to obtain diversified interference object samples. These interference objects mainly include weeds, fallen leaves, soil blocks and other objects that may affect the accuracy of the identification of the beet armyworm. For the collected interference object image data, the category of each interference object is accurately marked, for example, it is explicitly marked whether the object in an image is a weed or a fallen leaf, etc.

[0021] S302, extracting interference feature vectors in the interference object image data; In this embodiment, the YOLOv11 model is used to extract the interference feature vectors in the interference object image data. Specifically, first, the collected interference object image data is input into the pre-trained YOLOv11 model. The YOLOv11 model performs multi-level and multi-scale feature extraction on the interference objects in the image. Through convolutional layers, pooling layers and other operations, the image data is gradually converted into high-dimensional feature representations. During the extraction process, the model will focus on key feature information such as the shape, texture and color of the interference object, and encode these information into feature vectors. For example, for a weed image, the model will extract the edge features of its leaves, the texture features of the leaf veins, and the color distribution features of the leaves, etc., and combine these features to form a unique feature vector; for a fallen leaf image, it will extract features such as shape outline, surface wrinkles and color changes to generate a corresponding feature vector. Finally, the obtained interference feature vectors can accurately and comprehensively describe the features of the interference object.

[0022] S303, perform clustering analysis based on the interference feature vectors to obtain a plurality of interference category clusters, each of which corresponds to an interference category; In step S303, it needs to be explained in detail that when performing clustering analysis based on the interference feature vectors, the K-means algorithm is used to initialize the cluster center, and the elbow rule is used to determine the optimal cluster number K. The specific process is as follows: first, calculate the Euclidean distance between all interference feature vectors to construct a distance matrix; then randomly select an initial cluster center, and in turn select the vector farthest from the selected center as the new center until K center points are selected; then assign each interference feature vector to the cluster to which the nearest cluster center belongs; continuously update the cluster center position through iterative optimization until the sum of squared errors (SSE) of the cluster samples converges or the maximum iteration number is reached. For example, when K=5, the weed category cluster may contain feature vectors with long leaf blades and clear leaf veins, the fallen leaf category cluster gathers feature vectors with curled shape and brown color, and the soil block category cluster presents feature vectors with low reflectivity and high texture roughness. Each interference category cluster obtained finally corresponds to an interference category, such as a weed cluster, a fallen leaf cluster, and a soil block cluster.

[0023] S304, store the feature vectors of each interference category cluster to form an anti-interference feature library; In step S304, it needs to be explained in detail that when storing the feature vectors of each interference category cluster, a structured database management system is used for efficient organization. Specifically, an independent data table is created for each interference category cluster, which contains statistical information such as feature vector dimension, mean vector, and covariance matrix, and records the interference category label (such as weed, fallen leaf, etc.) and sample number corresponding to the cluster. For example, the data table of the weed cluster contains a 256-dimensional mean vector array, a 128x128 covariance matrix, and a "weed" category field and a count of 5000 samples. To speed up subsequent similarity calculations, the feature vectors are normalized and a spatial index structure (such as KD tree) is established, so that the nearest neighbor cluster can be quickly retrieved when filtering interference targets. In addition, the database design supports a dynamic update mechanism, when new interference samples are added, the parameters of the category cluster are automatically adjusted through incremental clustering algorithm to ensure the timeliness and accuracy of the anti-interference feature library.

[0024] S305, introduce new interference samples, retrain and update the anti-interference feature library.

[0025] In step S305, when a new interference sample is introduced, the new sample needs to be pre-processed first, including image enhancement (such as contrast adjustment, noise filtering) and standardization operation, to ensure that it is consistent with the dimension and scale of the existing feature library. Then, the pre-trained YOLOv11 model is used to extract the feature vector of the new sample, and the Euclidean distance similarity algorithm is used to calculate the distance between the new sample and the cluster center in the anti-interference feature library. If the similarity of the new sample to a certain cluster exceeds the preset threshold (such as 0.85), the new sample is assigned to the cluster and the statistical information (such as mean vector, covariance matrix) of the cluster is updated; if the similarity is lower than the threshold, the incremental clustering process is triggered, and the new cluster center is initialized by K-means algorithm, and all samples are re-assigned to optimize the cluster structure. For example, when a batch of withered leaf samples affected by pests is added, the system will compare the feature vectors of the samples with the existing categories such as fallen leaf cluster and weed cluster. If the texture features of the samples are significantly different from the fallen leaf cluster (such as higher edge sharpness, steeper color gradient), a new "withered leaf" sub-cluster may be split out. The updated feature library is synchronized to the spatiotemporal attention enhanced LSTM-GRU hybrid network, which dynamically adjusts the attention weight mechanism to make the model more accurate in distinguishing between sweet potato moth activity areas and interference in subsequent prediction, thereby improving the anti-interference ability in complex environments.

[0026] As an optional embodiment of the present application, the interference target in the aligned image data set is filtered out by the anti-interference feature library in step S4, which comprises: S401, extracting image feature vectors in the aligned image data set; In step S401, for the aligned image data set, similar feature extraction methods as in step S302 are used to extract corresponding feature vectors for different modal data. Specifically, for the hyperspectral image part, the 3D-CNN network is used to deeply mine the spectral information of each pixel point, and a 256-dimensional feature vector reflecting the physiological state of plants (such as chlorophyll content, water status, etc.) is extracted; for the RGB image, the pre-trained ResNet50 network is used to extract a 512-dimensional feature vector containing crown shape, color distribution, etc.; for the LiDAR point cloud data, the PointNet network is used to extract a 256-dimensional feature vector related to the three-dimensional structure of the plant, such as plant height, crown volume, etc. These feature vectors extracted from different modalities together constitute the comprehensive feature representation of the image data set.

[0027] S402, calculating the Euclidean distance similarity of the image feature vector and the features in the anti-interference feature library; In step S402, the image feature vector extracted in step S401 is compared with the anti-interference feature library constructed in step S303 one by one, and the similarity is quantified by calculating the Euclidean distance between them. As a classical similarity measurement method, the Euclidean distance can effectively reflect the distance relationship of the feature vectors in the multi-dimensional space. In the specific calculation, for each image feature vector, the Euclidean distance between it and all feature vectors in the anti-interference feature library is calculated, and the minimum distance is selected as the similarity index of the image feature vector and the feature library. If the minimum distance is less than the preset threshold, it is considered that the image feature vector is highly similar to a certain feature in the feature library, which may correspond to the beet armyworm; otherwise, if the minimum distance is greater than the threshold, it indicates that the image feature vector is significantly different from the features in the feature library, which may be an interference target such as weeds, fallen leaves, etc. Through this calculation method based on the Euclidean distance similarity, interference targets can be efficiently and accurately filtered out from the aligned image data set.

[0028] S403, set the similarity threshold, if the similarity is lower than the threshold, it is determined as an interference target and is filtered out.

[0029] In step S403, in actual application, the threshold can be determined by cross-validation method: the labeled sample data set is divided into training set and test set, different thresholds (such as 0.6 to 0.9, step 0.05) are traversed on the training set by grid search, the prediction accuracy of the filtered data set on the test set is calculated under each threshold, and the threshold that makes the accuracy optimal is selected as the final set value. For example, when the threshold is set to 0.75, the beet armyworm recognition accuracy of the model on the test set reaches 92%, and the misjudgment rate (the proportion of misjudging the real target as the interference target) is controlled within 5%, at this time the threshold can be identified as a reasonable threshold. In addition, the threshold setting also needs to consider environmental interference factors, such as rainy days or high temperature weather, the temperature distribution of the infrared feature map may fluctuate abnormally, at this time the threshold can be temporarily adjusted (such as increased to 0.85) to enhance the anti-interference ability of the model. Through this dynamic threshold adjustment mechanism, the accuracy and robustness of interference target filtering can be ensured.

[0030] As an optional embodiment of the present application, optionally, in step S5, the pest migration and diffusion matrix is constructed by: S501, collect wind speed, effective migration time, crop continuity coefficient and atmospheric vertical velocity coefficient to form multi-modal data; The wind speed data is collected in real time by a meteorological station deployed around the farmland, with a measurement range of 0-30 m / s and an accuracy of 0.1 m / s, which can accurately reflect the airflow conditions during the migration of the beet armyworm; the effective migration duration is calculated in combination with the diurnal rhythm model, and is determined according to the sunrise and sunset time of the target area and the peak activity period of the beet armyworm (usually 2-4 hours after sunset), for example, in the summer in the 30° north latitude area, the effective migration duration can be set to 3.5 hours; the crop continuity coefficient is obtained by remote sensing image interpretation, and the soybean planting area ratio is counted in 100m x 100m grid units, when the soybean planting ratio in the grid is more than 70%, the continuity coefficient is set to 1.2, and when it is less than 30%, the continuity coefficient is set to 0.8; the atmospheric vertical velocity coefficient is calculated in combination with meteorological reanalysis data (such as ERA5 data), and the vertical velocity field of the 500 hPa height layer is extracted, when the vertical upward velocity is more than 0.5 m / s, the coefficient is set to 1.1, and when the sinking velocity is more than 0.3 m / s, the coefficient is set to 0.9. These multi-modal data are transmitted in real time to the edge computing node through the Internet of Things terminal, and data cleaning (such as removing wind speed outliers, correcting meteorological station clock deviation) and space-time alignment (unified to UTC time standard) are performed.

[0031] S502, calculate the migration and diffusion distance based on the multi-modal data; S503, according to the migration and diffusion distance and the spatial position relationship between the migration starting point and the potential migration target area, construct a pest migration and diffusion matrix.

[0032] Migration and diffusion matrix construction: Diffusion distance = (wind speed x migration duration x crop continuity coefficient x atmospheric vertical velocity coefficient (Example: south wind 5 (wind speed 8 m / s), migration duration 6 hours, crop continuity 80% (coefficient 0.8), atmospheric vertical velocity 0.6 m / s (coefficient 1.2), then diffusion distance = 8 x 6 x 3600 x 0.8 x 1.2 = 165888 m ≈ 166 km); The construction of the pest migration and diffusion matrix needs to consider the migration and diffusion distance and spatial position relationship. Specifically, taking the migration starting point as the origin, the potential migration range is divided into multiple concentric circular regions according to the migration and diffusion distance calculated in step S502; at the same time, the latitude and longitude coordinates of the potential migration target regions (such as adjacent farmland, ecological protection zones, etc.) are obtained by combining the geographic information system (GIS), and the spatial distance and azimuth angle of each target region from the origin are calculated. The migration and diffusion distance is divided into different levels (such as short distance <5km, medium distance 5-20km, long distance >20km), and cross analysis is performed with the spatial position information (azimuth angle, target region type) to form a matrix containing migration probability, direction preference and influence range. For example, when the wind speed is 3m / s and the effective migration time is 3.5 hours, the migration and diffusion distance may cover the medium distance range, and the corresponding cell in the matrix for the azimuth angle (such as northeast direction) needs to be marked with a high migration probability (such as 0.8) and the possible affected target region (such as another piece of field 5km away); if the atmospheric vertical velocity coefficient is 1.1 (updraft), the matrix needs to correct the upper limit of the migration distance and expand the influence range. In addition, the matrix needs to be dynamically updated: according to the real-time collected wind speed and direction data, the migration probability value is adjusted every 10 minutes; combined with the crop contiguity coefficient, the migration risk level of contiguous planting areas (coefficient >1.2) is improved. Through this matrix construction method, the migration possibility of the beet armyworm from the starting point to different target regions can be intuitively displayed.

[0033] The pest migration and diffusion matrix is essentially a mathematical function or calculation framework that integrates four key environmental and ecological factors that affect the migration ability of the beet armyworm, and estimates its possible diffusion distance and direction through quantitative formulas. It is not a simple two-dimensional table, but a multi-parameter driven dynamic calculation model.

[0034] The core expression is as follows: Migration and diffusion distance (D) = f (wind speed, migration time, crop landscape, atmospheric conditions) In order to more clearly show its composition, the various components of the matrix, its quantitative method and biological mechanism are sorted out as follows: Comprehensive calculation model and example of the matrix Based on the above matrix components, a comprehensive calculation model is constructed: Migration and diffusion distance (D) = Ws x T x Cc x Va Where Ws represents wind speed, T represents effective migration time, Cc represents crop contiguity coefficient, and Va represents atmospheric vertical velocity coefficient; Calculation example: Suppose the following conditions are monitored in the evening: wind speed (Ws): south wind 5, about 8 m / s; effective migration time (T): 5 hours (18000 seconds) of adult active migration are observed; crop continuity coefficient (C_c): there is a large-scale cabbage planting area in the upwind direction, the continuity is 85%, and the coefficient is taken as 1.1; atmospheric vertical velocity coefficient (V_a): there is weak upward airflow, the vertical velocity is 0.4 m / s, and the coefficient is taken as 1.1; then the predicted migration and diffusion distance is: D = 8 m / s x 18000 s x 1.1 x 1.1 = 174,240 meters ≈ 174 km; Conclusion and output: The model predicts that under this meteorological and landscape condition, the adult beet armyworms in the source area are likely to migrate about 174 kilometers to the north. The system will combine this result with the wind direction to draw a potential migration path sector on the electronic map and mark the high-risk farmland area within 174 kilometers radius downstream to achieve early warning across regions. The core role and advantage of this matrix in this embodiment is from qualitative to quantitative: it changes the traditional qualitative experience of "possible migration" to the quantitative prediction of "how far can it fly and where will it fly", solving the industry pain points.

[0035] As an optional embodiment of the present application, optionally, the growth period dynamic weight mechanism in step S5 comprises: S504, defining a feature saliency calculation method for different modal data based on multi-modal data to obtain saliency scores of different modalities; S505, obtaining a regulation coefficient of each growth period of the crop based on field test data of the crop; In step S505, the field test data is obtained by setting standard test fields at different growth stages of crops (such as seedling stage, branching stage, podding stage, etc.), continuously monitoring and recording the population density of Spodoptera exigua, the damage degree (such as leaf consumption rate, hole number), crop growth index (such as plant height, leaf area index) and environmental parameters (such as temperature, humidity) at each stage. Taking soybean as an example, at the seedling stage (20-30 days after sowing), the larvae of Spodoptera exigua mainly feed on tender leaves, and the correlation between the leaf consumption rate recorded in the test field and the population density reaches 0.85 at this stage, indicating that the growth inhibition effect of the pest on crops at this stage is significant, and therefore the adjustment coefficient at the seedling stage is set to 1.2 (higher than that at other stages); at the podding stage (60-80 days after sowing), the soybean leaves have basically matured, and the direct impact of Spodoptera exigua damage on yield is reduced, and therefore the adjustment coefficient at this stage is set to 0.8. The calculation of the adjustment coefficient uses the weighted average method: the pest damage loss rate at each growth stage (calculated by comparing the yield of the test field) is compared with the average loss rate during the whole growth period to obtain the relative weight value. For example, if the loss rate at the seedling stage is 30% (the average loss rate during the whole growth period is 20%), the adjustment coefficient is 30% / 20%=1.5, but to avoid overestimating the impact of a single stage, the final coefficient needs to be corrected by expert experience (such as limited to the range of 0.5-1.5). In addition, the adjustment coefficient needs to be dynamically adjusted according to the climate zoning: in the warm and humid southern region, the number of generations of Spodoptera exigua is large throughout the year, and the interval between the seedling stage and the podding stage is shortened, so the adjustment coefficient can be appropriately smoothed (such as 1.1 at the seedling stage and 0.9 at the podding stage); in the cold and dry northern region, the growth period is prolonged, and the difference in the coefficient needs to be more obvious (such as 1.3 at the seedling stage and 0.7 at the podding stage). Through this adjustment coefficient setting based on field tests, the differential impact of pest damage on crops at different growth stages can be accurately quantified.

[0036] S506, calculating dynamic weights using significance scores and adjustment coefficients.

[0037] In step S506, the calculation of dynamic weights is as follows: the expression for calculating dynamic weights in step S506. Through this dynamic weight mechanism, the model can automatically adjust the contribution of each modality according to real-time data and crop growth status, thereby improving the accuracy and adaptability of migration prediction. In addition, the dynamic weights need to be updated regularly: as the crop growth process advances or environmental conditions change significantly (such as sudden strong wind weather), the significance scores and adjustment coefficients need to be recalculated to ensure the timeliness and accuracy of the weights.

[0038] The expression for calculating dynamic weights in step S506 is: wherein, represents the dynamic weight of the th modality data at the th time point, represents the dynamic weight of the At this moment The significance scores of the modal data (such as the significance score of wind speed, the significance score of crop contiguousness coefficient, etc.) are calculated in step S504. Indicates the first At present, crops are in their growth period The adjustment coefficients for different growth stages (such as seedling stage, jointing stage, etc.) are obtained by fitting field trial data in step S505. This represents the total number of categories in the multimodal data. The denominator of the above expression is the sum of the "significance score × adjustment coefficient" of all modal data, used to normalize the weights and ensure that the sum of all modal weights is 1.

[0039] As an optional embodiment of the present invention, optionally, in step S6, the filtered aligned image dataset is input into the spatiotemporal attention-enhanced LSTM-GRU hybrid network, and the output prediction results of the beet armyworm include: S601. Process the temporal data in the aligned image dataset through the LSTM layer of the hybrid network to extract the long-term dependency features of the beet armyworm behavior. In step S601, it is necessary to explain in detail that the internal structure of LSTM (Long Short-Term Memory) network includes mechanisms such as input gates, forget gates, and output gates, which can effectively capture long-term dependencies in the data. When processing time-series data in the aligned image dataset, the LSTM layer analyzes the image data frame by frame, and uses these gating mechanisms to filter out key information related to the behavior of the beet armyworm, such as changes in flight trajectory during migration and dwell time in different environments, thereby extracting data representations that can reflect the long-term dependency characteristics of the beet armyworm's behavior.

[0040] S602. Apply GRU layer and spatiotemporal attention mechanism to weight long-term dependency features to obtain weighted features; In step S602, it needs to be specified that the GRU (Gated Recurrent Unit) dynamically regulates information flow through update gate and reset gate, and can efficiently fuse the long-term dependence features output by the LSTM layer. The spatio-temporal attention mechanism further introduces weight distribution in the spatial dimension (such as different regions in the image) and the temporal dimension (such as different time points): in space, the model focuses on the regions in the image that are strongly related to the behavior of the beet armyworm (such as the position of the insect body and the damage trace on the crop leaves); in time, the model emphasizes the feature contribution of key time nodes (such as the start time of migration and the time of environmental mutation). In specific implementation, the spatio-temporal attention module calculates the attention weight of each spatio-temporal position, dynamically weights the features output by the GRU layer, and generates weighted features with spatio-temporal sensitivity. This weighting mechanism enables the model to capture both the long-term evolution of the beet armyworm behavior and the influence of key spatio-temporal events, thereby improving the accuracy of the prediction. For example, when a sudden strong wind is detected, the spatio-temporal attention mechanism automatically increases the feature weight of the windward region at that time period, enabling the model to more accurately predict the deflection of the migration path.

[0041] In step S603, based on the weighted features, the output layer calculates the beet armyworm occurrence probability, the migration path vector, and the damage peak intensity to form the final prediction result.

[0042] In step S603, it needs to be specified that the output layer adopts a fully connected neural network structure to map the weighted features to three-dimensional prediction targets: first, the beet armyworm occurrence probability (value range 0-1) is calculated through the Sigmoid activation function, reflecting the possibility of insect infestation in the target area; second, the migration path vector is generated using the hyperbolic tangent function (tanh), with the direction component representing the migration main direction (latitude and longitude coordinate difference) and the modulus component representing the predicted migration distance; finally, the damage peak intensity is calculated through the ReLU function, quantifying the maximum damage degree of the crop caused by the insect (such as the peak rate of leaf consumption). For example, when the model outputs an occurrence probability of 0.85, a path vector of (0.3, 0.7), and a damage intensity of 0.6, it indicates that there is an 85% probability of insect infestation in the area, the migration direction is biased towards the northeast quadrant, and the maximum damage rate of the crop may reach 60%. To enhance the interpretability of the prediction results, the system simultaneously generates a visual report: a heat map of the migration path is superimposed on the electronic map, with color depth representing the damage intensity distribution, and the predicted parameters of key time nodes are marked. This multi-dimensional output mechanism can push warning information to farmers through mobile terminals, achieving precise connection from regional prevention and control to field management. In addition, the model supports dynamic correction function: when the new monitoring data (such as real-time wind speed and crop growth) deviates from the prediction conditions by more than a threshold value, the system will automatically trigger the re-prediction process to ensure the timeliness of the prediction results.

[0043] As an optional embodiment of the present application, optionally, the method further comprises: S7, recommending a corresponding control method based on the occurrence degree of the beet armyworm; Based on the migration path of the beet armyworm and the beet armyworm damage peak, the application time is recommended.

[0044] It should be noted that in step S7, based on the identified larval stage, the "beet armyworm resistance database" is called, the acetylcholinesterase activity of 1-3 instar larvae is low, and biological pesticides (Bacillus thuringiensis) are recommended; the cuticle of 4-6 instar larvae is thickened, and low-toxicity chemical pesticides (chlorantraniliprole) are recommended, and the efficacy is increased by 35%-45%; then, according to the predicted high-risk plot coordinates (such as "North Latitude 30.5°, East Longitude 114.3°, Area 20 mu"), a drone flight path (using a "chessboard spraying" mode) is generated, and the spraying amount (such as 5g / mu of chlorantraniliprole, with a total dose of 100g) is output, realizing seamless connection of "accurate prediction-accurate application".

[0045] At the same time, combined with the prediction result of the migration path of the beet armyworm, the system analyzes the time window of the migration to the target area. If it is predicted that the migration path passes through a certain farmer's plot, and the damage peak appears 3 days later (such as the system showing "migration arrival time: 72 hours later, damage intensity peak 0.7"), the recommended application time is 24 hours before the migration arrives (i.e. 48 hours after application), at which time the pesticide can form a protective barrier to avoid the outbreak of the pest. For multiple plots on the migration path, the system will generate a priority list in order of damage intensity to guide farmers to reasonably allocate control resources. In addition, the system also supports dynamic adjustment of the application time according to meteorological data: if it is predicted that it will rain on the application day (probability>60%), it will be automatically delayed to 24 hours after the rain, to ensure the adhesion effect of the pesticide. Through this "occurrence degree-migration path-damage peak" three-dimensional linkage control recommendation mechanism, the pesticide use amount can be significantly reduced (verified by a pilot, with an average reduction of 28% of the pesticide use amount), and the pest control efficiency can be improved (the success rate of control is increased from 72% to 89%).

[0046] Embodiment 2 As shown in Figure 2 A beet armyworm prediction system based on YOLOv11, the system is used to implement the beet armyworm prediction method based on YOLOv11 described above; the system comprises: A multi-modal perception module for collecting multi-modal data; The module is mainly composed of a multi-rotor unmanned aerial vehicle carrying a hyperspectral camera, an RGB industrial camera, a miniature LiDAR sensor, and a ground Internet of Things node integrating soil temperature and humidity sensors, photosynthetically active radiation sensors, etc. The unmanned aerial vehicle flies according to the set flight parameters (such as adjustable flight height 5-15 m, ground sampling distance 0.05-0.2 m), synchronously collects hyperspectral, RGB image, LiDAR point cloud data, and records parameters such as solar elevation angle and atmospheric transmittance. The ground Internet of Things node real-time collects environmental data such as soil temperature and humidity, photosynthetically active radiation, and atmospheric temperature. The unmanned aerial vehicle and the ground node realize data synchronization through the 5G edge computing module, ensuring that the time error of multi-modal data is ≤0.5s.

[0047] The image recognition module is connected with the multi-modal perception module, and is used for recognizing visible light feature maps and infrared feature maps of the beet armyworm by using YOLOv11 in infrared light environment and visible light environment respectively and combining the diurnal activity characteristics of the beet armyworm, and constructing a beet armyworm image data set. The dynamic preprocessing module is connected with the image recognition module, and is used for preprocessing the constructed beet armyworm image data set by using a cross-scale data alignment algorithm, aligning the data obtained in the infrared light environment and the visible light environment on different scales, and obtaining an aligned image data set. The dynamic preprocessing module first performs data quality evaluation. The hyperspectral, LiDAR, and RGB image data are respectively evaluated in terms of signal-to-noise ratio, point cloud density, and definition, and are processed by using corresponding methods, such as wavelet threshold denoising algorithm for bands with SNR<30dB in the hyperspectral.

[0048] The spatio-temporal prediction model module is connected with the dynamic preprocessing module, and is used for outputting a prediction result of the beet armyworm based on the aligned image data set. The prediction result of the beet armyworm includes the occurrence degree of the beet armyworm, the migration path of the beet armyworm, and the damage peak value of the beet armyworm. The spatio-temporal prediction model module mainly consists of a dynamic weight Cross-Attention fusion module, a phenotype-environment mechanism feedback module, and a cross-scene adaptive learning unit. First, the feature vectors of hyperspectral, RGB, LiDAR, and environmental data are extracted through various modal feature encoding methods. Then, the dynamic weight Cross-Attention fusion module is used to calculate the fusion weight of each modal data according to the feature saliency and the growth period adjustment coefficient. After multiplying the encoding features of each modality with the dynamic weight, the Cross-Attention layer is inputted for cross-modal feature interaction, and the fusion feature vector is outputted. Next, the phenotype-environment mechanism feedback module is used to analyze the deviation between the data-driven predicted phenotype value and the theoretical value calculated by the mechanism model, and the fusion features are corrected according to the deviation coefficient. Finally, the cross-scene adaptive learning unit uses the domain adversarial neural network to take different regions of the noctuid dataset as the source domain and the target domain, and minimizes the cross-regional distribution difference through the gradient reversal layer to improve the model's generalization ability under different soil and climate conditions, and realizes the spatio-temporal prediction of noctuids.

[0049] The control decision module is connected with the spatio-temporal prediction model module, and is used for recommending corresponding control methods based on the occurrence degree of the beet armyworm; the spraying time is recommended based on the migration path of the beet armyworm and the beet armyworm damage peak value; The control decision module makes decisions based on the prediction results of the feedback adjustment and visualization subsystem. When the phenotype prediction deviation is >5%, the system automatically sends instructions to the unmanned aerial vehicle to increase the patrol frequency from 7 times / week to 3 times / week; when the deviation is <2%, the patrol frequency is reduced to 10 times / week. According to the predicted soil nitrogen content deviation, canopy water content, and other phenotype information, automatic fertilization suggestions (such as a 5 kg / ha urea adjustment for every 1% deviation) and irrigation suggestions (such as irrigation within 3 days when the water content is <60%) are generated for cultivation optimization suggestions. For breeding test fields, the top 20% of materials are selected according to the yield, lodging resistance, and other phenotype prediction results to generate a recommended list, providing decision support for precision control and intelligent breeding.

[0050] The result output module is connected with the control decision module, and is used for outputting the control method and the spraying time.

[0051] It should be noted that the result output module supports output in text form and has a visual display function. After receiving the prevention and control method and application time information from the prevention and control decision module, the module presents the information in the form of intuitive charts, maps, etc. For example, for the prevention and control method, the specific measures to be taken under different occurrence degrees are displayed in the form of a flowchart, including detailed information such as the type of pesticide used, the concentration, the application method, etc.; for the application time, the recommended application time of each plot is marked on the electronic map, and the application time arrangement of the entire region is displayed in the form of a time axis, facilitating the overall planning of farmers and agricultural management personnel.

[0052] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for predicting beet armyworm based on YOLOv11, characterized in that, The method includes: S1. Using YOLOv11, the visible light feature map and infrared feature map of the beet armyworm were identified in infrared and visible light environments, respectively, and combined with the diurnal activity characteristics of the beet armyworm, to construct an image dataset of the beet armyworm. S2. The constructed beet armyworm image dataset is preprocessed using a cross-scale data alignment algorithm. Data acquired under infrared and visible light environments are aligned at different scales to obtain an aligned image dataset. S3. Collect image data of interfering objects to construct an anti-interference feature library; S4. Use the anti-interference feature library to filter out the interference targets in the aligned image dataset; S5. Construct a spatiotemporal attention-enhanced LSTM-GRU hybrid network and introduce pest migration and diffusion matrix and dynamic weighting mechanism during the reproductive period into it. S6. Input the filtered aligned image dataset into the spatiotemporal attention-enhanced LSTM-GRU hybrid network and output the prediction results of beet armyworm. The prediction results of beet armyworm include the occurrence degree of beet armyworm, the migration path of beet armyworm, and the peak damage of beet armyworm.

2. The method for predicting beet armyworm based on YOLOv11 as described in claim 1, characterized in that, The construction of the anti-interference feature library in step S3 includes: S301. Under visible and infrared light environments, acquire image data of interfering objects and label the types of interfering objects in the image data; S302. Extract the interference feature vector from the interference image data; S303. Based on the interference feature vector, perform cluster analysis to obtain multiple interference category clusters, each interference category cluster corresponding to one interference category; S304. Store the feature vectors of each interference category cluster to form an anti-interference feature library; S305. Introduce new interference samples, retrain and update the anti-interference feature library.

3. The method for predicting beet armyworm based on YOLOv11 as described in claim 1, characterized in that, The construction of the pest migration and diffusion matrix in step S5 includes: S501, collected wind speed, effective migration time, crop contiguousness coefficient and atmospheric vertical velocity coefficient, constitute multimodal data; S502. Calculate the migration and diffusion distance based on the multimodal data; S503. Based on the migration and diffusion distance and the spatial relationship between the migration starting point and the potential migration target area, construct the pest migration and diffusion matrix.

4. A method for predicting beet armyworm based on YOLOv11 as described in claim 1 or 3, characterized in that, The dynamic weighting mechanism for reproductive period in step S5 includes: S504. Based on the multimodal data, define feature saliency calculation methods for different modal data to obtain saliency scores for different modalities; S505. Obtain the regulation coefficients of crops at each growth stage based on field trial data of crops; S506. Calculate the dynamic weight using the significance score and adjustment coefficient.

5. The method for predicting beet armyworm based on YOLOv11 as described in claim 4, characterized in that, The expression for calculating the dynamic weights in step S506 is as follows: in, Indicates the first At this moment Dynamic weights of modal data, Indicates the first At this moment Significance scores of modal data Indicates the first At present, crops are in their growth period The adjustment coefficient at that time This represents the total number of categories in multimodal data.

6. The method for predicting beet armyworm based on YOLOv11 as described in claim 1, characterized in that, The beet armyworm image dataset in step S1 includes: S101. Perform a convolution operation on the visible light feature map and the infrared feature map to obtain the feature information of the visible light feature map and the feature information of the infrared feature map, respectively. S102. Calculate the attention weights of the visible light feature map and the infrared feature map at different channels and spatial locations based on the feature information of the visible light feature map and the feature information of the infrared feature map, respectively. S103. Based on the attention weight, the visible light feature map and the infrared feature map are weighted and fused to obtain the fused feature map, and all the fused feature maps are combined to form a beet armyworm image dataset.

7. The method for predicting beet armyworm based on YOLOv11 as described in claim 1, characterized in that, In step S4, filtering out interference targets in the aligned image dataset using the anti-interference feature library includes: S401. Extract the image feature vector from the aligned image dataset; S402. Calculate the Euclidean distance similarity between the image feature vector and the features in the anti-interference feature library; S403. Set a similarity threshold. If the similarity is lower than the threshold, it is determined to be an interfering target and filtered out.

8. The method for predicting beet armyworm based on YOLOv11 as described in claim 1, characterized in that, In step S6, the filtered aligned image dataset is input into the spatiotemporal attention-enhanced LSTM-GRU hybrid network, and the output prediction results for the beet armyworm include: S601. Process the temporal data in the aligned image dataset through the LSTM layer of the hybrid network to extract long-term dependency features of beet armyworm behavior; S602. Apply GRU layer and spatiotemporal attention mechanism to weight long-term dependency features to obtain weighted features; S603. Based on the weighted features, the output layer calculates the probability of beet armyworm occurrence, migration path vector, and peak damage intensity to form the final prediction result.

9. The method for predicting beet armyworm based on YOLOv11 as described in claim 1, characterized in that, The method further includes: S7. Recommend corresponding control methods based on the severity of the beet armyworm outbreak; Based on the reported migration path of the beet armyworm and the peak damage caused by the beet armyworm, the recommended application time is [not specified].

10. A beet armyworm prediction system based on YOLOv11, characterized in that, The system is used to implement the beet armyworm prediction method based on YOLOv11 as described in any one of claims 1 to 9; the system includes: A multimodal sensing module is used to collect multimodal data; The image recognition module, connected to the multimodal sensing module, is used to identify the visible light feature map and infrared feature map of the beet armyworm in infrared light environment and visible light environment respectively, combined with the diurnal activity characteristics of the beet armyworm, and to construct the beet armyworm image dataset. A dynamic preprocessing module, connected to the image recognition module, is used to preprocess the constructed beet armyworm image dataset using a cross-scale data alignment algorithm, aligning data acquired under infrared and visible light environments at different scales to obtain an aligned image dataset. The spatiotemporal prediction model module, connected to the dynamic preprocessing module, is used to output the prediction results of beet armyworm based on the aligned image dataset. The prediction results of beet armyworm include the occurrence degree of beet armyworm, the migration path of beet armyworm, and the peak damage of beet armyworm. The prevention and control decision module, connected to the spatiotemporal prediction model module, is used to recommend corresponding prevention and control methods based on the severity of beet armyworm occurrence; and to recommend pesticide application times based on the beet armyworm migration path and the peak damage caused by the beet armyworm. The result output module is connected to the prevention and control decision module and is used to output the prevention and control method and the application time.

Images